Merge pull request #2254 from kohya-ss/dev

Merge the changes from the sd3 branch into main

.ai/claude.prompt.md

@@ -0,0 +1,9 @@
+## About This File
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## 1. Project Context
+Here is the essential context for our project. Please read and understand it thoroughly.
+
+### Project Overview
+@./context/01-overview.md

.ai/context/01-overview.md

@@ -0,0 +1,101 @@
+This file provides the overview and guidance for developers working with the codebase, including setup instructions, architecture details, and common commands.
+
+## Project Architecture
+
+### Core Training Framework
+The codebase is built around a **strategy pattern architecture** that supports multiple diffusion model families:
+
+- **`library/strategy_base.py`**: Base classes for tokenization, text encoding, latent caching, and training strategies
+- **`library/strategy_*.py`**: Model-specific implementations for SD, SDXL, SD3, FLUX, etc.
+- **`library/train_util.py`**: Core training utilities shared across all model types
+- **`library/config_util.py`**: Configuration management with TOML support
+
+### Model Support Structure
+Each supported model family has a consistent structure:
+- **Training script**: `{model}_train.py` (full fine-tuning), `{model}_train_network.py` (LoRA/network training)
+- **Model utilities**: `library/{model}_models.py`, `library/{model}_train_utils.py`, `library/{model}_utils.py`
+- **Networks**: `networks/lora_{model}.py`, `networks/oft_{model}.py` for adapter training
+
+### Supported Models
+- **Stable Diffusion 1.x**: `train*.py`, `library/train_util.py`, `train_db.py` (for DreamBooth)
+- **SDXL**: `sdxl_train*.py`, `library/sdxl_*`
+- **SD3**: `sd3_train*.py`, `library/sd3_*`
+- **FLUX.1**: `flux_train*.py`, `library/flux_*`
+
+### Key Components
+
+#### Memory Management
+- **Block swapping**: CPU-GPU memory optimization via `--blocks_to_swap` parameter, works with custom offloading. Only available for models with transformer architectures like SD3 and FLUX.1.
+- **Custom offloading**: `library/custom_offloading_utils.py` for advanced memory management
+- **Gradient checkpointing**: Memory reduction during training
+
+#### Training Features
+- **LoRA training**: Low-rank adaptation networks in `networks/lora*.py`
+- **ControlNet training**: Conditional generation control
+- **Textual Inversion**: Custom embedding training
+- **Multi-resolution training**: Bucket-based aspect ratio handling
+- **Validation loss**: Real-time training monitoring, only for LoRA training
+
+#### Configuration System
+Dataset configuration uses TOML files with structured validation:
+```toml
+[datasets.sample_dataset]
+  resolution = 1024
+  batch_size = 2
+  
+  [[datasets.sample_dataset.subsets]]
+    image_dir = "path/to/images"
+    caption_extension = ".txt"
+```
+
+## Common Development Commands
+
+### Training Commands Pattern
+All training scripts follow this general pattern:
+```bash
+accelerate launch --mixed_precision bf16 {script_name}.py \
+  --pretrained_model_name_or_path model.safetensors \
+  --dataset_config config.toml \
+  --output_dir output \
+  --output_name model_name \
+  [model-specific options]
+```
+
+### Memory Optimization
+For low VRAM environments, use block swapping:
+```bash
+# Add to any training command for memory reduction
+--blocks_to_swap 10  # Swap 10 blocks to CPU (adjust number as needed)
+```
+
+### Utility Scripts
+Located in `tools/` directory:
+- `tools/merge_lora.py`: Merge LoRA weights into base models
+- `tools/cache_latents.py`: Pre-cache VAE latents for faster training
+- `tools/cache_text_encoder_outputs.py`: Pre-cache text encoder outputs
+
+## Development Notes
+
+### Strategy Pattern Implementation
+When adding support for new models, implement the four core strategies:
+1. `TokenizeStrategy`: Text tokenization handling
+2. `TextEncodingStrategy`: Text encoder forward pass
+3. `LatentsCachingStrategy`: VAE encoding/caching
+4. `TextEncoderOutputsCachingStrategy`: Text encoder output caching
+
+### Testing Approach
+- Unit tests focus on utility functions and model loading
+- Integration tests validate training script syntax and basic execution
+- Most tests use mocks to avoid requiring actual model files
+- Add tests for new model support in `tests/test_{model}_*.py`
+
+### Configuration System
+- Use `config_util.py` dataclasses for type-safe configuration
+- Support both command-line arguments and TOML file configuration
+- Validate configuration early in training scripts to prevent runtime errors
+
+### Memory Management
+- Always consider VRAM limitations when implementing features
+- Use gradient checkpointing for large models
+- Implement block swapping for models with transformer architectures
+- Cache intermediate results (latents, text embeddings) when possible

.ai/gemini.prompt.md

@@ -0,0 +1,9 @@
+## About This File
+
+This file provides guidance to Gemini CLI (https://github.com/google-gemini/gemini-cli) when working with code in this repository.
+
+## 1. Project Context
+Here is the essential context for our project. Please read and understand it thoroughly.
+
+### Project Overview
+@./context/01-overview.md

.github/FUNDING.yml

@@ -0,0 +1,3 @@
+# These are supported funding model platforms
+
+github: kohya-ss

.github/workflows/tests.yml

@@ -0,0 +1,51 @@
+name: Test with pytest
+
+on: 
+  push:
+    branches:
+      - main
+      - dev
+      - sd3
+  pull_request:
+    branches:
+      - main
+      - dev
+      - sd3
+
+# CKV2_GHA_1: "Ensure top-level permissions are not set to write-all"
+permissions: read-all
+
+jobs:
+  build:
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: ["3.10"] # Python versions to test
+        pytorch-version: ["2.4.0", "2.6.0"] # PyTorch versions to test
+
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          # https://woodruffw.github.io/zizmor/audits/#artipacked
+          persist-credentials: false
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip' 
+
+      - name: Install and update pip, setuptools, wheel
+        run: |
+          # Setuptools, wheel for compiling some packages
+          python -m pip install --upgrade pip setuptools wheel
+
+      - name: Install dependencies
+        run: |
+          # Pre-install torch to pin version (requirements.txt has dependencies like transformers which requires pytorch)
+          pip install dadaptation==3.2 torch==${{ matrix.pytorch-version }} torchvision pytest==8.3.4
+          pip install -r requirements.txt
+
+      - name: Test with pytest
+        run: pytest # See pytest.ini for configuration
+

.github/workflows/typos.yml

@@ -1,21 +1,29 @@
 ---
-# yamllint disable rule:line-length
 name: Typos
 
-on:  # yamllint disable-line rule:truthy
+on: 
   push:
+    branches:
+      - main
+      - dev
   pull_request:
     types:
       - opened
       - synchronize
       - reopened
 
+# CKV2_GHA_1: "Ensure top-level permissions are not set to write-all"
+permissions: read-all
+
 jobs:
   build:
     runs-on: ubuntu-latest
 
     steps:
       - uses: actions/checkout@v4
+        with:
+          # https://woodruffw.github.io/zizmor/audits/#artipacked
+          persist-credentials: false
 
       - name: typos-action
-        uses: crate-ci/typos@v1.16.26
+        uses: crate-ci/typos@v1.28.1

.gitignore

@@ -6,3 +6,8 @@ venv
 build
 .vscode
 wandb
+CLAUDE.md
+GEMINI.md
+.claude
+.gemini
+MagicMock

README-ja.md

@@ -1,39 +1,164 @@
-SDXLがサポートされました。sdxlブランチはmainブランチにマージされました。リポジトリを更新したときにはUpgradeの手順を実行してください。また accelerate のバージョンが上がっていますので、accelerate config を再度実行してください。
+# sd-scripts
 
-SDXL学習については[こちら](./README.md#sdxl-training)をご覧ください（英語です）。
+[English](./README.md) / [日本語](./README-ja.md)
 
-## リポジトリについて
-Stable Diffusionの学習、画像生成、その他のスクリプトを入れたリポジトリです。
+## 目次
 
-[README in English](./README.md) ←更新情報はこちらにあります
+<details>
+<summary>クリックすると展開します</summary>
 
-GUIやPowerShellスクリプトなど、より使いやすくする機能が[bmaltais氏のリポジトリ](https://github.com/bmaltais/kohya_ss)で提供されています（英語です）のであわせてご覧ください。bmaltais氏に感謝します。
+- [はじめに](#はじめに)
+  - [スポンサー](#スポンサー)
+  - [スポンサー募集のお知らせ](#スポンサー募集のお知らせ)
+  - [更新履歴](#更新履歴)
+  - [サポートモデル](#サポートモデル)
+  - [機能](#機能)
+- [ドキュメント](#ドキュメント)
+  - [学習ドキュメント（英語および日本語）](#学習ドキュメント英語および日本語)
+  - [その他のドキュメント](#その他のドキュメント)
+  - [旧ドキュメント（日本語）](#旧ドキュメント日本語)
+- [AIコーディングエージェントを使う開発者の方へ](#aiコーディングエージェントを使う開発者の方へ)
+- [Windows環境でのインストール](#windows環境でのインストール)
+  - [Windowsでの動作に必要なプログラム](#windowsでの動作に必要なプログラム)
+  - [インストール手順](#インストール手順)
+  - [requirements.txtとPyTorchについて](#requirementstxtとpytorchについて)
+  - [xformersのインストール（オプション）](#xformersのインストールオプション)
+- [Linux/WSL2環境でのインストール](#linuxwsl2環境でのインストール)
+  - [DeepSpeedのインストール（実験的、LinuxまたはWSL2のみ）](#deepspeedのインストール実験的linuxまたはwsl2のみ)
+- [アップグレード](#アップグレード)
+  - [PyTorchのアップグレード](#pytorchのアップグレード)
+- [謝意](#謝意)
+- [ライセンス](#ライセンス)
 
-以下のスクリプトがあります。
+</details>
 
-* DreamBooth、U-NetおよびText Encoderの学習をサポート
-* fine-tuning、同上
-* LoRAの学習をサポート
+## はじめに
+
+Stable Diffusion等の画像生成モデルの学習、モデルによる画像生成、その他のスクリプトを入れたリポジトリです。
+
+### スポンサー
+
+このプロジェクトを支援してくださる企業・団体の皆様に深く感謝いたします。
+
+<a href="https://aihub.co.jp/">
+  <img src="./images/logo_aihub.png" alt="AiHUB株式会社" title="AiHUB株式会社" height="100px">
+</a>
+
+### スポンサー募集のお知らせ
+
+このプロジェクトがお役に立ったなら、ご支援いただけると嬉しく思います。 [GitHub Sponsors](https://github.com/sponsors/kohya-ss/)で受け付けています。
+
+### 更新履歴
+
+- **Version 0.10.0 (2026-01-19):**
+  - `sd3`ブランチを`main`ブランチにマージしました。このバージョンからFLUX.1およびSD3/SD3.5等のモデルが`main`ブランチでサポートされます。
+  - ドキュメントにはまだ不備があるため、お気づきの点はIssue等でお知らせください。
+  - `sd3`ブランチは当面、`dev`ブランチと同期して開発ブランチとして維持します。
+
+### サポートモデル
+
+* **Stable Diffusion 1.x/2.x**
+* **SDXL**
+* **SD3/SD3.5**
+* **FLUX.1**
+* **LUMINA**
+* **HunyuanImage-2.1**
+
+### 機能
+
+* LoRA学習
+* fine-tuning（DreamBooth）：HunyuanImage-2.1以外のモデル
+* Textual Inversion学習：SD/SDXL
 * 画像生成
-* モデル変換（Stable Diffision ckpt/safetensorsとDiffusersの相互変換）
+* その他、モデル変換やタグ付け、LoRAマージなどのユーティリティ
 
-## 使用法について
+## ドキュメント
+
+### 学習ドキュメント（英語および日本語）
+
+日本語は折りたたまれているか、別のドキュメントにあります。
+
+* [LoRA学習の概要](./docs/train_network.md)
+* [データセット設定](./docs/config_README-ja.md) / [英語版](./docs/config_README-en.md)
+* [高度な学習オプション](./docs/train_network_advanced.md)
+* [SDXL学習](./docs/sdxl_train_network.md)
+* [SD3学習](./docs/sd3_train_network.md)
+* [FLUX.1学習](./docs/flux_train_network.md)
+* [LUMINA学習](./docs/lumina_train_network.md)
+* [HunyuanImage-2.1学習](./docs/hunyuan_image_train_network.md)
+* [Fine-tuning](./docs/fine_tune.md)
+* [Textual Inversion学習](./docs/train_textual_inversion.md)
+* [ControlNet-LLLite学習](./docs/train_lllite_README-ja.md) / [英語版](./docs/train_lllite_README.md)
+* [Validation](./docs/validation.md)
+* [マスク損失学習](./docs/masked_loss_README-ja.md) / [英語版](./docs/masked_loss_README.md)
+
+### その他のドキュメント
+
+* [画像生成スクリプト](./docs/gen_img_README-ja.md) / [英語版](./docs/gen_img_README.md)
+* [WD14 Taggerによる画像タグ付け](./docs/wd14_tagger_README-ja.md) / [英語版](./docs/wd14_tagger_README-en.md)
+
+### 旧ドキュメント（日本語）
 
 * [学習について、共通編](./docs/train_README-ja.md) : データ整備やオプションなど
-    * [データセット設定](./docs/config_README-ja.md)
 * [DreamBoothの学習について](./docs/train_db_README-ja.md)
-* [fine-tuningのガイド](./docs/fine_tune_README_ja.md):
-* [LoRAの学習について](./docs/train_network_README-ja.md)
-* [Textual Inversionの学習について](./docs/train_ti_README-ja.md)
-* [画像生成スクリプト](./docs/gen_img_README-ja.md)
-* note.com [モデル変換スクリプト](https://note.com/kohya_ss/n/n374f316fe4ad)
 
-## Windowsでの動作に必要なプログラム
+## AIコーディングエージェントを使う開発者の方へ
 
-Python 3.10.6およびGitが必要です。
+This repository provides recommended instructions to help AI agents like Claude and Gemini understand our project context and coding standards.
 
-- Python 3.10.6: https://www.python.org/ftp/python/3.10.6/python-3.10.6-amd64.exe
-- git: https://git-scm.com/download/win
+To use them, you need to opt-in by creating your own configuration file in the project root.
+
+**Quick Setup:**
+
+1.  Create a `CLAUDE.md` and/or `GEMINI.md` file in the project root.
+2.  Add the following line to your `CLAUDE.md` to import the repository's recommended prompt:
+
+    ```markdown
+    @./.ai/claude.prompt.md
+    ```
+
+    or for Gemini:
+
+    ```markdown
+    @./.ai/gemini.prompt.md
+    ```
+
+3.  You can now add your own personal instructions below the import line (e.g., `Always respond in Japanese.`).
+
+This approach ensures that you have full control over the instructions given to your agent while benefiting from the shared project context. Your `CLAUDE.md` and `GEMINI.md` are already listed in `.gitignore`, so they won't be committed to the repository.
+
+このリポジトリでは、AIコーディングエージェント（例：Claude、Geminiなど）がプロジェクトのコンテキストやコーディング標準を理解できるようにするための推奨プロンプトを提供しています。
+
+それらを使用するには、プロジェクトディレクトリに設定ファイルを作成して明示的に有効にする必要があります。
+
+**簡単なセットアップ手順:**
+
+1.  プロジェクトルートに `CLAUDE.md` や `GEMINI.md` ファイルを作成します。
+2.  `CLAUDE.md` に以下の行を追加して、リポジトリの推奨プロンプトをインポートします。
+
+    ```markdown
+    @./.ai/claude.prompt.md
+    ```
+
+    またはGeminiの場合:
+
+    ```markdown
+    @./.ai/gemini.prompt.md
+    ``` 
+3.  インポート行の下に、独自の指示を追加できます（例：`常に日本語で応答してください。`）。
+
+この方法により、エージェントに与える指示を各開発者が管理しつつ、リポジトリの推奨コンテキストを活用できます。`CLAUDE.md` および `GEMINI.md` は `.gitignore` に登録されているため、リポジトリにコミットされることはありません。
+
+## Windows環境でのインストール
+
+### Windowsでの動作に必要なプログラム
+
+Python 3.10.xおよびGitが必要です。
+
+- Python 3.10.x: https://www.python.org/downloads/windows/ からWindows installer (64-bit)をダウンロード
+- git: https://git-scm.com/download/win から最新版をダウンロード
+
+Python 3.11.x、3.12.xでも恐らく動作します（未テスト）。
 
 PowerShellを使う場合、venvを使えるようにするためには以下の手順でセキュリティ設定を変更してください。
 （venvに限らずスクリプトの実行が可能になりますので注意してください。）
@@ -42,13 +167,7 @@ PowerShellを使う場合、venvを使えるようにするためには以下の
 - 「Set-ExecutionPolicy Unrestricted」と入力し、Yと答えます。
 - 管理者のPowerShellを閉じます。
 
-## Windows環境でのインストール
-
-スクリプトはPyTorch 2.0.1でテストしています。PyTorch 1.12.1でも動作すると思われます。
-
-以下の例ではPyTorchは2.0.1／CUDA 11.8版をインストールします。CUDA 11.6版やPyTorch 1.12.1を使う場合は適宜書き換えください。
-
-（なお、python -m venv～の行で「python」とだけ表示された場合、py -m venv～のようにpythonをpyに変更してください。）
+### インストール手順
 
 PowerShellを使う場合、通常の（管理者ではない）PowerShellを開き以下を順に実行します。
 
@@ -59,21 +178,22 @@ cd sd-scripts
 python -m venv venv
 .\venv\Scripts\activate
 
-pip install torch==2.0.1+cu118 torchvision==0.15.2+cu118 --index-url https://download.pytorch.org/whl/cu118
+pip install torch==2.6.0 torchvision==0.21.0 --index-url https://download.pytorch.org/whl/cu124
 pip install --upgrade -r requirements.txt
-pip install xformers==0.0.20
 
 accelerate config
 ```
 
 コマンドプロンプトでも同一です。
 
-（注:``python -m venv venv`` のほうが ``python -m venv --system-site-packages venv`` より安全そうなため書き換えました。globalなpythonにパッケージがインストールしてあると、後者だといろいろと問題が起きます。）
+（なお、python -m venv～の行で「python」とだけ表示された場合、py -m venv～のようにpythonをpyに変更してください。）
+
+注：`bitsandbytes`、`prodigyopt`、`lion-pytorch` は `requirements.txt` に含まれています。
+
+この例ではCUDA 12.4版をインストールします。異なるバージョンのCUDAを使用する場合は、適切なバージョンのPyTorchをインストールしてください。たとえばCUDA 12.1版の場合は `pip install torch==2.6.0 torchvision==0.21.0 --index-url https://download.pytorch.org/whl/cu121` としてください。
 
 accelerate configの質問には以下のように答えてください。（bf16で学習する場合、最後の質問にはbf16と答えてください。）
 
-※0.15.0から日本語環境では選択のためにカーソルキーを押すと落ちます（……）。数字キーの0、1、2……で選択できますので、そちらを使ってください。
-
 ```txt
 - This machine
 - No distributed training
@@ -87,39 +207,36 @@ accelerate configの質問には以下のように答えてください。（bf1
 ※場合によって ``ValueError: fp16 mixed precision requires a GPU`` というエラーが出ることがあるようです。この場合、6番目の質問（
 ``What GPU(s) (by id) should be used for training on this machine as a comma-separated list? [all]:``）に「0」と答えてください。（id `0`のGPUが使われます。）
 
-### オプション：`bitsandbytes`（8bit optimizer）を使う
+### requirements.txtとPyTorchについて
 
-`bitsandbytes`はオプションになりました。Linuxでは通常通りpipでインストールできます（0.41.1または以降のバージョンを推奨）。
+PyTorchは環境によってバージョンが異なるため、requirements.txtには含まれていません。前述のインストール手順を参考に、環境に合わせてPyTorchをインストールしてください。
 
-Windowsでは0.35.0または0.41.1を推奨します。
+スクリプトはPyTorch 2.6.0でテストしています。PyTorch 2.6.0以降が必要です。
 
-- `bitsandbytes` 0.35.0: 安定しているとみられるバージョンです。AdamW8bitは使用できますが、他のいくつかの8bit optimizer、学習時の`full_bf16`オプションは使用できません。
-- `bitsandbytes` 0.41.1: Lion8bit、PagedAdamW8bit、PagedLion8bitをサポートします。`full_bf16`が使用できます。
+RTX 50シリーズGPUの場合、PyTorch 2.8.0とCUDA 12.8/12.9を使用してください。`requirements.txt`はこのバージョンでも動作します。
 
-注：`bitsandbytes` 0.35.0から0.41.0までのバージョンには問題があるようです。 https://github.com/TimDettmers/bitsandbytes/issues/659
+### xformersのインストール（オプション）
 
-以下の手順に従い、`bitsandbytes`をインストールしてください。
+xformersをインストールするには、仮想環境を有効にした状態で以下のコマンドを実行してください。
 
-### 0.35.0を使う場合
-
-PowerShellの例です。コマンドプロンプトではcpの代わりにcopyを使ってください。
-
-```powershell
-cd sd-scripts
-.\venv\Scripts\activate
-pip install bitsandbytes==0.35.0
-
-cp .\bitsandbytes_windows\*.dll .\venv\Lib\site-packages\bitsandbytes\
-cp .\bitsandbytes_windows\cextension.py .\venv\Lib\site-packages\bitsandbytes\cextension.py
-cp .\bitsandbytes_windows\main.py .\venv\Lib\site-packages\bitsandbytes\cuda_setup\main.py
+```bash
+pip install xformers --index-url https://download.pytorch.org/whl/cu124
 ```
 
-### 0.41.1を使う場合
+必要に応じてCUDAバージョンを変更してください。一部のGPUアーキテクチャではxformersが利用できない場合があります。
 
-jllllll氏の配布されている[こちら](https://github.com/jllllll/bitsandbytes-windows-webui) または他の場所から、Windows用のwhlファイルをインストールしてください。
+## Linux/WSL2環境でのインストール
 
-```powershell
-python -m pip install bitsandbytes==0.41.1 --prefer-binary --extra-index-url=https://jllllll.github.io/bitsandbytes-windows-webui
+LinuxまたはWSL2環境でのインストール手順はWindows環境とほぼ同じです。`venv\Scripts\activate` の部分を `source venv/bin/activate` に変更してください。
+
+※NVIDIAドライバやCUDAツールキットなどは事前にインストールしておいてください。
+
+### DeepSpeedのインストール（実験的、LinuxまたはWSL2のみ）
+
+DeepSpeedをインストールするには、仮想環境を有効にした状態で以下のコマンドを実行してください。
+
+```bash
+pip install deepspeed==0.16.7
 ```
 
 ## アップグレード
@@ -135,6 +252,10 @@ pip install --use-pep517 --upgrade -r requirements.txt
 
 コマンドが成功すれば新しいバージョンが使用できます。
 
+### PyTorchのアップグレード
+
+PyTorchをアップグレードする場合は、[Windows環境でのインストール](#windows環境でのインストール)のセクションの`pip install`コマンドを参考にしてください。
+
 ## 謝意
 
 LoRAの実装は[cloneofsimo氏のリポジトリ](https://github.com/cloneofsimo/lora)を基にしたものです。感謝申し上げます。
@@ -150,5 +271,3 @@ Conv2d 3x3への拡大は [cloneofsimo氏](https://github.com/cloneofsimo/lora)
 [bitsandbytes](https://github.com/TimDettmers/bitsandbytes): MIT
 
 [BLIP](https://github.com/salesforce/BLIP): BSD-3-Clause
-
-

README.md

@@ -1,51 +1,132 @@
-__SDXL is now supported. The sdxl branch has been merged into the main branch. If you update the repository, please follow the upgrade instructions. Also, the version of accelerate has been updated, so please run accelerate config again.__ The documentation for SDXL training is [here](./README.md#sdxl-training).
+# sd-scripts
 
-This repository contains training, generation and utility scripts for Stable Diffusion.
+[English](./README.md) / [日本語](./README-ja.md)
 
-[__Change History__](#change-history) is moved to the bottom of the page. 
-更新履歴は[ページ末尾](#change-history)に移しました。
+## Table of Contents
+<details>
+<summary>Click to expand</summary>
 
-[日本語版READMEはこちら](./README-ja.md)
+- [Introduction](#introduction)
+  - [Supported Models](#supported-models)
+  - [Features](#features)
+  - [Sponsors](#sponsors)
+  - [Support the Project](#support-the-project)
+- [Documentation](#documentation)
+  - [Training Documentation (English and Japanese)](#training-documentation-english-and-japanese)
+  - [Other Documentation (English and Japanese)](#other-documentation-english-and-japanese)
+- [For Developers Using AI Coding Agents](#for-developers-using-ai-coding-agents)
+- [Windows Installation](#windows-installation)
+  - [Windows Required Dependencies](#windows-required-dependencies)
+  - [Installation Steps](#installation-steps)
+  - [About requirements.txt and PyTorch](#about-requirementstxt-and-pytorch)
+  - [xformers installation (optional)](#xformers-installation-optional)
+- [Linux/WSL2 Installation](#linuxwsl2-installation)
+  - [DeepSpeed installation (experimental, Linux or WSL2 only)](#deepspeed-installation-experimental-linux-or-wsl2-only)
+- [Upgrade](#upgrade)
+  - [Upgrade PyTorch](#upgrade-pytorch)
+- [Credits](#credits)
+- [License](#license)
 
-For easier use (GUI and PowerShell scripts etc...), please visit [the repository maintained by bmaltais](https://github.com/bmaltais/kohya_ss). Thanks to @bmaltais!
+</details>
 
-This repository contains the scripts for:
+## Introduction
+
+This repository contains training, generation and utility scripts for Stable Diffusion and other image generation models.
+
+### Sponsors
+
+We are grateful to the following companies for their generous sponsorship:
+
+<a href="https://aihub.co.jp/top-en">
+  <img src="./images/logo_aihub.png" alt="AiHUB Inc." title="AiHUB Inc." height="100px">
+</a>
+
+### Support the Project
+
+If you find this project helpful, please consider supporting its development via [GitHub Sponsors](https://github.com/sponsors/kohya-ss/). Your support is greatly appreciated!
+
+### Change History
+
+- **Version 0.10.0 (2026-01-19):**
+    - `sd3` branch is merged to `main` branch. From this version, FLUX.1 and SD3/SD3.5 etc. are supported in the `main` branch.
+    - There are still some missing parts in the documentation, so please let us know if you find any issues via Issues etc.
+    - The `sd3` branch will be maintained as a development branch synchronized with `dev` for the time being.
+
+### Supported Models
+
+* **Stable Diffusion 1.x/2.x**
+* **SDXL**
+* **SD3/SD3.5**
+* **FLUX.1**
+* **LUMINA**
+* **HunyuanImage-2.1**
+
+### Features
 
-* DreamBooth training, including U-Net and Text Encoder
-* Fine-tuning (native training), including U-Net and Text Encoder
 * LoRA training
-* Textual Inversion training
+* Fine-tuning (native training, DreamBooth): except for HunyuanImage-2.1
+* Textual Inversion training: SD/SDXL
 * Image generation
-* Model conversion (supports 1.x and 2.x, Stable Diffision ckpt/safetensors and Diffusers)
+* Other utilities such as model conversion, image tagging, LoRA merging, etc.
 
-## About requirements.txt
+## Documentation
 
-These files do not contain requirements for PyTorch. Because the versions of them depend on your environment. Please install PyTorch at first (see installation guide below.) 
+### Training Documentation (English and Japanese)
 
-The scripts are tested with Pytorch 2.0.1. 1.12.1 is not tested but should work.
+* [LoRA Training Overview](./docs/train_network.md)
+* [Dataset config](./docs/config_README-en.md) / [Japanese version](./docs/config_README-ja.md)
+* [Advanced Training](./docs/train_network_advanced.md)
+* [SDXL Training](./docs/sdxl_train_network.md)
+* [SD3 Training](./docs/sd3_train_network.md)
+* [FLUX.1 Training](./docs/flux_train_network.md)
+* [LUMINA Training](./docs/lumina_train_network.md)
+* [HunyuanImage-2.1 Training](./docs/hunyuan_image_train_network.md)
+* [Fine-tuning](./docs/fine_tune.md)
+* [Textual Inversion Training](./docs/train_textual_inversion.md)
+* [ControlNet-LLLite Training](./docs/train_lllite_README.md) / [Japanese version](./docs/train_lllite_README-ja.md)
+* [Validation](./docs/validation.md)
+* [Masked Loss Training](./docs/masked_loss_README.md) / [Japanese version](./docs/masked_loss_README-ja.md)
 
-## Links to usage documentation
+### Other Documentation (English and Japanese)
 
-Most of the documents are written in Japanese.
+* [Image generation](./docs/gen_img_README.md) / [Japanese version](./docs/gen_img_README-ja.md)
+* [Tagging images with WD14 Tagger](./docs/wd14_tagger_README-en.md) / [Japanese version](./docs/wd14_tagger_README-ja.md)
 
-[English translation by darkstorm2150 is here](https://github.com/darkstorm2150/sd-scripts#links-to-usage-documentation). Thanks to darkstorm2150!
+## For Developers Using AI Coding Agents
 
-* [Training guide - common](./docs/train_README-ja.md) : data preparation, options etc... 
-  * [Chinese version](./docs/train_README-zh.md)
-* [Dataset config](./docs/config_README-ja.md) 
-* [DreamBooth training guide](./docs/train_db_README-ja.md)
-* [Step by Step fine-tuning guide](./docs/fine_tune_README_ja.md):
-* [training LoRA](./docs/train_network_README-ja.md)
-* [training Textual Inversion](./docs/train_ti_README-ja.md)
-* [Image generation](./docs/gen_img_README-ja.md)
-* note.com [Model conversion](https://note.com/kohya_ss/n/n374f316fe4ad)
+This repository provides recommended instructions to help AI agents like Claude and Gemini understand our project context and coding standards.
 
-## Windows Required Dependencies
+To use them, you need to opt-in by creating your own configuration file in the project root.
 
-Python 3.10.6 and Git:
+**Quick Setup:**
 
-- Python 3.10.6: https://www.python.org/ftp/python/3.10.6/python-3.10.6-amd64.exe
-- git: https://git-scm.com/download/win
+1.  Create a `CLAUDE.md` and/or `GEMINI.md` file in the project root.
+2.  Add the following line to your `CLAUDE.md` to import the repository's recommended prompt:
+
+    ```markdown
+    @./.ai/claude.prompt.md
+    ```
+
+    or for Gemini:
+
+    ```markdown
+    @./.ai/gemini.prompt.md
+    ```
+
+3.  You can now add your own personal instructions below the import line (e.g., `Always respond in Japanese.`).
+
+This approach ensures that you have full control over the instructions given to your agent while benefiting from the shared project context. Your `CLAUDE.md` and `GEMINI.md` are already listed in `.gitignore`, so they won't be committed to the repository.
+
+## Windows Installation
+
+### Windows Required Dependencies
+
+Python 3.10.x and Git:
+
+- Python 3.10.x: Download Windows installer (64-bit) from https://www.python.org/downloads/windows/
+- git: Download latest installer from https://git-scm.com/download/win
+
+Python 3.11.x, and 3.12.x will work but not tested.
 
 Give unrestricted script access to powershell so venv can work:
 
@@ -53,7 +134,7 @@ Give unrestricted script access to powershell so venv can work:
 - Type `Set-ExecutionPolicy Unrestricted` and answer A
 - Close admin powershell window
 
-## Windows Installation
+### Installation Steps
 
 Open a regular Powershell terminal and type the following inside:
 
@@ -64,20 +145,18 @@ cd sd-scripts
 python -m venv venv
 .\venv\Scripts\activate
 
-pip install torch==2.0.1+cu118 torchvision==0.15.2+cu118 --index-url https://download.pytorch.org/whl/cu118
+pip install torch==2.6.0 torchvision==0.21.0 --index-url https://download.pytorch.org/whl/cu124
 pip install --upgrade -r requirements.txt
-pip install xformers==0.0.20
 
 accelerate config
 ```
 
-__Note:__ Now bitsandbytes is optional. Please install any version of bitsandbytes as needed. Installation instructions are in the following section.
+If `python -m venv` shows only `python`, change `python` to `py`.
+
+Note: `bitsandbytes`, `prodigyopt` and `lion-pytorch` are included in the requirements.txt. If you'd like to use another version, please install it manually.
+
+This installation is for CUDA 12.4. If you use a different version of CUDA, please install the appropriate version of PyTorch. For example, if you use CUDA 12.1, please install `pip install torch==2.6.0 torchvision==0.21.0 --index-url https://download.pytorch.org/whl/cu121`.
 
-<!-- 
-cp .\bitsandbytes_windows\*.dll .\venv\Lib\site-packages\bitsandbytes\
-cp .\bitsandbytes_windows\cextension.py .\venv\Lib\site-packages\bitsandbytes\cextension.py
-cp .\bitsandbytes_windows\main.py .\venv\Lib\site-packages\bitsandbytes\cuda_setup\main.py
--->
 Answers to accelerate config:
 
 ```txt
@@ -90,46 +169,43 @@ Answers to accelerate config:
 - fp16
 ```
 
-note: Some user reports ``ValueError: fp16 mixed precision requires a GPU`` is occurred in training. In this case, answer `0` for the 6th question: 
+If you'd like to use bf16, please answer `bf16` to the last question.
+
+Note: Some user reports ``ValueError: fp16 mixed precision requires a GPU`` is occurred in training. In this case, answer `0` for the 6th question: 
 ``What GPU(s) (by id) should be used for training on this machine as a comma-separated list? [all]:`` 
 
 (Single GPU with id `0` will be used.)
 
-### Optional: Use `bitsandbytes` (8bit optimizer)
+## About requirements.txt and PyTorch
 
-For 8bit optimizer, you need to install `bitsandbytes`. For Linux, please install `bitsandbytes` as usual (0.41.1 or later is recommended.)
+The file does not contain requirements for PyTorch. Because the version of PyTorch depends on the environment, it is not included in the file. Please install PyTorch first according to the environment. See installation instructions below.
 
-For Windows, there are several versions of `bitsandbytes`:
+The scripts are tested with PyTorch 2.6.0. PyTorch 2.6.0 or later is required.
 
-- `bitsandbytes` 0.35.0: Stable version. AdamW8bit is available. `full_bf16` is not available.
-- `bitsandbytes` 0.41.1: Lion8bit, PagedAdamW8bit and PagedLion8bit are available. `full_bf16` is available.
+For RTX 50 series GPUs, PyTorch 2.8.0 with CUDA 12.8/12.9 should be used. `requirements.txt` will work with this version.
 
-Note: `bitsandbytes`above 0.35.0 till 0.41.0 seems to have an issue: https://github.com/TimDettmers/bitsandbytes/issues/659
+### xformers installation (optional)
 
-Follow the instructions below to install `bitsandbytes` for Windows.
+To install xformers, run the following command in your activated virtual environment:
 
-### bitsandbytes 0.35.0 for Windows
-
-Open a regular Powershell terminal and type the following inside:
-
-```powershell
-cd sd-scripts
-.\venv\Scripts\activate
-pip install bitsandbytes==0.35.0
-
-cp .\bitsandbytes_windows\*.dll .\venv\Lib\site-packages\bitsandbytes\
-cp .\bitsandbytes_windows\cextension.py .\venv\Lib\site-packages\bitsandbytes\cextension.py
-cp .\bitsandbytes_windows\main.py .\venv\Lib\site-packages\bitsandbytes\cuda_setup\main.py
+```bash
+pip install xformers --index-url https://download.pytorch.org/whl/cu124
 ```
 
-This will install `bitsandbytes` 0.35.0 and copy the necessary files to the `bitsandbytes` directory.
+Please change the CUDA version in the URL according to your environment if necessary. xformers may not be available for some GPU architectures.
 
-### bitsandbytes 0.41.1 for Windows
+## Linux/WSL2 Installation
 
-Install the Windows version whl file from [here](https://github.com/jllllll/bitsandbytes-windows-webui) or other sources, like:
+Linux or WSL2 installation steps are almost the same as Windows. Just change `venv\Scripts\activate` to `source venv/bin/activate`.
 
-```powershell
-python -m pip install bitsandbytes==0.41.1 --prefer-binary --extra-index-url=https://jllllll.github.io/bitsandbytes-windows-webui
+Note: Please make sure that NVIDIA driver and CUDA toolkit are installed in advance.
+
+### DeepSpeed installation (experimental, Linux or WSL2 only)
+  
+To install DeepSpeed, run the following command in your activated virtual environment:
+
+```bash
+pip install deepspeed==0.16.7 
 ```
 
 ## Upgrade
@@ -145,6 +221,10 @@ pip install --use-pep517 --upgrade -r requirements.txt
 
 Once the commands have completed successfully you should be ready to use the new version.
 
+### Upgrade PyTorch
+
+If you want to upgrade PyTorch, you can upgrade it with `pip install` command in [Windows Installation](#windows-installation) section.
+
 ## Credits
 
 The implementation for LoRA is based on [cloneofsimo's repo](https://github.com/cloneofsimo/lora). Thank you for great work!
@@ -160,200 +240,3 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
 [bitsandbytes](https://github.com/TimDettmers/bitsandbytes): MIT
 
 [BLIP](https://github.com/salesforce/BLIP): BSD-3-Clause
-
-
-## SDXL training
-
-The documentation in this section will be moved to a separate document later.
-
-### Training scripts for SDXL
-
-- `sdxl_train.py` is a script for SDXL fine-tuning. The usage is almost the same as `fine_tune.py`, but it also supports DreamBooth dataset.
-  - `--full_bf16` option is added. Thanks to KohakuBlueleaf!
-    - This option enables the full bfloat16 training (includes gradients). This option is useful to reduce the GPU memory usage. 
-    - The full bfloat16 training might be unstable. Please use it at your own risk.
-  - The different learning rates for each U-Net block are now supported in sdxl_train.py. Specify with `--block_lr` option. Specify 23 values separated by commas like `--block_lr 1e-3,1e-3 ... 1e-3`.
-    - 23 values correspond to `0: time/label embed, 1-9: input blocks 0-8, 10-12: mid blocks 0-2, 13-21: output blocks 0-8, 22: out`.
-- `prepare_buckets_latents.py` now supports SDXL fine-tuning.
-
-- `sdxl_train_network.py` is a script for LoRA training for SDXL. The usage is almost the same as `train_network.py`.
-
-- Both scripts has following additional options:
-  - `--cache_text_encoder_outputs` and `--cache_text_encoder_outputs_to_disk`: Cache the outputs of the text encoders. This option is useful to reduce the GPU memory usage. This option cannot be used with options for shuffling or dropping the captions.
-  - `--no_half_vae`: Disable the half-precision (mixed-precision) VAE. VAE for SDXL seems to produce NaNs in some cases. This option is useful to avoid the NaNs.
-
-- `--weighted_captions` option is not supported yet for both scripts.
-
-- `sdxl_train_textual_inversion.py` is a script for Textual Inversion training for SDXL. The usage is almost the same as `train_textual_inversion.py`.
-  - `--cache_text_encoder_outputs` is not supported.
-  - There are two options for captions:
-    1. Training with captions. All captions must include the token string. The token string is replaced with multiple tokens.
-    2. Use `--use_object_template` or `--use_style_template` option. The captions are generated from the template. The existing captions are ignored.
-  - See below for the format of the embeddings.
-
-- `--min_timestep` and `--max_timestep` options are added to each training script. These options can be used to train U-Net with different timesteps. The default values are 0 and 1000.
-
-### Utility scripts for SDXL
-
-- `tools/cache_latents.py` is added. This script can be used to cache the latents to disk in advance. 
-  - The options are almost the same as `sdxl_train.py'. See the help message for the usage.
-  - Please launch the script as follows:
-    `accelerate launch  --num_cpu_threads_per_process 1 tools/cache_latents.py ...`
-  - This script should work with multi-GPU, but it is not tested in my environment.
-
-- `tools/cache_text_encoder_outputs.py` is added. This script can be used to cache the text encoder outputs to disk in advance. 
-  - The options are almost the same as `cache_latents.py` and `sdxl_train.py`. See the help message for the usage.
-
-- `sdxl_gen_img.py` is added. This script can be used to generate images with SDXL, including LoRA, Textual Inversion and ControlNet-LLLite. See the help message for the usage.
-
-### Tips for SDXL training
-
-- The default resolution of SDXL is 1024x1024.
-- The fine-tuning can be done with 24GB GPU memory with the batch size of 1. For 24GB GPU, the following options are recommended __for the fine-tuning with 24GB GPU memory__:
-  - Train U-Net only.
-  - Use gradient checkpointing.
-  - Use `--cache_text_encoder_outputs` option and caching latents.
-  - Use Adafactor optimizer. RMSprop 8bit or Adagrad 8bit may work. AdamW 8bit doesn't seem to work.
-- The LoRA training can be done with 8GB GPU memory (10GB recommended). For reducing the GPU memory usage, the following options are recommended:
-  - Train U-Net only.
-  - Use gradient checkpointing.
-  - Use `--cache_text_encoder_outputs` option and caching latents.
-  - Use one of 8bit optimizers or Adafactor optimizer.
-  - Use lower dim (4 to 8 for 8GB GPU).
-- `--network_train_unet_only` option is highly recommended for SDXL LoRA. Because SDXL has two text encoders, the result of the training will be unexpected.
-- PyTorch 2 seems to use slightly less GPU memory than PyTorch 1.
-- `--bucket_reso_steps` can be set to 32 instead of the default value 64. Smaller values than 32 will not work for SDXL training.
-
-Example of the optimizer settings for Adafactor with the fixed learning rate:
-```toml
-optimizer_type = "adafactor"
-optimizer_args = [ "scale_parameter=False", "relative_step=False", "warmup_init=False" ]
-lr_scheduler = "constant_with_warmup"
-lr_warmup_steps = 100
-learning_rate = 4e-7 # SDXL original learning rate
-```
-
-### Format of Textual Inversion embeddings for SDXL
-
-```python
-from safetensors.torch import save_file
-
-state_dict = {"clip_g": embs_for_text_encoder_1280, "clip_l": embs_for_text_encoder_768}
-save_file(state_dict, file)
-```
-
-### ControlNet-LLLite
-
-ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [documentation](./docs/train_lllite_README.md) for details.
-
-
-## Change History
-
-### Jan 15, 2024 / 2024/1/15: v0.8.0
-
-- Diffusers, Accelerate, Transformers and other related libraries have been updated. Please update the libraries with [Upgrade](#upgrade).
-  - Some model files (Text Encoder without position_id) based on the latest Transformers can be loaded.
-- `torch.compile` is supported (experimental). PR [#1024](https://github.com/kohya-ss/sd-scripts/pull/1024) Thanks to p1atdev!
-  - This feature works only on Linux or WSL.
-  - Please specify `--torch_compile` option in each training script.
-  - You can select the backend with `--dynamo_backend` option. The default is `"inductor"`. `inductor` or `eager` seems to work.
-  - Please use `--spda` option instead of `--xformers` option.
-  - PyTorch 2.1 or later is recommended.
-  - Please see [PR](https://github.com/kohya-ss/sd-scripts/pull/1024) for details.
-- The session name for wandb can be specified with `--wandb_run_name` option. PR [#1032](https://github.com/kohya-ss/sd-scripts/pull/1032) Thanks to hopl1t!
-- IPEX library is updated. PR [#1030](https://github.com/kohya-ss/sd-scripts/pull/1030) Thanks to Disty0!
-- Fixed a bug that Diffusers format model cannot be saved.
-
-- Diffusers、Accelerate、Transformers 等の関連ライブラリを更新しました。[Upgrade](#upgrade) を参照し更新をお願いします。
-  - 最新の Transformers を前提とした一部のモデルファイル（Text Encoder が position_id を持たないもの）が読み込めるようになりました。
-- `torch.compile` がサポートされしました（実験的）。 PR [#1024](https://github.com/kohya-ss/sd-scripts/pull/1024) p1atdev 氏に感謝します。
-  - Linux または WSL でのみ動作します。
-  - 各学習スクリプトで `--torch_compile` オプションを指定してください。
-  - `--dynamo_backend` オプションで使用される backend を選択できます。デフォルトは `"inductor"` です。 `inductor` または `eager` が動作するようです。
-  - `--xformers` オプションとは互換性がありません。 代わりに `--spda` オプションを使用してください。
-  - PyTorch 2.1以降を推奨します。
-  - 詳細は [PR](https://github.com/kohya-ss/sd-scripts/pull/1024) をご覧ください。
-- wandb 保存時のセッション名が各学習スクリプトの `--wandb_run_name` オプションで指定できるようになりました。 PR [#1032](https://github.com/kohya-ss/sd-scripts/pull/1032) hopl1t 氏に感謝します。
-- IPEX ライブラリが更新されました。[PR #1030](https://github.com/kohya-ss/sd-scripts/pull/1030) Disty0 氏に感謝します。
-- Diffusers 形式でのモデル保存ができなくなっていた不具合を修正しました。
-
-
-Please read [Releases](https://github.com/kohya-ss/sd-scripts/releases) for recent updates.
-最近の更新情報は [Release](https://github.com/kohya-ss/sd-scripts/releases) をご覧ください。
-
-### Naming of LoRA
-
-The LoRA supported by `train_network.py` has been named to avoid confusion. The documentation has been updated. The following are the names of LoRA types in this repository.
-
-1. __LoRA-LierLa__ : (LoRA for __Li__ n __e__ a __r__  __La__ yers)
-
-    LoRA for Linear layers and Conv2d layers with 1x1 kernel
-
-2. __LoRA-C3Lier__ : (LoRA for __C__ olutional layers with __3__ x3 Kernel and  __Li__ n __e__ a __r__ layers)
-
-    In addition to 1., LoRA for Conv2d layers with 3x3 kernel 
-    
-LoRA-LierLa is the default LoRA type for `train_network.py` (without `conv_dim` network arg). LoRA-LierLa can be used with [our extension](https://github.com/kohya-ss/sd-webui-additional-networks) for AUTOMATIC1111's Web UI, or with the built-in LoRA feature of the Web UI.
-
-To use LoRA-C3Lier with Web UI, please use our extension.
-
-### LoRAの名称について
-
-`train_network.py` がサポートするLoRAについて、混乱を避けるため名前を付けました。ドキュメントは更新済みです。以下は当リポジトリ内の独自の名称です。
-
-1. __LoRA-LierLa__ : (LoRA for __Li__ n __e__ a __r__  __La__ yers、リエラと読みます)
-
-    Linear 層およびカーネルサイズ 1x1 の Conv2d 層に適用されるLoRA
-
-2. __LoRA-C3Lier__ : (LoRA for __C__ olutional layers with __3__ x3 Kernel and  __Li__ n __e__ a __r__ layers、セリアと読みます)
-
-    1.に加え、カーネルサイズ 3x3 の Conv2d 層に適用されるLoRA
-
-LoRA-LierLa は[Web UI向け拡張](https://github.com/kohya-ss/sd-webui-additional-networks)、またはAUTOMATIC1111氏のWeb UIのLoRA機能で使用することができます。
-
-LoRA-C3Lierを使いWeb UIで生成するには拡張を使用してください。
-
-## Sample image generation during training
-  A prompt file might look like this, for example
-
-```
-# prompt 1
-masterpiece, best quality, (1girl), in white shirts, upper body, looking at viewer, simple background --n low quality, worst quality, bad anatomy,bad composition, poor, low effort --w 768 --h 768 --d 1 --l 7.5 --s 28
-
-# prompt 2
-masterpiece, best quality, 1boy, in business suit, standing at street, looking back --n (low quality, worst quality), bad anatomy,bad composition, poor, low effort --w 576 --h 832 --d 2 --l 5.5 --s 40
-```
-
-  Lines beginning with `#` are comments. You can specify options for the generated image with options like `--n` after the prompt. The following can be used.
-
-  * `--n` Negative prompt up to the next option.
-  * `--w` Specifies the width of the generated image.
-  * `--h` Specifies the height of the generated image.
-  * `--d` Specifies the seed of the generated image.
-  * `--l` Specifies the CFG scale of the generated image.
-  * `--s` Specifies the number of steps in the generation.
-
-  The prompt weighting such as `( )` and `[ ]` are working.
-
-## サンプル画像生成
-プロンプトファイルは例えば以下のようになります。
-
-```
-# prompt 1
-masterpiece, best quality, (1girl), in white shirts, upper body, looking at viewer, simple background --n low quality, worst quality, bad anatomy,bad composition, poor, low effort --w 768 --h 768 --d 1 --l 7.5 --s 28
-
-# prompt 2
-masterpiece, best quality, 1boy, in business suit, standing at street, looking back --n (low quality, worst quality), bad anatomy,bad composition, poor, low effort --w 576 --h 832 --d 2 --l 5.5 --s 40
-```
-
-  `#` で始まる行はコメントになります。`--n` のように「ハイフン二個＋英小文字」の形でオプションを指定できます。以下が使用可能できます。
-
-  * `--n` Negative prompt up to the next option.
-  * `--w` Specifies the width of the generated image.
-  * `--h` Specifies the height of the generated image.
-  * `--d` Specifies the seed of the generated image.
-  * `--l` Specifies the CFG scale of the generated image.
-  * `--s` Specifies the number of steps in the generation.
-
-  `( )` や `[ ]` などの重みづけも動作します。
-

XTI_hijack.py

@@ -1,11 +1,7 @@
 import torch
-try:
-    import intel_extension_for_pytorch as ipex
-    if torch.xpu.is_available():
-        from library.ipex import ipex_init
-        ipex_init()
-except Exception:
-    pass
+from library.device_utils import init_ipex
+init_ipex()
+
 from typing import Union, List, Optional, Dict, Any, Tuple
 from diffusers.models.unet_2d_condition import UNet2DConditionOutput
 

_typos.toml

@@ -2,6 +2,7 @@
 # Instruction:  https://github.com/marketplace/actions/typos-action#getting-started
 
 [default.extend-identifiers]
+ddPn08="ddPn08"
 
 [default.extend-words]
 NIN="NIN"
@@ -27,7 +28,10 @@ rik="rik"
 koo="koo"
 yos="yos"
 wn="wn"
-
+hime="hime"
+OT="OT"
+byt="byt"
+tak="tak"
 
 [files]
 extend-exclude = ["_typos.toml", "venv"]

docs/config_README-en.md

@@ -0,0 +1,386 @@
+First version: A.I Translation by Model: NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO, editing by Darkstorm2150
+
+Document is updated and maintained manually.
+
+# Config Readme
+
+This README is about the configuration files that can be passed with the `--dataset_config` option.
+
+## Overview
+
+By passing a configuration file, users can make detailed settings.
+
+* Multiple datasets can be configured
+   * For example, by setting `resolution` for each dataset, they can be mixed and trained.
+   * In training methods that support both the DreamBooth approach and the fine-tuning approach, datasets of the DreamBooth method and the fine-tuning method can be mixed.
+* Settings can be changed for each subset
+   * A subset is a partition of the dataset by image directory or metadata. Several subsets make up a dataset.
+   * Options such as `keep_tokens` and `flip_aug` can be set for each subset. On the other hand, options such as `resolution` and `batch_size` can be set for each dataset, and their values are common among subsets belonging to the same dataset. More details will be provided later.
+
+The configuration file format can be JSON or TOML. Considering the ease of writing, it is recommended to use [TOML](https://toml.io/ja/v1.0.0-rc.2). The following explanation assumes the use of TOML.
+
+
+Here is an example of a configuration file written in TOML.
+
+```toml
+[general]
+shuffle_caption = true
+caption_extension = '.txt'
+keep_tokens = 1
+
+# This is a DreamBooth-style dataset
+[[datasets]]
+resolution = 512
+batch_size = 4
+keep_tokens = 2
+
+  [[datasets.subsets]]
+  image_dir = 'C:\hoge'
+  class_tokens = 'hoge girl'
+  # This subset uses keep_tokens = 2 (the value of the parent datasets)
+
+  [[datasets.subsets]]
+  image_dir = 'C:\fuga'
+  class_tokens = 'fuga boy'
+  keep_tokens = 3
+
+  [[datasets.subsets]]
+  is_reg = true
+  image_dir = 'C:\reg'
+  class_tokens = 'human'
+  keep_tokens = 1
+
+# This is a fine-tuning dataset
+[[datasets]]
+resolution = [768, 768]
+batch_size = 2
+
+  [[datasets.subsets]]
+  image_dir = 'C:\piyo'
+  metadata_file = 'C:\piyo\piyo_md.json'
+  # This subset uses keep_tokens = 1 (the value of [general])
+```
+
+In this example, three directories are trained as a DreamBooth-style dataset at 512x512 (batch size 4), and one directory is trained as a fine-tuning dataset at 768x768 (batch size 2).
+
+## Settings for datasets and subsets
+
+Settings for datasets and subsets are divided into several registration locations.
+
+* `[general]`
+    * This is where options that apply to all datasets or all subsets are specified.
+    * If there are options with the same name in the dataset-specific or subset-specific settings, the dataset-specific or subset-specific settings take precedence.
+* `[[datasets]]`
+    * `datasets` is where settings for datasets are registered. This is where options that apply individually to each dataset are specified.
+	* If there are subset-specific settings, the subset-specific settings take precedence.
+* `[[datasets.subsets]]`
+    * `datasets.subsets` is where settings for subsets are registered. This is where options that apply individually to each subset are specified.
+
+Here is an image showing the correspondence between image directories and registration locations in the previous example.
+
+```
+C:\
+├─ hoge  ->  [[datasets.subsets]] No.1  ┐                        ┐
+├─ fuga  ->  [[datasets.subsets]] No.2  |->  [[datasets]] No.1   |->  [general]
+├─ reg   ->  [[datasets.subsets]] No.3  ┘                        |
+└─ piyo  ->  [[datasets.subsets]] No.4  -->  [[datasets]] No.2   ┘
+```
+
+The image directory corresponds to each `[[datasets.subsets]]`. Then, multiple `[[datasets.subsets]]` are combined to form one `[[datasets]]`. All `[[datasets]]` and `[[datasets.subsets]]` belong to `[general]`.
+
+The available options for each registration location may differ, but if the same option is specified, the value in the lower registration location will take precedence. You can check how the `keep_tokens` option is handled in the previous example for better understanding.
+
+Additionally, the available options may vary depending on the method that the learning approach supports.
+
+* Options specific to the DreamBooth method
+* Options specific to the fine-tuning method
+* Options available when using the caption dropout technique
+
+When using both the DreamBooth method and the fine-tuning method, they can be used together with a learning approach that supports both.
+When using them together, a point to note is that the method is determined based on the dataset, so it is not possible to mix DreamBooth method subsets and fine-tuning method subsets within the same dataset.
+In other words, if you want to use both methods together, you need to set up subsets of different methods belonging to different datasets.
+
+In terms of program behavior, if the `metadata_file` option exists, it is determined to be a subset of fine-tuning. Therefore, for subsets belonging to the same dataset, as long as they are either "all have the `metadata_file` option" or "all have no `metadata_file` option," there is no problem.
+
+Below, the available options will be explained. For options with the same name as the command-line argument, the explanation will be omitted in principle. Please refer to other READMEs.
+
+### Common options for all learning methods
+
+These are options that can be specified regardless of the learning method.
+
+#### Data set specific options
+
+These are options related to the configuration of the data set. They cannot be described in `datasets.subsets`.
+
+
+| Option Name | Example Setting | `[general]` | `[[datasets]]` |
+| ---- | ---- | ---- | ---- |
+| `batch_size` | `1` | o | o |
+| `bucket_no_upscale` | `true` | o | o |
+| `bucket_reso_steps` | `64` | o | o |
+| `enable_bucket` | `true` | o | o |
+| `max_bucket_reso` | `1024` | o | o |
+| `min_bucket_reso` | `128` | o | o |
+| `resolution` | `256`, `[512, 512]` | o | o |
+
+* `batch_size`
+    * This corresponds to the command-line argument `--train_batch_size`.
+* `max_bucket_reso`, `min_bucket_reso`
+    * Specify the maximum and minimum resolutions of the bucket. It must be divisible by `bucket_reso_steps`.
+
+These settings are fixed per dataset. That means that subsets belonging to the same dataset will share these settings. For example, if you want to prepare datasets with different resolutions, you can define them as separate datasets as shown in the example above, and set different resolutions for each.
+
+#### Options for Subsets
+
+These options are related to subset configuration.
+
+| Option Name | Example | `[general]` | `[[datasets]]` | `[[dataset.subsets]]` |
+| ---- | ---- | ---- | ---- | ---- |
+| `color_aug` | `false` | o | o | o |
+| `face_crop_aug_range` | `[1.0, 3.0]` | o | o | o |
+| `flip_aug` | `true` | o | o | o |
+| `keep_tokens` | `2` | o | o | o |
+| `num_repeats` | `10` | o | o | o |
+| `random_crop` | `false` | o | o | o |
+| `shuffle_caption` | `true` | o | o | o |
+| `caption_prefix` | `"masterpiece, best quality, "` | o | o | o |
+| `caption_suffix` | `", from side"` | o | o | o |
+| `caption_separator` |  (not specified) | o | o | o |
+| `keep_tokens_separator` | `“|||”` | o | o | o |
+| `secondary_separator` | `“;;;”` | o | o | o |
+| `enable_wildcard` | `true` | o | o | o |
+| `resize_interpolation` | (not specified) | o | o | o |
+
+* `num_repeats`
+    * Specifies the number of repeats for images in a subset. This is equivalent to `--dataset_repeats` in fine-tuning but can be specified for any training method.
+* `caption_prefix`, `caption_suffix`
+    * Specifies the prefix and suffix strings to be appended to the captions. Shuffling is performed with these strings included. Be cautious when using `keep_tokens`.
+* `caption_separator`
+    * Specifies the string to separate the tags. The default is `,`. This option is usually not necessary to set.
+* `keep_tokens_separator`
+    * Specifies the string to separate the parts to be fixed in the caption. For example, if you specify `aaa, bbb ||| ccc, ddd, eee, fff ||| ggg, hhh`, the parts `aaa, bbb` and `ggg, hhh` will remain, and the rest will be shuffled and dropped. The comma in between is not necessary. As a result, the prompt will be `aaa, bbb, eee, ccc, fff, ggg, hhh` or `aaa, bbb, fff, ccc, eee, ggg, hhh`, etc.
+* `secondary_separator`
+    * Specifies an additional separator. The part separated by this separator is treated as one tag and is shuffled and dropped. It is then replaced by `caption_separator`. For example, if you specify `aaa;;;bbb;;;ccc`, it will be replaced by `aaa,bbb,ccc` or dropped together.
+* `enable_wildcard`
+    * Enables wildcard notation. This will be explained later.
+* `resize_interpolation`
+    * Specifies the interpolation method used when resizing images. Normally, there is no need to specify this. The following options can be specified: `lanczos`, `nearest`, `bilinear`, `linear`, `bicubic`, `cubic`, `area`, `box`. By default (when not specified), `area` is used for downscaling, and `lanczos` is used for upscaling. If this option is specified, the same interpolation method will be used for both upscaling and downscaling. When `lanczos` or `box` is specified, PIL is used; for other options, OpenCV is used.
+
+### DreamBooth-specific options
+
+DreamBooth-specific options only exist as subsets-specific options.
+
+#### Subset-specific options
+
+Options related to the configuration of DreamBooth subsets.
+
+| Option Name | Example Setting | `[general]` | `[[datasets]]` | `[[dataset.subsets]]` |
+| ---- | ---- | ---- | ---- | ---- |
+| `image_dir` | `'C:\hoge'` | - | - | o (required) |
+| `caption_extension` | `".txt"` | o | o | o |
+| `class_tokens` | `"sks girl"` | - | - | o |
+| `cache_info` | `false` | o | o | o |
+| `is_reg` | `false` | - | - | o |
+
+Firstly, note that for `image_dir`, the path to the image files must be specified as being directly in the directory. Unlike the previous DreamBooth method, where images had to be placed in subdirectories, this is not compatible with that specification. Also, even if you name the folder something like "5_cat", the number of repeats of the image and the class name will not be reflected. If you want to set these individually, you will need to explicitly specify them using `num_repeats` and `class_tokens`.
+
+* `image_dir`
+    * Specifies the path to the image directory. This is a required option.
+    * Images must be placed directly under the directory.
+* `class_tokens`
+    * Sets the class tokens.
+    * Only used during training when a corresponding caption file does not exist. The determination of whether or not to use it is made on a per-image basis. If `class_tokens` is not specified and a caption file is not found, an error will occur.
+* `cache_info`
+    * Specifies whether to cache the image size and caption. If not specified, it is set to `false`. The cache is saved in `metadata_cache.json` in `image_dir`.
+    * Caching speeds up the loading of the dataset after the first time. It is effective when dealing with thousands of images or more.
+* `is_reg`
+    * Specifies whether the subset images are for normalization. If not specified, it is set to `false`, meaning that the images are not for normalization.
+
+### Fine-tuning method specific options
+
+The options for the fine-tuning method only exist for subset-specific options.
+
+#### Subset-specific options
+
+These options are related to the configuration of the fine-tuning method's subsets.
+
+| Option name | Example setting | `[general]` | `[[datasets]]` | `[[dataset.subsets]]` |
+| ---- | ---- | ---- | ---- | ---- |
+| `image_dir` | `'C:\hoge'` | - | - | o |
+| `metadata_file` | `'C:\piyo\piyo_md.json'` | - | - | o (required) |
+
+* `image_dir`
+    * Specify the path to the image directory. Unlike the DreamBooth method, specifying it is not mandatory, but it is recommended to do so.
+        * The case where it is not necessary to specify is when the `--full_path` is added to the command line when generating the metadata file.
+    * The images must be placed directly under the directory.
+* `metadata_file`
+    * Specify the path to the metadata file used for the subset. This is a required option.
+        * It is equivalent to the command-line argument `--in_json`.
+    * Due to the specification that a metadata file must be specified for each subset, it is recommended to avoid creating a metadata file with images from different directories as a single metadata file. It is strongly recommended to prepare a separate metadata file for each image directory and register them as separate subsets.
+
+### Options available when caption dropout method can be used
+
+The options available when the caption dropout method can be used exist only for subsets. Regardless of whether it's the DreamBooth method or fine-tuning method, if it supports caption dropout, it can be specified.
+
+#### Subset-specific options
+
+Options related to the setting of subsets that caption dropout can be used for.
+
+| Option Name | `[general]` | `[[datasets]]` | `[[dataset.subsets]]` |
+| ---- | ---- | ---- | ---- |
+| `caption_dropout_every_n_epochs` | o | o | o |
+| `caption_dropout_rate` | o | o | o |
+| `caption_tag_dropout_rate` | o | o | o |
+
+## Behavior when there are duplicate subsets
+
+In the case of the DreamBooth dataset, if there are multiple `image_dir` directories with the same content, they are considered to be duplicate subsets. For the fine-tuning dataset, if there are multiple `metadata_file` files with the same content, they are considered to be duplicate subsets. If duplicate subsets exist in the dataset, subsequent subsets will be ignored.
+
+However, if they belong to different datasets, they are not considered duplicates. For example, if you have subsets with the same `image_dir` in different datasets, they will not be considered duplicates. This is useful when you want to train with the same image but with different resolutions.
+
+```toml
+# If data sets exist separately, they are not considered duplicates and are both used for training.
+
+[[datasets]]
+resolution = 512
+
+  [[datasets.subsets]]
+  image_dir = 'C:\hoge'
+
+[[datasets]]
+resolution = 768
+
+  [[datasets.subsets]]
+  image_dir = 'C:\hoge'
+```
+
+## Command Line Argument and Configuration File
+
+There are options in the configuration file that have overlapping roles with command line argument options.
+
+The following command line argument options are ignored if a configuration file is passed:
+
+* `--train_data_dir`
+* `--reg_data_dir`
+* `--in_json`
+
+For the command line options listed below, if an option is specified in both the command line arguments and the configuration file, the value from the configuration file will be given priority. Unless otherwise noted, the option names are the same.
+
+| Command Line Argument Option   | Corresponding Configuration File Option |
+| ------------------------------- | --------------------------------------- |
+| `--bucket_no_upscale`           |                                       |
+| `--bucket_reso_steps`           |                                       |
+| `--caption_dropout_every_n_epochs` |                                       |
+| `--caption_dropout_rate`        |                                       |
+| `--caption_extension`           |                                       |
+| `--caption_tag_dropout_rate`    |                                       |
+| `--color_aug`                   |                                       |
+| `--dataset_repeats`             | `num_repeats`                          |
+| `--enable_bucket`               |                                       |
+| `--face_crop_aug_range`         |                                       |
+| `--flip_aug`                    |                                       |
+| `--keep_tokens`                 |                                       |
+| `--min_bucket_reso`              |                                       |
+| `--random_crop`                 |                                       |
+| `--resolution`                  |                                       |
+| `--shuffle_caption`             |                                       |
+| `--train_batch_size`            | `batch_size`                           |
+
+## Error Guide
+
+Currently, we are using an external library to check if the configuration file is written correctly, but the development has not been completed, and there is a problem that the error message is not clear. In the future, we plan to improve this problem.
+
+As a temporary measure, we will list common errors and their solutions. If you encounter an error even though it should be correct or if the error content is not understandable, please contact us as it may be a bug.
+
+* `voluptuous.error.MultipleInvalid: required key not provided @ ...`: This error occurs when a required option is not provided. It is highly likely that you forgot to specify the option or misspelled the option name.
+  * The error location is indicated by `...` in the error message. For example, if you encounter an error like `voluptuous.error.MultipleInvalid: required key not provided @ data['datasets'][0]['subsets'][0]['image_dir']`, it means that the `image_dir` option does not exist in the 0th `subsets` of the 0th `datasets` setting.
+* `voluptuous.error.MultipleInvalid: expected int for dictionary value @ ...`: This error occurs when the specified value format is incorrect. It is highly likely that the value format is incorrect. The `int` part changes depending on the target option. The example configurations in this README may be helpful.
+* `voluptuous.error.MultipleInvalid: extra keys not allowed @ ...`: This error occurs when there is an option name that is not supported. It is highly likely that you misspelled the option name or mistakenly included it.
+
+## Miscellaneous
+
+### Multi-line captions
+
+By setting `enable_wildcard = true`, multiple-line captions are also enabled. If the caption file consists of multiple lines, one line is randomly selected as the caption. 
+
+```txt
+1girl, hatsune miku, vocaloid, upper body, looking at viewer, microphone, stage
+a girl with a microphone standing on a stage
+detailed digital art of a girl with a microphone on a stage
+```
+
+It can be combined with wildcard notation.
+
+In metadata files, you can also specify multiple-line captions. In the `.json` metadata file, use `\n` to represent a line break. If the caption file consists of multiple lines, `merge_captions_to_metadata.py` will create a metadata file in this format.
+
+The tags in the metadata (`tags`) are added to each line of the caption.
+
+```json
+{
+    "/path/to/image.png": {
+        "caption": "a cartoon of a frog with the word frog on it\ntest multiline caption1\ntest multiline caption2",
+        "tags": "open mouth, simple background, standing, no humans, animal, black background, frog, animal costume, animal focus"
+    },
+    ...
+}
+```
+
+In this case, the actual caption will be `a cartoon of a frog with the word frog on it, open mouth, simple background ...`, `test multiline caption1, open mouth, simple background ...`, `test multiline caption2, open mouth, simple background ...`, etc.
+
+### Example of configuration file : `secondary_separator`, wildcard notation, `keep_tokens_separator`, etc.
+
+```toml
+[general]
+flip_aug = true
+color_aug = false
+resolution = [1024, 1024]
+
+[[datasets]]
+batch_size = 6
+enable_bucket = true
+bucket_no_upscale = true
+caption_extension = ".txt"
+keep_tokens_separator= "|||"
+shuffle_caption = true
+caption_tag_dropout_rate = 0.1
+secondary_separator = ";;;" # subset 側に書くこともできます / can be written in the subset side
+enable_wildcard = true # 同上 / same as above
+
+  [[datasets.subsets]]
+  image_dir = "/path/to/image_dir"
+  num_repeats = 1
+
+  # ||| の前後はカンマは不要です（自動的に追加されます） / No comma is required before and after ||| (it is added automatically)
+  caption_prefix = "1girl, hatsune miku, vocaloid |||" 
+  
+  # ||| の後はシャッフル、drop されず残ります / After |||, it is not shuffled or dropped and remains
+  # 単純に文字列として連結されるので、カンマなどは自分で入れる必要があります / It is simply concatenated as a string, so you need to put commas yourself
+  caption_suffix = ", anime screencap ||| masterpiece, rating: general"
+```
+
+### Example of caption, secondary_separator notation: `secondary_separator = ";;;"`
+
+```txt
+1girl, hatsune miku, vocaloid, upper body, looking at viewer, sky;;;cloud;;;day, outdoors
+```
+The part `sky;;;cloud;;;day` is replaced with `sky,cloud,day` without shuffling or dropping. When shuffling and dropping are enabled, it is processed as a whole (as one tag). For example, it becomes `vocaloid, 1girl, upper body, sky,cloud,day, outdoors, hatsune miku` (shuffled) or `vocaloid, 1girl, outdoors, looking at viewer, upper body, hatsune miku` (dropped).
+
+### Example of caption, enable_wildcard notation: `enable_wildcard = true`
+
+```txt
+1girl, hatsune miku, vocaloid, upper body, looking at viewer, {simple|white} background
+```
+`simple` or `white` is randomly selected, and it becomes `simple background` or `white background`.
+
+```txt
+1girl, hatsune miku, vocaloid, {{retro style}}
+```
+If you want to include `{` or `}` in the tag string, double them like `{{` or `}}` (in this example, the actual caption used for training is `{retro style}`).
+
+### Example of caption, `keep_tokens_separator` notation: `keep_tokens_separator = "|||"`
+
+```txt
+1girl, hatsune miku, vocaloid ||| stage, microphone, white shirt, smile ||| best quality, rating: general
+```
+It becomes `1girl, hatsune miku, vocaloid, microphone, stage, white shirt, best quality, rating: general` or `1girl, hatsune miku, vocaloid, white shirt, smile, stage, microphone, best quality, rating: general` etc.
+

docs/config_README-ja.md

@@ -1,5 +1,3 @@
-For non-Japanese speakers: this README is provided only in Japanese in the current state. Sorry for inconvenience. We will provide English version in the near future.
-
 `--dataset_config` で渡すことができる設定ファイルに関する説明です。
 
 ## 概要
@@ -120,6 +118,8 @@ DreamBooth の手法と fine tuning の手法の両方とも利用可能な学
 
 * `batch_size`
     * コマンドライン引数の `--train_batch_size` と同等です。
+* `max_bucket_reso`, `min_bucket_reso`
+    * bucketの最大、最小解像度を指定します。`bucket_reso_steps` で割り切れる必要があります。
 
 これらの設定はデータセットごとに固定です。
 つまり、データセットに所属するサブセットはこれらの設定を共有することになります。
@@ -140,12 +140,32 @@ DreamBooth の手法と fine tuning の手法の両方とも利用可能な学
 | `shuffle_caption` | `true` | o | o | o |
 | `caption_prefix` | `“masterpiece, best quality, ”` | o | o | o |
 | `caption_suffix` | `“, from side”` | o | o | o |
+| `caption_separator` | （通常は設定しません） | o | o | o |
+| `keep_tokens_separator` | `“|||”` | o | o | o |
+| `secondary_separator` | `“;;;”` | o | o | o |
+| `enable_wildcard` | `true` | o | o | o |
+| `resize_interpolation` |（通常は設定しません） | o | o | o |
 
 * `num_repeats`
     * サブセットの画像の繰り返し回数を指定します。fine tuning における `--dataset_repeats` に相当しますが、`num_repeats` はどの学習方法でも指定可能です。
 * `caption_prefix`, `caption_suffix`
     * キャプションの前、後に付与する文字列を指定します。シャッフルはこれらの文字列を含めた状態で行われます。`keep_tokens` を指定する場合には注意してください。
 
+* `caption_separator`
+    * タグを区切る文字列を指定します。デフォルトは `,` です。このオプションは通常は設定する必要はありません。
+
+* `keep_tokens_separator`
+    *  キャプションで固定したい部分を区切る文字列を指定します。たとえば `aaa, bbb ||| ccc, ddd, eee, fff ||| ggg, hhh` のように指定すると、`aaa, bbb` と `ggg, hhh` の部分はシャッフル、drop されず残ります。間のカンマは不要です。結果としてプロンプトは `aaa, bbb, eee, ccc, fff, ggg, hhh` や `aaa, bbb, fff, ccc, eee, ggg, hhh` などになります。
+
+* `secondary_separator`
+    * 追加の区切り文字を指定します。この区切り文字で区切られた部分は一つのタグとして扱われ、シャッフル、drop されます。その後、`caption_separator` に置き換えられます。たとえば `aaa;;;bbb;;;ccc` のように指定すると、`aaa,bbb,ccc` に置き換えられるか、まとめて drop されます。
+
+* `enable_wildcard`
+    * ワイルドカード記法および複数行キャプションを有効にします。ワイルドカード記法、複数行キャプションについては後述します。
+
+* `resize_interpolation`
+    * 画像のリサイズ時に使用する補間方法を指定します。通常は指定しなくて構いません。`lanczos`, `nearest`, `bilinear`, `linear`, `bicubic`, `cubic`, `area`, `box` が指定可能です。デフォルト（未指定時）は、縮小時は `area`、拡大時は `lanczos` になります。このオプションを指定すると、拡大時・縮小時とも同じ補間方法が使用されます。`lanczos`、`box`を指定するとPILが、それ以外を指定するとOpenCVが使用されます。
+
 ### DreamBooth 方式専用のオプション
 
 DreamBooth 方式のオプションは、サブセット向けオプションのみ存在します。
@@ -159,6 +179,7 @@ DreamBooth 方式のサブセットの設定に関わるオプションです。
 | `image_dir` | `‘C:\hoge’` | - | - | o（必須） |
 | `caption_extension` | `".txt"` | o | o | o |
 | `class_tokens` | `“sks girl”` | - | - | o |
+| `cache_info` | `false` | o | o | o | 
 | `is_reg` | `false` | - | - | o |
 
 まず注意点として、 `image_dir` には画像ファイルが直下に置かれているパスを指定する必要があります。従来の DreamBooth の手法ではサブディレクトリに画像を置く必要がありましたが、そちらとは仕様に互換性がありません。また、`5_cat` のようなフォルダ名にしても、画像の繰り返し回数とクラス名は反映されません。これらを個別に設定したい場合、`num_repeats` と `class_tokens` で明示的に指定する必要があることに注意してください。
@@ -169,6 +190,9 @@ DreamBooth 方式のサブセットの設定に関わるオプションです。
 * `class_tokens`
     * クラストークンを設定します。
     * 画像に対応する caption ファイルが存在しない場合にのみ学習時に利用されます。利用するかどうかの判定は画像ごとに行います。`class_tokens` を指定しなかった場合に caption ファイルも見つからなかった場合にはエラーになります。
+* `cache_info`
+    * 画像サイズ、キャプションをキャッシュするかどうかを指定します。指定しなかった場合は `false` になります。キャッシュは `image_dir` に `metadata_cache.json` というファイル名で保存されます。
+    * キャッシュを行うと、二回目以降のデータセット読み込みが高速化されます。数千枚以上の画像を扱う場合には有効です。
 * `is_reg`
     * サブセットの画像が正規化用かどうかを指定します。指定しなかった場合は `false` として、つまり正規化画像ではないとして扱います。
 
@@ -280,4 +304,89 @@ resolution = 768
 * `voluptuous.error.MultipleInvalid: expected int for dictionary value @ ...`: 指定する値の形式が不正というエラーです。値の形式が間違っている可能性が高いです。`int` の部分は対象となるオプションによって変わります。この README に載っているオプションの「設定例」が役立つかもしれません。
 * `voluptuous.error.MultipleInvalid: extra keys not allowed @ ...`: 対応していないオプション名が存在している場合に発生するエラーです。オプション名を間違って記述しているか、誤って紛れ込んでいる可能性が高いです。
 
+## その他
 
+### 複数行キャプション
+
+`enable_wildcard = true` を設定することで、複数行キャプションも同時に有効になります。キャプションファイルが複数の行からなる場合、ランダムに一つの行が選ばれてキャプションとして利用されます。
+
+```txt
+1girl, hatsune miku, vocaloid, upper body, looking at viewer, microphone, stage
+a girl with a microphone standing on a stage
+detailed digital art of a girl with a microphone on a stage
+```
+
+ワイルドカード記法と組み合わせることも可能です。
+
+メタデータファイルでも同様に複数行キャプションを指定することができます。メタデータの .json 内には、`\n` を使って改行を表現してください。キャプションファイルが複数行からなる場合、`merge_captions_to_metadata.py` を使うと、この形式でメタデータファイルが作成されます。
+
+メタデータのタグ (`tags`) は、キャプションの各行に追加されます。
+
+```json
+{
+    "/path/to/image.png": {
+        "caption": "a cartoon of a frog with the word frog on it\ntest multiline caption1\ntest multiline caption2",
+        "tags": "open mouth, simple background, standing, no humans, animal, black background, frog, animal costume, animal focus"
+    },
+    ...
+}
+```
+
+この場合、実際のキャプションは `a cartoon of a frog with the word frog on it, open mouth, simple background ...` または `test multiline caption1, open mouth, simple background ...`、 `test multiline caption2, open mouth, simple background ...` 等になります。
+
+### 設定ファイルの記述例：追加の区切り文字、ワイルドカード記法、`keep_tokens_separator` 等
+
+```toml
+[general]
+flip_aug = true
+color_aug = false
+resolution = [1024, 1024]
+
+[[datasets]]
+batch_size = 6
+enable_bucket = true
+bucket_no_upscale = true
+caption_extension = ".txt"
+keep_tokens_separator= "|||"
+shuffle_caption = true
+caption_tag_dropout_rate = 0.1
+secondary_separator = ";;;" # subset 側に書くこともできます / can be written in the subset side
+enable_wildcard = true # 同上 / same as above
+
+  [[datasets.subsets]]
+  image_dir = "/path/to/image_dir"
+  num_repeats = 1
+
+  # ||| の前後はカンマは不要です（自動的に追加されます） / No comma is required before and after ||| (it is added automatically)
+  caption_prefix = "1girl, hatsune miku, vocaloid |||" 
+  
+  # ||| の後はシャッフル、drop されず残ります / After |||, it is not shuffled or dropped and remains
+  # 単純に文字列として連結されるので、カンマなどは自分で入れる必要があります / It is simply concatenated as a string, so you need to put commas yourself
+  caption_suffix = ", anime screencap ||| masterpiece, rating: general"
+```
+
+### キャプション記述例、secondary_separator 記法：`secondary_separator = ";;;"` の場合
+
+```txt
+1girl, hatsune miku, vocaloid, upper body, looking at viewer, sky;;;cloud;;;day, outdoors
+```
+`sky;;;cloud;;;day` の部分はシャッフル、drop されず `sky,cloud,day` に置換されます。シャッフル、drop が有効な場合、まとめて（一つのタグとして）処理されます。つまり `vocaloid, 1girl, upper body, sky,cloud,day, outdoors, hatsune miku` （シャッフル）や `vocaloid, 1girl, outdoors, looking at viewer, upper body, hatsune miku` （drop されたケース）などになります。
+
+### キャプション記述例、ワイルドカード記法： `enable_wildcard = true` の場合
+
+```txt
+1girl, hatsune miku, vocaloid, upper body, looking at viewer, {simple|white} background
+```
+ランダムに `simple` または `white` が選ばれ、`simple background` または `white background` になります。
+
+```txt
+1girl, hatsune miku, vocaloid, {{retro style}}
+```
+タグ文字列に `{` や `}` そのものを含めたい場合は `{{` や `}}` のように二つ重ねてください（この例では実際に学習に用いられるキャプションは `{retro style}` になります）。
+
+### キャプション記述例、`keep_tokens_separator` 記法： `keep_tokens_separator = "|||"` の場合
+
+```txt
+1girl, hatsune miku, vocaloid ||| stage, microphone, white shirt, smile ||| best quality, rating: general
+```
+`1girl, hatsune miku, vocaloid, microphone, stage, white shirt, best quality, rating: general` や `1girl, hatsune miku, vocaloid, white shirt, smile, stage, microphone, best quality, rating: general` などになります。

docs/fine_tune.md

@@ -0,0 +1,347 @@
+# Fine-tuning Guide
+
+This document explains how to perform fine-tuning on various model architectures using the `*_train.py` scripts.
+
+<details>
+<summary>日本語</summary>
+
+# Fine-tuning ガイド
+
+このドキュメントでは、`*_train.py` スクリプトを用いた、各種モデルアーキテクチャのFine-tuningの方法について解説します。
+
+</details>
+
+### Difference between Fine-tuning and LoRA tuning
+
+This repository supports two methods for additional model training: **Fine-tuning** and **LoRA (Low-Rank Adaptation)**. Each method has distinct features and advantages.
+
+**Fine-tuning** is a method that retrains all (or most) of the weights of a pre-trained model.
+- **Pros**: It can improve the overall expressive power of the model and is suitable for learning styles or concepts that differ significantly from the original model.
+- **Cons**:
+    - It requires a large amount of VRAM and computational cost.
+    - The saved file size is large (same as the original model).
+    - It is prone to "overfitting," where the model loses the diversity of the original model if over-trained.
+- **Corresponding scripts**: Scripts named `*_train.py`, such as `sdxl_train.py`, `sd3_train.py`, `flux_train.py`, and `lumina_train.py`.
+
+**LoRA tuning** is a method that freezes the model's weights and only trains a small additional network called an "adapter."
+- **Pros**:
+    - It allows for fast training with low VRAM and computational cost.
+    - It is considered resistant to overfitting because it trains fewer weights.
+    - The saved file (LoRA network) is very small, ranging from tens to hundreds of MB, making it easy to manage.
+    - Multiple LoRAs can be used in combination.
+- **Cons**: Since it does not train the entire model, it may not achieve changes as significant as fine-tuning.
+- **Corresponding scripts**: Scripts named `*_train_network.py`, such as `sdxl_train_network.py`, `sd3_train_network.py`, and `flux_train_network.py`.
+
+| Feature | Fine-tuning | LoRA tuning |
+|:---|:---|:---|
+| **Training Target** | All model weights | Additional network (adapter) only |
+| **VRAM/Compute Cost**| High | Low |
+| **Training Time** | Long | Short |
+| **File Size** | Large (several GB) | Small (few MB to hundreds of MB) |
+| **Overfitting Risk** | High | Low |
+| **Suitable Use Case** | Major style changes, concept learning | Adding specific characters or styles |
+
+Generally, it is recommended to start with **LoRA tuning** if you want to add a specific character or style. **Fine-tuning** is a valid option for more fundamental style changes or aiming for a high-quality model.
+
+<details>
+<summary>日本語</summary>
+
+### Fine-tuningとLoRA学習の違い
+
+このリポジトリでは、モデルの追加学習手法として**Fine-tuning**と**LoRA (Low-Rank Adaptation)**学習の2種類をサポートしています。それぞれの手法には異なる特徴と利点があります。
+
+**Fine-tuning**は、事前学習済みモデルの重み全体（または大部分）を再学習する手法です。
+- **利点**: モデル全体の表現力を向上させることができ、元のモデルから大きく変化した画風やコンセプトの学習に適しています。
+- **欠点**:
+    - 学習には多くのVRAMと計算コストが必要です。
+    - 保存されるファイルサイズが大きくなります（元のモデルと同じサイズ）。
+    - 学習させすぎると、元のモデルが持っていた多様性が失われる「過学習（overfitting）」に陥りやすい傾向があります。
+- **対応スクリプト**: `sdxl_train.py`, `sd3_train.py`, `flux_train.py`, `lumina_train.py` など、`*_train.py` という命名規則のスクリプトが対応します。
+
+**LoRA学習**は、モデルの重みは凍結（固定）したまま、「アダプター」と呼ばれる小さな追加ネットワークのみを学習する手法です。
+- **利点**:
+    - 少ないVRAMと計算コストで高速に学習できます。
+    - 学習する重みが少ないため、過学習に強いとされています。
+    - 保存されるファイル（LoRAネットワーク）は数十〜数百MBと非常に小さく、管理が容易です。
+    - 複数のLoRAを組み合わせて使用することも可能です。
+- **欠点**: モデル全体を学習するわけではないため、Fine-tuningほどの大きな変化は期待できない場合があります。
+- **対応スクリプト**: `sdxl_train_network.py`, `sd3_train_network.py`, `flux_train_network.py` など、`*_train_network.py` という命名規則のスクリプトが対応します。
+
+| 特徴 | Fine-tuning | LoRA学習 |
+|:---|:---|:---|
+| **学習対象** | モデルの全重み | 追加ネットワーク（アダプター）のみ |
+| **VRAM/計算コスト**| 大 | 小 |
+| **学習時間** | 長 | 短 |
+| **ファイルサイズ** | 大（数GB） | 小（数MB〜数百MB） |
+| **過学習リスク** | 高 | 低 |
+| **適した用途** | 大規模な画風変更、コンセプト学習 | 特定のキャラ、画風の追加学習 |
+
+一般的に、特定のキャラクターや画風を追加したい場合は**LoRA学習**から試すことが推奨されます。より根本的な画風の変更や、高品質なモデルを目指す場合は**Fine-tuning**が有効な選択肢となります。
+
+</details>
+
+--- 
+
+### Fine-tuning for each architecture
+
+Fine-tuning updates the entire weights of the model, so it has different options and considerations than LoRA tuning. This section describes the fine-tuning scripts for major architectures.
+
+The basic command structure is common to all architectures.
+
+```bash
+accelerate launch --mixed_precision bf16 {script_name}.py \
+  --pretrained_model_name_or_path <path_to_model> \
+  --dataset_config <path_to_config.toml> \
+  --output_dir <output_directory> \
+  --output_name <model_output_name> \
+  --save_model_as safetensors \
+  --max_train_steps 10000 \
+  --learning_rate 1e-5 \
+  --optimizer_type AdamW8bit
+```
+
+<details>
+<summary>日本語</summary>
+
+### 各アーキテクチャのFine-tuning
+
+Fine-tuningはモデルの重み全体を更新するため、LoRA学習とは異なるオプションや考慮事項があります。ここでは主要なアーキテクチャごとのFine-tuningスクリプトについて説明します。
+
+基本的なコマンドの構造は、どのアーキテクチャでも共通です。
+
+```bash
+accelerate launch --mixed_precision bf16 {script_name}.py \
+  --pretrained_model_name_or_path <path_to_model> \
+  --dataset_config <path_to_config.toml> \
+  --output_dir <output_directory> \
+  --output_name <model_output_name> \
+  --save_model_as safetensors \
+  --max_train_steps 10000 \
+  --learning_rate 1e-5 \
+  --optimizer_type AdamW8bit
+```
+
+</details>
+
+#### SDXL (`sdxl_train.py`)
+
+Performs fine-tuning for SDXL models. It is possible to train both the U-Net and the Text Encoders.
+
+**Key Options:**
+
+- `--train_text_encoder`: Includes the weights of the Text Encoders (CLIP ViT-L and OpenCLIP ViT-bigG) in the training. Effective for significant style changes or strongly learning specific concepts.
+- `--learning_rate_te1`, `--learning_rate_te2`: Set individual learning rates for each Text Encoder.
+- `--block_lr`: Divides the U-Net into 23 blocks and sets a different learning rate for each block. This allows for advanced adjustments, such as strengthening or weakening the learning of specific layers. (Not available in LoRA tuning).
+
+**Command Example:**
+
+```bash
+accelerate launch --mixed_precision bf16 sdxl_train.py \
+  --pretrained_model_name_or_path "sd_xl_base_1.0.safetensors" \
+  --dataset_config "dataset_config.toml" \
+  --output_dir "output" \
+  --output_name "sdxl_finetuned" \
+  --train_text_encoder \
+  --learning_rate 1e-5 \
+  --learning_rate_te1 5e-6 \
+  --learning_rate_te2 2e-6
+```
+
+<details>
+<summary>日本語</summary>
+
+#### SDXL (`sdxl_train.py`)
+
+SDXLモデルのFine-tuningを行います。U-NetとText Encoderの両方を学習させることが可能です。
+
+**主要なオプション:**
+
+- `--train_text_encoder`: Text Encoder（CLIP ViT-LとOpenCLIP ViT-bigG）の重みを学習対象に含めます。画風を大きく変えたい場合や、特定の概念を強く学習させたい場合に有効です。
+- `--learning_rate_te1`, `--learning_rate_te2`: それぞれのText Encoderに個別の学習率を設定します。
+- `--block_lr`: U-Netを23個のブロックに分割し、ブロックごとに異なる学習率を設定できます。特定の層の学習を強めたり弱めたりする高度な調整が可能です。（LoRA学習では利用できません）
+
+**コマンド例:**
+
+```bash
+accelerate launch --mixed_precision bf16 sdxl_train.py \
+  --pretrained_model_name_or_path "sd_xl_base_1.0.safetensors" \
+  --dataset_config "dataset_config.toml" \
+  --output_dir "output" \
+  --output_name "sdxl_finetuned" \
+  --train_text_encoder \
+  --learning_rate 1e-5 \
+  --learning_rate_te1 5e-6 \
+  --learning_rate_te2 2e-6
+```
+
+</details>
+
+#### SD3 (`sd3_train.py`)
+
+Performs fine-tuning for Stable Diffusion 3 Medium models. SD3 consists of three Text Encoders (CLIP-L, CLIP-G, T5-XXL) and a MMDiT (equivalent to U-Net), which can be targeted for training.
+
+**Key Options:**
+
+- `--train_text_encoder`: Enables training for CLIP-L and CLIP-G.
+- `--train_t5xxl`: Enables training for T5-XXL. T5-XXL is a very large model and requires a lot of VRAM for training.
+- `--blocks_to_swap`: A memory optimization feature to reduce VRAM usage. It swaps some blocks of the MMDiT to CPU memory during training. Useful for using larger batch sizes in low VRAM environments. (Also available in LoRA tuning).
+- `--num_last_block_to_freeze`: Freezes the weights of the last N blocks of the MMDiT, excluding them from training. Useful for maintaining model stability while focusing on learning in the lower layers.
+
+**Command Example:**
+
+```bash
+accelerate launch --mixed_precision bf16 sd3_train.py \
+  --pretrained_model_name_or_path "sd3_medium.safetensors" \
+  --dataset_config "dataset_config.toml" \
+  --output_dir "output" \
+  --output_name "sd3_finetuned" \
+  --train_text_encoder \
+  --learning_rate 4e-6 \
+  --blocks_to_swap 10
+```
+
+<details>
+<summary>日本語</summary>
+
+#### SD3 (`sd3_train.py`)
+
+Stable Diffusion 3 MediumモデルのFine-tuningを行います。SD3は3つのText Encoder（CLIP-L, CLIP-G, T5-XXL）とMMDiT（U-Netに相当）で構成されており、これらを学習対象にできます。
+
+**主要なオプション:**
+
+- `--train_text_encoder`: CLIP-LとCLIP-Gの学習を有効にします。
+- `--train_t5xxl`: T5-XXLの学習を有効にします。T5-XXLは非常に大きなモデルのため、学習には多くのVRAMが必要です。
+- `--blocks_to_swap`: VRAM使用量を削減するためのメモリ最適化機能です。MMDiTの一部のブロックを学習中にCPUメモリに退避（スワップ）させます。VRAMが少ない環境で大きなバッチサイズを使いたい場合に有効です。（LoRA学習でも利用可能）
+- `--num_last_block_to_freeze`: MMDiTの最後のNブロックの重みを凍結し、学習対象から除外します。モデルの安定性を保ちつつ、下位層を中心に学習させたい場合に有効です。
+
+**コマンド例:**
+
+```bash
+accelerate launch --mixed_precision bf16 sd3_train.py \
+  --pretrained_model_name_or_path "sd3_medium.safetensors" \
+  --dataset_config "dataset_config.toml" \
+  --output_dir "output" \
+  --output_name "sd3_finetuned" \
+  --train_text_encoder \
+  --learning_rate 4e-6 \
+  --blocks_to_swap 10
+```
+
+</details>
+
+#### FLUX.1 (`flux_train.py`)
+
+Performs fine-tuning for FLUX.1 models. FLUX.1 is internally composed of two Transformer blocks (Double Blocks, Single Blocks).
+
+**Key Options:**
+
+- `--blocks_to_swap`: Similar to SD3, this feature swaps Transformer blocks to the CPU for memory optimization.
+- `--blockwise_fused_optimizers`: An experimental feature that aims to streamline training by applying individual optimizers to each block.
+
+**Command Example:**
+
+```bash
+accelerate launch --mixed_precision bf16 flux_train.py \
+  --pretrained_model_name_or_path "FLUX.1-dev.safetensors" \
+  --dataset_config "dataset_config.toml" \
+  --output_dir "output" \
+  --output_name "flux1_finetuned" \
+  --learning_rate 1e-5 \
+  --blocks_to_swap 18
+```
+
+<details>
+<summary>日本語</summary>
+
+#### FLUX.1 (`flux_train.py`)
+
+FLUX.1モデルのFine-tuningを行います。FLUX.1は内部的に2つのTransformerブロック（Double Blocks, Single Blocks）で構成されています。
+
+**主要なオプション:**
+
+- `--blocks_to_swap`: SD3と同様に、メモリ最適化のためにTransformerブロックをCPUにスワップする機能です。
+- `--blockwise_fused_optimizers`: 実験的な機能で、各ブロックに個別のオプティマイザを適用し、学習を効率化することを目指します。
+
+**コマンド例:**
+
+```bash
+accelerate launch --mixed_precision bf16 flux_train.py \
+  --pretrained_model_name_or_path "FLUX.1-dev.safetensors" \
+  --dataset_config "dataset_config.toml" \
+  --output_dir "output" \
+  --output_name "flux1_finetuned" \
+  --learning_rate 1e-5 \
+  --blocks_to_swap 18
+```
+
+</details>
+
+#### Lumina (`lumina_train.py`)
+
+Performs fine-tuning for Lumina-Next DiT models.
+
+**Key Options:**
+
+- `--use_flash_attn`: Enables Flash Attention to speed up computation.
+- `lumina_train.py` is relatively new, and many of its options are shared with other scripts. Training can be performed following the basic command pattern.
+
+**Command Example:**
+
+```bash
+accelerate launch --mixed_precision bf16 lumina_train.py \
+  --pretrained_model_name_or_path "Lumina-Next-DiT-B.safetensors" \
+  --dataset_config "dataset_config.toml" \
+  --output_dir "output" \
+  --output_name "lumina_finetuned" \
+  --learning_rate 1e-5
+```
+
+<details>
+<summary>日本語</summary>
+
+#### Lumina (`lumina_train.py`)
+
+Lumina-Next DiTモデルのFine-tuningを行います。
+
+**主要なオプション:**
+
+- `--use_flash_attn`: Flash Attentionを有効にし、計算を高速化します。
+- `lumina_train.py`は比較的新しく、オプションは他のスクリプトと共通化されている部分が多いです。基本的なコマンドパターンに従って学習を行えます。
+
+**コマンド例:**
+
+```bash
+accelerate launch --mixed_precision bf16 lumina_train.py \
+  --pretrained_model_name_or_path "Lumina-Next-DiT-B.safetensors" \
+  --dataset_config "dataset_config.toml" \
+  --output_dir "output" \
+  --output_name "lumina_finetuned" \
+  --learning_rate 1e-5
+```
+
+</details>
+
+--- 
+
+### Differences between Fine-tuning and LoRA tuning per architecture
+
+| Architecture | Key Features/Options Specific to Fine-tuning | Main Differences from LoRA tuning |
+|:---|:---|:---|
+| **SDXL** | `--block_lr` | Only fine-tuning allows for granular control over the learning rate for each U-Net block. |
+| **SD3** | `--train_text_encoder`, `--train_t5xxl`, `--num_last_block_to_freeze` | Only fine-tuning can train the entire Text Encoders. LoRA only trains the adapter parts. |
+| **FLUX.1** | `--blockwise_fused_optimizers` | Since fine-tuning updates the entire model's weights, more experimental optimizer options are available. |
+| **Lumina** | (Few specific options) | Basic training options are common, but fine-tuning differs in that it updates the entire model's foundation. |
+
+<details>
+<summary>日本語</summary>
+
+### アーキテクチャごとのFine-tuningとLoRA学習の違い
+
+| アーキテクチャ | Fine-tuning特有の主要機能・オプション | LoRA学習との主な違い |
+|:---|:---|:---|
+| **SDXL** | `--block_lr` | U-Netのブロックごとに学習率を細かく制御できるのはFine-tuningのみです。 |
+| **SD3** | `--train_text_encoder`, `--train_t5xxl`, `--num_last_block_to_freeze` | Text Encoder全体を学習対象にできるのはFine-tuningです。LoRAではアダプター部分のみ学習します。 |
+| **FLUX.1** | `--blockwise_fused_optimizers` | Fine-tuningではモデル全体の重みを更新するため、より実験的なオプティマイザの選択肢が用意されています。 |
+| **Lumina** | （特有のオプションは少ない） | 基本的な学習オプションは共通ですが、Fine-tuningはモデルの基盤全体を更新する点で異なります。 |
+
+</details>

docs/flux_train_network.md

@@ -0,0 +1,709 @@
+Status: reviewed
+
+# LoRA Training Guide for FLUX.1 using `flux_train_network.py` / `flux_train_network.py` を用いたFLUX.1モデルのLoRA学習ガイド
+
+This document explains how to train LoRA models for the FLUX.1 model using `flux_train_network.py` included in the `sd-scripts` repository.
+
+<details>
+<summary>日本語</summary>
+
+このドキュメントでは、`sd-scripts`リポジトリに含まれる`flux_train_network.py`を使用して、FLUX.1モデルに対するLoRA (Low-Rank Adaptation) モデルを学習する基本的な手順について解説します。
+
+</details>
+
+## 1. Introduction / はじめに
+
+`flux_train_network.py` trains additional networks such as LoRA on the FLUX.1 model, which uses a transformer-based architecture different from Stable Diffusion. Two text encoders, CLIP-L and T5-XXL, and a dedicated AutoEncoder are used.
+
+This guide assumes you know the basics of LoRA training. For common options see [train_network.py](train_network.md) and [sdxl_train_network.py](sdxl_train_network.md).
+
+**Prerequisites:**
+
+* The repository is cloned and the Python environment is ready.
+* A training dataset is prepared. See the dataset configuration guide.
+
+<details>
+<summary>日本語</summary>
+
+`flux_train_network.py`は、FLUX.1モデルに対してLoRAなどの追加ネットワークを学習させるためのスクリプトです。FLUX.1はStable Diffusionとは異なるアーキテクチャを持つ画像生成モデルであり、このスクリプトを使用することで、特定のキャラクターや画風を再現するLoRAモデルを作成できます。
+
+このガイドは、基本的なLoRA学習の手順を理解しているユーザーを対象としています。基本的な使い方や共通のオプションについては、[`train_network.py`のガイド](train_network.md)を参照してください。また一部のパラメータは [`sdxl_train_network.py`](sdxl_train_network.md) と同様のものがあるため、そちらも参考にしてください。
+
+**前提条件:**
+
+* `sd-scripts`リポジトリのクローンとPython環境のセットアップが完了していること。
+* 学習用データセットの準備が完了していること。（データセットの準備については[データセット設定ガイド](link/to/dataset/config/doc)を参照してください）
+
+</details>
+
+## 2. Differences from `train_network.py` / `train_network.py` との違い
+
+`flux_train_network.py` is based on `train_network.py` but adapted for FLUX.1. Main differences include:
+
+* **Target model:** FLUX.1 model (dev or schnell version).
+* **Model structure:** Unlike Stable Diffusion, FLUX.1 uses a Transformer-based architecture with two text encoders (CLIP-L and T5-XXL) and a dedicated AutoEncoder (AE) instead of VAE.
+* **Required arguments:** Additional arguments for FLUX.1 model, CLIP-L, T5-XXL, and AE model files.
+* **Incompatible options:** Some Stable Diffusion-specific arguments (e.g., `--v2`, `--clip_skip`, `--max_token_length`) are not used in FLUX.1 training.
+* **FLUX.1-specific arguments:** Additional arguments for FLUX.1-specific training parameters like timestep sampling and guidance scale.
+
+<details>
+<summary>日本語</summary>
+
+`flux_train_network.py`は`train_network.py`をベースに、FLUX.1モデルに対応するための変更が加えられています。主な違いは以下の通りです。
+
+* **対象モデル:** FLUX.1モデル（dev版またはschnell版）を対象とします。
+* **モデル構造:** Stable Diffusionとは異なり、FLUX.1はTransformerベースのアーキテクチャを持ちます。Text EncoderとしてCLIP-LとT5-XXLの二つを使用し、VAEの代わりに専用のAutoEncoder (AE) を使用します。
+* **必須の引数:** FLUX.1モデル、CLIP-L、T5-XXL、AEの各モデルファイルを指定する引数が追加されています。
+* **一部引数の非互換性:** Stable Diffusion向けの引数の一部（例: `--v2`, `--clip_skip`, `--max_token_length`）はFLUX.1の学習では使用されません。
+* **FLUX.1特有の引数:** タイムステップのサンプリング方法やガイダンススケールなど、FLUX.1特有の学習パラメータを指定する引数が追加されています。
+
+</details>
+
+## 3. Preparation / 準備
+
+Before starting training you need:
+
+1. **Training script:** `flux_train_network.py`
+2. **FLUX.1 model file:** Base FLUX.1 model `.safetensors` file (e.g., `flux1-dev.safetensors`).
+3. **Text Encoder model files:**
+   - CLIP-L model `.safetensors` file (e.g., `clip_l.safetensors`)
+   - T5-XXL model `.safetensors` file (e.g., `t5xxl.safetensors`)
+4. **AutoEncoder model file:** FLUX.1-compatible AE model `.safetensors` file (e.g., `ae.safetensors`).
+5. **Dataset definition file (.toml):** TOML format file describing training dataset configuration (e.g., `my_flux_dataset_config.toml`).
+
+### Downloading Required Models
+
+To train FLUX.1 models, you need to download the following model files:
+
+- **DiT, AE**: Download from the [black-forest-labs/FLUX.1 dev](https://huggingface.co/black-forest-labs/FLUX.1-dev) repository. Use `flux1-dev.safetensors` and `ae.safetensors`. The weights in the subfolder are in Diffusers format and cannot be used.
+- **Text Encoder 1 (T5-XXL), Text Encoder 2 (CLIP-L)**: Download from the [ComfyUI FLUX Text Encoders](https://huggingface.co/comfyanonymous/flux_text_encoders) repository. Please use `t5xxl_fp16.safetensors` for T5-XXL. Thanks to ComfyUI for providing these models.
+
+To train Chroma models, you need to download the Chroma model file from the following repository:
+
+- **Chroma Base**: Download from the [lodestones/Chroma1-Base](https://huggingface.co/lodestones/Chroma1-Base) repository. Use `Chroma.safetensors`.
+
+We have tested Chroma training with the weights from the [lodestones/Chroma](https://huggingface.co/lodestones/Chroma) repository. 
+
+AE and T5-XXL models are same as FLUX.1, so you can use the same files. CLIP-L model is not used for Chroma training, so you can omit the `--clip_l` argument.
+
+<details>
+<summary>日本語</summary>
+
+学習を開始する前に、以下のファイルが必要です。
+
+1. **学習スクリプト:** `flux_train_network.py`
+2. **FLUX.1モデルファイル:** 学習のベースとなるFLUX.1モデルの`.safetensors`ファイル（例: `flux1-dev.safetensors`）。
+3. **Text Encoderモデルファイル:**
+   - CLIP-Lモデルの`.safetensors`ファイル。例として`clip_l.safetensors`を使用します。
+   - T5-XXLモデルの`.safetensors`ファイル。例として`t5xxl.safetensors`を使用します。
+4. **AutoEncoderモデルファイル:** FLUX.1に対応するAEモデルの`.safetensors`ファイル。例として`ae.safetensors`を使用します。
+5. **データセット定義ファイル (.toml):** 学習データセットの設定を記述したTOML形式のファイル。（詳細は[データセット設定ガイド](link/to/dataset/config/doc)を参照してください）。例として`my_flux_dataset_config.toml`を使用します。
+
+**必要なモデルのダウンロード**
+
+FLUX.1モデルを学習するためには、以下のモデルファイルをダウンロードする必要があります。
+
+- **DiT, AE**: [black-forest-labs/FLUX.1 dev](https://huggingface.co/black-forest-labs/FLUX.1-dev) リポジトリからダウンロードします。`flux1-dev.safetensors`と`ae.safetensors`を使用してください。サブフォルダ内の重みはDiffusers形式であり、使用できません。
+- **Text Encoder 1 (T5-XXL), Text Encoder 2 (CLIP-L)**: [ComfyUI FLUX Text Encoders](https://huggingface.co/comfyanonymous/flux_text_encoders) リポジトリからダウンロードします。T5-XXLには`t5xxl_fp16.safetensors`を使用してください。これらのモデルを提供いただいたComfyUIに感謝します。
+
+Chromaモデルを学習する場合は、以下のリポジトリからChromaモデルファイルをダウンロードする必要があります。
+
+- **Chroma Base**: [lodestones/Chroma1-Base](https://huggingface.co/lodestones/Chroma1-Base) リポジトリからダウンロードします。`Chroma.safetensors`を使用してください。
+
+Chromaの学習のテストは [lodestones/Chroma](https://huggingface.co/lodestones/Chroma) リポジトリの重みを使用して行いました。
+
+AEとT5-XXLモデルはFLUX.1と同じものを使用できるため、同じファイルを使用します。CLIP-LモデルはChroma学習では使用されないため、`--clip_l`引数は省略できます。
+
+</details>
+
+## 4. Running the Training / 学習の実行
+
+Run `flux_train_network.py` from the terminal with FLUX.1 specific arguments. Here's a basic command example:
+
+```bash
+accelerate launch --num_cpu_threads_per_process 1 flux_train_network.py \
+  --pretrained_model_name_or_path="<path to FLUX.1 model>" \
+  --clip_l="<path to CLIP-L model>" \
+  --t5xxl="<path to T5-XXL model>" \
+  --ae="<path to AE model>" \
+  --dataset_config="my_flux_dataset_config.toml" \
+  --output_dir="<output directory>" \
+  --output_name="my_flux_lora" \
+  --save_model_as=safetensors \
+  --network_module=networks.lora_flux \
+  --network_dim=16 \
+  --network_alpha=1 \
+  --learning_rate=1e-4 \
+  --optimizer_type="AdamW8bit" \
+  --lr_scheduler="constant" \
+  --sdpa \
+  --max_train_epochs=10 \
+  --save_every_n_epochs=1 \
+  --mixed_precision="fp16" \
+  --gradient_checkpointing \
+  --guidance_scale=1.0 \
+  --timestep_sampling="flux_shift" \
+  --model_prediction_type="raw" \
+  --blocks_to_swap=18 \
+  --cache_text_encoder_outputs \
+  --cache_latents
+```
+
+### Training Chroma Models
+
+If you want to train a Chroma model, specify `--model_type=chroma`. Chroma does not use CLIP-L, so the `--clip_l` argument is not needed. T5XXL and AE are same as FLUX.1. The command would look like this:
+
+```bash
+accelerate launch --num_cpu_threads_per_process 1 flux_train_network.py \
+  --pretrained_model_name_or_path="<path to Chroma model>" \
+  --model_type=chroma \
+  --t5xxl="<path to T5-XXL model>" \
+  --ae="<path to AE model>" \
+  --dataset_config="my_flux_dataset_config.toml" \
+  --output_dir="<output directory>" \
+  --output_name="my_chroma_lora" \
+  --guidance_scale=0.0 \
+  --timestep_sampling="sigmoid" \
+  --apply_t5_attn_mask \
+  ...
+```
+
+Note that for Chroma models, `--guidance_scale=0.0` is required to disable guidance scale, and `--apply_t5_attn_mask` is needed to apply attention masks for T5XXL Text Encoder.
+
+The sample image generation during training requires specifying a negative prompt. Also, set `--g 0` to disable embedded guidance scale and `--l 4.0` to set the CFG scale. For example:
+
+```
+Japanese shrine in the summer forest. --n low quality, ugly, unfinished, out of focus, deformed, disfigure, blurry, smudged, restricted palette, flat colors --w 512 --h 512 --d 1 --l 4.0 --g 0.0 --s 20
+```
+
+<details>
+<summary>日本語</summary>
+
+学習は、ターミナルから`flux_train_network.py`を実行することで開始します。基本的なコマンドラインの構造は`train_network.py`と同様ですが、FLUX.1特有の引数を指定する必要があります。
+
+コマンドラインの例は英語のドキュメントを参照してください。
+
+#### Chromaモデルの学習
+
+Chromaモデルを学習したい場合は、`--model_type=chroma`を指定します。ChromaはCLIP-Lを使用しないため、`--clip_l`引数は不要です。T5XXLとAEはFLUX.1と同様です。
+
+コマンドラインの例は英語のドキュメントを参照してください。
+
+学習中のサンプル画像生成には、ネガティブプロンプトを指定してください。また `--g 0` を指定して埋め込みガイダンススケールを無効化し、`--l 4.0` を指定してCFGスケールを設定します。
+
+</details>
+
+### 4.1. Explanation of Key Options / 主要なコマンドライン引数の解説
+
+The script adds FLUX.1 specific arguments. For common arguments (like `--output_dir`, `--output_name`, `--network_module`, etc.), see the [`train_network.py` guide](train_network.md).
+
+#### Model-related [Required]
+
+* `--pretrained_model_name_or_path="<path to FLUX.1/Chroma model>"` **[Required]**
+  - Specifies the path to the base FLUX.1 or Chroma model `.safetensors` file. Diffusers format directories are not currently supported.
+* `--model_type=<model type>`
+  - Specifies the type of base model for training. Choose from `flux` or `chroma`. Default is `flux`.
+* `--clip_l="<path to CLIP-L model>"` **[Required when flux is selected]**
+  - Specifies the path to the CLIP-L Text Encoder model `.safetensors` file. Not needed when `--model_type=chroma`.
+* `--t5xxl="<path to T5-XXL model>"` **[Required]**
+  - Specifies the path to the T5-XXL Text Encoder model `.safetensors` file.
+* `--ae="<path to AE model>"` **[Required]**
+  - Specifies the path to the FLUX.1-compatible AutoEncoder model `.safetensors` file.
+
+#### FLUX.1 Training Parameters
+
+* `--guidance_scale=<float>`
+  - FLUX.1 dev version is distilled with specific guidance scale values, but for training, specify `1.0` to disable guidance scale. Default is `3.5`, so be sure to specify this. Usually ignored for schnell version.
+  - Chroma requires `--guidance_scale=0.0` to disable guidance scale.
+* `--timestep_sampling=<choice>`
+  - Specifies the sampling method for timesteps (noise levels) during training. Choose from `sigma`, `uniform`, `sigmoid`, `shift`, `flux_shift`. Default is `sigma`. Recommended is `flux_shift`. For Chroma models, `sigmoid` is recommended.
+* `--sigmoid_scale=<float>`
+  - Scale factor when `timestep_sampling` is set to `sigmoid`, `shift`, or `flux_shift`. Default and recommended value is `1.0`.
+* `--model_prediction_type=<choice>`
+  - Specifies what the model predicts. Choose from `raw` (use prediction as-is), `additive` (add to noise input), `sigma_scaled` (apply sigma scaling). Default is `sigma_scaled`. Recommended is `raw`.
+* `--discrete_flow_shift=<float>`
+  - Specifies the shift value for the scheduler used in Flow Matching. Default is `3.0`. This value is ignored when `timestep_sampling` is set to other than `shift`.
+
+#### Memory/Speed Related
+
+* `--fp8_base` 
+  - Enables training in FP8 format for FLUX.1, CLIP-L, and T5-XXL. This can significantly reduce VRAM usage, but the training results may vary. 
+* `--blocks_to_swap=<integer>` **[Experimental Feature]**
+  - Setting to reduce VRAM usage by swapping parts of the model (Transformer blocks) between CPU and GPU. Specify the number of blocks to swap as an integer (e.g., `18`). Larger values reduce VRAM usage but decrease training speed. Adjust according to your GPU's VRAM capacity. Can be used with `gradient_checkpointing`.
+  - Cannot be used with `--cpu_offload_checkpointing`.
+* `--cache_text_encoder_outputs`
+  - Caches the outputs of CLIP-L and T5-XXL. This reduces memory usage.
+* `--cache_latents`, `--cache_latents_to_disk`
+  - Caches the outputs of AE. Similar functionality to [sdxl_train_network.py](sdxl_train_network.md).
+
+#### Incompatible/Deprecated Arguments
+
+* `--v2`, `--v_parameterization`, `--clip_skip`: These are Stable Diffusion-specific arguments and are not used in FLUX.1 training.
+* `--max_token_length`: This is an argument for Stable Diffusion v1/v2. For FLUX.1, use `--t5xxl_max_token_length`.
+* `--split_mode`: Deprecated argument. Use `--blocks_to_swap` instead.
+
+<details>
+<summary>日本語</summary>
+
+[`train_network.py`のガイド](train_network.md)で説明されている引数に加え、以下のFLUX.1特有の引数を指定します。共通の引数（`--output_dir`, `--output_name`, `--network_module`, `--network_dim`, `--network_alpha`, `--learning_rate`など）については、上記ガイドを参照してください。
+
+コマンドラインの例と詳細な引数の説明は英語のドキュメントを参照してください。
+
+</details>
+
+### 4.2. Starting Training / 学習の開始
+
+Training begins once you run the command with the required options. Log checking is the same as in [`train_network.py`](train_network.md#32-starting-the-training--学習の開始).
+
+<details>
+<summary>日本語</summary>
+
+必要な引数を設定し、コマンドを実行すると学習が開始されます。基本的な流れやログの確認方法は[`train_network.py`のガイド](train_network.md#32-starting-the-training--学習の開始)と同様です。
+
+</details>
+
+## 5. Using the Trained Model / 学習済みモデルの利用
+
+After training, a LoRA model file is saved in `output_dir` and can be used in inference environments supporting FLUX.1 (e.g. ComfyUI + Flux nodes).
+
+<details>
+<summary>日本語</summary>
+
+学習が完了すると、指定した`output_dir`にLoRAモデルファイル（例: `my_flux_lora.safetensors`）が保存されます。このファイルは、FLUX.1モデルに対応した推論環境（例: ComfyUI + ComfyUI-FluxNodes）で使用できます。
+
+</details>
+
+## 6. Advanced Settings / 高度な設定
+
+### 6.1. VRAM Usage Optimization / VRAM使用量の最適化
+
+FLUX.1 is a relatively large model, so GPUs without sufficient VRAM require optimization. Here are settings to reduce VRAM usage (with `--fp8_base`):
+
+#### Recommended Settings by GPU Memory
+
+| GPU Memory | Recommended Settings |
+|------------|---------------------|
+| 24GB VRAM | Basic settings work fine (batch size 2) |
+| 16GB VRAM | Set batch size to 1 and use `--blocks_to_swap` |
+| 12GB VRAM | Use `--blocks_to_swap 16` and 8bit AdamW |
+| 10GB VRAM | Use `--blocks_to_swap 22`, recommend fp8 format for T5XXL |
+| 8GB VRAM | Use `--blocks_to_swap 28`, recommend fp8 format for T5XXL |
+
+#### Key VRAM Reduction Options
+
+- **`--fp8_base`**: Enables training in FP8 format.
+
+- **`--blocks_to_swap <number>`**: Swaps blocks between CPU and GPU to reduce VRAM usage. Higher numbers save more VRAM but reduce training speed. FLUX.1 supports up to 35 blocks for swapping.
+
+- **`--cpu_offload_checkpointing`**: Offloads gradient checkpoints to CPU. Can reduce VRAM usage by up to 1GB but decreases training speed by about 15%. Cannot be used with `--blocks_to_swap`. Chroma models do not support this option.
+
+- **Using Adafactor optimizer**: Can reduce VRAM usage more than 8bit AdamW:
+  ```
+  --optimizer_type adafactor --optimizer_args "relative_step=False" "scale_parameter=False" "warmup_init=False" --lr_scheduler constant_with_warmup --max_grad_norm 0.0
+  ```
+
+- **Using T5XXL fp8 format**: For GPUs with less than 10GB VRAM, using fp8 format T5XXL checkpoints is recommended. Download `t5xxl_fp8_e4m3fn.safetensors` from [comfyanonymous/flux_text_encoders](https://huggingface.co/comfyanonymous/flux_text_encoders) (use without `scaled`).
+
+- **FP8/FP16 Mixed Training [Experimental]**: Specify `--fp8_base_unet` to train the FLUX.1 model in FP8 format while training Text Encoders (CLIP-L/T5XXL) in BF16/FP16 format. This can further reduce VRAM usage.
+
+<details>
+<summary>日本語</summary>
+
+FLUX.1モデルは比較的大きなモデルであるため、十分なVRAMを持たないGPUでは工夫が必要です。VRAM使用量を削減するための設定の詳細は英語のドキュメントを参照してください。
+
+主要なVRAM削減オプション：
+- `--fp8_base`: FP8形式での学習を有効化
+- `--blocks_to_swap`: CPUとGPU間でブロックをスワップ
+- `--cpu_offload_checkpointing`: 勾配チェックポイントをCPUにオフロード  
+- Adafactorオプティマイザの使用
+- T5XXLのfp8形式の使用
+- FP8/FP16混合学習（実験的機能）
+
+</details>
+
+### 6.2. Important FLUX.1 LoRA Training Settings / FLUX.1 LoRA学習の重要な設定
+
+FLUX.1 training has many unknowns, and several settings can be specified with arguments:
+
+#### Timestep Sampling Methods
+
+The `--timestep_sampling` option specifies how timesteps (0-1) are sampled:
+
+- `sigma`: Sigma-based like SD3
+- `uniform`: Uniform random
+- `sigmoid`: Sigmoid of normal distribution random (similar to x-flux, AI-toolkit)
+- `shift`: Sigmoid value of normal distribution random with shift. The `--discrete_flow_shift` setting is used to shift the sigmoid value.
+- `flux_shift`: Shift sigmoid value of normal distribution random according to resolution (similar to FLUX.1 dev inference).
+
+`--discrete_flow_shift` only applies when `--timestep_sampling` is set to `shift`.
+
+#### Model Prediction Processing
+
+The `--model_prediction_type` option specifies how to interpret and process model predictions:
+
+- `raw`: Use as-is (similar to x-flux) **[Recommended]**
+- `additive`: Add to noise input
+- `sigma_scaled`: Apply sigma scaling (similar to SD3)
+
+#### Recommended Settings
+
+Based on experiments, the following settings work well:
+```
+--timestep_sampling shift --discrete_flow_shift 3.1582 --model_prediction_type raw --guidance_scale 1.0
+```
+
+For Chroma models, the following settings are recommended:
+```
+--timestep_sampling sigmoid --model_prediction_type raw --guidance_scale 0.0
+```
+
+**About Guidance Scale**: FLUX.1 dev version is distilled with specific guidance scale values, but for training, specify `--guidance_scale 1.0` to disable guidance scale. Chroma requires `--guidance_scale 0.0` to disable guidance scale because it is not distilled.
+
+<details>
+<summary>日本語</summary>
+
+FLUX.1の学習には多くの未知の点があり、いくつかの設定は引数で指定できます。詳細な説明とコマンドラインの例は英語のドキュメントを参照してください。
+
+主要な設定オプション：
+- タイムステップのサンプリング方法（`--timestep_sampling`）
+- モデル予測の処理方法（`--model_prediction_type`）
+- 推奨設定の組み合わせ
+
+</details>
+
+### 6.3. Layer-specific Rank Configuration / 各層に対するランク指定
+
+You can specify different ranks (network_dim) for each layer of FLUX.1. This allows you to emphasize or disable LoRA effects for specific layers.
+
+Specify the following network_args to set ranks for each layer. Setting 0 disables LoRA for that layer:
+
+| network_args | Target Layer |
+|--------------|--------------|
+| img_attn_dim | DoubleStreamBlock img_attn |
+| txt_attn_dim | DoubleStreamBlock txt_attn |
+| img_mlp_dim | DoubleStreamBlock img_mlp |
+| txt_mlp_dim | DoubleStreamBlock txt_mlp |
+| img_mod_dim | DoubleStreamBlock img_mod |
+| txt_mod_dim | DoubleStreamBlock txt_mod |
+| single_dim | SingleStreamBlock linear1 and linear2 |
+| single_mod_dim | SingleStreamBlock modulation |
+
+Example usage:
+```
+--network_args "img_attn_dim=4" "img_mlp_dim=8" "txt_attn_dim=2" "txt_mlp_dim=2" "img_mod_dim=2" "txt_mod_dim=2" "single_dim=4" "single_mod_dim=2"
+```
+
+To apply LoRA to FLUX conditioning layers, specify `in_dims` in network_args as a comma-separated list of 5 numbers:
+
+```
+--network_args "in_dims=[4,2,2,2,4]"
+```
+
+Each number corresponds to `img_in`, `time_in`, `vector_in`, `guidance_in`, `txt_in`. The example above applies LoRA to all conditioning layers with ranks of 4 for `img_in` and `txt_in`, and ranks of 2 for others.
+
+<details>
+<summary>日本語</summary>
+
+FLUX.1の各層に対して異なるランク（network_dim）を指定できます。これにより、特定の層に対してLoRAの効果を強調したり、無効化したりできます。
+
+詳細な設定方法とコマンドラインの例は英語のドキュメントを参照してください。
+
+</details>
+
+### 6.4. Block Selection for Training / 学習するブロックの指定
+
+You can specify which blocks to train using `train_double_block_indices` and `train_single_block_indices` in network_args. Indices are 0-based. Default is to train all blocks if omitted.
+
+Specify indices as integer lists like `0,1,5,8` or integer ranges like `0,1,4-5,7`:
+- Double blocks: 19 blocks, valid range 0-18
+- Single blocks: 38 blocks, valid range 0-37
+- Specify `all` to train all blocks
+- Specify `none` to skip training blocks
+
+Example usage:
+```
+--network_args "train_double_block_indices=0,1,8-12,18" "train_single_block_indices=3,10,20-25,37"
+```
+
+Or:
+```
+--network_args "train_double_block_indices=none" "train_single_block_indices=10-15"
+```
+
+<details>
+<summary>日本語</summary>
+
+FLUX.1 LoRA学習では、network_argsの`train_double_block_indices`と`train_single_block_indices`を指定することで、学習するブロックを指定できます。
+
+詳細な設定方法とコマンドラインの例は英語のドキュメントを参照してください。
+
+</details>
+
+### 6.5. Regular Expression-based Rank/LR Configuration / 正規表現によるランク・学習率の指定
+
+You can specify ranks (dims) and learning rates for LoRA modules using regular expressions. This allows for more flexible and fine-grained control than specifying by layer.
+
+These settings are specified via the `network_args` argument.
+
+*   `network_reg_dims`: Specify ranks for modules matching a regular expression. The format is a comma-separated string of `pattern=rank`.
+    *   Example: `--network_args "network_reg_dims=single.*_modulation.*=4,img_attn=8"`
+    *   This sets the rank to 4 for modules whose names contain `single` and contain `_modulation`, and to 8 for modules containing `img_attn`.
+*   `network_reg_lrs`: Specify learning rates for modules matching a regular expression. The format is a comma-separated string of `pattern=lr`.
+    *   Example: `--network_args "network_reg_lrs=single_blocks_(\d|10)_=1e-3,double_blocks=2e-3"`
+    *   This sets the learning rate to `1e-3` for modules whose names contain `single_blocks` followed by a digit (`0` to `9`) or `10`, and to `2e-3` for modules whose names contain `double_blocks`.
+
+**Notes:**
+
+*   Settings via `network_reg_dims` and `network_reg_lrs` take precedence over the global `--network_dim` and `--learning_rate` settings.
+*   If a module name matches multiple patterns, the setting from the last matching pattern in the string will be applied.
+*   These settings are applied after the block-specific training settings (`train_double_block_indices`, `train_single_block_indices`).
+
+<details>
+<summary>日本語</summary>
+
+正規表現を用いて、LoRAのモジュールごとにランク（dim）や学習率を指定することができます。これにより、層ごとの指定よりも柔軟できめ細やかな制御が可能になります。
+
+これらの設定は `network_args` 引数で指定します。
+
+*   `network_reg_dims`: 正規表現にマッチするモジュールに対してランクを指定します。`pattern=rank` という形式の文字列をカンマで区切って指定します。
+    *   例: `--network_args "network_reg_dims=single.*_modulation.*=4,img_attn=8"`
+    *   この例では、名前に `single` で始まり `_modulation` を含むモジュールのランクを4に、`img_attn` を含むモジュールのランクを8に設定します。
+*   `network_reg_lrs`: 正規表現にマッチするモジュールに対して学習率を指定します。`pattern=lr` という形式の文字列をカンマで区切って指定します。
+    *   例: `--network_args "network_reg_lrs=single_blocks_(\d|10)_=1e-3,double_blocks=2e-3"`
+    *   この例では、名前が `single_blocks` で始まり、後に数字（`0`から`9`）または`10`が続くモジュールの学習率を `1e-3` に、`double_blocks` を含むモジュールの学習率を `2e-3` に設定します。
+**注意点:**
+
+*   `network_reg_dims` および `network_reg_lrs` での設定は、全体設定である `--network_dim` や `--learning_rate` よりも優先されます。
+*   あるモジュール名が複数のパターンにマッチした場合、文字列の中で後方にあるパターンの設定が適用されます。
+*   これらの設定は、ブロック指定（`train_double_block_indices`, `train_single_block_indices`）が適用された後に行われます。
+
+</details>
+
+### 6.6. Text Encoder LoRA Support / Text Encoder LoRAのサポート
+
+FLUX.1 LoRA training supports training CLIP-L and T5XXL LoRA:
+
+- To train only FLUX.1: specify `--network_train_unet_only`
+- To train FLUX.1 and CLIP-L: omit `--network_train_unet_only`
+- To train FLUX.1, CLIP-L, and T5XXL: omit `--network_train_unet_only` and add `--network_args "train_t5xxl=True"`
+
+You can specify individual learning rates for CLIP-L and T5XXL with `--text_encoder_lr`. For example, `--text_encoder_lr 1e-4 1e-5` sets the first value for CLIP-L and the second for T5XXL. Specifying one value uses the same learning rate for both. If `--text_encoder_lr` is not specified, the default `--learning_rate` is used for both.
+
+<details>
+<summary>日本語</summary>
+
+FLUX.1 LoRA学習は、CLIP-LとT5XXL LoRAのトレーニングもサポートしています。
+
+詳細な設定方法とコマンドラインの例は英語のドキュメントを参照してください。
+
+</details>
+
+### 6.7. Multi-Resolution Training / マルチ解像度トレーニング
+
+You can define multiple resolutions in the dataset configuration file, with different batch sizes for each resolution.
+
+Configuration file example:
+```toml
+[general]
+# Common settings
+flip_aug = true
+color_aug = false
+keep_tokens_separator= "|||"
+shuffle_caption = false
+caption_tag_dropout_rate = 0
+caption_extension = ".txt"
+
+[[datasets]]
+# First resolution settings
+batch_size = 2
+enable_bucket = true
+resolution = [1024, 1024]
+
+  [[datasets.subsets]]
+  image_dir = "path/to/image/directory"
+  num_repeats = 1
+
+[[datasets]]
+# Second resolution settings
+batch_size = 3
+enable_bucket = true
+resolution = [768, 768]
+
+  [[datasets.subsets]]
+  image_dir = "path/to/image/directory"
+  num_repeats = 1
+```
+
+<details>
+<summary>日本語</summary>
+
+データセット設定ファイルで複数の解像度を定義できます。各解像度に対して異なるバッチサイズを指定することができます。
+
+設定ファイルの例は英語のドキュメントを参照してください。
+
+</details>
+
+### 6.8. Validation / 検証
+
+You can calculate validation loss during training using a validation dataset to evaluate model generalization performance.
+
+To set up validation, add a `validation_split` and optionally `validation_seed` to your dataset configuration TOML file. 
+
+```toml
+validation_seed = 42 # [Optional] Validation seed, otherwise uses training seed for validation split .
+enable_bucket = true
+resolution = [1024, 1024]
+
+[[datasets]]
+  [[datasets.subsets]]
+  # This directory will use 100% of the images for training
+  image_dir = "path/to/image/directory"
+
+[[datasets]]
+validation_split = 0.1 # Split between 0.0 and 1.0 where 1.0 will use the full subset as a validation dataset
+
+  [[datasets.subsets]]
+  # This directory will split 10% to validation and 90% to training
+  image_dir = "path/to/image/second-directory"
+
+[[datasets]]
+validation_split = 1.0 # Will use this full subset as a validation subset. 
+
+  [[datasets.subsets]]
+  # This directory will use the 100% to validation and 0% to training
+  image_dir = "path/to/image/full_validation"
+```
+
+**Notes:**
+
+* Validation loss calculation uses fixed timestep sampling and random seeds to reduce loss variation due to randomness for more stable evaluation.
+* Currently, validation loss is not supported when using Schedule-Free optimizers (`AdamWScheduleFree`, `RAdamScheduleFree`, `ProdigyScheduleFree`).
+
+<details>
+<summary>日本語</summary>
+
+学習中に検証データセットを使用して損失 (Validation Loss) を計算し、モデルの汎化性能を評価できます。
+
+詳細な設定方法とコマンドラインの例は英語のドキュメントを参照してください。
+
+</details>
+
+## 7. Additional Options / 追加オプション
+
+### 7.1. Other FLUX.1-specific Options / その他のFLUX.1特有のオプション
+
+- **T5 Attention Mask Application**: Specify `--apply_t5_attn_mask` to apply attention masks during T5XXL Text Encoder training and inference. Not recommended due to limited inference environment support. **For Chroma models, this option is required.**
+
+- **IP Noise Gamma**: Use `--ip_noise_gamma` and `--ip_noise_gamma_random_strength` to adjust Input Perturbation noise gamma values during training. See Stable Diffusion 3 training options for details.
+
+- **LoRA-GGPO Support**: Use LoRA-GGPO (Gradient Group Proportion Optimizer) to stabilize LoRA training:
+  ```bash
+  --network_args "ggpo_sigma=0.03" "ggpo_beta=0.01"
+  ```
+
+- **Q/K/V Projection Layer Splitting [Experimental]**: Specify `--network_args "split_qkv=True"` to individually split and apply LoRA to Q/K/V (and SingleStreamBlock Text) projection layers within Attention layers.
+
+<details>
+<summary>日本語</summary>
+
+その他のFLUX.1特有のオプション：
+- T5 Attention Maskの適用（Chromaモデルでは必須）
+- IPノイズガンマ
+- LoRA-GGPOサポート
+- Q/K/V射影層の分割（実験的機能）
+
+詳細な設定方法とコマンドラインの例は英語のドキュメントを参照してください。
+
+</details>
+
+### 7.2. Dataset-related Additional Options / データセット関連の追加オプション
+
+#### Interpolation Method for Resizing
+
+You can specify the interpolation method when resizing dataset images to training resolution. Specify `interpolation_type` in the `[[datasets]]` or `[general]` section of the dataset configuration TOML file.
+
+Available values: `bicubic` (default), `bilinear`, `lanczos`, `nearest`, `area`
+
+```toml
+[[datasets]]
+resolution = [1024, 1024]
+enable_bucket = true
+interpolation_type = "lanczos" # Example: Use Lanczos interpolation
+# ...
+```
+
+<details>
+<summary>日本語</summary>
+
+データセットの画像を学習解像度にリサイズする際の補間方法を指定できます。
+
+設定方法とオプションの詳細は英語のドキュメントを参照してください。
+
+</details>
+
+### 7.3. Other Training Options / その他の学習オプション
+
+- **`--controlnet_model_name_or_path`**: Specifies the path to a ControlNet model compatible with FLUX.1. This allows for training a LoRA that works in conjunction with ControlNet. This is an advanced feature and requires a compatible ControlNet model.
+
+- **`--loss_type`**: Specifies the loss function for training. The default is `l2`.
+  - `l1`: L1 loss.
+  - `l2`: L2 loss (mean squared error).
+  - `huber`: Huber loss.
+  - `smooth_l1`: Smooth L1 loss.
+
+- **`--huber_schedule`**, **`--huber_c`**, **`--huber_scale`**: These are parameters for Huber loss. They are used when `--loss_type` is set to `huber` or `smooth_l1`.
+
+- **`--t5xxl_max_token_length`**: Specifies the maximum token length for the T5-XXL text encoder. For details, refer to the [`sd3_train_network.md` guide](sd3_train_network.md).
+
+- **`--weighting_scheme`**, **`--logit_mean`**, **`--logit_std`**, **`--mode_scale`**: These options allow you to adjust the loss weighting for each timestep. For details, refer to the [`sd3_train_network.md` guide](sd3_train_network.md).
+
+- **`--fused_backward_pass`**: Fuses the backward pass and optimizer step to reduce VRAM usage. For details, refer to the [`sdxl_train_network.md` guide](sdxl_train_network.md).
+
+<details>
+<summary>日本語</summary>
+
+- **`--controlnet_model_name_or_path`**: FLUX.1互換のControlNetモデルへのパスを指定します。これにより、ControlNetと連携して動作するLoRAを学習できます。これは高度な機能であり、互換性のあるControlNetモデルが必要です。
+- **`--loss_type`**: 学習に用いる損失関数を指定します。デフォルトは `l2` です。
+  - `l1`: L1損失。
+  - `l2`: L2損失（平均二乗誤差）。
+  - `huber`: Huber損失。
+  - `smooth_l1`: Smooth L1損失。
+- **`--huber_schedule`**, **`--huber_c`**, **`--huber_scale`**: これらはHuber損失のパラメータです。`--loss_type` が `huber` または `smooth_l1` の場合に使用されます。
+- **`--t5xxl_max_token_length`**: T5-XXLテキストエンコーダの最大トークン長を指定します。詳細は [`sd3_train_network.md` ガイド](sd3_train_network.md) を参照してください。
+- **`--weighting_scheme`**, **`--logit_mean`**, **`--logit_std`**, **`--mode_scale`**: これらのオプションは、各タイムステップの損失の重み付けを調整するために使用されます。詳細は [`sd3_train_network.md` ガイド](sd3_train_network.md) を参照してください。
+- **`--fused_backward_pass`**: バックワードパスとオプティマイザステップを融合してVRAM使用量を削減します。詳細は [`sdxl_train_network.md` ガイド](sdxl_train_network.md) を参照してください。
+
+</details>
+
+## 8. Related Tools / 関連ツール
+
+Several related scripts are provided for models trained with `flux_train_network.py` and to assist with the training process:
+
+* **`networks/flux_extract_lora.py`**: Extracts LoRA models from the difference between trained and base models.
+* **`convert_flux_lora.py`**: Converts trained LoRA models to other formats like Diffusers (AI-Toolkit) format. When trained with Q/K/V split option, converting with this script can reduce model size.
+* **`networks/flux_merge_lora.py`**: Merges trained LoRA models into FLUX.1 base models.
+* **`flux_minimal_inference.py`**: Simple inference script for generating images with trained LoRA models. You can specify `flux` or `chroma` with the `--model_type` argument.
+
+<details>
+<summary>日本語</summary>
+
+`flux_train_network.py` で学習したモデルや、学習プロセスに役立つ関連スクリプトが提供されています：
+
+* **`networks/flux_extract_lora.py`**: 学習済みモデルとベースモデルの差分から LoRA モデルを抽出。
+* **`convert_flux_lora.py`**: 学習した LoRA モデルを Diffusers (AI-Toolkit) 形式など他の形式に変換。
+* **`networks/flux_merge_lora.py`**: 学習した LoRA モデルを FLUX.1 ベースモデルにマージ。
+* **`flux_minimal_inference.py`**: 学習した LoRA モデルを適用して画像を生成するシンプルな推論スクリプト。
+  `--model_type` 引数で `flux` または `chroma` を指定できます。
+
+</details>
+
+## 9. Others / その他
+
+`flux_train_network.py` includes many features common with `train_network.py`, such as sample image generation (`--sample_prompts`, etc.) and detailed optimizer settings. For these features, refer to the [`train_network.py` guide](train_network.md#5-other-features--その他の機能) or the script help (`python flux_train_network.py --help`).
+
+<details>
+<summary>日本語</summary>
+
+`flux_train_network.py`には、サンプル画像の生成 (`--sample_prompts`など) や詳細なオプティマイザ設定など、`train_network.py`と共通の機能も多く存在します。これらについては、[`train_network.py`のガイド](train_network.md#5-other-features--その他の機能)やスクリプトのヘルプ (`python flux_train_network.py --help`) を参照してください。
+
+</details>

docs/gen_img_README-ja.md

@@ -1,30 +1,24 @@
-SD 1.xおよび2.xのモデル、当リポジトリで学習したLoRA、ControlNet（v1.0のみ動作確認）などに対応した、Diffusersベースの推論（画像生成）スクリプトです。コマンドラインから用います。
+SD 1.x、2.x、およびSDXLのモデル、当リポジトリで学習したLoRA、ControlNet、ControlNet-LLLiteなどに対応した、独自の推論（画像生成）スクリプトです。コマンドラインから用います。
 
 # 概要
 
-* Diffusers (v0.10.2) ベースの推論（画像生成）スクリプト。
-* SD 1.xおよび2.x (base/v-parameterization)モデルに対応。
+* 独自の推論（画像生成）スクリプト。
+* SD 1.x、2.x (base/v-parameterization)、およびSDXLモデルに対応。
 * txt2img、img2img、inpaintingに対応。
 * 対話モード、およびファイルからのプロンプト読み込み、連続生成に対応。
 * プロンプト1行あたりの生成枚数を指定可能。
 * 全体の繰り返し回数を指定可能。
 * `fp16`だけでなく`bf16`にも対応。
-* xformersに対応し高速生成が可能。
-    * xformersにより省メモリ生成を行いますが、Automatic 1111氏のWeb UIほど最適化していないため、512*512の画像生成でおおむね6GB程度のVRAMを使用します。
+* xformers、SDPA（Scaled Dot-Product Attention）に対応。
 * プロンプトの225トークンへの拡張。ネガティブプロンプト、重みづけに対応。
-* Diffusersの各種samplerに対応（Web UIよりもsampler数は少ないです）。
+* Diffusersの各種samplerに対応。
 * Text Encoderのclip skip（最後からn番目の層の出力を用いる）に対応。
-* VAEの別途読み込み。
-* CLIP Guided Stable Diffusion、VGG16 Guided Stable Diffusion、Highres. fix、upscale対応。
-    * Highres. fixはWeb UIの実装を全く確認していない独自実装のため、出力結果は異なるかもしれません。
-* LoRA対応。適用率指定、複数LoRA同時利用、重みのマージに対応。
-    * Text EncoderとU-Netで別の適用率を指定することはできません。
-* Attention Coupleに対応。
-* ControlNet v1.0に対応。
+* VAEの別途読み込み、VAEのバッチ処理やスライスによる省メモリ化に対応。
+* Highres. fix（独自実装およびGradual Latent）、upscale対応。
+* LoRA、DyLoRA対応。適用率指定、複数LoRA同時利用、重みのマージに対応。
+* Attention Couple、Regional LoRAに対応。
+* ControlNet (v1.0/v1.1)、ControlNet-LLLiteに対応。
 * 途中でモデルを切り替えることはできませんが、バッチファイルを組むことで対応できます。
-* 個人的に欲しくなった機能をいろいろ追加。
-
-機能追加時にすべてのテストを行っているわけではないため、以前の機能に影響が出て一部機能が動かない可能性があります。何か問題があればお知らせください。
 
 # 基本的な使い方
 
@@ -33,18 +27,20 @@ SD 1.xおよび2.xのモデル、当リポジトリで学習したLoRA、Control
 以下のように入力してください。
 
 ```batchfile
-python gen_img_diffusers.py --ckpt <モデル名> --outdir <画像出力先> --xformers --fp16 --interactive
+python gen_img.py --ckpt <モデル名> --outdir <画像出力先> --xformers --fp16 --interactive
 ```
 
 `--ckpt`オプションにモデル（Stable Diffusionのcheckpointファイル、またはDiffusersのモデルフォルダ）、`--outdir`オプションに画像の出力先フォルダを指定します。
 
-`--xformers`オプションでxformersの使用を指定します（xformersを使わない場合は外してください）。`--fp16`オプションでfp16（単精度）での推論を行います。RTX 30系のGPUでは `--bf16`オプションでbf16（bfloat16）での推論を行うこともできます。
+`--xformers`オプションでxformersの使用を指定します。`--fp16`オプションでfp16（半精度）での推論を行います。RTX 30系以降のGPUでは `--bf16`オプションでbf16（bfloat16）での推論を行うこともできます。
 
 `--interactive`オプションで対話モードを指定しています。
 
 Stable Diffusion 2.0（またはそこからの追加学習モデル）を使う場合は`--v2`オプションを追加してください。v-parameterizationを使うモデル（`768-v-ema.ckpt`およびそこからの追加学習モデル）を使う場合はさらに`--v_parameterization`を追加してください。
 
-`--v2`の指定有無が間違っているとモデル読み込み時にエラーになります。`--v_parameterization`の指定有無が間違っていると茶色い画像が表示されます。
+SDXLモデルを使う場合は`--sdxl`オプションを追加してください。
+
+`--v2`や`--sdxl`の指定有無が間違っているとモデル読み込み時にエラーになります。`--v_parameterization`の指定有無が間違っていると茶色い画像が表示されます。
 
 `Type prompt:`と表示されたらプロンプトを入力してください。
 
@@ -59,7 +55,7 @@ Stable Diffusion 2.0（またはそこからの追加学習モデル）を使う
 以下のように入力します（実際には1行で入力します）。
 
 ```batchfile
-python gen_img_diffusers.py --ckpt <モデル名> --outdir <画像出力先> 
+python gen_img.py --ckpt <モデル名> --outdir <画像出力先> 
     --xformers --fp16 --images_per_prompt <生成枚数> --prompt "<プロンプト>"
 ```
 
@@ -72,7 +68,7 @@ python gen_img_diffusers.py --ckpt <モデル名> --outdir <画像出力先>
 以下のように入力します。
 
 ```batchfile
-python gen_img_diffusers.py --ckpt <モデル名> --outdir <画像出力先> 
+python gen_img.py --ckpt <モデル名> --outdir <画像出力先> 
     --xformers --fp16 --from_file <プロンプトファイル名>
 ```
 
@@ -96,13 +92,29 @@ python gen_img_diffusers.py --ckpt <モデル名> --outdir <画像出力先>
 
 - `--ckpt <モデル名>`：モデル名を指定します。`--ckpt`オプションは必須です。Stable Diffusionのcheckpointファイル、またはDiffusersのモデルフォルダ、Hugging FaceのモデルIDを指定できます。
 
+- `--v1`：Stable Diffusion 1.x系のモデルを使う場合に指定します。これがデフォルトの動作です。
+
 - `--v2`：Stable Diffusion 2.x系のモデルを使う場合に指定します。1.x系の場合には指定不要です。
 
+- `--sdxl`：Stable Diffusion XLモデルを使う場合に指定します。
+
 - `--v_parameterization`：v-parameterizationを使うモデルを使う場合に指定します（`768-v-ema.ckpt`およびそこからの追加学習モデル、Waifu Diffusion v1.5など）。
     
-    `--v2`の指定有無が間違っているとモデル読み込み時にエラーになります。`--v_parameterization`の指定有無が間違っていると茶色い画像が表示されます。
+    `--v2`や`--sdxl`の指定有無が間違っているとモデル読み込み時にエラーになります。`--v_parameterization`の指定有無が間違っていると茶色い画像が表示されます。
 
-- `--vae`：使用するVAEを指定します。未指定時はモデル内のVAEを使用します。
+- `--zero_terminal_snr`：noise schedulerのbetasを修正して、zero terminal SNRを強制します。
+
+- `--pyramid_noise_prob`：ピラミッドノイズを適用する確率を指定します。
+
+- `--pyramid_noise_discount_range`：ピラミッドノイズの割引率の範囲を指定します。
+
+- `--noise_offset_prob`：ノイズオフセットを適用する確率を指定します。
+
+- `--noise_offset_range`：ノイズオフセットの範囲を指定します。
+
+- `--vae`：使用する VAE を指定します。未指定時はモデル内の VAE を使用します。
+
+- `--tokenizer_cache_dir`：トークナイザーのキャッシュディレクトリを指定します（オフライン利用のため）。
 
 ## 画像生成と出力
 
@@ -112,6 +124,10 @@ python gen_img_diffusers.py --ckpt <モデル名> --outdir <画像出力先>
 
 - `--from_file <プロンプトファイル名>`：プロンプトが記述されたファイルを指定します。1行1プロンプトで記述してください。なお画像サイズやguidance scaleはプロンプトオプション（後述）で指定できます。
 
+- `--from_module <モジュールファイル>`：Pythonモジュールからプロンプトを読み込みます。モジュールは`get_prompter(args, pipe, networks)`関数を実装している必要があります。
+
+- `--prompter_module_args`：prompterモジュールに渡す追加の引数を指定します。
+
 - `--W <画像幅>`：画像の幅を指定します。デフォルトは`512`です。
 
 - `--H <画像高さ>`：画像の高さを指定します。デフォルトは`512`です。
@@ -120,30 +136,59 @@ python gen_img_diffusers.py --ckpt <モデル名> --outdir <画像出力先>
 
 - `--scale <ガイダンススケール>`：unconditionalガイダンススケールを指定します。デフォルトは`7.5`です。
 
-- `--sampler <サンプラー名>`：サンプラーを指定します。デフォルトは`ddim`です。Diffusersで提供されているddim、pndm、dpmsolver、dpmsolver+++、lms、euler、euler_a、が指定可能です（後ろの三つはk_lms、k_euler、k_euler_aでも指定できます）。
+- `--sampler <サンプラー名>`：サンプラーを指定します。デフォルトは`ddim`です。
+    `ddim`, `pndm`, `lms`, `euler`, `euler_a`, `heun`, `dpm_2`, `dpm_2_a`, `dpmsolver`, `dpmsolver++`, `dpmsingle`, `k_lms`, `k_euler`, `k_euler_a`, `k_dpm_2`, `k_dpm_2_a` が指定可能です。
 
 - `--outdir <画像出力先フォルダ>`：画像の出力先を指定します。
 
 - `--images_per_prompt <生成枚数>`：プロンプト1件当たりの生成枚数を指定します。デフォルトは`1`です。
 
-- `--clip_skip <スキップ数>`：CLIPの後ろから何番目の層を使うかを指定します。省略時は最後の層を使います。
+- `--clip_skip <スキップ数>`：CLIPの後ろから何番目の層を使うかを指定します。デフォルトはSD1/2の場合1、SDXLの場合2です。
 
 - `--max_embeddings_multiples <倍数>`：CLIPの入出力長をデフォルト（75）の何倍にするかを指定します。未指定時は75のままです。たとえば3を指定すると入出力長が225になります。
 
 - `--negative_scale` : uncoditioningのguidance scaleを個別に指定します。[gcem156氏のこちらの記事](https://note.com/gcem156/n/ne9a53e4a6f43)を参考に実装したものです。
 
+- `--emb_normalize_mode`：embedding正規化モードを指定します。"original"（デフォルト）、"abs"、"none"から選択できます。プロンプトの重みの正規化方法に影響します。
+
+- `--force_scheduler_zero_steps_offset`：スケジューラのステップオフセットを、スケジューラ設定の `steps_offset` の値に関わらず強制的にゼロにします。
+
+## SDXL固有のオプション
+
+SDXL モデル（`--sdxl`フラグ付き）を使用する場合、追加のコンディショニングオプションが利用できます：
+
+- `--original_height`：SDXL コンディショニング用の元の高さを指定します。これはモデルの対象解像度の理解に影響します。
+
+- `--original_width`：SDXL コンディショニング用の元の幅を指定します。これはモデルの対象解像度の理解に影響します。
+
+- `--original_height_negative`：SDXL ネガティブコンディショニング用の元の高さを指定します。
+
+- `--original_width_negative`：SDXL ネガティブコンディショニング用の元の幅を指定します。
+
+- `--crop_top`：SDXL コンディショニング用のクロップ上オフセットを指定します。
+
+- `--crop_left`：SDXL コンディショニング用のクロップ左オフセットを指定します。
+
 ## メモリ使用量や生成速度の調整
 
 - `--batch_size <バッチサイズ>`：バッチサイズを指定します。デフォルトは`1`です。バッチサイズが大きいとメモリを多く消費しますが、生成速度が速くなります。
 
-- `--vae_batch_size <VAEのバッチサイズ>`：VAEのバッチサイズを指定します。デフォルトはバッチサイズと同じです。
+- `--vae_batch_size <VAEのバッチサイズ>`：VAEのバッチサイズを指定します。デフォルトはバッチサイズと同じです。1未満の値を指定すると、バッチサイズに対する比率として扱われます。
     VAEのほうがメモリを多く消費するため、デノイジング後（stepが100%になった後）でメモリ不足になる場合があります。このような場合にはVAEのバッチサイズを小さくしてください。
 
+- `--vae_slices <スライス数>`：VAE処理時に画像をスライスに分割してVRAM使用量を削減します。None（デフォルト）で分割なし。16や32のような値が推奨されます。有効にすると処理が遅くなりますが、VRAM使用量が少なくなります。
+
+- `--no_half_vae`：VAE処理でfp16/bf16精度の使用を防ぎます。代わりにfp32を使用します。VAE関連の問題やアーティファクトが発生した場合に使用してください。
+
 - `--xformers`：xformersを使う場合に指定します。
 
-- `--fp16`：fp16（単精度）での推論を行います。`fp16`と`bf16`をどちらも指定しない場合はfp32（単精度）での推論を行います。
+- `--sdpa`：最適化のためにPyTorch 2のscaled dot-product attentionを使用します。
 
-- `--bf16`：bf16（bfloat16）での推論を行います。RTX 30系のGPUでのみ指定可能です。`--bf16`オプションはRTX 30系以外のGPUではエラーになります。`fp16`よりも`bf16`のほうが推論結果がNaNになる（真っ黒の画像になる）可能性が低いようです。
+- `--diffusers_xformers`：Diffusers経由でxformersを使用します（注：Hypernetworksと互換性がありません）。
+
+- `--fp16`：fp16（半精度）での推論を行います。`fp16`と`bf16`をどちらも指定しない場合はfp32（単精度）での推論を行います。
+
+- `--bf16`：bf16（bfloat16）での推論を行います。RTX 30系以降のGPUでのみ指定可能です。`--bf16`オプションはRTX 30系以外のGPUではエラーになります。SDXLでは`fp16`よりも`bf16`のほうが推論結果がNaNになる（真っ黒の画像になる）可能性が低いようです。
 
 ## 追加ネットワーク（LoRA等）の使用
 
@@ -157,12 +202,18 @@ python gen_img_diffusers.py --ckpt <モデル名> --outdir <画像出力先>
 
 - `--network_pre_calc`：使用する追加ネットワークの重みを生成ごとにあらかじめ計算します。プロンプトオプションの`--am`が使用できます。LoRA未使用時と同じ程度まで生成は高速化されますが、生成前に重みを計算する時間が必要で、またメモリ使用量も若干増加します。Regional LoRA使用時は無効になります 。
 
+- `--network_regional_mask_max_color_codes`：リージョナルマスクに使用する色コードの最大数を指定します。指定されていない場合、マスクはチャンネルごとに適用されます。Regional LoRAと組み合わせて、マスク内の色で定義できるリージョン数を制御するために使用されます。
+
+- `--network_args`：key=value形式でネットワークモジュールに渡す追加引数を指定します。例: `--network_args "alpha=1.0,dropout=0.1"`。
+
+- `--network_merge_n_models`：ネットワークマージを使用する場合、マージするモデル数を指定します（全ての読み込み済みネットワークをマージする代わりに）。
+
 # 主なオプションの指定例
 
 次は同一プロンプトで64枚をバッチサイズ4で一括生成する例です。
 
 ```batchfile
-python gen_img_diffusers.py --ckpt model.ckpt --outdir outputs 
+python gen_img.py --ckpt model.ckpt --outdir outputs 
     --xformers --fp16 --W 512 --H 704 --scale 12.5 --sampler k_euler_a 
     --steps 32 --batch_size 4 --images_per_prompt 64 
     --prompt "beautiful flowers --n monochrome"
@@ -171,7 +222,7 @@ python gen_img_diffusers.py --ckpt model.ckpt --outdir outputs
 次はファイルに書かれたプロンプトを、それぞれ10枚ずつ、バッチサイズ4で一括生成する例です。
 
 ```batchfile
-python gen_img_diffusers.py --ckpt model.ckpt --outdir outputs 
+python gen_img.py --ckpt model.ckpt --outdir outputs 
     --xformers --fp16 --W 512 --H 704 --scale 12.5 --sampler k_euler_a 
     --steps 32 --batch_size 4 --images_per_prompt 10 
     --from_file prompts.txt
@@ -180,7 +231,7 @@ python gen_img_diffusers.py --ckpt model.ckpt --outdir outputs
 Textual Inversion（後述）およびLoRAの使用例です。
 
 ```batchfile
-python gen_img_diffusers.py --ckpt model.safetensors 
+python gen_img.py --ckpt model.safetensors 
     --scale 8 --steps 48 --outdir txt2img --xformers 
     --W 512 --H 768 --fp16 --sampler k_euler_a 
     --textual_inversion_embeddings goodembed.safetensors negprompt.pt 
@@ -216,6 +267,22 @@ python gen_img_diffusers.py --ckpt model.safetensors
 
 - `--am`：追加ネットワークの重みを指定します。コマンドラインからの指定を上書きします。複数の追加ネットワークを使用する場合は`--am 0.8,0.5,0.3`のように __カンマ区切りで__ 指定します。
 
+- `--ow`：SDXLのoriginal_widthを指定します。
+
+- `--oh`：SDXLのoriginal_heightを指定します。
+
+- `--nw`：SDXLのoriginal_width_negativeを指定します。
+
+- `--nh`：SDXLのoriginal_height_negativeを指定します。
+
+- `--ct`：SDXLのcrop_topを指定します。
+
+- `--cl`：SDXLのcrop_leftを指定します。
+
+- `--c`：CLIPプロンプトを指定します。
+
+- `--f`：生成ファイル名を指定します。
+
 ※これらのオプションを指定すると、バッチサイズよりも小さいサイズでバッチが実行される場合があります（これらの値が異なると一括生成できないため）。（あまり気にしなくて大丈夫ですが、ファイルからプロンプトを読み込み生成する場合は、これらの値が同一のプロンプトを並べておくと効率が良くなります。）
 
 例：
@@ -225,6 +292,21 @@ python gen_img_diffusers.py --ckpt model.safetensors
 
 ![image](https://user-images.githubusercontent.com/52813779/235343446-25654172-fff4-4aaf-977a-20d262b51676.png)
 
+# プロンプトのワイルドカード (Dynamic Prompts)
+
+Dynamic Prompts (Wildcard) 記法に対応しています。Web UIの拡張機能等と完全に同じではありませんが、以下の機能が利用可能です。
+
+- `{A|B|C}` : A, B, C の中からランダムに1つを選択します。
+- `{e$$A|B|C}` : A, B, C のすべてを順に利用します（全列挙）。プロンプト内に複数の `{e$$...}` がある場合、すべての組み合わせが生成されます。
+  - 例：`{e$$red|blue} flower, {e$$1girl|2girls}` → `red flower, 1girl`, `red flower, 2girls`, `blue flower, 1girl`, `blue flower, 2girls` の4枚が生成されます。
+- `{n$$A|B|C}` : A, B, C の中から n 個をランダムに選択して結合します。
+  - 例：`{2$$A|B|C}` → `A, B` や `B, C` など。
+- `{n-m$$A|B|C}` : A, B, C の中から n 個から m 個をランダムに選択して結合します。
+- `{$$sep$$A|B|C}` : 選択された項目を sep で結合します（デフォルトは `, `）。
+  - 例：`{2$$ and $$A|B|C}` → `A and B` など。
+
+これらは組み合わせて利用可能です。
+
 # img2img
 
 ## オプション
@@ -235,12 +317,14 @@ python gen_img_diffusers.py --ckpt model.safetensors
 
 - `--sequential_file_name`：ファイル名を連番にするかどうかを指定します。指定すると生成されるファイル名が`im_000001.png`からの連番になります。
 
-- `--use_original_file_name`：指定すると生成ファイル名がオリジナルのファイル名と同じになります。
+- `--use_original_file_name`：指定すると生成ファイル名がオリジナルのファイル名の前に追加されます（img2imgモード用）。
+
+- `--clip_vision_strength`：指定した強度でimg2img用のCLIP Vision Conditioningを有効にします。CLIP Visionモデルを使用して入力画像からのコンディショニングを強化します。
 
 ## コマンドラインからの実行例
 
 ```batchfile
-python gen_img_diffusers.py --ckpt trinart_characters_it4_v1_vae_merged.ckpt 
+python gen_img.py --ckpt trinart_characters_it4_v1_vae_merged.ckpt 
     --outdir outputs --xformers --fp16 --scale 12.5 --sampler k_euler --steps 32 
     --image_path template.png --strength 0.8 
     --prompt "1girl, cowboy shot, brown hair, pony tail, brown eyes, 
@@ -281,10 +365,6 @@ img2img時にコマンドラインオプションの`--W`と`--H`で生成画像
 
 モデルとして、当リポジトリで学習したTextual Inversionモデル、およびWeb UIで学習したTextual Inversionモデル（画像埋め込みは非対応）を利用できます
 
-## Extended Textual Inversion
-
-`--textual_inversion_embeddings`の代わりに`--XTI_embeddings`オプションを指定してください。使用法は`--textual_inversion_embeddings`と同じです。
-
 ## Highres. fix
 
 AUTOMATIC1111氏のWeb UIにある機能の類似機能です（独自実装のためもしかしたらいろいろ異なるかもしれません）。最初に小さめの画像を生成し、その画像を元にimg2imgすることで、画像全体の破綻を防ぎつつ大きな解像度の画像を生成します。
@@ -299,6 +379,8 @@ img2imgと併用できません。
 
 - `--highres_fix_steps`：1st stageの画像のステップ数を指定します。デフォルトは`28`です。
 
+- `--highres_fix_strength`：1st stageのimg2img時のstrengthを指定します。省略時は`--strength`と同じ値になります。
+
 - `--highres_fix_save_1st`：1st stageの画像を保存するかどうかを指定します。
 
 - `--highres_fix_latents_upscaling`：指定すると2nd stageの画像生成時に1st stageの画像をlatentベースでupscalingします（bilinearのみ対応）。未指定時は画像をLANCZOS4でupscalingします。
@@ -306,12 +388,14 @@ img2imgと併用できません。
 - `--highres_fix_upscaler`：2nd stageに任意のupscalerを利用します。現在は`--highres_fix_upscaler tools.latent_upscaler` のみ対応しています。
 
 - `--highres_fix_upscaler_args`：`--highres_fix_upscaler`で指定したupscalerに渡す引数を指定します。
-    `tools.latent_upscaler`の場合は、`--highres_fix_upscaler_args "weights=D:\Work\SD\Models\others\etc\upscaler-v1-e100-220.safetensors"`のように重みファイルを指定します。 
+    `tools.latent_upscaler`の場合は、`--highres_fix_upscaler_args "weights=D:\Work\SD\Models\others\etc\upscaler-v1-e100-220.safetensors"`のように重みファイルを指定します。
+
+- `--highres_fix_disable_control_net`：Highres fixの2nd stageでControlNetを無効にします。デフォルトでは、ControlNetは両ステージで使用されます。
 
 コマンドラインの例です。
 
 ```batchfile
-python gen_img_diffusers.py  --ckpt trinart_characters_it4_v1_vae_merged.ckpt
+python gen_img.py  --ckpt trinart_characters_it4_v1_vae_merged.ckpt
     --n_iter 1 --scale 7.5 --W 1024 --H 1024 --batch_size 1 --outdir ../txt2img 
     --steps 48 --sampler ddim --fp16 
     --xformers 
@@ -319,6 +403,34 @@ python gen_img_diffusers.py  --ckpt trinart_characters_it4_v1_vae_merged.ckpt
     --highres_fix_scale 0.5 --highres_fix_steps 28 --strength 0.5
 ```
 
+## Deep Shrink
+
+Deep Shrinkは、異なるタイムステップで異なる深度のUNetを使用して生成プロセスを最適化する技術です。生成品質と効率を向上させることができます。
+
+以下のオプションがあります：
+
+- `--ds_depth_1`：第1フェーズでこの深度のDeep Shrinkを有効にします。有効な値は0から8です。
+
+- `--ds_timesteps_1`：このタイムステップまでDeep Shrink深度1を適用します。デフォルトは650です。
+
+- `--ds_depth_2`：Deep Shrinkの第2フェーズの深度を指定します。
+
+- `--ds_timesteps_2`：このタイムステップまでDeep Shrink深度2を適用します。デフォルトは650です。
+
+- `--ds_ratio`：Deep Shrinkでのダウンサンプリングの比率を指定します。デフォルトは0.5です。
+
+これらのパラメータはプロンプトオプションでも指定できます：
+
+- `--dsd1`：プロンプトからDeep Shrink深度1を指定します。
+  
+- `--dst1`：プロンプトからDeep Shrinkタイムステップ1を指定します。
+  
+- `--dsd2`：プロンプトからDeep Shrink深度2を指定します。
+  
+- `--dst2`：プロンプトからDeep Shrinkタイムステップ2を指定します。
+  
+- `--dsr`：プロンプトからDeep Shrink比率を指定します。
+
 ## ControlNet
 
 現在はControlNet 1.0のみ動作確認しています。プリプロセスはCannyのみサポートしています。
@@ -333,19 +445,33 @@ python gen_img_diffusers.py  --ckpt trinart_characters_it4_v1_vae_merged.ckpt
 - `--control_net_preps`：ControlNetのプリプロセスを指定します。`--control_net_models`と同様に複数指定可能です。現在はcannyのみ対応しています。対象モデルでプリプロセスを使用しない場合は `none` を指定します。
    cannyの場合 `--control_net_preps canny_63_191`のように、閾値1と2を'_'で区切って指定できます。
 
-- `--control_net_weights`：ControlNetの適用時の重みを指定します（`1.0`で通常、`0.5`なら半分の影響力で適用）。`--control_net_models`と同様に複数指定可能です。
+- `--control_net_multipliers`：ControlNetの適用時の重みを指定します（`1.0`で通常、`0.5`なら半分の影響力で適用）。`--control_net_models`と同様に複数指定可能です。
 
 - `--control_net_ratios`：ControlNetを適用するstepの範囲を指定します。`0.5`の場合は、step数の半分までControlNetを適用します。`--control_net_models`と同様に複数指定可能です。
 
 コマンドラインの例です。
 
 ```batchfile
-python gen_img_diffusers.py --ckpt model_ckpt --scale 8 --steps 48 --outdir txt2img --xformers 
+python gen_img.py --ckpt model_ckpt --scale 8 --steps 48 --outdir txt2img --xformers 
     --W 512 --H 768 --bf16 --sampler k_euler_a 
-    --control_net_models diff_control_sd15_canny.safetensors --control_net_weights 1.0 
+    --control_net_models diff_control_sd15_canny.safetensors --control_net_multipliers 1.0 
     --guide_image_path guide.png --control_net_ratios 1.0 --interactive
 ```
 
+## ControlNet-LLLite
+
+ControlNet-LLLiteは、類似の誘導目的に使用できるControlNetの軽量な代替手段です。
+
+以下のオプションがあります：
+
+- `--control_net_lllite_models`：ControlNet-LLLiteモデルファイルを指定します。
+
+- `--control_net_multipliers`：ControlNet-LLLiteの倍率を指定します（重みに類似）。
+
+- `--control_net_ratios`：ControlNet-LLLiteを適用するステップの比率を指定します。
+
+注意：ControlNetとControlNet-LLLiteは同時に使用できません。
+
 ## Attention Couple + Reginal LoRA
 
 プロンプトをいくつかの部分に分割し、それぞれのプロンプトを画像内のどの領域に適用するかを指定できる機能です。個別のオプションはありませんが、`mask_path`とプロンプトで指定します。
@@ -370,70 +496,6 @@ ControlNetと組み合わせることも可能です（細かい位置指定に
 
 LoRAを指定すると、`--network_weights`で指定した複数のLoRAがそれぞれANDの各部分に対応します。現在の制約として、LoRAの数はANDの部分の数と同じである必要があります。
 
-## CLIP Guided Stable Diffusion
-
-DiffusersのCommunity Examplesの[こちらのcustom pipeline](https://github.com/huggingface/diffusers/blob/main/examples/community/README.md#clip-guided-stable-diffusion)からソースをコピー、変更したものです。
-
-通常のプロンプトによる生成指定に加えて、追加でより大規模のCLIPでプロンプトのテキストの特徴量を取得し、生成中の画像の特徴量がそのテキストの特徴量に近づくよう、生成される画像をコントロールします（私のざっくりとした理解です）。大きめのCLIPを使いますのでVRAM使用量はかなり増加し（VRAM 8GBでは512*512でも厳しいかもしれません）、生成時間も掛かります。
-
-なお選択できるサンプラーはDDIM、PNDM、LMSのみとなります。
-
-`--clip_guidance_scale`オプションにどの程度、CLIPの特徴量を反映するかを数値で指定します。先のサンプルでは100になっていますので、そのあたりから始めて増減すると良いようです。
-
-デフォルトではプロンプトの先頭75トークン（重みづけの特殊文字を除く）がCLIPに渡されます。プロンプトの`--c`オプションで、通常のプロンプトではなく、CLIPに渡すテキストを別に指定できます（たとえばCLIPはDreamBoothのidentifier（識別子）や「1girl」などのモデル特有の単語は認識できないと思われますので、それらを省いたテキストが良いと思われます）。
-
-コマンドラインの例です。
-
-```batchfile
-python gen_img_diffusers.py  --ckpt v1-5-pruned-emaonly.ckpt --n_iter 1 
-    --scale 2.5 --W 512 --H 512 --batch_size 1 --outdir ../txt2img --steps 36  
-    --sampler ddim --fp16 --opt_channels_last --xformers --images_per_prompt 1  
-    --interactive --clip_guidance_scale 100
-```
-
-## CLIP Image Guided Stable Diffusion
-
-テキストではなくCLIPに別の画像を渡し、その特徴量に近づくよう生成をコントロールする機能です。`--clip_image_guidance_scale`オプションで適用量の数値を、`--guide_image_path`オプションでguideに使用する画像（ファイルまたはフォルダ）を指定してください。
-
-コマンドラインの例です。
-
-```batchfile
-python gen_img_diffusers.py  --ckpt trinart_characters_it4_v1_vae_merged.ckpt
-    --n_iter 1 --scale 7.5 --W 512 --H 512 --batch_size 1 --outdir ../txt2img 
-    --steps 80 --sampler ddim --fp16 --opt_channels_last --xformers 
-    --images_per_prompt 1  --interactive  --clip_image_guidance_scale 100 
-    --guide_image_path YUKA160113420I9A4104_TP_V.jpg
-```
-
-### VGG16 Guided Stable Diffusion
-
-指定した画像に近づくように画像生成する機能です。通常のプロンプトによる生成指定に加えて、追加でVGG16の特徴量を取得し、生成中の画像が指定したガイド画像に近づくよう、生成される画像をコントロールします。img2imgでの使用をお勧めします（通常の生成では画像がぼやけた感じになります）。CLIP Guided Stable Diffusionの仕組みを流用した独自の機能です。またアイデアはVGGを利用したスタイル変換から拝借しています。
-
-なお選択できるサンプラーはDDIM、PNDM、LMSのみとなります。
-
-`--vgg16_guidance_scale`オプションにどの程度、VGG16特徴量を反映するかを数値で指定します。試した感じでは100くらいから始めて増減すると良いようです。`--guide_image_path`オプションでguideに使用する画像（ファイルまたはフォルダ）を指定してください。
-
-複数枚の画像を一括でimg2img変換し、元画像をガイド画像とする場合、`--guide_image_path`と`--image_path`に同じ値を指定すればOKです。
-
-コマンドラインの例です。
-
-```batchfile
-python gen_img_diffusers.py --ckpt wd-v1-3-full-pruned-half.ckpt 
-    --n_iter 1 --scale 5.5 --steps 60 --outdir ../txt2img 
-    --xformers --sampler ddim --fp16 --W 512 --H 704 
-    --batch_size 1 --images_per_prompt 1 
-    --prompt "picturesque, 1girl, solo, anime face, skirt, beautiful face 
-        --n lowres, bad anatomy, bad hands, error, missing fingers, 
-        cropped, worst quality, low quality, normal quality, 
-        jpeg artifacts, blurry, 3d, bad face, monochrome --d 1" 
-    --strength 0.8 --image_path ..\src_image
-    --vgg16_guidance_scale 100 --guide_image_path ..\src_image 
-```
-
-`--vgg16_guidance_layerPで特徴量取得に使用するVGG16のレイヤー番号を指定できます（デフォルトは20でconv4-2のReLUです）。上の層ほど画風を表現し、下の層ほどコンテンツを表現するといわれています。
-
-![image](https://user-images.githubusercontent.com/52813779/235343813-3c1f0d7a-4fb3-4274-98e4-b92d76b551df.png)
-
 # その他のオプション
 
 - `--no_preview` : 対話モードでプレビュー画像を表示しません。OpenCVがインストールされていない場合や、出力されたファイルを直接確認する場合に指定してください。
@@ -450,5 +512,26 @@ python gen_img_diffusers.py --ckpt wd-v1-3-full-pruned-half.ckpt
 
 - `--opt_channels_last` : 推論時にテンソルのチャンネルを最後に配置します。場合によっては高速化されることがあります。
 
-- `--network_show_meta` : 追加ネットワークのメタデータを表示します。
+- `--shuffle_prompts`：繰り返し時にプロンプトの順序をシャッフルします。`--from_file`で複数のプロンプトを使用する場合に便利です。
+
+- `--network_show_meta`：追加ネットワークのメタデータを表示します。
+
+--- 
+
+# Gradual Latent について
+
+latentのサイズを徐々に大きくしていくHires fixです。
+
+- `--gradual_latent_timesteps` : latentのサイズを大きくし始めるタイムステップを指定します。デフォルトは None で、Gradual Latentを使用しません。750 くらいから始めてみてください。
+- `--gradual_latent_ratio` : latentの初期サイズを指定します。デフォルトは 0.5 で、デフォルトの latent サイズの半分のサイズから始めます。
+- `--gradual_latent_ratio_step`: latentのサイズを大きくする割合を指定します。デフォルトは 0.125 で、latentのサイズを 0.625, 0.75, 0.875, 1.0 と徐々に大きくします。
+- `--gradual_latent_ratio_every_n_steps`: latentのサイズを大きくする間隔を指定します。デフォルトは 3 で、3ステップごとに latent のサイズを大きくします。
+- `--gradual_latent_s_noise`：Gradual LatentのS_noiseパラメータを指定します。デフォルトは1.0です。
+- `--gradual_latent_unsharp_params`：Gradual Latentのアンシャープマスクパラメータをksize,sigma,strength,target-x形式で指定します（target-x: 1=True, 0=False）。推奨値：`3,0.5,0.5,1`または`3,1.0,1.0,0`。
+
+それぞれのオプションは、プロンプトオプション、`--glt`、`--glr`、`--gls`、`--gle` でも指定できます。
+
+サンプラーに手を加えているため、__サンプラーに `euler_a` を指定してください。__ 他のサンプラーでは動作しません。
+
+SD 1.5 のほうが効果があります。SDXL ではかなり微妙です。
 

docs/gen_img_README.md

@@ -0,0 +1,552 @@
+<!-- filepath: d:\\Work\\SD\\dev\\sd-scripts\\docs\\gen_img_README-en.md -->
+This is an inference (image generation) script that supports SD 1.x and 2.x models, LoRA trained with this repository, ControlNet (only v1.0 has been confirmed to work), etc. It is used from the command line.
+
+# Overview
+
+* Inference (image generation) script.
+* Supports SD 1.x, 2.x (base/v-parameterization), and SDXL models.
+* Supports txt2img, img2img, and inpainting.
+* Supports interactive mode, prompt reading from files, and continuous generation.
+* The number of images generated per prompt line can be specified.
+* The total number of repetitions can be specified.
+* Supports not only `fp16` but also `bf16`.
+* Supports xformers and SDPA (Scaled Dot-Product Attention).
+* Extension of prompts to 225 tokens. Supports negative prompts and weighting.
+* Supports various samplers from Diffusers.
+* Supports clip skip (uses the output of the nth layer from the end) of Text Encoder.
+* Separate loading of VAE, supports VAE batch processing and slicing for memory saving.
+* Highres. fix (original implementation and Gradual Latent), upscale support.
+* LoRA, DyLoRA support. Supports application rate specification, simultaneous use of multiple LoRAs, and weight merging.
+* Supports Attention Couple, Regional LoRA.
+* Supports ControlNet (v1.0/v1.1), ControlNet-LLLite.
+* It is not possible to switch models midway, but it can be handled by creating a batch file.
+
+# Basic Usage
+
+## Image Generation in Interactive Mode
+
+Enter as follows:
+
+```batchfile
+python gen_img.py --ckpt <model_name> --outdir <image_output_destination> --xformers --fp16 --interactive
+```
+
+Specify the model (Stable Diffusion checkpoint file or Diffusers model folder) in the `--ckpt` option and the image output destination folder in the `--outdir` option.
+
+Specify the use of xformers with the `--xformers` option (remove it if you do not use xformers). The `--fp16` option performs inference in fp16 (single precision). For RTX 30 series GPUs, you can also perform inference in bf16 (bfloat16) with the `--bf16` option.
+
+The `--interactive` option specifies interactive mode.
+
+If you are using Stable Diffusion 2.0 (or a model with additional training from it), add the `--v2` option. If you are using a model that uses v-parameterization (`768-v-ema.ckpt` and models with additional training from it), add `--v_parameterization` as well.
+
+If the `--v2` specification is incorrect, an error will occur when loading the model. If the `--v_parameterization` specification is incorrect, a brown image will be displayed.
+
+When `Type prompt:` is displayed, enter the prompt.
+
+![image](https://user-images.githubusercontent.com/52813779/235343115-f3b8ac82-456d-4aab-9724-0cc73c4534aa.png)
+
+*If the image is not displayed and an error occurs, headless (no screen display function) OpenCV may be installed. Install normal OpenCV with `pip install opencv-python`. Alternatively, stop image display with the `--no_preview` option.
+
+Select the image window and press any key to close the window and enter the next prompt. Press Ctrl+Z and then Enter in the prompt to close the script.
+
+## Batch Generation of Images with a Single Prompt
+
+Enter as follows (actually entered on one line):
+
+```batchfile
+python gen_img.py --ckpt <model_name> --outdir <image_output_destination> \
+    --xformers --fp16 --images_per_prompt <number_of_images_to_generate> --prompt "<prompt>"
+```
+
+Specify the number of images to generate per prompt with the `--images_per_prompt` option. Specify the prompt with the `--prompt` option. If it contains spaces, enclose it in double quotes.
+
+You can specify the batch size with the `--batch_size` option (described later).
+
+## Batch Generation by Reading Prompts from a File
+
+Enter as follows:
+
+```batchfile
+python gen_img.py --ckpt <model_name> --outdir <image_output_destination> \
+    --xformers --fp16 --from_file <prompt_file_name>
+```
+
+Specify the file containing the prompts with the `--from_file` option. Write one prompt per line. You can specify the number of images to generate per line with the `--images_per_prompt` option.
+
+## Using Negative Prompts and Weighting
+
+If you write `--n` in the prompt options (specified like `--x` in the prompt, described later), the following will be a negative prompt.
+
+Also, weighting with `()` and `[]`, `(xxx:1.3)`, etc., similar to AUTOMATIC1111's Web UI, is possible (the implementation is copied from Diffusers' [Long Prompt Weighting Stable Diffusion](https://github.com/huggingface/diffusers/blob/main/examples/community/README.md#long-prompt-weighting-stable-diffusion)).
+
+It can be specified similarly for prompt specification from the command line and prompt reading from files.
+
+![image](https://user-images.githubusercontent.com/52813779/235343128-e79cd768-ec59-46f5-8395-fce9bdc46208.png)
+
+# Main Options
+
+Specify from the command line.
+
+## Model Specification
+
+- `--ckpt <model_name>`: Specifies the model name. The `--ckpt` option is mandatory. You can specify a Stable Diffusion checkpoint file, a Diffusers model folder, or a Hugging Face model ID.
+
+- `--v1`: Specify when using Stable Diffusion 1.x series models. This is the default behavior.
+
+- `--v2`: Specify when using Stable Diffusion 2.x series models. Not required for 1.x series.
+
+- `--sdxl`: Specify when using Stable Diffusion XL models.
+
+- `--v_parameterization`: Specify when using models that use v-parameterization (`768-v-ema.ckpt` and models with additional training from it, Waifu Diffusion v1.5, etc.).
+
+    If the `--v2` or `--sdxl` specification is incorrect, an error will occur when loading the model. If the `--v_parameterization` specification is incorrect, a brown image will be displayed.
+
+- `--zero_terminal_snr`: Modifies the noise scheduler betas to enforce zero terminal SNR.
+
+- `--pyramid_noise_prob`: Specifies the probability of applying pyramid noise.
+
+- `--pyramid_noise_discount_range`: Specifies the discount range for pyramid noise.
+
+- `--noise_offset_prob`: Specifies the probability of applying noise offset.
+
+- `--noise_offset_range`: Specifies the range of noise offset.
+
+- `--vae`: Specifies the VAE to use. If not specified, the VAE in the model will be used.
+
+- `--tokenizer_cache_dir`: Specifies the cache directory for the tokenizer (for offline usage).
+
+## Image Generation and Output
+
+- `--interactive`: Operates in interactive mode. Images are generated when prompts are entered.
+
+- `--prompt <prompt>`: Specifies the prompt. If it contains spaces, enclose it in double quotes.
+
+- `--from_file <prompt_file_name>`: Specifies the file containing the prompts. Write one prompt per line. Image size and guidance scale can be specified with prompt options (described later).
+
+- `--from_module <module_file>`: Loads prompts from a Python module. The module should implement a `get_prompter(args, pipe, networks)` function.
+
+- `--prompter_module_args`: Specifies additional arguments to pass to the prompter module.
+
+- `--W <image_width>`: Specifies the width of the image. The default is `512`.
+
+- `--H <image_height>`: Specifies the height of the image. The default is `512`.
+
+- `--steps <number_of_steps>`: Specifies the number of sampling steps. The default is `50`.
+
+- `--scale <guidance_scale>`: Specifies the unconditional guidance scale. The default is `7.5`.
+
+- `--sampler <sampler_name>`: Specifies the sampler. The default is `ddim`.
+    `ddim`, `pndm`, `lms`, `euler`, `euler_a`, `heun`, `dpm_2`, `dpm_2_a`, `dpmsolver`, `dpmsolver++`, `dpmsingle`, `k_lms`, `k_euler`, `k_euler_a`, `k_dpm_2`, `k_dpm_2_a` can be specified.
+
+- `--outdir <image_output_destination_folder>`: Specifies the output destination for images.
+
+- `--images_per_prompt <number_of_images_to_generate>`: Specifies the number of images to generate per prompt. The default is `1`.
+
+- `--clip_skip <number_of_skips>`: Specifies which layer from the end of CLIP to use. Default is 1 for SD1/2, 2 for SDXL.
+
+- `--max_embeddings_multiples <multiplier>`: Specifies how many times the CLIP input/output length should be multiplied by the default (75). If not specified, it remains 75. For example, specifying 3 makes the input/output length 225.
+
+- `--negative_scale`: Specifies the guidance scale for unconditioning individually. Implemented with reference to [this article by gcem156](https://note.com/gcem156/n/ne9a53e4a6f43).
+
+- `--emb_normalize_mode`: Specifies the embedding normalization mode. Options are "original" (default), "abs", and "none". This affects how prompt weights are normalized.
+
+- `--force_scheduler_zero_steps_offset`: Forces the scheduler step offset to zero regardless of the `steps_offset` value in the scheduler configuration.
+
+## SDXL-Specific Options
+
+When using SDXL models (with `--sdxl` flag), additional conditioning options are available:
+
+- `--original_height`: Specifies the original height for SDXL conditioning. This affects the model's understanding of the target resolution.
+
+- `--original_width`: Specifies the original width for SDXL conditioning. This affects the model's understanding of the target resolution.
+
+- `--original_height_negative`: Specifies the original height for SDXL negative conditioning.
+
+- `--original_width_negative`: Specifies the original width for SDXL negative conditioning.
+
+- `--crop_top`: Specifies the crop top offset for SDXL conditioning.
+
+- `--crop_left`: Specifies the crop left offset for SDXL conditioning.
+
+## Adjusting Memory Usage and Generation Speed
+
+- `--batch_size <batch_size>`: Specifies the batch size. The default is `1`. A larger batch size consumes more memory but speeds up generation.
+
+- `--vae_batch_size <VAE_batch_size>`: Specifies the VAE batch size. The default is the same as the batch size.
+    Since VAE consumes more memory, memory shortages may occur after denoising (after the step reaches 100%). In such cases, reduce the VAE batch size.
+
+- `--vae_slices <number_of_slices>`: Splits the image into slices for VAE processing to reduce VRAM usage. None (default) for no splitting. Values like 16 or 32 are recommended. Enabling this is slower but uses less VRAM.
+
+- `--no_half_vae`: Prevents using fp16/bf16 precision for VAE processing. Uses fp32 instead. Use this if you encounter VAE-related issues or artifacts.
+
+- `--xformers`: Specify when using xformers.
+
+- `--sdpa`: Use scaled dot-product attention in PyTorch 2 for optimization.
+
+- `--diffusers_xformers`: Use xformers via Diffusers (note: incompatible with Hypernetworks).
+
+- `--fp16`: Performs inference in fp16 (single precision). If neither `fp16` nor `bf16` is specified, inference is performed in fp32 (single precision).
+
+- `--bf16`: Performs inference in bf16 (bfloat16). Can only be specified for RTX 30 series GPUs. The `--bf16` option will cause an error on GPUs other than the RTX 30 series. It seems that `bf16` is less likely to result in NaN (black image) inference results than `fp16`.
+
+## Using Additional Networks (LoRA, etc.)
+
+- `--network_module`: Specifies the additional network to use. For LoRA, specify `--network_module networks.lora`. To use multiple LoRAs, specify like `--network_module networks.lora networks.lora networks.lora`.
+
+- `--network_weights`: Specifies the weight file of the additional network to use. Specify like `--network_weights model.safetensors`. To use multiple LoRAs, specify like `--network_weights model1.safetensors model2.safetensors model3.safetensors`. The number of arguments should be the same as the number specified in `--network_module`.
+
+- `--network_mul`: Specifies how many times to multiply the weight of the additional network to use. The default is `1`. Specify like `--network_mul 0.8`. To use multiple LoRAs, specify like `--network_mul 0.4 0.5 0.7`. The number of arguments should be the same as the number specified in `--network_module`.
+
+- `--network_merge`: Merges the weights of the additional networks to be used in advance with the weights specified in `--network_mul`. Cannot be used simultaneously with `--network_pre_calc`. The prompt option `--am` and Regional LoRA can no longer be used, but generation will be accelerated to the same extent as when LoRA is not used.
+
+- `--network_pre_calc`: Calculates the weights of the additional network to be used in advance for each generation. The prompt option `--am` can be used. Generation is accelerated to the same extent as when LoRA is not used, but time is required to calculate the weights before generation, and memory usage also increases slightly. It is disabled when Regional LoRA is used.
+
+- `--network_regional_mask_max_color_codes`: Specifies the maximum number of color codes to use for regional masks. If not specified, masks are applied by channel. Used with Regional LoRA to control the number of regions that can be defined by colors in the mask.
+
+- `--network_args`: Specifies additional arguments to pass to the network module in key=value format. For example: `--network_args "alpha=1.0,dropout=0.1"`.
+
+- `--network_merge_n_models`: When using network merging, specifies the number of models to merge (instead of merging all loaded networks).
+
+# Examples of Main Option Specifications
+
+The following is an example of batch generating 64 images with the same prompt and a batch size of 4.
+
+```batchfile
+python gen_img.py --ckpt model.ckpt --outdir outputs \
+    --xformers --fp16 --W 512 --H 704 --scale 12.5 --sampler k_euler_a \
+    --steps 32 --batch_size 4 --images_per_prompt 64 \
+    --prompt "beautiful flowers --n monochrome"
+```
+
+The following is an example of batch generating 10 images each for prompts written in a file, with a batch size of 4.
+
+```batchfile
+python gen_img.py --ckpt model.ckpt --outdir outputs \
+    --xformers --fp16 --W 512 --H 704 --scale 12.5 --sampler k_euler_a \
+    --steps 32 --batch_size 4 --images_per_prompt 10 \
+    --from_file prompts.txt
+```
+
+Example of using Textual Inversion (described later) and LoRA.
+
+```batchfile
+python gen_img.py --ckpt model.safetensors \
+    --scale 8 --steps 48 --outdir txt2img --xformers \
+    --W 512 --H 768 --fp16 --sampler k_euler_a \
+    --textual_inversion_embeddings goodembed.safetensors negprompt.pt \
+    --network_module networks.lora networks.lora \
+    --network_weights model1.safetensors model2.safetensors \
+    --network_mul 0.4 0.8 \
+    --clip_skip 2 --max_embeddings_multiples 1 \
+    --batch_size 8 --images_per_prompt 1 --interactive
+```
+
+# Prompt Options
+
+In the prompt, you can specify various options from the prompt with "two hyphens + n alphabetic characters" like `--n`. It is valid whether specifying the prompt from interactive mode, command line, or file.
+
+Please put spaces before and after the prompt option specification `--n`.
+
+- `--n`: Specifies a negative prompt.
+
+- `--w`: Specifies the image width. Overrides the command line specification.
+
+- `--h`: Specifies the image height. Overrides the command line specification.
+
+- `--s`: Specifies the number of steps. Overrides the command line specification.
+
+- `--d`: Specifies the random seed for this image. If `--images_per_prompt` is specified, specify multiple seeds separated by commas, like "--d 1,2,3,4".
+    *For various reasons, the generated image may differ from the Web UI even with the same random seed.
+
+- `--l`: Specifies the guidance scale. Overrides the command line specification.
+
+- `--t`: Specifies the strength of img2img (described later). Overrides the command line specification.
+
+- `--nl`: Specifies the guidance scale for negative prompts (described later). Overrides the command line specification.
+
+- `--am`: Specifies the weight of the additional network. Overrides the command line specification. If using multiple additional networks, specify them separated by __commas__, like `--am 0.8,0.5,0.3`.
+
+- `--ow`: Specifies original_width for SDXL.
+
+- `--oh`: Specifies original_height for SDXL.
+
+- `--nw`: Specifies original_width_negative for SDXL.
+
+- `--nh`: Specifies original_height_negative for SDXL.
+
+- `--ct`: Specifies crop_top for SDXL.
+
+- `--cl`: Specifies crop_left for SDXL.
+
+- `--c`: Specifies the CLIP prompt.
+
+- `--f`: Specifies the generated file name.
+
+- `--glt`: Specifies the timestep to start increasing the size of the latent for Gradual Latent. Overrides the command line specification.
+
+- `--glr`: Specifies the initial size of the latent for Gradual Latent as a ratio. Overrides the command line specification.
+
+- `--gls`: Specifies the ratio to increase the size of the latent for Gradual Latent. Overrides the command line specification.
+
+- `--gle`: Specifies the interval to increase the size of the latent for Gradual Latent. Overrides the command line specification.
+
+*Specifying these options may cause the batch to be executed with a size smaller than the batch size (because they cannot be generated collectively if these values are different). (You don't have to worry too much, but when reading prompts from a file and generating, arranging prompts with the same values for these options will improve efficiency.)
+
+Example:
+```
+(masterpiece, best quality), 1girl, in shirt and plated skirt, standing at street under cherry blossoms, upper body, [from below], kind smile, looking at another, [goodembed] --n realistic, real life, (negprompt), (lowres:1.1), (worst quality:1.2), (low quality:1.1), bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, normal quality, jpeg artifacts, signature, watermark, username, blurry --w 960 --h 640 --s 28 --d 1
+```
+
+![image](https://user-images.githubusercontent.com/52813779/235343446-25654172-fff4-4aaf-977a-20d262b51676.png)
+
+# Wildcards in Prompts (Dynamic Prompts)
+
+Dynamic Prompts (Wildcard) notation is supported. While not exactly the same as the Web UI extension, the following features are available.
+
+- `{A|B|C}` : Randomly selects one from A, B, or C.
+- `{e$$A|B|C}` : Uses all of A, B, and C in order (enumeration). If there are multiple `{e$$...}` in the prompt, all combinations will be generated.
+  - Example: `{e$$red|blue} flower, {e$$1girl|2girls}` -> Generates 4 images: `red flower, 1girl`, `red flower, 2girls`, `blue flower, 1girl`, `blue flower, 2girls`.
+- `{n$$A|B|C}` : Randomly selects n items from A, B, C and combines them.
+  - Example: `{2$$A|B|C}` -> `A, B` or `B, C`, etc.
+- `{n-m$$A|B|C}` : Randomly selects between n and m items from A, B, C and combines them.
+- `{$$sep$$A|B|C}` : Combines selected items with `sep` (default is `, `).
+  - Example: `{2$$ and $$A|B|C}` -> `A and B`, etc.
+
+These can be used in combination.
+
+# img2img
+
+## Options
+
+- `--image_path`: Specifies the image to use for img2img. Specify like `--image_path template.png`. If a folder is specified, images in that folder will be used sequentially.
+
+- `--strength`: Specifies the strength of img2img. Specify like `--strength 0.8`. The default is `0.8`.
+
+- `--sequential_file_name`: Specifies whether to make file names sequential. If specified, the generated file names will be sequential starting from `im_000001.png`.
+
+- `--use_original_file_name`: If specified, the generated file name will be prepended with the original file name (for img2img mode).
+
+- `--clip_vision_strength`: Enables CLIP Vision Conditioning for img2img with the specified strength. Uses the CLIP Vision model to enhance conditioning from the input image.
+
+## Command Line Execution Example
+
+```batchfile
+python gen_img.py --ckpt trinart_characters_it4_v1_vae_merged.ckpt \
+    --outdir outputs --xformers --fp16 --scale 12.5 --sampler k_euler --steps 32 \
+    --image_path template.png --strength 0.8 \
+    --prompt "1girl, cowboy shot, brown hair, pony tail, brown eyes, \
+          sailor school uniform, outdoors \
+          --n lowres, bad anatomy, bad hands, error, missing fingers, cropped, \
+          worst quality, low quality, normal quality, jpeg artifacts, (blurry), \
+          hair ornament, glasses" \
+    --batch_size 8 --images_per_prompt 32
+```
+
+If a folder is specified in the `--image_path` option, images in that folder will be read sequentially. The number of images generated will be the number of prompts, not the number of images, so please match the number of images to img2img and the number of prompts by specifying the `--images_per_prompt` option.
+
+Files are read sorted by file name. Note that the sort order is string order (not `1.jpg -> 2.jpg -> 10.jpg` but `1.jpg -> 10.jpg -> 2.jpg`), so please pad the beginning with zeros (e.g., `01.jpg -> 02.jpg -> 10.jpg`).
+
+## Upscale using img2img
+
+If you specify the generated image size with the `--W` and `--H` command line options during img2img, the original image will be resized to that size before img2img.
+
+Also, if the original image for img2img was generated by this script, omitting the prompt will retrieve the prompt from the original image's metadata and use it as is. This allows you to perform only the 2nd stage operation of Highres. fix.
+
+## Inpainting during img2img
+
+You can specify an image and a mask image for inpainting (inpainting models are not supported, it simply performs img2img on the mask area).
+
+The options are as follows:
+
+- `--mask_image`: Specifies the mask image. Similar to `--img_path`, if a folder is specified, images in that folder will be used sequentially.
+
+The mask image is a grayscale image, and the white parts will be inpainted. It is recommended to gradient the boundaries to make it somewhat smooth.
+
+![image](https://user-images.githubusercontent.com/52813779/235343795-9eaa6d98-02ff-4f32-b089-80d1fc482453.png)
+
+# Other Features
+
+## Textual Inversion
+
+Specify the embeddings to use with the `--textual_inversion_embeddings` option (multiple specifications possible). By using the file name without the extension in the prompt, that embedding will be used (same usage as Web UI). It can also be used in negative prompts.
+
+As models, you can use Textual Inversion models trained with this repository and Textual Inversion models trained with Web UI (image embedding is not supported).
+
+## Highres. fix
+
+This is a similar feature to the one in AUTOMATIC1111's Web UI (it may differ in various ways as it is an original implementation). It first generates a smaller image and then uses that image as a base for img2img to generate a large resolution image while preventing the entire image from collapsing.
+
+The number of steps for the 2nd stage is calculated from the values of the `--steps` and `--strength` options (`steps*strength`).
+
+Cannot be used with img2img.
+
+The following options are available:
+
+- `--highres_fix_scale`: Enables Highres. fix and specifies the size of the image generated in the 1st stage as a magnification. If the final output is 1024x1024 and you want to generate a 512x512 image first, specify like `--highres_fix_scale 0.5`. Please note that this is the reciprocal of the specification in Web UI.
+
+- `--highres_fix_steps`: Specifies the number of steps for the 1st stage image. The default is `28`.
+
+- `--highres_fix_save_1st`: Specifies whether to save the 1st stage image.
+
+- `--highres_fix_latents_upscaling`: If specified, the 1st stage image will be upscaled on a latent basis during 2nd stage image generation (only bilinear is supported). If not specified, the image will be upscaled with LANCZOS4.
+
+- `--highres_fix_upscaler`: Uses an arbitrary upscaler for the 2nd stage. Currently, only `--highres_fix_upscaler tools.latent_upscaler` is supported.
+
+- `--highres_fix_upscaler_args`: Specifies the arguments to pass to the upscaler specified with `--highres_fix_upscaler`.
+    For `tools.latent_upscaler`, specify the weight file like `--highres_fix_upscaler_args "weights=D:\\Work\\SD\\Models\\others\\etc\\upscaler-v1-e100-220.safetensors"`.
+
+- `--highres_fix_disable_control_net`: Disables ControlNet for the 2nd stage of Highres fix. By default, ControlNet is used in both stages.
+
+Command line example:
+
+```batchfile
+python gen_img.py  --ckpt trinart_characters_it4_v1_vae_merged.ckpt\
+    --n_iter 1 --scale 7.5 --W 1024 --H 1024 --batch_size 1 --outdir ../txt2img \
+    --steps 48 --sampler ddim --fp16 \
+    --xformers \
+    --images_per_prompt 1  --interactive \
+    --highres_fix_scale 0.5 --highres_fix_steps 28 --strength 0.5
+```
+
+## Deep Shrink
+
+Deep Shrink is a technique that optimizes the generation process by using different depths of the UNet at different timesteps. It can improve generation quality and efficiency.
+
+The following options are available:
+
+- `--ds_depth_1`: Enables Deep Shrink with this depth for the first phase. Valid values are 0 to 8.
+
+- `--ds_timesteps_1`: Applies Deep Shrink depth 1 until this timestep. Default is 650.
+
+- `--ds_depth_2`: Specifies the depth for the second phase of Deep Shrink.
+
+- `--ds_timesteps_2`: Applies Deep Shrink depth 2 until this timestep. Default is 650.
+
+- `--ds_ratio`: Specifies the ratio for downsampling in Deep Shrink. Default is 0.5.
+
+These parameters can also be specified through prompt options:
+
+- `--dsd1`: Specifies Deep Shrink depth 1 from the prompt.
+  
+- `--dst1`: Specifies Deep Shrink timestep 1 from the prompt.
+  
+- `--dsd2`: Specifies Deep Shrink depth 2 from the prompt.
+  
+- `--dst2`: Specifies Deep Shrink timestep 2 from the prompt.
+  
+- `--dsr`: Specifies Deep Shrink ratio from the prompt.
+
+*Additional prompt options for Gradual Latent (requires `euler_a` sampler):*
+
+- `--glt`: Specifies the timestep to start increasing the size of the latent for Gradual Latent. Overrides the command line specification.
+
+- `--glr`: Specifies the initial size of the latent for Gradual Latent as a ratio. Overrides the command line specification.
+
+- `--gls`: Specifies the ratio to increase the size of the latent for Gradual Latent. Overrides the command line specification.
+
+- `--gle`: Specifies the interval to increase the size of the latent for Gradual Latent. Overrides the command line specification.
+
+## ControlNet
+
+Currently, only ControlNet 1.0 has been confirmed to work. Only Canny is supported for preprocessing.
+
+The following options are available:
+
+- `--control_net_models`: Specifies the ControlNet model file.
+    If multiple are specified, they will be switched and used for each step (differs from the implementation of the ControlNet extension in Web UI). Supports both diff and normal.
+
+- `--guide_image_path`: Specifies the hint image to use for ControlNet. Similar to `--img_path`, if a folder is specified, images in that folder will be used sequentially. For models other than Canny, please perform preprocessing beforehand.
+
+- `--control_net_preps`: Specifies the preprocessing for ControlNet. Multiple specifications are possible, similar to `--control_net_models`. Currently, only canny is supported. If preprocessing is not used for the target model, specify `none`.
+   For canny, you can specify thresholds 1 and 2 separated by `_`, like `--control_net_preps canny_63_191`.
+
+- `--control_net_weights`: Specifies the weight when applying ControlNet (`1.0` for normal, `0.5` for half influence). Multiple specifications are possible, similar to `--control_net_models`.
+
+- `--control_net_ratios`: Specifies the range of steps to apply ControlNet. If `0.5`, ControlNet is applied up to half the number of steps. Multiple specifications are possible, similar to `--control_net_models`.
+
+Command line example:
+
+```batchfile
+python gen_img.py --ckpt model_ckpt --scale 8 --steps 48 --outdir txt2img --xformers \
+    --W 512 --H 768 --bf16 --sampler k_euler_a \
+    --control_net_models diff_control_sd15_canny.safetensors --control_net_weights 1.0 \
+    --guide_image_path guide.png --control_net_ratios 1.0 --interactive
+```
+
+## ControlNet-LLLite
+
+ControlNet-LLLite is a lightweight alternative to ControlNet that can be used for similar guidance purposes.
+
+The following options are available:
+
+- `--control_net_lllite_models`: Specifies the ControlNet-LLLite model files.
+
+- `--control_net_multipliers`: Specifies the multiplier for ControlNet-LLLite (similar to weights).
+
+- `--control_net_ratios`: Specifies the ratio of steps to apply ControlNet-LLLite.
+
+Note that ControlNet and ControlNet-LLLite cannot be used at the same time.
+
+## Attention Couple + Regional LoRA
+
+This is a feature that allows you to divide the prompt into several parts and specify which region in the image each prompt should be applied to. There are no individual options, but it is specified with `mask_path` and the prompt.
+
+First, define multiple parts using ` AND ` in the prompt. Region specification can be done for the first three parts, and subsequent parts are applied to the entire image. Negative prompts are applied to the entire image.
+
+In the following, three parts are defined with AND.
+
+```
+shs 2girls, looking at viewer, smile AND bsb 2girls, looking back AND 2girls --n bad quality, worst quality
+```
+
+Next, prepare a mask image. The mask image is a color image, and each RGB channel corresponds to the part separated by AND in the prompt. Also, if the value of a certain channel is all 0, it is applied to the entire image.
+
+In the example above, the R channel corresponds to `shs 2girls, looking at viewer, smile`, the G channel to `bsb 2girls, looking back`, and the B channel to `2girls`. If you use a mask image like the following, since there is no specification for the B channel, `2girls` will be applied to the entire image.
+
+![image](https://user-images.githubusercontent.com/52813779/235343061-b4dc9392-3dae-4831-8347-1e9ae5054251.png)
+
+The mask image is specified with `--mask_path`. Currently, only one image is supported. It is automatically resized and applied to the specified image size.
+
+It can also be combined with ControlNet (combination with ControlNet is recommended for detailed position specification).
+
+If LoRA is specified, multiple LoRAs specified with `--network_weights` will correspond to each part of AND. As a current constraint, the number of LoRAs must be the same as the number of AND parts.
+
+# Other Options
+
+- `--no_preview`: Does not display preview images in interactive mode. Specify this if OpenCV is not installed or if you want to check the output files directly.
+
+- `--n_iter`: Specifies the number of times to repeat generation. The default is 1. Specify this when you want to perform generation multiple times when reading prompts from a file.
+
+- `--tokenizer_cache_dir`: Specifies the cache directory for the tokenizer. (Work in progress)
+
+- `--seed`: Specifies the random seed. When generating one image, it is the seed for that image. When generating multiple images, it is the seed for the random numbers used to generate the seeds for each image (when generating multiple images with `--from_file`, specifying the `--seed` option will make each image have the same seed when executed multiple times).
+
+- `--iter_same_seed`: When there is no random seed specification in the prompt, the same seed is used for all repetitions of `--n_iter`. Used to unify and compare seeds between multiple prompts specified with `--from_file`.
+
+- `--shuffle_prompts`: Shuffles the order of prompts in iteration. Useful when using `--from_file` with multiple prompts.
+
+- `--diffusers_xformers`: Uses Diffuser's xformers.
+
+- `--opt_channels_last`: Arranges tensor channels last during inference. May speed up in some cases.
+
+- `--network_show_meta`: Displays the metadata of the additional network.
+
+
+---
+
+# About Gradual Latent
+
+Gradual Latent is a Hires fix that gradually increases the size of the latent.  `gen_img.py`, `sdxl_gen_img.py`, and `gen_img.py` have the following options.
+
+- `--gradual_latent_timesteps`: Specifies the timestep to start increasing the size of the latent. The default is None, which means Gradual Latent is not used. Please try around 750 at first.
+- `--gradual_latent_ratio`: Specifies the initial size of the latent. The default is 0.5, which means it starts with half the default latent size.
+- `--gradual_latent_ratio_step`: Specifies the ratio to increase the size of the latent. The default is 0.125, which means the latent size is gradually increased to 0.625, 0.75, 0.875, 1.0.
+- `--gradual_latent_ratio_every_n_steps`: Specifies the interval to increase the size of the latent. The default is 3, which means the latent size is increased every 3 steps.
+- `--gradual_latent_s_noise`: Specifies the s_noise parameter for Gradual Latent. Default is 1.0.
+- `--gradual_latent_unsharp_params`: Specifies unsharp mask parameters for Gradual Latent in the format: ksize,sigma,strength,target-x (target-x: 1=True, 0=False). Recommended values: `3,0.5,0.5,1` or `3,1.0,1.0,0`.
+
+Each option can also be specified with prompt options, `--glt`, `--glr`, `--gls`, `--gle`.
+
+__Please specify `euler_a` for the sampler.__ Because the source code of the sampler is modified. It will not work with other samplers.
+
+It is more effective with SD 1.5. It is quite subtle with SDXL.

docs/hunyuan_image_train_network.md

@@ -0,0 +1,525 @@
+Status: reviewed
+
+# LoRA Training Guide for HunyuanImage-2.1 using `hunyuan_image_train_network.py` / `hunyuan_image_train_network.py` を用いたHunyuanImage-2.1モデルのLoRA学習ガイド
+
+This document explains how to train LoRA models for the HunyuanImage-2.1 model using `hunyuan_image_train_network.py` included in the `sd-scripts` repository.
+
+<details>
+<summary>日本語</summary>
+
+このドキュメントでは、`sd-scripts`リポジトリに含まれる`hunyuan_image_train_network.py`を使用して、HunyuanImage-2.1モデルに対するLoRA (Low-Rank Adaptation) モデルを学習する基本的な手順について解説します。
+
+</details>
+
+## 1. Introduction / はじめに
+
+`hunyuan_image_train_network.py` trains additional networks such as LoRA on the HunyuanImage-2.1 model, which uses a transformer-based architecture (DiT) different from Stable Diffusion. Two text encoders, Qwen2.5-VL and byT5, and a dedicated VAE are used.
+
+This guide assumes you know the basics of LoRA training. For common options see [train_network.py](train_network.md) and [sdxl_train_network.py](sdxl_train_network.md).
+
+**Prerequisites:**
+
+* The repository is cloned and the Python environment is ready.
+* A training dataset is prepared. See the dataset configuration guide.
+
+<details>
+<summary>日本語</summary>
+
+`hunyuan_image_train_network.py`はHunyuanImage-2.1モデルに対してLoRAなどの追加ネットワークを学習させるためのスクリプトです。HunyuanImage-2.1はStable Diffusionとは異なるDiT (Diffusion Transformer) アーキテクチャを持つ画像生成モデルであり、このスクリプトを使用することで、特定のキャラクターや画風を再現するLoRAモデルを作成できます。
+
+このガイドは、基本的なLoRA学習の手順を理解しているユーザーを対象としています。基本的な使い方や共通のオプションについては、[`train_network.py`のガイド](train_network.md)を参照してください。また一部のパラメータは [`sdxl_train_network.py`](sdxl_train_network.md) や [`flux_train_network.py`](flux_train_network.md) と同様のものがあるため、そちらも参考にしてください。
+
+**前提条件:**
+
+* `sd-scripts`リポジトリのクローンとPython環境のセットアップが完了していること。
+* 学習用データセットの準備が完了していること。（データセットの準備については[データセット設定ガイド](config_README-ja.md)を参照してください）
+
+</details>
+
+## 2. Differences from `train_network.py` / `train_network.py` との違い
+
+`hunyuan_image_train_network.py` is based on `train_network.py` but adapted for HunyuanImage-2.1. Main differences include:
+
+* **Target model:** HunyuanImage-2.1 model.
+* **Model structure:** HunyuanImage-2.1 uses a Transformer-based architecture (DiT). It uses two text encoders (Qwen2.5-VL and byT5) and a dedicated VAE.
+* **Required arguments:** Additional arguments for the DiT model, Qwen2.5-VL, byT5, and VAE model files.
+* **Incompatible options:** Some Stable Diffusion-specific arguments (e.g., `--v2`, `--clip_skip`, `--max_token_length`) are not used.
+* **HunyuanImage-2.1-specific arguments:** Additional arguments for specific training parameters like flow matching.
+
+<details>
+<summary>日本語</summary>
+
+`hunyuan_image_train_network.py`は`train_network.py`をベースに、HunyuanImage-2.1モデルに対応するための変更が加えられています。主な違いは以下の通りです。
+
+* **対象モデル:** HunyuanImage-2.1モデルを対象とします。
+* **モデル構造:** HunyuanImage-2.1はDiTベースのアーキテクチャを持ちます。Text EncoderとしてQwen2.5-VLとbyT5の二つを使用し、専用のVAEを使用します。
+* **必須の引数:** DiTモデル、Qwen2.5-VL、byT5、VAEの各モデルファイルを指定する引数が追加されています。
+* **一部引数の非互換性:** Stable Diffusion向けの引数の一部（例: `--v2`, `--clip_skip`, `--max_token_length`）は使用されません。
+* **HunyuanImage-2.1特有の引数:** Flow Matchingなど、特有の学習パラメータを指定する引数が追加されています。
+
+</details>
+
+## 3. Preparation / 準備
+
+Before starting training you need:
+
+1. **Training script:** `hunyuan_image_train_network.py`
+2. **HunyuanImage-2.1 DiT model file:** Base DiT model `.safetensors` file.
+3. **Text Encoder model files:**
+   - Qwen2.5-VL model file (`--text_encoder`).
+   - byT5 model file (`--byt5`).
+4. **VAE model file:** HunyuanImage-2.1-compatible VAE model `.safetensors` file (`--vae`).
+5. **Dataset definition file (.toml):** TOML format file describing training dataset configuration.
+
+### Downloading Required Models
+
+To train HunyuanImage-2.1 models, you need to download the following model files:
+
+- **DiT Model**: Download from the [Tencent HunyuanImage-2.1](https://huggingface.co/tencent/HunyuanImage-2.1/) repository. Use `dit/hunyuanimage2.1.safetensors`.
+- **Text Encoders and VAE**: Download from the [Comfy-Org/HunyuanImage_2.1_ComfyUI](https://huggingface.co/Comfy-Org/HunyuanImage_2.1_ComfyUI) repository:
+  - Qwen2.5-VL: `split_files/text_encoders/qwen_2.5_vl_7b.safetensors`
+  - byT5: `split_files/text_encoders/byt5_small_glyphxl_fp16.safetensors`
+  - VAE: `split_files/vae/hunyuan_image_2.1_vae_fp16.safetensors`
+
+<details>
+<summary>日本語</summary>
+
+学習を開始する前に、以下のファイルが必要です。
+
+1. **学習スクリプト:** `hunyuan_image_train_network.py`
+2. **HunyuanImage-2.1 DiTモデルファイル:** 学習のベースとなるDiTモデルの`.safetensors`ファイル。
+3. **Text Encoderモデルファイル:**
+   - Qwen2.5-VLモデルファイル (`--text_encoder`)。
+   - byT5モデルファイル (`--byt5`)。
+4. **VAEモデルファイル:** HunyuanImage-2.1に対応するVAEモデルの`.safetensors`ファイル (`--vae`)。
+5. **データセット定義ファイル (.toml):** 学習データセットの設定を記述したTOML形式のファイル。（詳細は[データセット設定ガイド](config_README-ja.md)を参照してください）。
+
+**必要なモデルのダウンロード**
+
+HunyuanImage-2.1モデルを学習するためには、以下のモデルファイルをダウンロードする必要があります：
+
+- **DiTモデル**: [Tencent HunyuanImage-2.1](https://huggingface.co/tencent/HunyuanImage-2.1/) リポジトリから `dit/hunyuanimage2.1.safetensors` をダウンロードします。
+- **Text EncoderとVAE**: [Comfy-Org/HunyuanImage_2.1_ComfyUI](https://huggingface.co/Comfy-Org/HunyuanImage_2.1_ComfyUI) リポジトリから以下をダウンロードします：
+  - Qwen2.5-VL: `split_files/text_encoders/qwen_2.5_vl_7b.safetensors`
+  - byT5: `split_files/text_encoders/byt5_small_glyphxl_fp16.safetensors`
+  - VAE: `split_files/vae/hunyuan_image_2.1_vae_fp16.safetensors`
+
+</details>
+
+## 4. Running the Training / 学習の実行
+
+Run `hunyuan_image_train_network.py` from the terminal with HunyuanImage-2.1 specific arguments. Here's a basic command example:
+
+```bash
+accelerate launch --num_cpu_threads_per_process 1 hunyuan_image_train_network.py \
+  --pretrained_model_name_or_path="<path to HunyuanDiT model>" \
+  --text_encoder="<path to Qwen2.5-VL model>" \
+  --byt5="<path to byT5 model>" \
+  --vae="<path to VAE model>" \
+  --dataset_config="my_hunyuan_dataset_config.toml" \
+  --output_dir="<output directory>" \
+  --output_name="my_hunyuan_lora" \
+  --save_model_as=safetensors \
+  --network_module=networks.lora_hunyuan_image \
+  --network_dim=16 \
+  --network_alpha=1 \
+  --network_train_unet_only \
+  --learning_rate=1e-4 \
+  --optimizer_type="AdamW8bit" \
+  --lr_scheduler="constant" \
+  --attn_mode="torch" \
+  --split_attn \
+  --max_train_epochs=10 \
+  --save_every_n_epochs=1 \
+  --mixed_precision="bf16" \
+  --gradient_checkpointing \
+  --model_prediction_type="raw" \
+  --discrete_flow_shift=5.0 \
+  --blocks_to_swap=18 \
+  --cache_text_encoder_outputs \
+  --cache_latents
+```
+
+**HunyuanImage-2.1 training does not support LoRA modules for Text Encoders, so `--network_train_unet_only` is required.**
+
+<details>
+<summary>日本語</summary>
+
+学習は、ターミナルから`hunyuan_image_train_network.py`を実行することで開始します。基本的なコマンドラインの構造は`train_network.py`と同様ですが、HunyuanImage-2.1特有の引数を指定する必要があります。
+
+コマンドラインの例は英語のドキュメントを参照してください。
+
+</details>
+
+### 4.1. Explanation of Key Options / 主要なコマンドライン引数の解説
+
+The script adds HunyuanImage-2.1 specific arguments. For common arguments (like `--output_dir`, `--output_name`, `--network_module`, etc.), see the [`train_network.py` guide](train_network.md).
+
+#### Model-related [Required]
+
+* `--pretrained_model_name_or_path="<path to HunyuanDiT model>"` **[Required]**
+  - Specifies the path to the base DiT model `.safetensors` file.
+* `--text_encoder="<path to Qwen2.5-VL model>"` **[Required]**
+  - Specifies the path to the Qwen2.5-VL Text Encoder model file. Should be `bfloat16`.
+* `--byt5="<path to byT5 model>"` **[Required]**
+  - Specifies the path to the byT5 Text Encoder model file. Should be `float16`.
+* `--vae="<path to VAE model>"` **[Required]**
+  - Specifies the path to the HunyuanImage-2.1-compatible VAE model `.safetensors` file.
+
+#### HunyuanImage-2.1 Training Parameters
+
+* `--network_train_unet_only` **[Required]**
+  - Specifies that only the DiT model will be trained. LoRA modules for Text Encoders are not supported.
+* `--discrete_flow_shift=<float>`
+  - Specifies the shift value for the scheduler used in Flow Matching. Default is `5.0`.
+* `--model_prediction_type=<choice>`
+  - Specifies what the model predicts. Choose from `raw`, `additive`, `sigma_scaled`. Default and recommended is `raw`.
+* `--timestep_sampling=<choice>`
+  - Specifies the sampling method for timesteps (noise levels) during training. Choose from `sigma`, `uniform`, `sigmoid`, `shift`, `flux_shift`. Default is `sigma`.
+* `--sigmoid_scale=<float>`
+  - Scale factor when `timestep_sampling` is set to `sigmoid`, `shift`, or `flux_shift`. Default is `1.0`.
+
+#### Memory/Speed Related
+
+* `--attn_mode=<choice>`
+  - Specifies the attention implementation to use. Options are `torch`, `xformers`, `flash`, `sageattn`. Default is `torch` (use scaled dot product attention). Each library must be installed separately other than `torch`. If using `xformers`, also specify `--split_attn` if the batch size is more than 1.
+* `--split_attn`
+  - Splits the batch during attention computation to process one item at a time, reducing VRAM usage by avoiding attention mask computation. Can improve speed when using `torch`. Required when using `xformers` with batch size greater than 1.
+* `--fp8_scaled`
+  - Enables training the DiT model in scaled FP8 format. This can significantly reduce VRAM usage (can run with as little as 8GB VRAM when combined with `--blocks_to_swap`), but the training results may vary. This is a newer alternative to the unsupported `--fp8_base` option. See [Musubi Tuner's documentation](https://github.com/kohya-ss/musubi-tuner/blob/main/docs/advanced_config.md#fp8-weight-optimization-for-models--%E3%83%A2%E3%83%87%E3%83%AB%E3%81%AE%E9%87%8D%E3%81%BF%E3%81%AEfp8%E3%81%B8%E3%81%AE%E6%9C%80%E9%81%A9%E5%8C%96) for details.
+* `--fp8_vl`
+  - Use FP8 for the VLM (Qwen2.5-VL) text encoder.
+* `--text_encoder_cpu`
+  - Runs the text encoders on CPU to reduce VRAM usage. This is useful when VRAM is insufficient (less than 12GB). Encoding one text may take a few minutes (depending on CPU). It is highly recommended to use this option with `--cache_text_encoder_outputs_to_disk` to avoid repeated encoding every time training starts. **In addition, increasing `--num_cpu_threads_per_process` in the `accelerate launch` command, like `--num_cpu_threads_per_process=8` or `16`, can speed up encoding in some environments.**
+* `--blocks_to_swap=<integer>` **[Experimental Feature]**
+  - Setting to reduce VRAM usage by swapping parts of the model (Transformer blocks) between CPU and GPU. Specify the number of blocks to swap as an integer (e.g., `18`). Larger values reduce VRAM usage but decrease training speed. Adjust according to your GPU's VRAM capacity. Can be used with `gradient_checkpointing`.
+* `--cache_text_encoder_outputs`
+  - Caches the outputs of Qwen2.5-VL and byT5. This reduces memory usage.
+* `--cache_latents`, `--cache_latents_to_disk`
+  - Caches the outputs of VAE. Similar functionality to [sdxl_train_network.py](sdxl_train_network.md).
+* `--vae_chunk_size=<integer>`
+  - Enables chunked processing in the VAE to reduce VRAM usage during encoding and decoding. Specify the chunk size as an integer (e.g., `16`). Larger values use more VRAM but are faster. Default is `None` (no chunking). This option is useful when VRAM is limited (e.g., 8GB or 12GB).
+
+<details>
+<summary>日本語</summary>
+
+[`train_network.py`のガイド](train_network.md)で説明されている引数に加え、以下のHunyuanImage-2.1特有の引数を指定します。共通の引数（`--output_dir`, `--output_name`, `--network_module`, `--network_dim`, `--network_alpha`, `--learning_rate`など）については、上記ガイドを参照してください。
+
+コマンドラインの例と詳細な引数の説明は英語のドキュメントを参照してください。
+
+</details>
+
+## 5. Using the Trained Model / 学習済みモデルの利用
+
+After training, a LoRA model file is saved in `output_dir` and can be used in inference environments supporting HunyuanImage-2.1.
+
+<details>
+<summary>日本語</summary>
+
+学習が完了すると、指定した`output_dir`にLoRAモデルファイル（例: `my_hunyuan_lora.safetensors`）が保存されます。このファイルは、HunyuanImage-2.1モデルに対応した推論環境で使用できます。
+
+</details>
+
+## 6. Advanced Settings / 高度な設定
+
+### 6.1. VRAM Usage Optimization / VRAM使用量の最適化
+
+HunyuanImage-2.1 is a large model, so GPUs without sufficient VRAM require optimization.
+
+#### Recommended Settings by GPU Memory
+
+Based on testing with the pull request, here are recommended VRAM optimization settings:
+
+| GPU Memory | Recommended Settings |
+|------------|---------------------|
+| 40GB+ VRAM | Standard settings (no special optimization needed) |
+| 24GB VRAM  | `--fp8_scaled --blocks_to_swap 9` |
+| 12GB VRAM  | `--fp8_scaled --blocks_to_swap 32` |
+| 8GB VRAM   | `--fp8_scaled --blocks_to_swap 37` |
+
+#### Key VRAM Reduction Options
+
+- **`--fp8_scaled`**: Enables training the DiT in scaled FP8 format. This is the recommended FP8 option for HunyuanImage-2.1, replacing the unsupported `--fp8_base` option. Essential for <40GB VRAM environments.
+- **`--fp8_vl`**: Use FP8 for the VLM (Qwen2.5-VL) text encoder.
+- **`--blocks_to_swap <number>`**: Swaps blocks between CPU and GPU to reduce VRAM usage. Higher numbers save more VRAM but reduce training speed. Up to 37 blocks can be swapped for HunyuanImage-2.1.
+- **`--cpu_offload_checkpointing`**: Offloads gradient checkpoints to CPU. Can reduce VRAM usage but decreases training speed. Cannot be used with `--blocks_to_swap`.
+- **Using Adafactor optimizer**: Can reduce VRAM usage more than 8bit AdamW:
+  ```
+  --optimizer_type adafactor --optimizer_args "relative_step=False" "scale_parameter=False" "warmup_init=False" --lr_scheduler constant_with_warmup --max_grad_norm 0.0
+  ```
+
+<details>
+<summary>日本語</summary>
+
+HunyuanImage-2.1は大きなモデルであるため、十分なVRAMを持たないGPUでは工夫が必要です。
+
+#### GPU別推奨設定
+
+Pull Requestのテスト結果に基づく推奨VRAM最適化設定：
+
+| GPU Memory | 推奨設定 |
+|------------|---------|
+| 40GB+ VRAM | 標準設定（特別な最適化不要） |
+| 24GB VRAM  | `--fp8_scaled --blocks_to_swap 9` |
+| 12GB VRAM  | `--fp8_scaled --blocks_to_swap 32` |
+| 8GB VRAM   | `--fp8_scaled --blocks_to_swap 37` |
+
+主要なVRAM削減オプション：
+- `--fp8_scaled`: DiTをスケールされたFP8形式で学習（推奨されるFP8オプション、40GB VRAM未満の環境では必須）
+- `--fp8_vl`: VLMテキストエンコーダにFP8を使用
+- `--blocks_to_swap`: CPUとGPU間でブロックをスワップ（最大37ブロック）
+- `--cpu_offload_checkpointing`: 勾配チェックポイントをCPUにオフロード
+- Adafactorオプティマイザの使用
+
+</details>
+
+### 6.2. Important HunyuanImage-2.1 LoRA Training Settings / HunyuanImage-2.1 LoRA学習の重要な設定
+
+HunyuanImage-2.1 training has several settings that can be specified with arguments:
+
+#### Timestep Sampling Methods
+
+The `--timestep_sampling` option specifies how timesteps (0-1) are sampled:
+
+- `sigma`: Sigma-based like SD3 (Default)
+- `uniform`: Uniform random
+- `sigmoid`: Sigmoid of normal distribution random
+- `shift`: Sigmoid value of normal distribution random with shift.
+- `flux_shift`: Shift sigmoid value of normal distribution random according to resolution.
+
+#### Model Prediction Processing
+
+The `--model_prediction_type` option specifies how to interpret and process model predictions:
+
+- `raw`: Use as-is **[Recommended, Default]**
+- `additive`: Add to noise input
+- `sigma_scaled`: Apply sigma scaling
+
+#### Recommended Settings
+
+Based on experiments, the default settings work well:
+```
+--model_prediction_type raw --discrete_flow_shift 5.0
+```
+
+<details>
+<summary>日本語</summary>
+
+HunyuanImage-2.1の学習には、引数で指定できるいくつかの設定があります。詳細な説明とコマンドラインの例は英語のドキュメントを参照してください。
+
+主要な設定オプション：
+- タイムステップのサンプリング方法（`--timestep_sampling`）
+- モデル予測の処理方法（`--model_prediction_type`）
+- 推奨設定の組み合わせ
+
+</details>
+
+### 6.3. Regular Expression-based Rank/LR Configuration / 正規表現によるランク・学習率の指定
+
+You can specify ranks (dims) and learning rates for LoRA modules using regular expressions. This allows for more flexible and fine-grained control.
+
+These settings are specified via the `network_args` argument.
+
+*   `network_reg_dims`: Specify ranks for modules matching a regular expression. The format is a comma-separated string of `pattern=rank`.
+    *   Example: `--network_args "network_reg_dims=attn.*.q_proj=4,attn.*.k_proj=4"`
+*   `network_reg_lrs`: Specify learning rates for modules matching a regular expression. The format is a comma-separated string of `pattern=lr`.
+    *   Example: `--network_args "network_reg_lrs=down_blocks.1=1e-4,up_blocks.2=2e-4"`
+
+**Notes:**
+
+*   To find the correct module names for the patterns, you may need to inspect the model structure.
+*   Settings via `network_reg_dims` and `network_reg_lrs` take precedence over the global `--network_dim` and `--learning_rate` settings.
+*   If a module name matches multiple patterns, the setting from the last matching pattern in the string will be applied.
+
+<details>
+<summary>日本語</summary>
+
+正規表現を用いて、LoRAのモジュールごとにランク（dim）や学習率を指定することができます。これにより、柔軟できめ細やかな制御が可能になります。
+
+これらの設定は `network_args` 引数で指定します。
+
+*   `network_reg_dims`: 正規表現にマッチするモジュールに対してランクを指定します。
+*   `network_reg_lrs`: 正規表現にマッチするモジュールに対して学習率を指定します。
+
+**注意点:**
+
+*   パターンのための正確なモジュール名を見つけるには、モデルの構造を調べる必要があるかもしれません。
+*   `network_reg_dims` および `network_reg_lrs` での設定は、全体設定である `--network_dim` や `--learning_rate` よりも優先されます。
+*   あるモジュール名が複数のパターンにマッチした場合、文字列の中で後方にあるパターンの設定が適用されます。
+
+</details>
+
+### 6.4. Multi-Resolution Training / マルチ解像度トレーニング
+
+You can define multiple resolutions in the dataset configuration file, with different batch sizes for each resolution.
+
+**Note:** This feature is available, but it is **not recommended** as the HunyuanImage-2.1 base model was not trained with multi-resolution capabilities. Using it may lead to unexpected results.
+
+Configuration file example:
+```toml
+[general]
+shuffle_caption = true
+caption_extension = ".txt"
+
+[[datasets]]
+batch_size = 2
+enable_bucket = true
+resolution = [1024, 1024]
+
+  [[datasets.subsets]]
+  image_dir = "path/to/image/directory"
+  num_repeats = 1
+
+[[datasets]]
+batch_size = 1
+enable_bucket = true
+resolution = [1280, 768]
+
+  [[datasets.subsets]]
+  image_dir = "path/to/another/directory"
+  num_repeats = 1
+```
+
+<details>
+<summary>日本語</summary>
+
+データセット設定ファイルで複数の解像度を定義できます。各解像度に対して異なるバッチサイズを指定することができます。
+
+**注意:** この機能は利用可能ですが、HunyuanImage-2.1のベースモデルはマルチ解像度で学習されていないため、**非推奨**です。使用すると予期しない結果になる可能性があります。
+
+設定ファイルの例は英語のドキュメントを参照してください。
+
+</details>
+
+### 6.5. Validation / 検証
+
+You can calculate validation loss during training using a validation dataset to evaluate model generalization performance. This feature works the same as in other training scripts. For details, please refer to the [Validation Guide](validation.md).
+
+<details>
+<summary>日本語</summary>
+
+学習中に検証データセットを使用して損失 (Validation Loss) を計算し、モデルの汎化性能を評価できます。この機能は他の学習スクリプトと同様に動作します。詳細は[検証ガイド](validation.md)を参照してください。
+
+</details>
+
+## 7. Other Training Options / その他の学習オプション
+
+- **`--ip_noise_gamma`**: Use `--ip_noise_gamma` and `--ip_noise_gamma_random_strength` to adjust Input Perturbation noise gamma values during training. See Stable Diffusion 3 training options for details.
+
+- **`--loss_type`**: Specifies the loss function for training. The default is `l2`.
+  - `l1`: L1 loss.
+  - `l2`: L2 loss (mean squared error).
+  - `huber`: Huber loss.
+  - `smooth_l1`: Smooth L1 loss.
+
+- **`--huber_schedule`**, **`--huber_c`**, **`--huber_scale`**: These are parameters for Huber loss. They are used when `--loss_type` is `huber` or `smooth_l1`.
+
+- **`--weighting_scheme`**, **`--logit_mean`**, **`--logit_std`**, **`--mode_scale`**: These options allow you to adjust the loss weighting for each timestep. For details, refer to the [`sd3_train_network.md` guide](sd3_train_network.md).
+
+- **`--fused_backward_pass`**: Fuses the backward pass and optimizer step to reduce VRAM usage.
+
+<details>
+<summary>日本語</summary>
+
+- **`--ip_noise_gamma`**: Input Perturbationノイズのガンマ値を調整します。
+- **`--loss_type`**: 学習に用いる損失関数を指定します。
+- **`--huber_schedule`**, **`--huber_c`**, **`--huber_scale`**: Huber損失のパラメータです。
+- **`--weighting_scheme`**, **`--logit_mean`**, **`--logit_std`**, **`--mode_scale`**: 各タイムステップの損失の重み付けを調整します。
+- **`--fused_backward_pass`**: バックワードパスとオプティマイザステップを融合してVRAM使用量を削減します。
+
+</details>
+
+## 8. Using the Inference Script / 推論スクリプトの使用法
+
+The `hunyuan_image_minimal_inference.py` script allows you to generate images using trained LoRA models. Here's a basic usage example:
+
+```bash
+python hunyuan_image_minimal_inference.py \
+  --dit "<path to hunyuanimage2.1.safetensors>" \
+  --text_encoder "<path to qwen_2.5_vl_7b.safetensors>" \
+  --byt5 "<path to byt5_small_glyphxl_fp16.safetensors>" \
+  --vae "<path to hunyuan_image_2.1_vae_fp16.safetensors>" \
+  --lora_weight "<path to your trained LoRA>" \
+  --lora_multiplier 1.0 \
+  --attn_mode "torch" \
+  --prompt "A cute cartoon penguin in a snowy landscape" \
+  --image_size 2048 2048 \
+  --infer_steps 50 \
+  --guidance_scale 3.5 \
+  --flow_shift 5.0 \
+  --seed 542017 \
+  --save_path "output_image.png"
+```
+
+**Key Options:**
+- `--fp8_scaled`: Use scaled FP8 format for reduced VRAM usage during inference
+- `--blocks_to_swap`: Swap blocks to CPU to reduce VRAM usage
+- `--image_size`: Resolution in **height width**  (inference is most stable at 2560x1536, 2304x1792, 2048x2048, 1792x2304, 1536x2560 according to the official repo)
+- `--guidance_scale`: CFG scale (default: 3.5)
+- `--flow_shift`: Flow matching shift parameter (default: 5.0)
+- `--text_encoder_cpu`: Run the text encoders on CPU to reduce VRAM usage
+- `--vae_chunk_size`: Chunk size for VAE decoding to reduce memory usage (default: None, no chunking). 16 is recommended if enabled.
+- `--apg_start_step_general` and `--apg_start_step_ocr`: Start steps for APG (Adaptive Projected Guidance) if using APG during inference. `5` and `38` are the official recommended values for 50 steps. If this value exceeds `--infer_steps`, APG will not be applied.
+- `--guidance_rescale`: Rescales the guidance for steps before APG starts. Default is `0.0` (no rescaling). If you use this option, a value around `0.5` might be good starting point.
+- `--guidance_rescale_apg`: Rescales the guidance for APG. Default is `0.0` (no rescaling). This option doesn't seem to have a large effect, but if you use it, a value around `0.5` might be a good starting point.
+
+`--split_attn` is not supported (since inference is done one at a time). `--fp8_vl` is not supported, please use CPU for the text encoder if VRAM is insufficient.
+
+<details>
+<summary>日本語</summary>
+
+`hunyuan_image_minimal_inference.py`スクリプトを使用して、学習したLoRAモデルで画像を生成できます。基本的な使用例は英語のドキュメントを参照してください。
+
+**主要なオプション:**
+- `--fp8_scaled`: VRAM使用量削減のためのスケールFP8形式
+- `--blocks_to_swap`: VRAM使用量削減のためのブロックスワップ
+- `--image_size`: 解像度（2048x2048で最も安定）
+- `--guidance_scale`: CFGスケール（推奨: 3.5）
+- `--flow_shift`: Flow Matchingシフトパラメータ（デフォルト: 5.0）
+- `--text_encoder_cpu`: テキストエンコーダをCPUで実行してVRAM使用量削減
+- `--vae_chunk_size`: VAEデコーディングのチャンクサイズ（デフォルト: None、チャンク処理なし）。有効にする場合は16を推奨。
+- `--apg_start_step_general` と `--apg_start_step_ocr`: 推論中にAPGを使用する場合の開始ステップ。50ステップの場合、公式推奨値はそれぞれ5と38です。この値が`--infer_steps`を超えると、APGは適用されません。
+- `--guidance_rescale`: APG開始前のステップに対するガイダンスのリスケーリング。デフォルトは0.0（リスケーリングなし）。使用する場合、0.5程度から始めて調整してください。
+- `--guidance_rescale_apg`: APGに対するガイダンスのリスケーリング。デフォルトは0.0（リスケーリングなし）。このオプションは大きな効果はないようですが、使用する場合は0.5程度から始めて調整してください。
+
+`--split_attn`はサポートされていません（1件ずつ推論するため）。`--fp8_vl`もサポートされていません。VRAMが不足する場合はテキストエンコーダをCPUで実行してください。
+
+</details>
+
+## 9. Related Tools / 関連ツール
+
+### `networks/convert_hunyuan_image_lora_to_comfy.py`
+
+A script to convert LoRA models to ComfyUI-compatible format. The formats differ slightly, so conversion is necessary. You can convert from the sd-scripts format to ComfyUI format with:
+
+```bash
+python networks/convert_hunyuan_image_lora_to_comfy.py path/to/source.safetensors path/to/destination.safetensors
+```
+
+Using the `--reverse` option allows conversion in the opposite direction (ComfyUI format to sd-scripts format). However, reverse conversion is only possible for LoRAs converted by this script. LoRAs created with other training tools cannot be converted.
+
+<details>
+<summary>日本語</summary>
+
+**`networks/convert_hunyuan_image_lora_to_comfy.py`**
+
+LoRAモデルをComfyUI互換形式に変換するスクリプト。わずかに形式が異なるため、変換が必要です。以下の指定で、sd-scriptsの形式からComfyUI形式に変換できます。
+
+```bash
+python networks/convert_hunyuan_image_lora_to_comfy.py path/to/source.safetensors path/to/destination.safetensors
+```
+
+`--reverse`オプションを付けると、逆変換（ComfyUI形式からsd-scripts形式）も可能です。ただし、逆変換ができるのはこのスクリプトで変換したLoRAに限ります。他の学習ツールで作成したLoRAは変換できません。
+
+</details>
+
+## 10. Others / その他
+
+`hunyuan_image_train_network.py` includes many features common with `train_network.py`, such as sample image generation (`--sample_prompts`, etc.) and detailed optimizer settings. For these features, refer to the [`train_network.py` guide](train_network.md#5-other-features--その他の機能) or the script help (`python hunyuan_image_train_network.py --help`).
+
+<details>
+<summary>日本語</summary>
+
+`hunyuan_image_train_network.py`には、サンプル画像の生成 (`--sample_prompts`など) や詳細なオプティマイザ設定など、`train_network.py`と共通の機能も多く存在します。これらについては、[`train_network.py`のガイド](train_network.md#5-other-features--その他の機能)やスクリプトのヘルプ (`python hunyuan_image_train_network.py --help`) を参照してください。
+
+</details>

docs/lumina_train_network.md

@@ -0,0 +1,319 @@
+# LoRA Training Guide for Lumina Image 2.0 using `lumina_train_network.py` / `lumina_train_network.py` を用いたLumina Image 2.0モデルのLoRA学習ガイド
+
+This document explains how to train LoRA (Low-Rank Adaptation) models for Lumina Image 2.0 using `lumina_train_network.py` in the `sd-scripts` repository.
+
+## 1. Introduction / はじめに
+
+`lumina_train_network.py` trains additional networks such as LoRA for Lumina Image 2.0 models. Lumina Image 2.0 adopts a Next-DiT (Next-generation Diffusion Transformer) architecture, which differs from previous Stable Diffusion models. It uses a single text encoder (Gemma2) and a dedicated AutoEncoder (AE).
+
+This guide assumes you already understand the basics of LoRA training. For common usage and options, see [the train_network.py guide](./train_network.md). Some parameters are similar to those in [`sd3_train_network.py`](sd3_train_network.md) and [`flux_train_network.py`](flux_train_network.md).
+
+**Prerequisites:**
+
+* The `sd-scripts` repository has been cloned and the Python environment is ready.
+* A training dataset has been prepared. See the [Dataset Configuration Guide](./config_README-en.md).
+* Lumina Image 2.0 model files for training are available.
+
+<details>
+<summary>日本語</summary>
+
+`lumina_train_network.py`は、Lumina Image 2.0モデルに対してLoRAなどの追加ネットワークを学習させるためのスクリプトです。Lumina Image 2.0は、Next-DiT (Next-generation Diffusion Transformer) と呼ばれる新しいアーキテクチャを採用しており、従来のStable Diffusionモデルとは構造が異なります。テキストエンコーダーとしてGemma2を単体で使用し、専用のAutoEncoder (AE) を使用します。
+
+このガイドは、基本的なLoRA学習の手順を理解しているユーザーを対象としています。基本的な使い方や共通のオプションについては、`train_network.py`のガイド（作成中）を参照してください。また一部のパラメータは [`sd3_train_network.py`](sd3_train_network.md) や [`flux_train_network.py`](flux_train_network.md) と同様のものがあるため、そちらも参考にしてください。
+
+**前提条件:**
+
+*   `sd-scripts`リポジトリのクローンとPython環境のセットアップが完了していること。
+*   学習用データセットの準備が完了していること。（データセットの準備については[データセット設定ガイド](./config_README-en.md)を参照してください）
+*   学習対象のLumina Image 2.0モデルファイルが準備できていること。
+</details>
+
+## 2. Differences from `train_network.py` / `train_network.py` との違い
+
+`lumina_train_network.py` is based on `train_network.py` but modified for Lumina Image 2.0. Main differences are:
+
+* **Target models:** Lumina Image 2.0 models.
+* **Model structure:** Uses Next-DiT (Transformer based) instead of U-Net and employs a single text encoder (Gemma2). The AutoEncoder (AE) is not compatible with SDXL/SD3/FLUX.
+* **Arguments:** Options exist to specify the Lumina Image 2.0 model, Gemma2 text encoder and AE. With a single `.safetensors` file, these components are typically provided separately.
+* **Incompatible arguments:** Stable Diffusion v1/v2 options such as `--v2`, `--v_parameterization` and `--clip_skip` are not used.
+* **Lumina specific options:** Additional parameters for timestep sampling, model prediction type, discrete flow shift, and system prompt.
+
+<details>
+<summary>日本語</summary>
+`lumina_train_network.py`は`train_network.py`をベースに、Lumina Image 2.0モデルに対応するための変更が加えられています。主な違いは以下の通りです。
+
+*   **対象モデル:** Lumina Image 2.0モデルを対象とします。
+*   **モデル構造:** U-Netの代わりにNext-DiT (Transformerベース) を使用します。Text EncoderとしてGemma2を単体で使用し、専用のAutoEncoder (AE) を使用します。
+*   **引数:** Lumina Image 2.0モデル、Gemma2 Text Encoder、AEを指定する引数があります。通常、これらのコンポーネントは個別に提供されます。
+*   **一部引数の非互換性:** Stable Diffusion v1/v2向けの引数（例: `--v2`, `--v_parameterization`, `--clip_skip`）はLumina Image 2.0の学習では使用されません。
+*   **Lumina特有の引数:** タイムステップのサンプリング、モデル予測タイプ、離散フローシフト、システムプロンプトに関する引数が追加されています。
+</details>
+
+## 3. Preparation / 準備
+
+The following files are required before starting training:
+
+1. **Training script:** `lumina_train_network.py`
+2. **Lumina Image 2.0 model file:** `.safetensors` file for the base model.
+3. **Gemma2 text encoder file:** `.safetensors` file for the text encoder.
+4. **AutoEncoder (AE) file:** `.safetensors` file for the AE.
+5. **Dataset definition file (.toml):** Dataset settings in TOML format. (See the [Dataset Configuration Guide](./config_README-en.md). In this document we use `my_lumina_dataset_config.toml` as an example.
+
+
+**Model Files:**
+* Lumina Image 2.0: `lumina-image-2.safetensors` ([full precision link](https://huggingface.co/rockerBOO/lumina-image-2/blob/main/lumina-image-2.safetensors)) or `lumina_2_model_bf16.safetensors` ([bf16 link](https://huggingface.co/Comfy-Org/Lumina_Image_2.0_Repackaged/blob/main/split_files/diffusion_models/lumina_2_model_bf16.safetensors))
+* Gemma2 2B (fp16): `gemma-2-2b.safetensors` ([link](https://huggingface.co/Comfy-Org/Lumina_Image_2.0_Repackaged/blob/main/split_files/text_encoders/gemma_2_2b_fp16.safetensors))
+* AutoEncoder: `ae.safetensors` ([link](https://huggingface.co/Comfy-Org/Lumina_Image_2.0_Repackaged/blob/main/split_files/vae/ae.safetensors)) (same as FLUX)
+
+
+<details>
+<summary>日本語</summary>
+学習を開始する前に、以下のファイルが必要です。
+
+1.  **学習スクリプト:** `lumina_train_network.py`
+2.  **Lumina Image 2.0モデルファイル:** 学習のベースとなるLumina Image 2.0モデルの`.safetensors`ファイル。
+3.  **Gemma2テキストエンコーダーファイル:** Gemma2テキストエンコーダーの`.safetensors`ファイル。
+4.  **AutoEncoder (AE) ファイル:** AEの`.safetensors`ファイル。
+5.  **データセット定義ファイル (.toml):** 学習データセットの設定を記述したTOML形式のファイル。（詳細は[データセット設定ガイド](./config_README-en.md)を参照してください）。
+    *   例として`my_lumina_dataset_config.toml`を使用します。
+
+**モデルファイル** は英語ドキュメントの通りです。
+
+</details>
+
+## 4. Running the Training / 学習の実行
+
+Execute `lumina_train_network.py` from the terminal to start training. The overall command-line format is the same as `train_network.py`, but Lumina Image 2.0 specific options must be supplied.
+
+Example command:
+
+```bash
+accelerate launch --num_cpu_threads_per_process 1 lumina_train_network.py \
+  --pretrained_model_name_or_path="lumina-image-2.safetensors" \
+  --gemma2="gemma-2-2b.safetensors" \
+  --ae="ae.safetensors" \
+  --dataset_config="my_lumina_dataset_config.toml" \
+  --output_dir="./output" \
+  --output_name="my_lumina_lora" \
+  --save_model_as=safetensors \
+  --network_module=networks.lora_lumina \
+  --network_dim=8 \
+  --network_alpha=8 \
+  --learning_rate=1e-4 \
+  --optimizer_type="AdamW" \
+  --lr_scheduler="constant" \
+  --timestep_sampling="nextdit_shift" \
+  --discrete_flow_shift=6.0 \
+  --model_prediction_type="raw" \
+  --system_prompt="You are an assistant designed to generate high-quality images based on user prompts." \
+  --max_train_epochs=10 \
+  --save_every_n_epochs=1 \
+  --mixed_precision="bf16" \
+  --gradient_checkpointing \
+  --cache_latents \
+  --cache_text_encoder_outputs
+```
+
+*(Write the command on one line or use `\` or `^` for line breaks.)*
+
+<details>
+<summary>日本語</summary>
+学習は、ターミナルから`lumina_train_network.py`を実行することで開始します。基本的なコマンドラインの構造は`train_network.py`と同様ですが、Lumina Image 2.0特有の引数を指定する必要があります。
+
+以下に、基本的なコマンドライン実行例を示します。
+
+```bash
+accelerate launch --num_cpu_threads_per_process 1 lumina_train_network.py \
+  --pretrained_model_name_or_path="lumina-image-2.safetensors" \
+  --gemma2="gemma-2-2b.safetensors" \
+  --ae="ae.safetensors" \
+  --dataset_config="my_lumina_dataset_config.toml" \
+  --output_dir="./output" \
+  --output_name="my_lumina_lora" \
+  --save_model_as=safetensors \
+  --network_module=networks.lora_lumina \
+  --network_dim=8 \
+  --network_alpha=8 \
+  --learning_rate=1e-4 \
+  --optimizer_type="AdamW" \
+  --lr_scheduler="constant" \
+  --timestep_sampling="nextdit_shift" \
+  --discrete_flow_shift=6.0 \
+  --model_prediction_type="raw" \
+  --system_prompt="You are an assistant designed to generate high-quality images based on user prompts." \
+  --max_train_epochs=10 \
+  --save_every_n_epochs=1 \
+  --mixed_precision="bf16" \
+  --gradient_checkpointing \
+  --cache_latents \
+  --cache_text_encoder_outputs
+```
+
+※実際には1行で書くか、適切な改行文字（`\` または `^`）を使用してください。
+</details>
+
+### 4.1. Explanation of Key Options / 主要なコマンドライン引数の解説
+
+Besides the arguments explained in the [train_network.py guide](train_network.md), specify the following Lumina Image 2.0 options. For shared options (`--output_dir`, `--output_name`, etc.), see that guide.
+
+#### Model Options / モデル関連
+
+* `--pretrained_model_name_or_path="<path to Lumina model>"` **required** – Path to the Lumina Image 2.0 model.
+* `--gemma2="<path to Gemma2 model>"` **required** – Path to the Gemma2 text encoder `.safetensors` file.
+* `--ae="<path to AE model>"` **required** – Path to the AutoEncoder `.safetensors` file.
+
+#### Lumina Image 2.0 Training Parameters / Lumina Image 2.0 学習パラメータ
+
+* `--gemma2_max_token_length=<integer>` – Max token length for Gemma2. Default is 256.
+* `--timestep_sampling=<choice>` – Timestep sampling method. Options: `sigma`, `uniform`, `sigmoid`, `shift`, `nextdit_shift`. Default `shift`. **Recommended: `nextdit_shift`**
+* `--discrete_flow_shift=<float>` – Discrete flow shift for the Euler Discrete Scheduler. Default `6.0`.
+* `--model_prediction_type=<choice>` – Model prediction processing method. Options: `raw`, `additive`, `sigma_scaled`. Default `raw`. **Recommended: `raw`**
+* `--system_prompt=<string>` – System prompt to prepend to all prompts. Recommended: `"You are an assistant designed to generate high-quality images based on user prompts."` or `"You are an assistant designed to generate high-quality images with the highest degree of image-text alignment based on textual prompts."`
+* `--use_flash_attn` – Use Flash Attention. Requires `pip install flash-attn` (may not be supported in all environments). If installed correctly, it speeds up training. 
+* `--use_sage_attn` – Use Sage Attention for the model.
+* `--sample_batch_size=<integer>` – Batch size to use for sampling, defaults to `--training_batch_size` value. Sample batches are bucketed by width, height, guidance scale, and seed.
+* `--sigmoid_scale=<float>` – Scale factor for sigmoid timestep sampling. Default `1.0`.
+
+#### Memory and Speed / メモリ・速度関連
+
+* `--blocks_to_swap=<integer>` **[experimental]** – Swap a number of Transformer blocks between CPU and GPU. More blocks reduce VRAM but slow training. Cannot be used with `--cpu_offload_checkpointing`.
+* `--cache_text_encoder_outputs` – Cache Gemma2 outputs to reduce memory usage.
+* `--cache_latents`, `--cache_latents_to_disk` – Cache AE outputs.
+* `--fp8_base` – Use FP8 precision for the base model.
+
+#### Network Arguments / ネットワーク引数
+
+For Lumina Image 2.0, you can specify different dimensions for various components:
+
+* `--network_args` can include:
+  * `"attn_dim=4"` – Attention dimension
+  * `"mlp_dim=4"` – MLP dimension  
+  * `"mod_dim=4"` – Modulation dimension
+  * `"refiner_dim=4"` – Refiner blocks dimension
+  * `"embedder_dims=[4,4,4]"` – Embedder dimensions for x, t, and caption embedders
+
+#### Incompatible or Deprecated Options / 非互換・非推奨の引数
+
+* `--v2`, `--v_parameterization`, `--clip_skip` – Options for Stable Diffusion v1/v2 that are not used for Lumina Image 2.0.
+
+<details>
+<summary>日本語</summary>
+
+[`train_network.py`のガイド](train_network.md)で説明されている引数に加え、以下のLumina Image 2.0特有の引数を指定します。共通の引数については、上記ガイドを参照してください。
+
+#### モデル関連
+
+*   `--pretrained_model_name_or_path="<path to Lumina model>"` **[必須]**
+    *   学習のベースとなるLumina Image 2.0モデルの`.safetensors`ファイルのパスを指定します。
+*   `--gemma2="<path to Gemma2 model>"` **[必須]**
+    *   Gemma2テキストエンコーダーの`.safetensors`ファイルのパスを指定します。
+*   `--ae="<path to AE model>"` **[必須]**
+    *   AutoEncoderの`.safetensors`ファイルのパスを指定します。
+
+#### Lumina Image 2.0 学習パラメータ
+
+*   `--gemma2_max_token_length=<integer>` – Gemma2で使用するトークンの最大長を指定します。デフォルトは256です。
+*   `--timestep_sampling=<choice>` – タイムステップのサンプリング方法を指定します。`sigma`, `uniform`, `sigmoid`, `shift`, `nextdit_shift`から選択します。デフォルトは`shift`です。**推奨: `nextdit_shift`**
+*   `--discrete_flow_shift=<float>` – Euler Discrete Schedulerの離散フローシフトを指定します。デフォルトは`6.0`です。
+*   `--model_prediction_type=<choice>` – モデル予測の処理方法を指定します。`raw`, `additive`, `sigma_scaled`から選択します。デフォルトは`raw`です。**推奨: `raw`**
+*   `--system_prompt=<string>` – 全てのプロンプトに前置するシステムプロンプトを指定します。推奨: `"You are an assistant designed to generate high-quality images based on user prompts."` または `"You are an assistant designed to generate high-quality images with the highest degree of image-text alignment based on textual prompts."`
+*   `--use_flash_attn` – Flash Attentionを使用します。`pip install flash-attn`でインストールが必要です（環境によってはサポートされていません）。正しくインストールされている場合は、指定すると学習が高速化されます。
+*   `--use_sage_attn` – Sage Attentionを使用します。
+*   `--sample_batch_size=<integer>` – サンプリングに使用するバッチサイズ。デフォルトは `--training_batch_size` の値です。サンプルバッチは、幅、高さ、ガイダンススケール、シードによってバケット化されます。
+*   `--sigmoid_scale=<float>` – sigmoidタイムステップサンプリングのスケール係数を指定します。デフォルトは`1.0`です。
+
+#### メモリ・速度関連
+
+*   `--blocks_to_swap=<integer>` **[実験的機能]** – TransformerブロックをCPUとGPUでスワップしてVRAMを節約します。`--cpu_offload_checkpointing`とは併用できません。
+*   `--cache_text_encoder_outputs` – Gemma2の出力をキャッシュしてメモリ使用量を削減します。
+*   `--cache_latents`, `--cache_latents_to_disk` – AEの出力をキャッシュします。
+*   `--fp8_base` – ベースモデルにFP8精度を使用します。
+
+#### ネットワーク引数
+
+Lumina Image 2.0では、各コンポーネントに対して異なる次元を指定できます：
+
+*   `--network_args` には以下を含めることができます：
+    *   `"attn_dim=4"` – アテンション次元
+    *   `"mlp_dim=4"` – MLP次元
+    *   `"mod_dim=4"` – モジュレーション次元
+    *   `"refiner_dim=4"` – リファイナーブロック次元
+    *   `"embedder_dims=[4,4,4]"` – x、t、キャプションエンベッダーのエンベッダー次元
+
+#### 非互換・非推奨の引数
+
+*   `--v2`, `--v_parameterization`, `--clip_skip` – Stable Diffusion v1/v2向けの引数のため、Lumina Image 2.0学習では使用されません。
+</details>
+
+### 4.2. Starting Training / 学習の開始
+
+After setting the required arguments, run the command to begin training. The overall flow and how to check logs are the same as in the [train_network.py guide](train_network.md#32-starting-the-training--学習の開始).
+
+## 5. Using the Trained Model / 学習済みモデルの利用
+
+When training finishes, a LoRA model file (e.g. `my_lumina_lora.safetensors`) is saved in the directory specified by `output_dir`. Use this file with inference environments that support Lumina Image 2.0, such as ComfyUI with appropriate nodes.
+
+### Inference with scripts in this repository / このリポジトリのスクリプトを使用した推論
+
+The inference script is also available. The script is `lumina_minimal_inference.py`. See `--help` for options. 
+
+```
+python lumina_minimal_inference.py --pretrained_model_name_or_path path/to/lumina.safetensors  --gemma2_path path/to/gemma.safetensors" --ae_path  path/to/flux_ae.safetensors  --output_dir path/to/output_dir --offload --seed 1234 --prompt "Positive prompt" --system_prompt "You are an assistant designed to generate high-quality images based on user prompts."  --negative_prompt "negative prompt"  
+```
+
+`--add_system_prompt_to_negative_prompt` option can be used to add the system prompt to the negative prompt.
+
+`--lora_weights` option can be used to specify the LoRA weights file, and optional multiplier (like `path;1.0`).
+
+## 6. Others / その他
+
+`lumina_train_network.py` shares many features with `train_network.py`, such as sample image generation (`--sample_prompts`, etc.) and detailed optimizer settings. For these, see the [train_network.py guide](train_network.md#5-other-features--その他の機能) or run `python lumina_train_network.py --help`.
+
+### 6.1. Recommended Settings / 推奨設定
+
+Based on the contributor's recommendations, here are the suggested settings for optimal training:
+
+**Key Parameters:**
+* `--timestep_sampling="nextdit_shift"`
+* `--discrete_flow_shift=6.0`
+* `--model_prediction_type="raw"`
+* `--mixed_precision="bf16"`
+
+**System Prompts:**
+* General purpose: `"You are an assistant designed to generate high-quality images based on user prompts."`
+* High image-text alignment: `"You are an assistant designed to generate high-quality images with the highest degree of image-text alignment based on textual prompts."`
+
+**Sample Prompts:**
+Sample prompts can include CFG truncate (`--ctr`) and Renorm CFG (`-rcfg`) parameters:
+* `--ctr 0.25 --rcfg 1.0` (default values)
+
+<details>
+<summary>日本語</summary>
+
+必要な引数を設定し、コマンドを実行すると学習が開始されます。基本的な流れやログの確認方法は[`train_network.py`のガイド](train_network.md#32-starting-the-training--学習の開始)と同様です。
+
+学習が完了すると、指定した`output_dir`にLoRAモデルファイル（例: `my_lumina_lora.safetensors`）が保存されます。このファイルは、Lumina Image 2.0モデルに対応した推論環境（例: ComfyUI + 適切なノード）で使用できます。
+
+当リポジトリ内の推論スクリプトを用いて推論することも可能です。スクリプトは`lumina_minimal_inference.py`です。オプションは`--help`で確認できます。記述例は英語版のドキュメントをご確認ください。
+
+`lumina_train_network.py`には、サンプル画像の生成 (`--sample_prompts`など) や詳細なオプティマイザ設定など、`train_network.py`と共通の機能も多く存在します。これらについては、[`train_network.py`のガイド](train_network.md#5-other-features--その他の機能)やスクリプトのヘルプ (`python lumina_train_network.py --help`) を参照してください。
+
+### 6.1. 推奨設定
+
+コントリビューターの推奨に基づく、最適な学習のための推奨設定：
+
+**主要パラメータ:**
+* `--timestep_sampling="nextdit_shift"`
+* `--discrete_flow_shift=6.0`
+* `--model_prediction_type="raw"`
+* `--mixed_precision="bf16"`
+
+**システムプロンプト:**
+* 汎用目的: `"You are an assistant designed to generate high-quality images based on user prompts."`
+* 高い画像-テキスト整合性: `"You are an assistant designed to generate high-quality images with the highest degree of image-text alignment based on textual prompts."`
+
+**サンプルプロンプト:**
+サンプルプロンプトには CFG truncate (`--ctr`) と Renorm CFG (`--rcfg`) パラメータを含めることができます：
+* `--ctr 0.25 --rcfg 1.0` (デフォルト値)
+
+</details>

docs/masked_loss_README-ja.md

@@ -0,0 +1,57 @@
+## マスクロスについて
+
+マスクロスは、入力画像のマスクで指定された部分だけ損失計算することで、画像の一部分だけを学習することができる機能です。
+たとえばキャラクタを学習したい場合、キャラクタ部分だけをマスクして学習することで、背景を無視して学習することができます。
+
+マスクロスのマスクには、二種類の指定方法があります。
+
+- マスク画像を用いる方法
+- 透明度（アルファチャネル）を使用する方法
+
+なお、サンプルは [ずんずんPJイラスト/3Dデータ](https://zunko.jp/con_illust.html) の「AI画像モデル用学習データ」を使用しています。
+
+### マスク画像を用いる方法
+
+学習画像それぞれに対応するマスク画像を用意する方法です。学習画像と同じファイル名のマスク画像を用意し、それを学習画像と別のディレクトリに保存します。
+
+- 学習画像
+  ![image](https://github.com/kohya-ss/sd-scripts/assets/52813779/607c5116-5f62-47de-8b66-9c4a597f0441)
+- マスク画像
+  ![image](https://github.com/kohya-ss/sd-scripts/assets/52813779/53e9b0f8-a4bf-49ed-882d-4026f84e8450)
+
+```.toml
+[[datasets.subsets]]
+image_dir = "/path/to/a_zundamon"
+caption_extension = ".txt"
+conditioning_data_dir = "/path/to/a_zundamon_mask"
+num_repeats = 8
+```
+
+マスク画像は、学習画像と同じサイズで、学習する部分を白、無視する部分を黒で描画します。グレースケールにも対応しています（127 ならロス重みが 0.5 になります）。なお、正確にはマスク画像の R チャネルが用いられます。
+
+DreamBooth 方式の dataset で、`conditioning_data_dir` で指定したディレクトリにマスク画像を保存してください。ControlNet のデータセットと同じですので、詳細は [ControlNet-LLLite](train_lllite_README-ja.md#データセットの準備) を参照してください。
+
+### 透明度（アルファチャネル）を使用する方法
+
+学習画像の透明度（アルファチャネル）がマスクとして使用されます。透明度が 0 の部分は無視され、255 の部分は学習されます。半透明の場合は、その透明度に応じてロス重みが変化します（127 ならおおむね 0.5）。
+
+![image](https://github.com/kohya-ss/sd-scripts/assets/52813779/0baa129b-446a-4aac-b98c-7208efb0e75e)
+
+※それぞれの画像は透過PNG
+
+学習時のスクリプトのオプションに `--alpha_mask` を指定するか、dataset の設定ファイルの subset で、`alpha_mask` を指定してください。たとえば、以下のようになります。
+
+```toml
+[[datasets.subsets]]
+image_dir = "/path/to/image/dir"
+caption_extension = ".txt"
+num_repeats = 8
+alpha_mask = true
+```
+
+## 学習時の注意事項
+
+- 現時点では DreamBooth 方式の dataset のみ対応しています。
+- マスクは latents のサイズ、つまり 1/8 に縮小されてから適用されます。そのため、細かい部分（たとえばアホ毛やイヤリングなど）はうまく学習できない可能性があります。マスクをわずかに拡張するなどの工夫が必要かもしれません。
+- マスクロスを用いる場合、学習対象外の部分をキャプションに含める必要はないかもしれません。（要検証）
+- `alpha_mask` の場合、マスクの有無を切り替えると latents キャッシュが自動的に再生成されます。

docs/masked_loss_README.md

@@ -0,0 +1,56 @@
+## Masked Loss
+
+Masked loss is a feature that allows you to train only part of an image by calculating the loss only for the part specified by the mask of the input image. For example, if you want to train a character, you can train only the character part by masking it, ignoring the background.
+
+There are two ways to specify the mask for masked loss.
+
+- Using a mask image
+- Using transparency (alpha channel) of the image
+
+The sample uses the "AI image model training data" from [ZunZunPJ Illustration/3D Data](https://zunko.jp/con_illust.html).
+
+### Using a mask image
+
+This is a method of preparing a mask image corresponding to each training image. Prepare a mask image with the same file name as the training image and save it in a different directory from the training image.
+
+- Training image
+  ![image](https://github.com/kohya-ss/sd-scripts/assets/52813779/607c5116-5f62-47de-8b66-9c4a597f0441)
+- Mask image
+  ![image](https://github.com/kohya-ss/sd-scripts/assets/52813779/53e9b0f8-a4bf-49ed-882d-4026f84e8450)
+
+```.toml
+[[datasets.subsets]]
+image_dir = "/path/to/a_zundamon"
+caption_extension = ".txt"
+conditioning_data_dir = "/path/to/a_zundamon_mask"
+num_repeats = 8
+```
+
+The mask image is the same size as the training image, with the part to be trained drawn in white and the part to be ignored in black. It also supports grayscale (127 gives a loss weight of 0.5). The R channel of the mask image is used currently.
+
+Use the dataset in the DreamBooth method, and save the mask image in the directory specified by `conditioning_data_dir`. It is the same as the ControlNet dataset, so please refer to [ControlNet-LLLite](train_lllite_README.md#Preparing-the-dataset) for details.
+
+### Using transparency (alpha channel) of the image
+
+The transparency (alpha channel) of the training image is used as a mask. The part with transparency 0 is ignored, the part with transparency 255 is trained. For semi-transparent parts, the loss weight changes according to the transparency (127 gives a weight of about 0.5).
+
+![image](https://github.com/kohya-ss/sd-scripts/assets/52813779/0baa129b-446a-4aac-b98c-7208efb0e75e)
+
+※Each image is a transparent PNG
+
+Specify `--alpha_mask` in the training script options or specify `alpha_mask` in the subset of the dataset configuration file. For example, it will look like this.
+
+```toml
+[[datasets.subsets]]
+image_dir = "/path/to/image/dir"
+caption_extension = ".txt"
+num_repeats = 8
+alpha_mask = true
+```
+
+## Notes on training
+
+- At the moment, only the dataset in the DreamBooth method is supported.
+- The mask is applied after the size is reduced to 1/8, which is the size of the latents. Therefore, fine details (such as ahoge or earrings) may not be learned well. Some dilations of the mask may be necessary.
+- If using masked loss, it may not be necessary to include parts that are not to be trained in the caption. (To be verified)
+- In the case of `alpha_mask`, the latents cache is automatically regenerated when the enable/disable state of the mask is switched.

docs/sd3_train_network.md

@@ -0,0 +1,355 @@
+# LoRA Training Guide for Stable Diffusion 3/3.5 using `sd3_train_network.py` / `sd3_train_network.py` を用いたStable Diffusion 3/3.5モデルのLoRA学習ガイド
+
+This document explains how to train LoRA (Low-Rank Adaptation) models for Stable Diffusion 3 (SD3) and Stable Diffusion 3.5 (SD3.5) using `sd3_train_network.py` in the `sd-scripts` repository.
+
+## 1. Introduction / はじめに
+
+`sd3_train_network.py` trains additional networks such as LoRA for SD3/3.5 models. SD3 adopts a new architecture called MMDiT (Multi-Modal Diffusion Transformer), so its structure differs from previous Stable Diffusion models. With this script you can create LoRA models specialized for SD3/3.5.
+
+This guide assumes you already understand the basics of LoRA training. For common usage and options, see the [train_network.py guide](train_network.md). Some parameters are the same as those in [`sdxl_train_network.py`](sdxl_train_network.md).
+
+**Prerequisites:**
+
+* The `sd-scripts` repository has been cloned and the Python environment is ready.
+* A training dataset has been prepared. See the [Dataset Configuration Guide](link/to/dataset/config/doc).
+* SD3/3.5 model files for training are available.
+
+<details>
+<summary>日本語</summary>
+
+`sd3_train_network.py`は、Stable Diffusion 3/3.5モデルに対してLoRAなどの追加ネットワークを学習させるためのスクリプトです。SD3は、MMDiT (Multi-Modal Diffusion Transformer) と呼ばれる新しいアーキテクチャを採用しており、従来のStable Diffusionモデルとは構造が異なります。このスクリプトを使用することで、SD3/3.5モデルに特化したLoRAモデルを作成できます。
+
+このガイドは、基本的なLoRA学習の手順を理解しているユーザーを対象としています。基本的な使い方や共通のオプションについては、[`train_network.py`のガイド](train_network.md)を参照してください。また一部のパラメータは [`sdxl_train_network.py`](sdxl_train_network.md) と同様のものがあるため、そちらも参考にしてください。
+
+**前提条件:**
+
+*   `sd-scripts`リポジトリのクローンとPython環境のセットアップが完了していること。
+*   学習用データセットの準備が完了していること。（データセットの準備については[データセット設定ガイド](link/to/dataset/config/doc)を参照してください）
+*   学習対象のSD3/3.5モデルファイルが準備できていること。
+</details>
+
+## 2. Differences from `train_network.py` / `train_network.py` との違い
+
+`sd3_train_network.py` is based on `train_network.py` but modified for SD3/3.5. Main differences are:
+
+* **Target models:** Stable Diffusion 3 and 3.5 Medium/Large.
+* **Model structure:** Uses MMDiT (Transformer based) instead of U-Net and employs three text encoders: CLIP-L, CLIP-G and T5-XXL. The VAE is not compatible with SDXL.
+* **Arguments:** Options exist to specify the SD3/3.5 model, text encoders and VAE. With a single `.safetensors` file, these paths are detected automatically, so separate paths are optional.
+* **Incompatible arguments:** Stable Diffusion v1/v2 options such as `--v2`, `--v_parameterization` and `--clip_skip` are not used.
+* **SD3 specific options:** Additional parameters for attention masks, dropout rates, positional embedding adjustments (for SD3.5), timestep sampling and loss weighting.
+
+<details>
+<summary>日本語</summary>
+`sd3_train_network.py`は`train_network.py`をベースに、SD3/3.5モデルに対応するための変更が加えられています。主な違いは以下の通りです。
+
+*   **対象モデル:** Stable Diffusion 3, 3.5 Medium / Large モデルを対象とします。
+*   **モデル構造:** U-Netの代わりにMMDiT (Transformerベース) を使用します。Text EncoderとしてCLIP-L, CLIP-G, T5-XXLの三つを使用します。VAEはSDXLと互換性がありません。
+*   **引数:** SD3/3.5モデル、Text Encoder群、VAEを指定する引数があります。ただし、単一ファイルの`.safetensors`形式であれば、内部で自動的に分離されるため、個別のパス指定は必須ではありません。
+*   **一部引数の非互換性:** Stable Diffusion v1/v2向けの引数（例: `--v2`, `--v_parameterization`, `--clip_skip`）はSD3/3.5の学習では使用されません。
+*   **SD3特有の引数:** Text Encoderのアテンションマスクやドロップアウト率、Positional Embeddingの調整（SD3.5向け）、タイムステップのサンプリングや損失の重み付けに関する引数が追加されています。
+</details>
+
+## 3. Preparation / 準備
+
+The following files are required before starting training:
+
+1. **Training script:** `sd3_train_network.py`
+2. **SD3/3.5 model file:** `.safetensors` file for the base model and paths to each text encoder. Single-file format can also be used.
+3. **Dataset definition file (.toml):** Dataset settings in TOML format. (See the [Dataset Configuration Guide](link/to/dataset/config/doc).) In this document we use `my_sd3_dataset_config.toml` as an example.
+
+<details>
+<summary>日本語</summary>
+学習を開始する前に、以下のファイルが必要です。
+
+1.  **学習スクリプト:** `sd3_train_network.py`
+2.  **SD3/3.5モデルファイル:** 学習のベースとなるSD3/3.5モデルの`.safetensors`ファイル。またText Encoderをそれぞれ対応する引数でパスを指定します。
+    * 単一ファイル形式も使用可能です。
+3.  **データセット定義ファイル (.toml):** 学習データセットの設定を記述したTOML形式のファイル。（詳細は[データセット設定ガイド](link/to/dataset/config/doc)を参照してください）。
+    *   例として`my_sd3_dataset_config.toml`を使用します。
+</details>
+
+## 4. Running the Training / 学習の実行
+
+Execute `sd3_train_network.py` from the terminal to start training. The overall command-line format is the same as `train_network.py`, but SD3/3.5 specific options must be supplied.
+
+Example command:
+
+```bash
+accelerate launch --num_cpu_threads_per_process 1 sd3_train_network.py \
+  --pretrained_model_name_or_path="<path to SD3 model>" \
+  --clip_l="<path to CLIP-L model>" \
+  --clip_g="<path to CLIP-G model>" \
+  --t5xxl="<path to T5-XXL model>" \
+  --dataset_config="my_sd3_dataset_config.toml" \
+  --output_dir="<output directory for training results>" \
+  --output_name="my_sd3_lora" \
+  --save_model_as=safetensors \
+  --network_module=networks.lora \
+  --network_dim=16 \
+  --network_alpha=1 \
+  --learning_rate=1e-4 \
+  --optimizer_type="AdamW8bit" \
+  --lr_scheduler="constant" \
+  --sdpa \
+  --max_train_epochs=10 \
+  --save_every_n_epochs=1 \
+  --mixed_precision="fp16" \
+  --gradient_checkpointing \
+  --weighting_scheme="uniform" \
+  --blocks_to_swap=32
+```
+
+*(Write the command on one line or use `\` or `^` for line breaks.)*
+
+<details>
+<summary>日本語</summary>
+
+学習は、ターミナルから`sd3_train_network.py`を実行することで開始します。基本的なコマンドラインの構造は`train_network.py`と同様ですが、SD3/3.5特有の引数を指定する必要があります。
+
+以下に、基本的なコマンドライン実行例を示します。
+
+```bash
+accelerate launch --num_cpu_threads_per_process 1 sd3_train_network.py
+ --pretrained_model_name_or_path="<path to SD3 model>"
+ --clip_l="<path to CLIP-L model>"
+ --clip_g="<path to CLIP-G model>"
+ --t5xxl="<path to T5-XXL model>"
+ --dataset_config="my_sd3_dataset_config.toml"
+ --output_dir="<output directory for training results>"
+ --output_name="my_sd3_lora"
+ --save_model_as=safetensors
+ --network_module=networks.lora
+ --network_dim=16
+ --network_alpha=1
+ --learning_rate=1e-4
+ --optimizer_type="AdamW8bit"
+ --lr_scheduler="constant"
+ --sdpa
+ --max_train_epochs=10
+ --save_every_n_epochs=1
+ --mixed_precision="fp16"
+ --gradient_checkpointing
+ --weighting_scheme="uniform"
+ --blocks_to_swap=32
+```
+
+※実際には1行で書くか、適切な改行文字（`\` または `^`）を使用してください。
+
+</details>
+
+### 4.1. Explanation of Key Options / 主要なコマンドライン引数の解説
+
+Besides the arguments explained in the [train_network.py guide](train_network.md), specify the following SD3/3.5 options. For shared options (`--output_dir`, `--output_name`, etc.), see that guide.
+
+#### Model Options / モデル関連
+
+* `--pretrained_model_name_or_path="<path to SD3 model>"` **required** – Path to the SD3/3.5 model.
+* `--clip_l`, `--clip_g`, `--t5xxl`, `--vae` – Skip these if the base model is a single file; otherwise specify each `.safetensors` path. `--vae` is usually unnecessary unless you use a different VAE.
+
+#### SD3/3.5 Training Parameters / SD3/3.5 学習パラメータ
+
+* `--t5xxl_max_token_length=<integer>` – Max token length for T5-XXL. Default `256`.
+* `--apply_lg_attn_mask` – Apply an attention mask to CLIP-L/CLIP-G outputs.
+* `--apply_t5_attn_mask` – Apply an attention mask to T5-XXL outputs.
+* `--clip_l_dropout_rate`, `--clip_g_dropout_rate`, `--t5_dropout_rate` – Dropout rates for the text encoders. Default `0.0`.
+* `--pos_emb_random_crop_rate=<float>` **[SD3.5]** – Probability of randomly cropping the positional embedding.
+* `--enable_scaled_pos_embed` **[SD3.5][experimental]** – Scale positional embeddings when training with multiple resolutions.
+* `--training_shift=<float>` – Shift applied to the timestep distribution. Default `1.0`.
+* `--weighting_scheme=<choice>` – Weighting method for loss by timestep. Default `uniform`.
+* `--logit_mean=<float>` – Mean value for `logit_normal` weighting scheme. Default `0.0`.
+* `--logit_std=<float>` – Standard deviation for `logit_normal` weighting scheme. Default `1.0`.
+* `--mode_scale=<float>` – Scale factor for `mode` weighting scheme. Default `1.29`.
+
+#### Memory and Speed / メモリ・速度関連
+
+* `--blocks_to_swap=<integer>` **[experimental]** – Swap a number of Transformer blocks between CPU and GPU. More blocks reduce VRAM but slow training. Cannot be used with `--cpu_offload_checkpointing`.
+* `--cache_text_encoder_outputs` – Caches the outputs of the text encoders to reduce VRAM usage and speed up training. This is particularly effective for SD3, which uses three text encoders. Recommended when not training the text encoder LoRA. For more details, see the [`sdxl_train_network.py` guide](sdxl_train_network.md).
+* `--cache_text_encoder_outputs_to_disk` – Caches the text encoder outputs to disk when the above option is enabled.
+* `--t5xxl_device=<device>` **[not supported yet]** – Specifies the device for T5-XXL model. If not specified, uses accelerator's device.
+* `--t5xxl_dtype=<dtype>` **[not supported yet]** – Specifies the dtype for T5-XXL model. If not specified, uses default dtype from mixed precision.
+* `--save_clip` **[not supported yet]** – Saves CLIP models to checkpoint (unified checkpoint format not yet supported).
+* `--save_t5xxl` **[not supported yet]** – Saves T5-XXL model to checkpoint (unified checkpoint format not yet supported).
+
+#### Incompatible or Deprecated Options / 非互換・非推奨の引数
+
+* `--v2`, `--v_parameterization`, `--clip_skip` – Options for Stable Diffusion v1/v2 that are not used for SD3/3.5.
+
+<details>
+<summary>日本語</summary>
+
+[`train_network.py`のガイド](train_network.md)で説明されている引数に加え、以下のSD3/3.5特有の引数を指定します。共通の引数については、上記ガイドを参照してください。
+
+#### モデル関連
+
+*   `--pretrained_model_name_or_path="<path to SD3 model>"` **[必須]**
+    *   学習のベースとなるSD3/3.5モデルの`.safetensors`ファイルのパスを指定します。
+*   `--clip_l`, `--clip_g`, `--t5xxl`, `--vae`:
+    *   ベースモデルが単一ファイル形式の場合、これらの指定は不要です（自動的にモデル内部から読み込まれます）。
+    *   Text Encoderが別ファイルとして提供されている場合は、それぞれの`.safetensors`ファイルのパスを指定します。`--vae` はベースモデルに含まれているため、通常は指定する必要はありません（明示的に異なるVAEを使用する場合のみ指定）。
+
+#### SD3/3.5 学習パラメータ
+
+*   `--t5xxl_max_token_length=<integer>` – T5-XXLで使用するトークンの最大長を指定します。デフォルトは`256`です。
+*   `--apply_lg_attn_mask` – CLIP-L/CLIP-Gの出力にパディング用のマスクを適用します。
+*   `--apply_t5_attn_mask` – T5-XXLの出力にパディング用のマスクを適用します。
+*   `--clip_l_dropout_rate`, `--clip_g_dropout_rate`, `--t5_dropout_rate` – 各Text Encoderのドロップアウト率を指定します。デフォルトは`0.0`です。
+*   `--pos_emb_random_crop_rate=<float>` **[SD3.5向け]** – Positional Embeddingにランダムクロップを適用する確率を指定します。
+*   `--enable_scaled_pos_embed` **[SD3.5向け][実験的機能]** – マルチ解像度学習時に解像度に応じてPositional Embeddingをスケーリングします。
+*   `--training_shift=<float>` – タイムステップ分布を調整するためのシフト値です。デフォルトは`1.0`です。
+*   `--weighting_scheme=<choice>` – タイムステップに応じた損失の重み付け方法を指定します。デフォルトは`uniform`です。
+*   `--logit_mean=<float>` – `logit_normal`重み付けスキームの平均値です。デフォルトは`0.0`です。
+*   `--logit_std=<float>` – `logit_normal`重み付けスキームの標準偏差です。デフォルトは`1.0`です。
+*   `--mode_scale=<float>` – `mode`重み付けスキームのスケール係数です。デフォルトは`1.29`です。
+
+#### メモリ・速度関連
+
+*   `--blocks_to_swap=<integer>` **[実験的機能]** – TransformerブロックをCPUとGPUでスワップしてVRAMを節約します。`--cpu_offload_checkpointing`とは併用できません。
+*   `--cache_text_encoder_outputs` – Text Encoderの出力をキャッシュし、VRAM使用量削減と学習高速化を図ります。SD3は3つのText Encoderを持つため特に効果的です。Text EncoderのLoRAを学習しない場合に推奨されます。詳細は[`sdxl_train_network.py`のガイド](sdxl_train_network.md)を参照してください。
+*   `--cache_text_encoder_outputs_to_disk` – 上記オプションと併用し、Text Encoderの出力をディスクにキャッシュします。
+*   `--t5xxl_device=<device>` **[未サポート]** – T5-XXLモデルのデバイスを指定します。指定しない場合はacceleratorのデバイスを使用します。
+*   `--t5xxl_dtype=<dtype>` **[未サポート]** – T5-XXLモデルのdtypeを指定します。指定しない場合はデフォルトのdtype（mixed precisionから）を使用します。
+*   `--save_clip` **[未サポート]** – CLIPモデルをチェックポイントに保存します（統合チェックポイント形式は未サポート）。
+*   `--save_t5xxl` **[未サポート]** – T5-XXLモデルをチェックポイントに保存します（統合チェックポイント形式は未サポート）。
+
+#### 非互換・非推奨の引数
+
+*   `--v2`, `--v_parameterization`, `--clip_skip` – Stable Diffusion v1/v2向けの引数のため、SD3/3.5学習では使用されません。
+
+</details>
+
+### 4.2. Starting Training / 学習の開始
+
+After setting the required arguments, run the command to begin training. The overall flow and how to check logs are the same as in the [train_network.py guide](train_network.md#32-starting-the-training--学習の開始).
+
+<details>
+<summary>日本語</summary>
+
+必要な引数を設定したら、コマンドを実行して学習を開始します。全体の流れやログの確認方法は、[train_network.pyのガイド](train_network.md#32-starting-the-training--学習の開始)と同様です。
+
+</details>
+
+## 5. LoRA Target Modules / LoRAの学習対象モジュール
+
+When training LoRA with `sd3_train_network.py`, the following modules are targeted by default:
+
+*   **MMDiT (replaces U-Net)**:
+    *   `qkv` (Query, Key, Value) matrices and `proj_out` (output projection) in the attention blocks.
+*   **final_layer**:
+    *   The output layer at the end of MMDiT.
+
+By using `--network_args`, you can apply more detailed controls, such as setting different ranks (dimensions) for each module.
+
+### Specify rank for each layer in SD3 LoRA / 各層のランクを指定する
+
+You can specify the rank for each layer in SD3 by specifying the following network_args. If you specify `0`, LoRA will not be applied to that layer.
+
+When network_args is not specified, the default value (`network_dim`) is applied, same as before.
+
+|network_args|target layer|
+|---|---|
+|context_attn_dim|attn in context_block|
+|context_mlp_dim|mlp in context_block|
+|context_mod_dim|adaLN_modulation in context_block|
+|x_attn_dim|attn in x_block|
+|x_mlp_dim|mlp in x_block|
+|x_mod_dim|adaLN_modulation in x_block|
+
+`"verbose=True"` is also available for debugging. It shows the rank of each layer.
+
+example: 
+```
+--network_args "context_attn_dim=2" "context_mlp_dim=3" "context_mod_dim=4" "x_attn_dim=5" "x_mlp_dim=6" "x_mod_dim=7" "verbose=True"
+```
+
+You can apply LoRA to the conditioning layers of SD3 by specifying `emb_dims` in network_args. When specifying, be sure to specify 6 numbers in `[]` as a comma-separated list.
+
+example: 
+```
+--network_args "emb_dims=[2,3,4,5,6,7]"
+```
+
+Each number corresponds to `context_embedder`, `t_embedder`, `x_embedder`, `y_embedder`, `final_layer_adaLN_modulation`, `final_layer_linear`. The above example applies LoRA to all conditioning layers, with rank 2 for `context_embedder`, 3 for `t_embedder`, 4 for `context_embedder`, 5 for `y_embedder`, 6 for `final_layer_adaLN_modulation`, and 7 for `final_layer_linear`.
+
+If you specify `0`, LoRA will not be applied to that layer. For example, `[4,0,0,4,0,0]` applies LoRA only to `context_embedder` and `y_embedder`.
+
+### Specify blocks to train in SD3 LoRA training
+
+You can specify the blocks to train in SD3 LoRA training by specifying `train_block_indices` in network_args. The indices are 0-based. The default (when omitted) is to train all blocks. The indices are specified as a list of integers or a range of integers, like `0,1,5,8` or `0,1,4-5,7`. 
+
+The number of blocks depends on the model. The valid range is 0-(the number of blocks - 1). `all` is also available to train all blocks, `none` is also available to train no blocks.
+
+example: 
+```
+--network_args "train_block_indices=1,2,6-8" 
+```
+
+<details>
+<summary>日本語</summary>
+
+`sd3_train_network.py`でLoRAを学習させる場合、デフォルトでは以下のモジュールが対象となります。
+
+*   **MMDiT (U-Netの代替)**:
+    *   Attentionブロック内の`qkv`（Query, Key, Value）行列と、`proj_out`（出力Projection）。
+*   **final_layer**:
+    *   MMDiTの最後にある出力層。
+
+`--network_args` を使用することで、モジュールごとに異なるランク（次元数）を設定するなど、より詳細な制御が可能です。
+
+### SD3 LoRAで各層のランクを指定する
+
+各層のランクを指定するには、`--network_args`オプションを使用します。`0`を指定すると、その層にはLoRAが適用されません。
+
+network_argsが指定されない場合、デフォルト値（`network_dim`）が適用されます。
+
+|network_args|target layer|
+|---|---|
+|context_attn_dim|attn in context_block|
+|context_mlp_dim|mlp in context_block|
+|context_mod_dim|adaLN_modulation in context_block|
+|x_attn_dim|attn in x_block|
+|x_mlp_dim|mlp in x_block|
+|x_mod_dim|adaLN_modulation in x_block|
+
+`"verbose=True"`を指定すると、各層のランクが表示されます。
+
+例：
+
+```bash
+--network_args "context_attn_dim=2" "context_mlp_dim=3" "context_mod_dim=4" "x_attn_dim=5" "x_mlp_dim=6" "x_mod_dim=7" "verbose=True"
+```
+
+また、`emb_dims`を指定することで、SD3の条件付け層にLoRAを適用することもできます。指定する際は、必ず`[]`内にカンマ区切りで6つの数字を指定してください。
+
+```bash
+--network_args "emb_dims=[2,3,4,5,6,7]"
+```
+
+各数字は、`context_embedder`、`t_embedder`、`x_embedder`、`y_embedder`、`final_layer_adaLN_modulation`、`final_layer_linear`に対応しています。上記の例では、すべての条件付け層にLoRAを適用し、`context_embedder`に2、`t_embedder`に3、`x_embedder`に4、`y_embedder`に5、`final_layer_adaLN_modulation`に6、`final_layer_linear`に7のランクを設定しています。
+
+`0`を指定すると、その層にはLoRAが適用されません。例えば、`[4,0,0,4,0,0]`と指定すると、`context_embedder`と`y_embedder`のみにLoRAが適用されます。
+
+</details>
+
+
+## 6. Using the Trained Model / 学習済みモデルの利用
+
+When training finishes, a LoRA model file (e.g. `my_sd3_lora.safetensors`) is saved in the directory specified by `output_dir`. Use this file with inference environments that support SD3/3.5, such as ComfyUI.
+
+<details>
+<summary>日本語</summary>
+
+学習が完了すると、指定した`output_dir`にLoRAモデルファイル（例: `my_sd3_lora.safetensors`）が保存されます。このファイルは、SD3/3.5モデルに対応した推論環境（例: ComfyUIなど）で使用できます。
+
+</details>
+
+
+## 7. Others / その他
+
+`sd3_train_network.py` shares many features with `train_network.py`, such as sample image generation (`--sample_prompts`, etc.) and detailed optimizer settings. For these, see the [train_network.py guide](train_network.md#5-other-features--その他の機能) or run `python sd3_train_network.py --help`.
+
+<details>
+<summary>日本語</summary>
+
+`sd3_train_network.py`には、サンプル画像の生成 (`--sample_prompts`など) や詳細なオプティマイザ設定など、`train_network.py`と共通の機能も多く存在します。これらについては、[`train_network.py`のガイド](train_network.md#5-other-features--その他の機能)やスクリプトのヘルプ (`python sd3_train_network.py --help`) を参照してください。
+
+</details>

docs/sdxl_train_network.md

@@ -0,0 +1,321 @@
+# How to Use the SDXL LoRA Training Script `sdxl_train_network.py` / SDXL LoRA学習スクリプト `sdxl_train_network.py` の使い方
+
+This document explains the basic procedure for training a LoRA (Low-Rank Adaptation) model for SDXL (Stable Diffusion XL) using `sdxl_train_network.py` included in the `sd-scripts` repository.
+
+<details>
+<summary>日本語</summary>
+このドキュメントでは、`sd-scripts` リポジトリに含まれる `sdxl_train_network.py` を使用して、SDXL (Stable Diffusion XL) モデルに対する LoRA (Low-Rank Adaptation) モデルを学習する基本的な手順について解説します。
+</details>
+
+## 1. Introduction / はじめに
+
+`sdxl_train_network.py` is a script for training additional networks such as LoRA for SDXL models. The basic usage is common with `train_network.py` (see [How to Use the LoRA Training Script `train_network.py`](train_network.md)), but SDXL model-specific settings are required.
+
+This guide focuses on SDXL LoRA training, explaining the main differences from `train_network.py` and SDXL-specific configuration items.
+
+**Prerequisites:**
+
+* You have cloned the `sd-scripts` repository and set up the Python environment.
+* Your training dataset is ready. (Please refer to the [Dataset Preparation Guide](link/to/dataset/doc) for dataset preparation)
+* You have read [How to Use the LoRA Training Script `train_network.py`](train_network.md).
+
+<details>
+<summary>日本語</summary>
+`sdxl_train_network.py` は、SDXL モデルに対して LoRA などの追加ネットワークを学習させるためのスクリプトです。基本的な使い方は `train_network.py` ([LoRA学習スクリプト `train_network.py` の使い方](train_network.md) 参照) と共通ですが、SDXL モデル特有の設定が必要となります。
+
+このガイドでは、SDXL LoRA 学習に焦点を当て、`train_network.py` との主な違いや SDXL 特有の設定項目を中心に説明します。
+
+**前提条件:**
+
+*   `sd-scripts` リポジトリのクローンと Python 環境のセットアップが完了していること。
+*   学習用データセットの準備が完了していること。（データセットの準備については[データセット準備ガイド](link/to/dataset/doc)を参照してください）
+*   [LoRA学習スクリプト `train_network.py` の使い方](train_network.md) を一読していること。
+</details>
+
+## 2. Preparation / 準備
+
+Before starting training, you need the following files:
+
+1. **Training Script:** `sdxl_train_network.py`
+2. **Dataset Definition File (.toml):** A TOML format file describing the training dataset configuration.
+
+### About the Dataset Definition File
+
+The basic format of the dataset definition file (`.toml`) is the same as for `train_network.py`. Please refer to the [Dataset Configuration Guide](link/to/dataset/config/doc) and [How to Use the LoRA Training Script `train_network.py`](train_network.md#about-the-dataset-definition-file).
+
+For SDXL, it is common to use high-resolution datasets and the aspect ratio bucketing feature (`enable_bucket = true`).
+
+In this example, we'll use a file named `my_sdxl_dataset_config.toml`.
+
+<details>
+<summary>日本語</summary>
+学習を開始する前に、以下のファイルが必要です。
+
+1.  **学習スクリプト:** `sdxl_train_network.py`
+2.  **データセット定義ファイル (.toml):** 学習データセットの設定を記述した TOML 形式のファイル。
+
+### データセット定義ファイルについて
+
+データセット定義ファイル (`.toml`) の基本的な書き方は `train_network.py` と共通です。[データセット設定ガイド](link/to/dataset/config/doc) および [LoRA学習スクリプト `train_network.py` の使い方](train_network.md#データセット定義ファイルについて) を参照してください。
+
+SDXL では、高解像度のデータセットや、アスペクト比バケツ機能 (`enable_bucket = true`) の利用が一般的です。
+
+ここでは、例として `my_sdxl_dataset_config.toml` という名前のファイルを使用することにします。
+</details>
+
+## 3. Running the Training / 学習の実行
+
+Training starts by running `sdxl_train_network.py` from the terminal.
+
+Here's a basic command line execution example for SDXL LoRA training:
+
+```bash
+accelerate launch --num_cpu_threads_per_process 1 sdxl_train_network.py 
+ --pretrained_model_name_or_path="<SDXL base model path>" 
+ --dataset_config="my_sdxl_dataset_config.toml" 
+ --output_dir="<output directory for training results>" 
+ --output_name="my_sdxl_lora" 
+ --save_model_as=safetensors 
+ --network_module=networks.lora 
+ --network_dim=32 
+ --network_alpha=16 
+ --learning_rate=1e-4 
+ --unet_lr=1e-4 
+ --text_encoder_lr1=1e-5 
+ --text_encoder_lr2=1e-5 
+ --optimizer_type="AdamW8bit" 
+ --lr_scheduler="constant" 
+ --max_train_epochs=10 
+ --save_every_n_epochs=1 
+ --mixed_precision="bf16" 
+ --gradient_checkpointing 
+ --cache_text_encoder_outputs 
+ --cache_latents
+```
+
+Comparing with the execution example of `train_network.py`, the following points are different:
+
+* The script to execute is `sdxl_train_network.py`.
+* You specify an SDXL base model for `--pretrained_model_name_or_path`.
+* `--text_encoder_lr` is split into `--text_encoder_lr1` and `--text_encoder_lr2` (since SDXL has two Text Encoders).
+* `--mixed_precision` is recommended to be `bf16` or `fp16`.
+* `--cache_text_encoder_outputs` and `--cache_latents` are recommended to reduce VRAM usage.
+
+Next, we'll explain the main command line arguments that differ from `train_network.py`. For common arguments, please refer to [How to Use the LoRA Training Script `train_network.py`](train_network.md#31-main-command-line-arguments).
+
+<details>
+<summary>日本語</summary>
+学習は、ターミナルから `sdxl_train_network.py` を実行することで開始します。
+
+以下に、SDXL LoRA 学習における基本的なコマンドライン実行例を示します。
+
+```bash
+accelerate launch --num_cpu_threads_per_process 1 sdxl_train_network.py 
+ --pretrained_model_name_or_path="<SDXLベースモデルのパス>" 
+ --dataset_config="my_sdxl_dataset_config.toml" 
+ --output_dir="<学習結果の出力先ディレクトリ>" 
+ --output_name="my_sdxl_lora" 
+ --save_model_as=safetensors 
+ --network_module=networks.lora 
+ --network_dim=32 
+ --network_alpha=16 
+ --learning_rate=1e-4 
+ --unet_lr=1e-4 
+ --text_encoder_lr1=1e-5 
+ --text_encoder_lr2=1e-5 
+ --optimizer_type="AdamW8bit" 
+ --lr_scheduler="constant" 
+ --max_train_epochs=10 
+ --save_every_n_epochs=1 
+ --mixed_precision="bf16" 
+ --gradient_checkpointing 
+ --cache_text_encoder_outputs 
+ --cache_latents
+```
+
+`train_network.py` の実行例と比較すると、以下の点が異なります。
+
+*   実行するスクリプトが `sdxl_train_network.py` になります。
+*   `--pretrained_model_name_or_path` には SDXL のベースモデルを指定します。
+*   `--text_encoder_lr` が `--text_encoder_lr1` と `--text_encoder_lr2` に分かれています（SDXL は2つの Text Encoder を持つため）。
+*   `--mixed_precision` は `bf16` または `fp16` が推奨されます。
+*   `--cache_text_encoder_outputs` や `--cache_latents` は VRAM 使用量を削減するために推奨されます。
+
+次に、`train_network.py` との差分となる主要なコマンドライン引数について解説します。共通の引数については、[LoRA学習スクリプト `train_network.py` の使い方](train_network.md#31-主要なコマンドライン引数) を参照してください。
+</details>
+
+### 3.1. Main Command Line Arguments (Differences) / 主要なコマンドライン引数（差分）
+
+#### Model Related / モデル関連
+
+* `--pretrained_model_name_or_path="<model path>"` **[Required]**
+  * Specifies the **SDXL model** to be used as the base for training. You can specify a Hugging Face Hub model ID (e.g., `"stabilityai/stable-diffusion-xl-base-1.0"`), a local Diffusers format model directory, or a path to a `.safetensors` file.
+* `--v2`, `--v_parameterization`
+  * These arguments are for SD1.x/2.x. When using `sdxl_train_network.py`, since an SDXL model is assumed, these **typically do not need to be specified**.
+
+#### Dataset Related / データセット関連
+
+* `--dataset_config="<config file path>"`
+  * This is common with `train_network.py`.
+  * For SDXL, it is common to use high-resolution data and the bucketing feature (specify `enable_bucket = true` in the `.toml` file).
+
+#### Output & Save Related / 出力・保存関連
+
+* These are common with `train_network.py`.
+
+#### LoRA Parameters / LoRA パラメータ
+
+* These are common with `train_network.py`.
+
+#### Training Parameters / 学習パラメータ
+
+* `--learning_rate=1e-4`
+  * Overall learning rate. This becomes the default value if `unet_lr`, `text_encoder_lr1`, and `text_encoder_lr2` are not specified.
+* `--unet_lr=1e-4`
+  * Learning rate for LoRA modules in the U-Net part. If not specified, the value of `--learning_rate` is used.
+* `--text_encoder_lr1=1e-5`
+  * Learning rate for LoRA modules in **Text Encoder 1 (OpenCLIP ViT-G/14)**. If not specified, the value of `--learning_rate` is used. A smaller value than U-Net is recommended.
+* `--text_encoder_lr2=1e-5`
+  * Learning rate for LoRA modules in **Text Encoder 2 (CLIP ViT-L/14)**. If not specified, the value of `--learning_rate` is used. A smaller value than U-Net is recommended.
+* `--optimizer_type="AdamW8bit"`
+  * Common with `train_network.py`.
+* `--lr_scheduler="constant"`
+  * Common with `train_network.py`.
+* `--lr_warmup_steps`
+  * Common with `train_network.py`.
+* `--max_train_steps`, `--max_train_epochs`
+  * Common with `train_network.py`.
+* `--mixed_precision="bf16"`
+  * Mixed precision training setting. For SDXL, `bf16` or `fp16` is recommended. Choose the one supported by your GPU. This reduces VRAM usage and improves training speed.
+* `--gradient_accumulation_steps=1`
+  * Common with `train_network.py`.
+* `--gradient_checkpointing`
+  * Common with `train_network.py`. Recommended to enable for SDXL due to its high memory consumption.
+* `--cache_latents`
+  * Caches VAE outputs in memory (or on disk when `--cache_latents_to_disk` is specified). By skipping VAE computation, this reduces VRAM usage and speeds up training. Image augmentations (`--color_aug`, `--flip_aug`, `--random_crop`, etc.) are disabled. This option is recommended for SDXL training.
+* `--cache_latents_to_disk`
+  * Used with `--cache_latents`, caches to disk. When loading the dataset for the first time, VAE outputs are cached to disk. This is recommended when you have a large number of training images, as it allows you to skip VAE computation on subsequent training runs.
+* `--cache_text_encoder_outputs`
+  * Caches Text Encoder outputs in memory (or on disk when `--cache_text_encoder_outputs_to_disk` is specified). By skipping Text Encoder computation, this reduces VRAM usage and speeds up training. Caption augmentations (`--shuffle_caption`, `--caption_dropout_rate`, etc.) are disabled.
+  * **Note:** When using this option, LoRA modules for Text Encoder cannot be trained (`--network_train_unet_only` must be specified).
+* `--cache_text_encoder_outputs_to_disk`
+  * Used with `--cache_text_encoder_outputs`, caches to disk.
+* `--no_half_vae`
+  * Runs VAE in `float32` even when using mixed precision (`fp16`/`bf16`). Since SDXL's VAE can be unstable in `float16`, enable this when using `fp16`.
+* `--clip_skip`
+  * Not normally used for SDXL. No need to specify.
+* `--fused_backward_pass`
+  * Fuses gradient computation and optimizer steps to reduce VRAM usage. Available for SDXL. (Currently only supports the `Adafactor` optimizer)
+
+#### Others / その他
+
+* `--seed`, `--logging_dir`, `--log_prefix`, etc. are common with `train_network.py`.
+
+<details>
+<summary>日本語</summary>
+#### モデル関連
+
+*   `--pretrained_model_name_or_path="<モデルのパス>"` **[必須]**
+    *   学習のベースとなる **SDXL モデル**を指定します。Hugging Face Hub のモデル ID (例: `"stabilityai/stable-diffusion-xl-base-1.0"`) や、ローカルの Diffusers 形式モデルのディレクトリ、`.safetensors` ファイルのパスを指定できます。
+*   `--v2`, `--v_parameterization`
+    *   これらの引数は SD1.x/2.x 用です。`sdxl_train_network.py` を使用する場合、SDXL モデルであることが前提となるため、通常は**指定する必要はありません**。
+
+#### データセット関連
+
+*   `--dataset_config="<設定ファイルのパス>"`
+    *   `train_network.py` と共通です。
+    *   SDXL では高解像度データやバケツ機能 (`.toml` で `enable_bucket = true` を指定) の利用が一般的です。
+
+#### 出力・保存関連
+
+*   `train_network.py` と共通です。
+
+#### LoRA パラメータ
+
+*   `train_network.py` と共通です。
+
+#### 学習パラメータ
+
+*   `--learning_rate=1e-4`
+    *   全体の学習率。`unet_lr`, `text_encoder_lr1`, `text_encoder_lr2` が指定されない場合のデフォルト値となります。
+*   `--unet_lr=1e-4`
+    *   U-Net 部分の LoRA モジュールに対する学習率。指定しない場合は `--learning_rate` の値が使用されます。
+*   `--text_encoder_lr1=1e-5`
+    *   **Text Encoder 1 (OpenCLIP ViT-G/14) の LoRA モジュール**に対する学習率。指定しない場合は `--learning_rate` の値が使用されます。U-Net より小さめの値が推奨されます。
+*   `--text_encoder_lr2=1e-5`
+    *   **Text Encoder 2 (CLIP ViT-L/14) の LoRA モジュール**に対する学習率。指定しない場合は `--learning_rate` の値が使用されます。U-Net より小さめの値が推奨されます。
+*   `--optimizer_type="AdamW8bit"`
+    *   `train_network.py` と共通です。
+*   `--lr_scheduler="constant"`
+    *   `train_network.py` と共通です。
+*   `--lr_warmup_steps`
+    *   `train_network.py` と共通です。
+*   `--max_train_steps`, `--max_train_epochs`
+    *   `train_network.py` と共通です。
+*   `--mixed_precision="bf16"`
+    *   混合精度学習の設定。SDXL では `bf16` または `fp16` の使用が推奨されます。GPU が対応している方を選択してください。VRAM 使用量を削減し、学習速度を向上させます。
+*   `--gradient_accumulation_steps=1`
+    *   `train_network.py` と共通です。
+*   `--gradient_checkpointing`
+    *   `train_network.py` と共通です。SDXL はメモリ消費が大きいため、有効にすることが推奨されます。
+*   `--cache_latents`
+    *   VAE の出力をメモリ（または `--cache_latents_to_disk` 指定時はディスク）にキャッシュします。VAE の計算を省略できるため、VRAM 使用量を削減し、学習を高速化できます。画像に対する Augmentation (`--color_aug`, `--flip_aug`, `--random_crop` 等) が無効になります。SDXL 学習では推奨されるオプションです。
+*   `--cache_latents_to_disk`
+    *   `--cache_latents` と併用し、キャッシュ先をディスクにします。データセットを最初に読み込む際に、VAE の出力をディスクにキャッシュします。二回目以降の学習で VAE の計算を省略できるため、学習データの枚数が多い場合に推奨されます。
+*   `--cache_text_encoder_outputs`
+    *   Text Encoder の出力をメモリ（または `--cache_text_encoder_outputs_to_disk` 指定時はディスク）にキャッシュします。Text Encoder の計算を省略できるため、VRAM 使用量を削減し、学習を高速化できます。キャプションに対する Augmentation (`--shuffle_caption`, `--caption_dropout_rate` 等) が無効になります。
+    *   **注意:** このオプションを使用する場合、Text Encoder の LoRA モジュールは学習できません (`--network_train_unet_only` の指定が必須です)。
+*   `--cache_text_encoder_outputs_to_disk`
+    *   `--cache_text_encoder_outputs` と併用し、キャッシュ先をディスクにします。
+*   `--no_half_vae`
+    *   混合精度 (`fp16`/`bf16`) 使用時でも VAE を `float32` で動作させます。SDXL の VAE は `float16` で不安定になることがあるため、`fp16` 指定時には有効にしてください。
+*   `--clip_skip`
+    *   SDXL では通常使用しません。指定は不要です。
+*   `--fused_backward_pass`
+    *   勾配計算とオプティマイザのステップを融合し、VRAM使用量を削減します。SDXLで利用可能です。（現在 `Adafactor` オプティマイザのみ対応）
+
+#### その他
+
+*   `--seed`, `--logging_dir`, `--log_prefix` などは `train_network.py` と共通です。
+</details>
+
+### 3.2. Starting the Training / 学習の開始
+
+After setting the necessary arguments, execute the command to start training. The training progress will be displayed on the console. The basic flow is the same as with `train_network.py`.
+
+<details>
+<summary>日本語</summary>
+必要な引数を設定し、コマンドを実行すると学習が開始されます。学習の進行状況はコンソールに出力されます。基本的な流れは `train_network.py` と同じです。
+</details>
+
+## 4. Using the Trained Model / 学習済みモデルの利用
+
+When training is complete, a LoRA model file (`.safetensors`, etc.) with the name specified by `output_name` will be saved in the directory specified by `output_dir`.
+
+This file can be used with GUI tools that support SDXL, such as AUTOMATIC1111/stable-diffusion-webui and ComfyUI.
+
+<details>
+<summary>日本語</summary>
+学習が完了すると、`output_dir` で指定したディレクトリに、`output_name` で指定した名前の LoRA モデルファイル (`.safetensors` など) が保存されます。
+
+このファイルは、AUTOMATIC1111/stable-diffusion-webui 、ComfyUI などの SDXL に対応した GUI ツールで利用できます。
+</details>
+
+## 5. Supplement: Main Differences from `train_network.py` / 補足: `train_network.py` との主な違い
+
+* **Target Model:** `sdxl_train_network.py` is exclusively for SDXL models.
+* **Text Encoder:** Since SDXL has two Text Encoders, there are differences in learning rate specifications (`--text_encoder_lr1`, `--text_encoder_lr2`), etc.
+* **Caching Features:** `--cache_text_encoder_outputs` is particularly effective for SDXL and is recommended.
+* **Recommended Settings:** Due to high VRAM usage, mixed precision (`bf16` or `fp16`), `gradient_checkpointing`, and caching features (`--cache_latents`, `--cache_text_encoder_outputs`) are recommended. When using `fp16`, it is recommended to run the VAE in `float32` with `--no_half_vae`.
+
+For other detailed options, please refer to the script's help (`python sdxl_train_network.py --help`) and other documents in the repository.
+
+<details>
+<summary>日本語</summary>
+*   **対象モデル:** `sdxl_train_network.py` は SDXL モデル専用です。
+*   **Text Encoder:** SDXL は 2 つの Text Encoder を持つため、学習率の指定 (`--text_encoder_lr1`, `--text_encoder_lr2`) などが異なります。
+*   **キャッシュ機能:** `--cache_text_encoder_outputs` は SDXL で特に効果が高く、推奨されます。
+*   **推奨設定:** VRAM 使用量が大きいため、`bf16` または `fp16` の混合精度、`gradient_checkpointing`、キャッシュ機能 (`--cache_latents`, `--cache_text_encoder_outputs`) の利用が推奨されます。`fp16` 指定時は、VAE は `--no_half_vae` で `float32` 動作を推奨します。
+
+その他の詳細なオプションについては、スクリプトのヘルプ (`python sdxl_train_network.py --help`) やリポジトリ内の他のドキュメントを参照してください。
+</details>

docs/train_README-ja.md

@@ -648,7 +648,7 @@ masterpiece, best quality, 1boy, in business suit, standing at street, looking b
 
     詳細については各自お調べください。
 
-    任意のスケジューラを使う場合、任意のオプティマイザと同様に、`--scheduler_args`でオプション引数を指定してください。
+    任意のスケジューラを使う場合、任意のオプティマイザと同様に、`--lr_scheduler_args`でオプション引数を指定してください。
 
 ### オプティマイザの指定について
 

docs/train_README-zh.md

@@ -582,7 +582,7 @@ masterpiece, best quality, 1boy, in business suit, standing at street, looking b
 
     有关详细信息，请自行研究。
 
-    要使用任何调度程序，请像使用任何优化器一样使用“--scheduler_args”指定可选参数。
+    要使用任何调度程序，请像使用任何优化器一样使用“--lr_scheduler_args”指定可选参数。
 ### 关于指定优化器
 
 使用 --optimizer_args 选项指定优化器选项参数。可以以key=value的格式指定多个值。此外，您可以指定多个值，以逗号分隔。例如，要指定 AdamW 优化器的参数，``--optimizer_args weight_decay=0.01 betas=.9,.999``。

docs/train_SDXL-en.md

@@ -0,0 +1,84 @@
+## SDXL training
+
+The documentation will be moved to the training documentation in the future. The following is a brief explanation of the training scripts for SDXL.
+
+### Training scripts for SDXL
+
+- `sdxl_train.py` is a script for SDXL fine-tuning. The usage is almost the same as `fine_tune.py`, but it also supports DreamBooth dataset.
+  - `--full_bf16` option is added. Thanks to KohakuBlueleaf!
+    - This option enables the full bfloat16 training (includes gradients). This option is useful to reduce the GPU memory usage. 
+    - The full bfloat16 training might be unstable. Please use it at your own risk.
+  - The different learning rates for each U-Net block are now supported in sdxl_train.py. Specify with `--block_lr` option. Specify 23 values separated by commas like `--block_lr 1e-3,1e-3 ... 1e-3`.
+    - 23 values correspond to `0: time/label embed, 1-9: input blocks 0-8, 10-12: mid blocks 0-2, 13-21: output blocks 0-8, 22: out`.
+- `prepare_buckets_latents.py` now supports SDXL fine-tuning.
+
+- `sdxl_train_network.py` is a script for LoRA training for SDXL. The usage is almost the same as `train_network.py`.
+
+- Both scripts has following additional options:
+  - `--cache_text_encoder_outputs` and `--cache_text_encoder_outputs_to_disk`: Cache the outputs of the text encoders. This option is useful to reduce the GPU memory usage. This option cannot be used with options for shuffling or dropping the captions.
+  - `--no_half_vae`: Disable the half-precision (mixed-precision) VAE. VAE for SDXL seems to produce NaNs in some cases. This option is useful to avoid the NaNs.
+
+- `--weighted_captions` option is not supported yet for both scripts.
+
+- `sdxl_train_textual_inversion.py` is a script for Textual Inversion training for SDXL. The usage is almost the same as `train_textual_inversion.py`.
+  - `--cache_text_encoder_outputs` is not supported.
+  - There are two options for captions:
+    1. Training with captions. All captions must include the token string. The token string is replaced with multiple tokens.
+    2. Use `--use_object_template` or `--use_style_template` option. The captions are generated from the template. The existing captions are ignored.
+  - See below for the format of the embeddings.
+
+- `--min_timestep` and `--max_timestep` options are added to each training script. These options can be used to train U-Net with different timesteps. The default values are 0 and 1000.
+
+### Utility scripts for SDXL
+
+- `tools/cache_latents.py` is added. This script can be used to cache the latents to disk in advance. 
+  - The options are almost the same as `sdxl_train.py'. See the help message for the usage.
+  - Please launch the script as follows:
+    `accelerate launch  --num_cpu_threads_per_process 1 tools/cache_latents.py ...`
+  - This script should work with multi-GPU, but it is not tested in my environment.
+
+- `tools/cache_text_encoder_outputs.py` is added. This script can be used to cache the text encoder outputs to disk in advance. 
+  - The options are almost the same as `cache_latents.py` and `sdxl_train.py`. See the help message for the usage.
+
+- `sdxl_gen_img.py` is added. This script can be used to generate images with SDXL, including LoRA, Textual Inversion and ControlNet-LLLite. See the help message for the usage.
+
+### Tips for SDXL training
+
+- The default resolution of SDXL is 1024x1024.
+- The fine-tuning can be done with 24GB GPU memory with the batch size of 1. For 24GB GPU, the following options are recommended __for the fine-tuning with 24GB GPU memory__:
+  - Train U-Net only.
+  - Use gradient checkpointing.
+  - Use `--cache_text_encoder_outputs` option and caching latents.
+  - Use Adafactor optimizer. RMSprop 8bit or Adagrad 8bit may work. AdamW 8bit doesn't seem to work.
+- The LoRA training can be done with 8GB GPU memory (10GB recommended). For reducing the GPU memory usage, the following options are recommended:
+  - Train U-Net only.
+  - Use gradient checkpointing.
+  - Use `--cache_text_encoder_outputs` option and caching latents.
+  - Use one of 8bit optimizers or Adafactor optimizer.
+  - Use lower dim (4 to 8 for 8GB GPU).
+- `--network_train_unet_only` option is highly recommended for SDXL LoRA. Because SDXL has two text encoders, the result of the training will be unexpected.
+- PyTorch 2 seems to use slightly less GPU memory than PyTorch 1.
+- `--bucket_reso_steps` can be set to 32 instead of the default value 64. Smaller values than 32 will not work for SDXL training.
+
+Example of the optimizer settings for Adafactor with the fixed learning rate:
+```toml
+optimizer_type = "adafactor"
+optimizer_args = [ "scale_parameter=False", "relative_step=False", "warmup_init=False" ]
+lr_scheduler = "constant_with_warmup"
+lr_warmup_steps = 100
+learning_rate = 4e-7 # SDXL original learning rate
+```
+
+### Format of Textual Inversion embeddings for SDXL
+
+```python
+from safetensors.torch import save_file
+
+state_dict = {"clip_g": embs_for_text_encoder_1280, "clip_l": embs_for_text_encoder_768}
+save_file(state_dict, file)
+```
+
+### ControlNet-LLLite
+
+ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [documentation](./docs/train_lllite_README.md) for details.
+

docs/train_lllite_README-ja.md

@@ -21,9 +21,13 @@ ComfyUIのカスタムノードを用意しています。: https://github.com/k
 ## モデルの学習
 
 ### データセットの準備
-通常のdatasetに加え、`conditioning_data_dir` で指定したディレクトリにconditioning imageを格納してください。conditioning imageは学習用画像と同じbasenameを持つ必要があります。また、conditioning imageは学習用画像と同じサイズに自動的にリサイズされます。conditioning imageにはキャプションファイルは不要です。
+DreamBooth 方式の dataset で、`conditioning_data_dir` で指定したディレクトリにconditioning imageを格納してください。
 
-たとえば DreamBooth 方式でキャプションファイルを用いる場合の設定ファイルは以下のようになります。
+（finetuning 方式の dataset はサポートしていません。）
+
+conditioning imageは学習用画像と同じbasenameを持つ必要があります。また、conditioning imageは学習用画像と同じサイズに自動的にリサイズされます。conditioning imageにはキャプションファイルは不要です。
+
+たとえば、キャプションにフォルダ名ではなくキャプションファイルを用いる場合の設定ファイルは以下のようになります。
 
 ```toml
 [[datasets.subsets]]

docs/train_lllite_README.md

@@ -26,7 +26,9 @@ Due to the limitations of the inference environment, only CrossAttention (attn1
 
 ### Preparing the dataset
 
-In addition to the normal dataset, please store the conditioning image in the directory specified by `conditioning_data_dir`. The conditioning image must have the same basename as the training image. The conditioning image will be automatically resized to the same size as the training image. The conditioning image does not require a caption file.
+In addition to the normal DreamBooth method dataset, please store the conditioning image in the directory specified by `conditioning_data_dir`. The conditioning image must have the same basename as the training image. The conditioning image will be automatically resized to the same size as the training image. The conditioning image does not require a caption file.
+
+(We do not support the finetuning method dataset.)
 
 ```toml
 [[datasets.subsets]]
@@ -183,7 +185,7 @@ for img_file in img_files:
 
 ### Creating a dataset configuration file
 
-You can use the command line arguments of `sdxl_train_control_net_lllite.py` to specify the conditioning image directory. However, if you want to use a `.toml` file, specify the conditioning image directory in `conditioning_data_dir`.
+You can use the command line argument `--conditioning_data_dir` of `sdxl_train_control_net_lllite.py` to specify the conditioning image directory. However, if you want to use a `.toml` file, specify the conditioning image directory in `conditioning_data_dir`.
 
 ```toml
 [general]

docs/train_network.md

@@ -0,0 +1,367 @@
+# How to use the LoRA training script `train_network.py` / LoRA学習スクリプト `train_network.py` の使い方
+
+This document explains the basic procedures for training LoRA (Low-Rank Adaptation) models using `train_network.py` included in the `sd-scripts` repository.
+
+<details>
+<summary>日本語</summary>
+このドキュメントでは、`sd-scripts` リポジトリに含まれる `train_network.py` を使用して LoRA (Low-Rank Adaptation) モデルを学習する基本的な手順について解説します。
+</details>
+
+## 1. Introduction / はじめに
+
+`train_network.py` is a script for training additional networks such as LoRA on Stable Diffusion models (v1.x, v2.x). It allows for additional training on the original model with a low computational cost, enabling the creation of models that reproduce specific characters or art styles.
+
+This guide focuses on LoRA training and explains the basic configuration items.
+
+**Prerequisites:**
+
+* The `sd-scripts` repository has been cloned and the Python environment has been set up.
+* The training dataset has been prepared. (For dataset preparation, please refer to [this guide](link/to/dataset/doc))
+
+<details>
+<summary>日本語</summary>
+
+`train_network.py` は、Stable Diffusion モデル（v1.x, v2.x）に対して、LoRA などの追加ネットワークを学習させるためのスクリプトです。少ない計算コストで元のモデルに追加学習を行い、特定のキャラクターや画風を再現するモデルを作成できます。
+
+このガイドでは、LoRA 学習に焦点を当て、基本的な設定項目を中心に説明します。
+
+**前提条件:**
+
+*   `sd-scripts` リポジトリのクローンと Python 環境のセットアップが完了していること。
+*   学習用データセットの準備が完了していること。（データセットの準備については[こちら](link/to/dataset/doc)を参照してください）
+</details>
+
+## 2. Preparation / 準備
+
+Before starting training, you will need the following files:
+
+1. **Training script:** `train_network.py`
+2. **Dataset definition file (.toml):** A file in TOML format that describes the configuration of the training dataset.
+
+### About the Dataset Definition File / データセット定義ファイルについて
+
+The dataset definition file (`.toml`) contains detailed settings such as the directory of images to use, repetition count, caption settings, resolution buckets (optional), etc.
+
+For more details on how to write the dataset definition file, please refer to the [Dataset Configuration Guide](./config_README-en.md).
+
+In this guide, we will use a file named `my_dataset_config.toml` as an example.
+
+<details>
+<summary>日本語</summary>
+
+学習を開始する前に、以下のファイルが必要です。
+
+1.  **学習スクリプト:** `train_network.py`
+2.  **データセット定義ファイル (.toml):** 学習データセットの設定を記述した TOML 形式のファイル。
+
+**データセット定義ファイルについて**
+
+データセット定義ファイル (`.toml`) には、使用する画像のディレクトリ、繰り返し回数、キャプションの設定、Aspect Ratio Bucketing（任意）などの詳細な設定を記述します。
+
+データセット定義ファイルの詳しい書き方については、[データセット設定ガイド](./config_README-ja.md)を参照してください。
+
+ここでは、例として `my_dataset_config.toml` という名前のファイルを使用することにします。
+</details>
+
+## 3. Running the Training / 学習の実行
+
+Training is started by executing `train_network.py` from the terminal. When executing, various training settings are specified as command-line arguments.
+
+Below is a basic command-line execution example:
+
+```bash
+accelerate launch --num_cpu_threads_per_process 1 train_network.py 
+ --pretrained_model_name_or_path="<path to Stable Diffusion model>" 
+ --dataset_config="my_dataset_config.toml" 
+ --output_dir="<output directory for training results>" 
+ --output_name="my_lora" 
+ --save_model_as=safetensors 
+ --network_module=networks.lora 
+ --network_dim=16 
+ --network_alpha=1 
+ --learning_rate=1e-4 
+ --optimizer_type="AdamW8bit" 
+ --lr_scheduler="constant" 
+ --sdpa  
+ --max_train_epochs=10 
+ --save_every_n_epochs=1 
+ --mixed_precision="fp16" 
+ --gradient_checkpointing 
+```
+
+In reality, you need to write this in a single line, but it's shown with line breaks for readability (on Linux or Mac, you can add `\` at the end of each line to break lines). For Windows, either write it in a single line without breaks or add `^` at the end of each line.
+
+Next, we'll explain the main command-line arguments.
+
+<details>
+<summary>日本語</summary>
+
+学習は、ターミナルから `train_network.py` を実行することで開始します。実行時には、学習に関する様々な設定をコマンドライン引数として指定します。
+
+以下に、基本的なコマンドライン実行例を示します。
+
+実際には1行で書く必要がありますが、見やすさのために改行しています（Linux や Mac では `\` を行末に追加することで改行できます）。Windows の場合は、改行せずに1行で書くか、`^` を行末に追加してください。
+
+次に、主要なコマンドライン引数について解説します。
+</details>
+
+### 3.1. Main Command-Line Arguments / 主要なコマンドライン引数
+
+#### Model Related / モデル関連
+
+* `--pretrained_model_name_or_path="<path to model>"` **[Required]**
+  * Specifies the Stable Diffusion model to be used as the base for training. You can specify the path to a local `.ckpt` or `.safetensors` file, or a directory containing a Diffusers format model. You can also specify a Hugging Face Hub model ID (e.g., `"stabilityai/stable-diffusion-2-1-base"`).
+* `--v2`
+  * Specify this when the base model is Stable Diffusion v2.x.
+* `--v_parameterization`
+  * Specify this when training with a v-prediction model (such as v2.x 768px models).
+
+#### Dataset Related / データセット関連
+
+* `--dataset_config="<path to configuration file>"`
+  * Specifies the path to a `.toml` file describing the dataset configuration. (For details on dataset configuration, see [here](link/to/dataset/config/doc))
+  * It's also possible to specify dataset settings from the command line, but using a `.toml` file is recommended as it becomes lengthy.
+
+#### Output and Save Related / 出力・保存関連
+
+* `--output_dir="<output directory>"` **[Required]**
+  * Specifies the directory where trained LoRA models, sample images, logs, etc. will be output.
+* `--output_name="<output filename>"` **[Required]**
+  * Specifies the filename of the trained LoRA model (excluding the extension).
+* `--save_model_as="safetensors"`
+  * Specifies the format for saving the model. You can choose from `safetensors` (recommended), `ckpt`, or `pt`. The default is `safetensors`.
+* `--save_every_n_epochs=1`
+  * Saves the model every specified number of epochs. If not specified, only the final model will be saved.
+* `--save_every_n_steps=1000`
+  * Saves the model every specified number of steps. If both epoch and step saving are specified, both will be saved.
+
+#### LoRA Parameters / LoRA パラメータ
+
+* `--network_module=networks.lora` **[Required]**
+  * Specifies the type of network to train. For LoRA, specify `networks.lora`.
+* `--network_dim=16` **[Required]**
+  * Specifies the rank (dimension) of LoRA. Higher values increase expressiveness but also increase file size and computational cost. Values between 4 and 128 are commonly used. There is no default (module dependent).
+* `--network_alpha=1`
+  * Specifies the alpha value for LoRA. This parameter is related to learning rate scaling. It is generally recommended to set it to about half the value of `network_dim`, but it can also be the same value as `network_dim`. The default is 1. Setting it to the same value as `network_dim` will result in behavior similar to older versions.
+* `--network_args`
+  * Used to specify additional parameters specific to the LoRA module. For example, to use Conv2d (3x3) LoRA (LoRA-C3Lier), specify the following in `--network_args`. Use `conv_dim` to specify the rank for Conv2d (3x3) and `conv_alpha` for alpha.
+    ```
+    --network_args "conv_dim=4" "conv_alpha=1"
+    ```
+
+    If alpha is omitted as shown below, it defaults to 1.
+    ```
+    --network_args "conv_dim=4"
+    ```
+
+#### Training Parameters / 学習パラメータ
+
+* `--learning_rate=1e-4`
+  * Specifies the learning rate. For LoRA training (when alpha value is 1), relatively higher values (e.g., from `1e-4` to `1e-3`) are often used.
+* `--unet_lr=1e-4`
+  * Used to specify a separate learning rate for the LoRA modules in the U-Net part. If not specified, the value of `--learning_rate` is used.
+* `--text_encoder_lr=1e-5`
+  * Used to specify a separate learning rate for the LoRA modules in the Text Encoder part. If not specified, the value of `--learning_rate` is used. A smaller value than that for U-Net is recommended.
+* `--optimizer_type="AdamW8bit"`
+  * Specifies the optimizer to use for training. Options include `AdamW8bit` (requires `bitsandbytes`), `AdamW`, `Lion` (requires `lion-pytorch`), `DAdaptation` (requires `dadaptation`), and `Adafactor`. `AdamW8bit` is memory-efficient and widely used.
+* `--lr_scheduler="constant"`
+  * Specifies the learning rate scheduler. This is the method for changing the learning rate as training progresses. Options include `constant` (no change), `cosine` (cosine curve), `linear` (linear decay), `constant_with_warmup` (constant with warmup), and `cosine_with_restarts`. `constant`, `cosine`, and `constant_with_warmup` are commonly used.
+* `--lr_warmup_steps=500`
+  * Specifies the number of warmup steps for the learning rate scheduler. This is the period during which the learning rate gradually increases at the start of training. Valid when the `lr_scheduler` supports warmup.
+* `--max_train_steps=10000`
+  * Specifies the total number of training steps. If `max_train_epochs` is specified, that takes precedence.
+* `--max_train_epochs=12`
+  * Specifies the number of training epochs. If this is specified, `max_train_steps` is ignored.
+* `--sdpa`
+  * Uses Scaled Dot-Product Attention. This can reduce memory usage and improve training speed for LoRA training.
+* `--mixed_precision="fp16"`
+  * Specifies the mixed precision training setting. Options are `no` (disabled), `fp16` (half precision), and `bf16` (bfloat16). If your GPU supports it, specifying `fp16` or `bf16` can improve training speed and reduce memory usage.
+* `--gradient_accumulation_steps=1`
+  * Specifies the number of steps to accumulate gradients. This effectively increases the batch size to `train_batch_size * gradient_accumulation_steps`. Set a larger value if GPU memory is insufficient. Usually `1` is fine.
+
+#### Others / その他
+
+* `--seed=42`
+  * Specifies the random seed. Set this if you want to ensure reproducibility of the training.
+* `--logging_dir="<log directory>"`
+  * Specifies the directory to output logs for TensorBoard, etc. If not specified, logs will not be output.
+* `--log_prefix="<prefix>"`
+  * Specifies the prefix for the subdirectory name created within `logging_dir`.
+* `--gradient_checkpointing`
+  * Enables Gradient Checkpointing. This can significantly reduce memory usage but slightly decreases training speed. Useful when memory is limited.
+* `--clip_skip=1`
+  * Specifies how many layers to skip from the last layer of the Text Encoder. Specifying `2` will use the output from the second-to-last layer. `None` or `1` means no skip (uses the last layer). Check the recommended value for the model you are training.
+
+<details>
+<summary>日本語</summary>
+
+#### モデル関連
+
+*   `--pretrained_model_name_or_path="<モデルのパス>"` **[必須]**
+    *   学習のベースとなる Stable Diffusion モデルを指定します。ローカルの `.ckpt` または `.safetensors` ファイルのパス、あるいは Diffusers 形式モデルのディレクトリを指定できます。Hugging Face Hub のモデル ID (例: `"stabilityai/stable-diffusion-2-1-base"`) も指定可能です。
+*   `--v2`
+    *   ベースモデルが Stable Diffusion v2.x の場合に指定します。
+*   `--v_parameterization`
+    *   v-prediction モデル（v2.x の 768px モデルなど）で学習する場合に指定します。
+
+#### データセット関連
+
+*   `--dataset_config="<設定ファイルのパス>"` 
+    *   データセット設定を記述した `.toml` ファイルのパスを指定します。（データセット設定の詳細は[こちら](link/to/dataset/config/doc)）
+    *   コマンドラインからデータセット設定を指定することも可能ですが、長くなるため `.toml` ファイルを使用することを推奨します。
+
+#### 出力・保存関連
+
+*   `--output_dir="<出力先ディレクトリ>"` **[必須]**
+    *   学習済み LoRA モデルやサンプル画像、ログなどが出力されるディレクトリを指定します。
+*   `--output_name="<出力ファイル名>"` **[必須]**
+    *   学習済み LoRA モデルのファイル名（拡張子を除く）を指定します。
+*   `--save_model_as="safetensors"`
+    *   モデルの保存形式を指定します。`safetensors` (推奨), `ckpt`, `pt` から選択できます。デフォルトは `safetensors` です。
+*   `--save_every_n_epochs=1`
+    *   指定したエポックごとにモデルを保存します。省略するとエポックごとの保存は行われません（最終モデルのみ保存）。
+*   `--save_every_n_steps=1000`
+    *   指定したステップごとにモデルを保存します。エポック指定 (`save_every_n_epochs`) と同時に指定された場合、両方とも保存されます。
+
+#### LoRA パラメータ
+
+*   `--network_module=networks.lora` **[必須]**
+    *   学習するネットワークの種別を指定します。LoRA の場合は `networks.lora` を指定します。
+*   `--network_dim=16` **[必須]**
+    *   LoRA のランク (rank / 次元数) を指定します。値が大きいほど表現力は増しますが、ファイルサイズと計算コストが増加します。一般的には 4〜128 程度の値が使われます。デフォルトは指定されていません（モジュール依存）。
+*   `--network_alpha=1`
+    *   LoRA のアルファ値 (alpha) を指定します。学習率のスケーリングに関係するパラメータで、一般的には `network_dim` の半分程度の値を指定することが推奨されますが、`network_dim` と同じ値を指定する場合もあります。デフォルトは 1 です。`network_dim` と同じ値に設定すると、旧バージョンと同様の挙動になります。
+
+* `--network_args`
+    *   LoRA モジュールに特有の追加パラメータを指定するために使用します。例えば、Conv2d (3x3) の LoRA  (LoRA-C3Lier) を使用する場合は`--network_args` に以下のように指定してください。`conv_dim` で Conv2d (3x3) の rank を、`conv_alpha` で alpha を指定します。
+        ```
+        --network_args "conv_dim=4" "conv_alpha=1"
+        ```
+        以下のように alpha を省略した時は1になります。
+        ```
+        --network_args "conv_dim=4"
+        ```
+
+#### 学習パラメータ
+
+*   `--learning_rate=1e-4`
+    *   学習率を指定します。LoRA 学習では（アルファ値が1の場合）比較的高めの値（例: `1e-4`から`1e-3`）が使われることが多いです。
+*   `--unet_lr=1e-4`
+    *   U-Net 部分の LoRA モジュールに対する学習率を個別に指定する場合に使用します。指定しない場合は `--learning_rate` の値が使用されます。
+*   `--text_encoder_lr=1e-5`
+    *   Text Encoder 部分の LoRA モジュールに対する学習率を個別に指定する場合に使用します。指定しない場合は `--learning_rate` の値が使用されます。U-Net よりも小さめの値が推奨されます。
+*   `--optimizer_type="AdamW8bit"`
+    *   学習に使用するオプティマイザを指定します。`AdamW8bit` (要 `bitsandbytes`), `AdamW`, `Lion` (要 `lion-pytorch`), `DAdaptation` (要 `dadaptation`), `Adafactor` などが選択可能です。`AdamW8bit` はメモリ効率が良く、広く使われています。
+*   `--lr_scheduler="constant"`
+    *   学習率スケジューラを指定します。学習の進行に合わせて学習率を変化させる方法です。`constant` (変化なし), `cosine` (コサインカーブ), `linear` (線形減衰), `constant_with_warmup` (ウォームアップ付き定数), `cosine_with_restarts` などが選択可能です。`constant`や`cosine` 、 `constant_with_warmup` がよく使われます。
+*   `--lr_warmup_steps=500`
+    *   学習率スケジューラのウォームアップステップ数を指定します。学習開始時に学習率を徐々に上げていく期間です。`lr_scheduler` がウォームアップをサポートする場合に有効です。
+*   `--max_train_steps=10000`
+    *   学習の総ステップ数を指定します。`max_train_epochs` が指定されている場合はそちらが優先されます。
+*   `--max_train_epochs=12`
+    *   学習のエポック数を指定します。これを指定すると `max_train_steps` は無視されます。
+*   `--sdpa`
+    *   Scaled Dot-Product Attention を使用します。LoRA の学習において、メモリ使用量を削減し、学習速度を向上させることができます。
+*   `--mixed_precision="fp16"`
+    *   混合精度学習の設定を指定します。`no` (無効), `fp16` (半精度), `bf16` (bfloat16) から選択できます。GPU が対応している場合は `fp16` または `bf16` を指定することで、学習速度の向上とメモリ使用量の削減が期待できます。
+*   `--gradient_accumulation_steps=1`
+    *   勾配を累積するステップ数を指定します。実質的なバッチサイズを `train_batch_size * gradient_accumulation_steps` に増やす効果があります。GPU メモリが足りない場合に大きな値を設定します。通常は `1` で問題ありません。
+
+#### その他
+
+*   `--seed=42`
+    *   乱数シードを指定します。学習の再現性を確保したい場合に設定します。
+*   `--logging_dir="<ログディレクトリ>"`
+    *   TensorBoard などのログを出力するディレクトリを指定します。指定しない場合、ログは出力されません。
+*   `--log_prefix="<プレフィックス>"`
+    *   `logging_dir` 内に作成されるサブディレクトリ名の接頭辞を指定します。
+*   `--gradient_checkpointing`
+    *   Gradient Checkpointing を有効にします。メモリ使用量を大幅に削減できますが、学習速度は若干低下します。メモリが厳しい場合に有効です。
+*   `--clip_skip=1`
+    *   Text Encoder の最後の層から数えて何層スキップするかを指定します。`2` を指定すると最後から 2 層目の出力を使用します。`None` または `1` はスキップなし（最後の層を使用）を意味します。学習対象のモデルの推奨する値を確認してください。
+</details>
+
+### 3.2. Starting the Training / 学習の開始
+
+After setting the necessary arguments and executing the command, training will begin. The progress of the training will be output to the console. If `logging_dir` is specified, you can visually check the training status (loss, learning rate, etc.) with TensorBoard.
+
+```bash
+tensorboard --logdir <directory specified by logging_dir>
+```
+
+<details>
+<summary>日本語</summary>
+
+必要な引数を設定し、コマンドを実行すると学習が開始されます。学習の進行状況はコンソールに出力されます。`logging_dir` を指定した場合は、TensorBoard などで学習状況（損失や学習率など）を視覚的に確認できます。
+</details>
+
+## 4. Using the Trained Model / 学習済みモデルの利用
+
+Once training is complete, a LoRA model file (`.safetensors` or `.ckpt`) with the name specified by `output_name` will be saved in the directory specified by `output_dir`.
+
+This file can be used with GUI tools such as AUTOMATIC1111/stable-diffusion-webui, ComfyUI, etc.
+
+<details>
+<summary>日本語</summary>
+
+学習が完了すると、`output_dir` で指定したディレクトリに、`output_name` で指定した名前の LoRA モデルファイル (`.safetensors` または `.ckpt`) が保存されます。
+
+このファイルは、AUTOMATIC1111/stable-diffusion-webui 、ComfyUI などの GUI ツールで利用できます。
+</details>
+
+## 5. Other Features / その他の機能
+
+`train_network.py` has many other options not introduced here.
+
+* Sample image generation (`--sample_prompts`, `--sample_every_n_steps`, etc.)
+* More detailed optimizer settings (`--optimizer_args`, etc.)
+* Caption preprocessing (`--shuffle_caption`, `--keep_tokens`, etc.)
+* Additional network settings (`--network_args`, etc.)
+
+For these features, please refer to the script's help (`python train_network.py --help`) or other documents in the repository.
+
+<details>
+<summary>日本語</summary>
+
+`train_network.py` には、ここで紹介した以外にも多くのオプションがあります。
+
+*   サンプル画像の生成 (`--sample_prompts`, `--sample_every_n_steps` など)
+*   より詳細なオプティマイザ設定 (`--optimizer_args` など)
+*   キャプションの前処理 (`--shuffle_caption`, `--keep_tokens` など)
+*   ネットワークの追加設定 (`--network_args` など)
+
+これらの機能については、スクリプトのヘルプ (`python train_network.py --help`) やリポジトリ内の他のドキュメントを参照してください。
+</details>
+
+## 6. Additional Information / 追加情報
+
+### Naming of LoRA
+
+The LoRA supported by `train_network.py` has been named to avoid confusion. The documentation has been updated. The following are the names of LoRA types in this repository.
+
+1. __LoRA-LierLa__ : (LoRA for __Li__ n __e__ a __r__  __La__ yers)
+
+    LoRA for Linear layers and Conv2d layers with 1x1 kernel
+
+2. __LoRA-C3Lier__ : (LoRA for __C__ olutional layers with __3__ x3 Kernel and  __Li__ n __e__ a __r__ layers)
+
+    In addition to 1., LoRA for Conv2d layers with 3x3 kernel 
+    
+LoRA-LierLa is the default LoRA type for `train_network.py` (without `conv_dim` network arg). 
+
+<details>
+<summary>日本語</summary>
+
+`train_network.py` がサポートするLoRAについて、混乱を避けるため名前を付けました。ドキュメントは更新済みです。以下は当リポジトリ内の独自の名称です。
+
+1. __LoRA-LierLa__ : (LoRA for __Li__ n __e__ a __r__  __La__ yers、リエラと読みます)
+
+    Linear 層およびカーネルサイズ 1x1 の Conv2d 層に適用されるLoRA
+
+2. __LoRA-C3Lier__ : (LoRA for __C__ olutional layers with __3__ x3 Kernel and  __Li__ n __e__ a __r__ layers、セリアと読みます)
+
+    1.に加え、カーネルサイズ 3x3 の Conv2d 層に適用されるLoRA
+
+デフォルトではLoRA-LierLaが使われます。LoRA-C3Lierを使う場合は `--network_args` に `conv_dim` を指定してください。
+
+</details>

docs/train_network_README-ja.md

@@ -102,6 +102,8 @@ accelerate launch --num_cpu_threads_per_process 1 train_network.py
   * Text Encoderに関連するLoRAモジュールに、通常の学習率（--learning_rateオプションで指定）とは異なる学習率を使う時に指定します。Text Encoderのほうを若干低めの学習率（5e-5など）にしたほうが良い、という話もあるようです。
 * `--network_args`
   * 複数の引数を指定できます。後述します。
+* `--alpha_mask`
+  * 画像のアルファ値をマスクとして使用します。透過画像を学習する際に使用します。[PR #1223](https://github.com/kohya-ss/sd-scripts/pull/1223)
 
 `--network_train_unet_only` と `--network_train_text_encoder_only` の両方とも未指定時（デフォルト）はText EncoderとU-Netの両方のLoRAモジュールを有効にします。
 
@@ -181,16 +183,16 @@ python networks\extract_lora_from_dylora.py --model "foldername/dylora-model.saf
 
 詳細は[PR #355](https://github.com/kohya-ss/sd-scripts/pull/355) をご覧ください。
 
-SDXLは現在サポートしていません。
-
 フルモデルの25個のブロックの重みを指定できます。最初のブロックに該当するLoRAは存在しませんが、階層別LoRA適用等との互換性のために25個としています。またconv2d3x3に拡張しない場合も一部のブロックにはLoRAが存在しませんが、記述を統一するため常に25個の値を指定してください。
 
+SDXL では down/up 9 個、middle 3 個の値を指定してください。
+
 `--network_args` で以下の引数を指定してください。
 
 - `down_lr_weight` : U-Netのdown blocksの学習率の重みを指定します。以下が指定可能です。
-  - ブロックごとの重み : `"down_lr_weight=0,0,0,0,0,0,1,1,1,1,1,1"` のように12個の数値を指定します。
+  - ブロックごとの重み : `"down_lr_weight=0,0,0,0,0,0,1,1,1,1,1,1"` のように12個（SDXL では 9 個）の数値を指定します。
   - プリセットからの指定 : `"down_lr_weight=sine"` のように指定します（サインカーブで重みを指定します）。sine, cosine, linear, reverse_linear, zeros が指定可能です。また `"down_lr_weight=cosine+.25"` のように `+数値` を追加すると、指定した数値を加算します（0.25~1.25になります）。
-- `mid_lr_weight` : U-Netのmid blockの学習率の重みを指定します。`"down_lr_weight=0.5"` のように数値を一つだけ指定します。
+- `mid_lr_weight` : U-Netのmid blockの学習率の重みを指定します。`"down_lr_weight=0.5"` のように数値を一つだけ指定します（SDXL の場合は 3 個）。
 - `up_lr_weight` : U-Netのup blocksの学習率の重みを指定します。down_lr_weightと同様です。
 - 指定を省略した部分は1.0として扱われます。また重みを0にするとそのブロックのLoRAモジュールは作成されません。
 - `block_lr_zero_threshold` : 重みがこの値以下の場合、LoRAモジュールを作成しません。デフォルトは0です。
@@ -215,6 +217,9 @@ network_args = [ "block_lr_zero_threshold=0.1", "down_lr_weight=sine+.5", "mid_l
 
 フルモデルの25個のブロックのdim (rank)を指定できます。階層別学習率と同様に一部のブロックにはLoRAが存在しない場合がありますが、常に25個の値を指定してください。
 
+SDXL では 23 個の値を指定してください。一部のブロックにはLoRA が存在しませんが、`sdxl_train.py` の[階層別学習率](./train_SDXL-en.md) との互換性のためです。
+対応は、`0: time/label embed, 1-9: input blocks 0-8, 10-12: mid blocks 0-2, 13-21: output blocks 0-8, 22: out` です。
+
 `--network_args` で以下の引数を指定してください。
 
 - `block_dims` : 各ブロックのdim (rank)を指定します。`"block_dims=2,2,2,2,4,4,4,4,6,6,6,6,8,6,6,6,6,4,4,4,4,2,2,2,2"` のように25個の数値を指定します。

docs/train_network_README-zh.md

@@ -101,6 +101,8 @@ LoRA的模型将会被保存在通过`--output_dir`选项指定的文件夹中
   * 当在Text Encoder相关的LoRA模块中使用与常规学习率（由`--learning_rate`选项指定）不同的学习率时，应指定此选项。可能最好将Text Encoder的学习率稍微降低（例如5e-5）。
 * `--network_args`
   * 可以指定多个参数。将在下面详细说明。
+* `--alpha_mask`
+  * 使用图像的 Alpha 值作为遮罩。这在学习透明图像时使用。[PR #1223](https://github.com/kohya-ss/sd-scripts/pull/1223)
 
 当未指定`--network_train_unet_only`和`--network_train_text_encoder_only`时（默认情况），将启用Text Encoder和U-Net的两个LoRA模块。
 

docs/train_network_advanced.md

@@ -0,0 +1,561 @@
+# Advanced Settings: Detailed Guide for SDXL LoRA Training Script `sdxl_train_network.py` / 高度な設定: SDXL LoRA学習スクリプト `sdxl_train_network.py` 詳細ガイド
+
+This document describes the advanced options available when training LoRA models for SDXL (Stable Diffusion XL) with `sdxl_train_network.py` in the `sd-scripts` repository. For the basics, please read [How to Use the LoRA Training Script `train_network.py`](train_network.md) and [How to Use the SDXL LoRA Training Script `sdxl_train_network.py`](sdxl_train_network.md).
+
+This guide targets experienced users who want to fine tune settings in detail.
+
+**Prerequisites:**
+
+* You have cloned the `sd-scripts` repository and prepared a Python environment.
+* A training dataset and its `.toml` configuration are ready (see the dataset configuration guide).
+* You are familiar with running basic LoRA training commands.
+
+## 1. Command Line Options / コマンドライン引数 詳細解説
+
+`sdxl_train_network.py` inherits the functionality of `train_network.py` and adds SDXL-specific features. Major options are grouped and explained below. For common arguments, see the other guides mentioned above.
+
+### 1.1. Model Loading
+
+* `--pretrained_model_name_or_path=\"<model path>\"` **[Required]**: specify the base SDXL model. Supports a Hugging Face model ID, a local Diffusers directory or a `.safetensors` file.
+* `--vae=\"<VAE path>\"`: optionally use a different VAE. Specify when using a VAE other than the one included in the SDXL model. Can specify `.ckpt` or `.safetensors` files.
+* `--no_half_vae`: keep the VAE in float32 even with fp16/bf16 training. The VAE for SDXL can become unstable with `float16`, so it is recommended to enable this when `fp16` is specified. Usually unnecessary for `bf16`.
+* `--fp8_base` / `--fp8_base_unet`: **Experimental**: load the base model (U-Net, Text Encoder) or just the U-Net in FP8 to reduce VRAM (requires PyTorch 2.1+). For details, refer to the relevant section in TODO add document later (this is an SD3 explanation but also applies to SDXL).
+
+### 1.2. Dataset Settings
+
+* `--dataset_config=\"<path to config>\"`: specify a `.toml` dataset config. High resolution data and aspect ratio buckets (specify `enable_bucket = true` in `.toml`) are common for SDXL. The resolution steps for aspect ratio buckets (`bucket_reso_steps`) must be multiples of 32 for SDXL. For details on writing `.toml` files, refer to the [Dataset Configuration Guide](link/to/dataset/config/doc).
+
+### 1.3. Output and Saving
+
+Options match `train_network.py`:
+
+* `--output_dir`, `--output_name` (both required)
+* `--save_model_as` (recommended `safetensors`), `ckpt`, `pt`, `diffusers`, `diffusers_safetensors`
+* `--save_precision=\"fp16\"`, `\"bf16\"`, `\"float\"`: Specifies the precision for saving the model. If not specified, the model is saved with the training precision (`fp16`, `bf16`, etc.).
+* `--save_every_n_epochs=N`, `--save_every_n_steps=N`: Saves the model every N epochs/steps.
+* `--save_last_n_epochs=M`, `--save_last_n_steps=M`: When saving at every epoch/step, only the latest M files are kept, and older ones are deleted.
+* `--save_state`, `--save_state_on_train_end`: Saves the training state (`state`), including Optimizer status, etc., when saving the model or at the end of training. Required for resuming training with the `--resume` option.
+* `--save_last_n_epochs_state=M`, `--save_last_n_steps_state=M`: Limits the number of saved `state` files to M. Overrides the `--save_last_n_epochs/steps` specification.
+* `--no_metadata`: Does not save metadata to the output model.
+* `--save_state_to_huggingface` and related options (e.g., `--huggingface_repo_id`): Options related to uploading models and states to Hugging Face Hub. See TODO add document for details.
+
+### 1.4. Network Parameters (LoRA)
+
+* `--network_module=networks.lora` **[Required]**
+* `--network_dim=N` **[Required]**: Specifies the rank (dimensionality) of LoRA. For SDXL, values like 32 or 64 are often tried, but adjustment is necessary depending on the dataset and purpose.
+* `--network_alpha=M`: LoRA alpha value. Generally around half of `network_dim` or the same value as `network_dim`. Default is 1.
+* `--network_dropout=P`: Dropout rate (0.0-1.0) within LoRA modules. Can be effective in suppressing overfitting. Default is None (no dropout).
+* `--network_args ...`: Allows advanced settings by specifying additional arguments to the network module in `key=value` format. For LoRA, the following advanced settings are available:
+    *   **Block-wise dimensions/alphas:**
+        *   Allows specifying different `dim` and `alpha` for each block of the U-Net. This enables adjustments to strengthen or weaken the influence of specific layers.
+        *   `block_dims`: Comma-separated dims for Linear and Conv2d 1x1 layers in U-Net (23 values for SDXL).
+        *   `block_alphas`: Comma-separated alpha values corresponding to the above.
+        *   `conv_block_dims`: Comma-separated dims for Conv2d 3x3 layers in U-Net.
+        *   `conv_block_alphas`: Comma-separated alpha values corresponding to the above.
+        *   Blocks not specified will use values from `--network_dim`/`--network_alpha` or `--conv_dim`/`--conv_alpha` (if they exist).
+        *   For details, refer to [Block-wise learning rate for LoRA](train_network.md#lora-の階層別学習率) (in train_network.md, applicable to SDXL) and the implementation ([lora.py](lora.py)).
+    *   **LoRA+:**
+        *   `loraplus_lr_ratio=R`: Sets the learning rate of LoRA's upward weights (UP) to R times the learning rate of downward weights (DOWN). Expected to improve learning speed. Paper recommends 16.
+        *   `loraplus_unet_lr_ratio=RU`: Specifies the LoRA+ learning rate ratio for the U-Net part individually.
+        *   `loraplus_text_encoder_lr_ratio=RT`: Specifies the LoRA+ learning rate ratio for the Text Encoder part individually (multiplied by the learning rates specified with `--text_encoder_lr1`, `--text_encoder_lr2`).
+        *   For details, refer to [README](../README.md#jan-17-2025--2025-01-17-version-090) and the implementation ([lora.py](lora.py)).
+* `--network_train_unet_only`: Trains only the LoRA modules of the U-Net. Specify this if not training Text Encoders. Required when using `--cache_text_encoder_outputs`.
+* `--network_train_text_encoder_only`: Trains only the LoRA modules of the Text Encoders. Specify this if not training the U-Net.
+* `--network_weights=\"<weight file>\"`: Starts training by loading pre-trained LoRA weights. Used for fine-tuning or resuming training. The difference from `--resume` is that this option only loads LoRA module weights, while `--resume` also restores Optimizer state, step count, etc.
+* `--dim_from_weights`: Automatically reads the LoRA dimension (`dim`) from the weight file specified by `--network_weights`. Specification of `--network_dim` becomes unnecessary.
+
+### 1.5. Training Parameters
+
+*   `--learning_rate=LR`: Sets the overall learning rate. This becomes the default value for each module (`unet_lr`, `text_encoder_lr1`, `text_encoder_lr2`). Values like `1e-3` or `1e-4` are often tried.
+*   `--unet_lr=LR_U`: Learning rate for the LoRA module of the U-Net part.
+*   `--text_encoder_lr1=LR_TE1`: Learning rate for the LoRA module of Text Encoder 1 (OpenCLIP ViT-G/14). Usually, a smaller value than U-Net (e.g., `1e-5`, `2e-5`) is recommended.
+*   `--text_encoder_lr2=LR_TE2`: Learning rate for the LoRA module of Text Encoder 2 (CLIP ViT-L/14). Usually, a smaller value than U-Net (e.g., `1e-5`, `2e-5`) is recommended.
+*   `--optimizer_type=\"...\"`: Specifies the optimizer to use. Options include `AdamW8bit` (memory-efficient, common), `Adafactor` (even more memory-efficient, proven in SDXL full model training), `Lion`, `DAdaptation`, `Prodigy`, etc. Each optimizer may require additional arguments (see `--optimizer_args`). `AdamW8bit` or `PagedAdamW8bit` (requires `bitsandbytes`) are common. `Adafactor` is memory-efficient but slightly complex to configure (relative step (`relative_step=True`) recommended, `adafactor` learning rate scheduler recommended). `DAdaptation`, `Prodigy` have automatic learning rate adjustment but cannot be used with LoRA+. Specify a learning rate around `1.0`. For details, see the `get_optimizer` function in [train_util.py](train_util.py).
+*   `--optimizer_args ...`: Specifies additional arguments to the optimizer in `key=value` format (e.g., `\"weight_decay=0.01\"` `\"betas=0.9,0.999\"`).
+*   `--lr_scheduler=\"...\"`: Specifies the learning rate scheduler. Options include `constant` (no change), `cosine` (cosine curve), `linear` (linear decay), `constant_with_warmup` (constant with warmup), `cosine_with_restarts`, etc. `constant`, `cosine`, and `constant_with_warmup` are commonly used. Some schedulers require additional arguments (see `--lr_scheduler_args`). If using optimizers with auto LR adjustment like `DAdaptation` or `Prodigy`, a scheduler is not needed (`constant` should be specified).
+*   `--lr_warmup_steps=N`: Number of warmup steps for the learning rate scheduler. The learning rate gradually increases during this period at the start of training. If N < 1, it's interpreted as a fraction of total steps.
+*   `--lr_scheduler_num_cycles=N` / `--lr_scheduler_power=P`: Parameters for specific schedulers (`cosine_with_restarts`, `polynomial`).
+*   `--max_train_steps=N` / `--max_train_epochs=N`: Specifies the total number of training steps or epochs. Epoch specification takes precedence.
+*   `--mixed_precision=\"bf16\"` / `\"fp16\"` / `\"no\"`: Mixed precision training settings. For SDXL, using `bf16` (if GPU supports it) or `fp16` is strongly recommended. Reduces VRAM usage and improves training speed.
+*   `--full_fp16` / `--full_bf16`: Performs gradient calculations entirely in half-precision/bf16. Can further reduce VRAM usage but may affect training stability. Use if VRAM is critically low.
+*   `--gradient_accumulation_steps=N`: Accumulates gradients for N steps before updating the optimizer. Effectively increases the batch size to `train_batch_size * N`, achieving the effect of a larger batch size with less VRAM. Default is 1.
+*   `--max_grad_norm=N`: Gradient clipping threshold. Clips gradients if their norm exceeds N. Default is 1.0. `0` disables it.
+*   `--gradient_checkpointing`: Significantly reduces memory usage but slightly decreases training speed. Recommended for SDXL due to high memory consumption.
+*   `--fused_backward_pass`: **Experimental**: Fuses gradient calculation and optimizer steps to reduce VRAM usage. Available for SDXL. Currently only supports `Adafactor` optimizer. Cannot be used with Gradient Accumulation.
+*   `--resume=\"<state directory>\"`: Resumes training from a saved state (saved with `--save_state`). Restores optimizer state, step count, etc.
+
+### 1.6. Caching
+
+Caching is effective for SDXL due to its high computational cost.
+
+*   `--cache_latents`: Caches VAE outputs (latents) in memory. Skips VAE computation, reducing VRAM usage and speeding up training. **Note:** Image augmentations (`color_aug`, `flip_aug`, `random_crop`, etc.) will be disabled.
+*   `--cache_latents_to_disk`: Used with `--cache_latents` to cache to disk. Particularly effective for large datasets or multiple training runs. Caches are generated on disk during the first run and loaded from there on subsequent runs.
+*   `--cache_text_encoder_outputs`: Caches Text Encoder outputs in memory. Skips Text Encoder computation, reducing VRAM usage and speeding up training. **Note:** Caption augmentations (`shuffle_caption`, `caption_dropout_rate`, etc.) will be disabled. **Also, when using this option, Text Encoder LoRA modules cannot be trained (requires `--network_train_unet_only`).**
+*   `--cache_text_encoder_outputs_to_disk`: Used with `--cache_text_encoder_outputs` to cache to disk.
+*   `--skip_cache_check`: Skips validation of cache file contents. File existence is checked, and if not found, caches are generated. Usually not needed unless intentionally re-caching for debugging, etc.
+
+### 1.7. Sample Image Generation
+
+Basic options are common with `train_network.py`.
+
+*   `--sample_every_n_steps=N` / `--sample_every_n_epochs=N`: Generates sample images every N steps/epochs.
+*   `--sample_at_first`: Generates sample images before training starts.
+*   `--sample_prompts=\"<prompt file>\"`: Specifies a file (`.txt`, `.toml`, `.json`) containing prompts for sample image generation. 
+*   `--sample_sampler=\"...\"`: Specifies the sampler (scheduler) for sample image generation. `euler_a`, `dpm++_2m_karras`, etc., are common. See `--help` for choices.
+
+#### Format of Prompt File
+
+A prompt file can contain multiple prompts with options, for example:
+
+```
+# prompt 1
+masterpiece, best quality, (1girl), in white shirts, upper body, looking at viewer, simple background --n low quality, worst quality, bad anatomy,bad composition, poor, low effort --w 768 --h 768 --d 1 --l 7.5 --s 28
+
+# prompt 2
+masterpiece, best quality, 1boy, in business suit, standing at street, looking back --n (low quality, worst quality), bad anatomy,bad composition, poor, low effort --w 576 --h 832 --d 2 --l 5.5 --s 40
+```
+
+  Lines beginning with `#` are comments. You can specify options for the generated image with options like `--n` after the prompt. The following can be used.
+
+  * `--n` Negative prompt up to the next option. Ignored when CFG scale is `1.0`.
+  * `--w` Specifies the width of the generated image.
+  * `--h` Specifies the height of the generated image.
+  * `--d` Specifies the seed of the generated image.
+  * `--l` Specifies the CFG scale of the generated image. For FLUX.1 models, the default is `1.0`, which means no CFG. For Chroma models, set to around `4.0` to enable CFG.
+  * `--g` Specifies the embedded guidance scale for the models with embedded guidance (FLUX.1), the default is `3.5`. Set to `0.0` for Chroma models.
+  * `--s` Specifies the number of steps in the generation.
+
+The prompt weighting such as `( )` and `[ ]` are working for SD/SDXL models, not working for other models like FLUX.1.
+
+### 1.8. Logging & Tracking
+
+*   `--logging_dir=\"<log directory>\"`: Specifies the directory for TensorBoard and other logs. If not specified, logs are not output.
+*   `--log_with=\"tensorboard\"` / `\"wandb\"` / `\"all\"`: Specifies the logging tool to use. If using `wandb`, `pip install wandb` is required.
+*   `--log_prefix=\"<prefix>\"`: Specifies the prefix for subdirectory names created within `logging_dir`.
+*   `--wandb_api_key=\"<API key>\"` / `--wandb_run_name=\"<run name>\"`: Options for Weights & Biases (wandb).
+*   `--log_tracker_name` / `--log_tracker_config`: Advanced tracker configuration options. Usually not needed.
+*   `--log_config`: Logs the training configuration used (excluding some sensitive information) at the start of training. Helps ensure reproducibility.
+
+### 1.9. Regularization and Advanced Techniques
+
+*   `--noise_offset=N`: Enables noise offset and specifies its value. Expected to improve bias in image brightness and contrast. Recommended to enable as SDXL base models are trained with this (e.g., 0.0357). Original technical explanation [here](https://www.crosslabs.org/blog/diffusion-with-offset-noise).
+*   `--noise_offset_random_strength`: Randomly varies noise offset strength between 0 and the specified value.
+*   `--adaptive_noise_scale=N`: Adjusts noise offset based on the mean absolute value of latents. Used with `--noise_offset`.
+*   `--multires_noise_iterations=N` / `--multires_noise_discount=D`: Enables multi-resolution noise. Adding noise of different frequency components is expected to improve detail reproduction. Specify iteration count N (around 6-10) and discount rate D (around 0.3). Technical explanation [here](https://wandb.ai/johnowhitaker/multires_noise/reports/Multi-Resolution-Noise-for-Diffusion-Model-Training--VmlldzozNjYyOTU2).
+*   `--ip_noise_gamma=G` / `--ip_noise_gamma_random_strength`: Enables Input Perturbation Noise. Adds small noise to input (latents) for regularization. Specify Gamma value (around 0.1). Strength can be randomized with `random_strength`.
+*   `--min_snr_gamma=N`: Applies Min-SNR Weighting Strategy. Adjusts loss weights for timesteps with high noise in early training to stabilize learning. `N=5` etc. are used.
+*   `--scale_v_pred_loss_like_noise_pred`: In v-prediction models, scales v-prediction loss similarly to noise prediction loss. **Not typically used for SDXL** as it's not a v-prediction model.
+*   `--v_pred_like_loss=N`: Adds v-prediction-like loss to noise prediction models. `N` specifies its weight. **Not typically used for SDXL**.
+*   `--debiased_estimation_loss`: Calculates loss using Debiased Estimation. Similar purpose to Min-SNR but a different approach.
+*   `--loss_type=\"l1\"` / `\"l2\"` / `\"huber\"` / `\"smooth_l1\"`: Specifies the loss function. Default is `l2` (MSE). `huber` and `smooth_l1` are robust to outliers.
+*   `--huber_schedule=\"constant\"` / `\"exponential\"` / `\"snr\"`: Scheduling method when using `huber` or `smooth_l1` loss. `snr` is recommended.
+*   `--huber_c=C` / `--huber_scale=S`: Parameters for `huber` or `smooth_l1` loss.
+*   `--masked_loss`: Limits loss calculation area based on a mask image. Requires specifying mask images (black and white) in `conditioning_data_dir` in dataset settings. See [About Masked Loss](masked_loss_README.md) for details.
+
+### 1.10. Distributed Training and Other Training Related Options
+
+*   `--seed=N`: Specifies the random seed. Set this to ensure training reproducibility.
+*   `--max_token_length=N` (`75`, `150`, `225`): Maximum token length processed by Text Encoders. For SDXL, typically `75` (default), `150`, or `225`. Longer lengths can handle more complex prompts but increase VRAM usage.
+*   `--clip_skip=N`: Uses the output from N layers skipped from the final layer of Text Encoders. **Not typically used for SDXL**.
+*   `--lowram` / `--highvram`: Options for memory usage optimization. `--lowram` is for environments like Colab where RAM < VRAM, `--highvram` is for environments with ample VRAM.
+*   `--persistent_data_loader_workers` / `--max_data_loader_n_workers=N`: Settings for DataLoader worker processes. Affects wait time between epochs and memory usage.
+*   `--config_file="<config file>"` / `--output_config`: Options to use/output a `.toml` file instead of command line arguments.
+*   **Accelerate/DeepSpeed related:** (`--ddp_timeout`, `--ddp_gradient_as_bucket_view`, `--ddp_static_graph`): Detailed settings for distributed training. Accelerate settings (`accelerate config`) are usually sufficient. DeepSpeed requires separate configuration.
+* `--initial_epoch=<integer>` – Sets the initial epoch number. `1` means first epoch (same as not specifying). Note: `initial_epoch`/`initial_step` doesn't affect the lr scheduler, which means lr scheduler will start from 0 without `--resume`.
+* `--initial_step=<integer>` – Sets the initial step number including all epochs. `0` means first step (same as not specifying). Overwrites `initial_epoch`.
+* `--skip_until_initial_step` – Skips training until `initial_step` is reached.
+
+### 1.11. Console and Logging / コンソールとログ
+
+* `--console_log_level`: Sets the logging level for the console output. Choose from `DEBUG`, `INFO`, `WARNING`, `ERROR`, `CRITICAL`.
+* `--console_log_file`: Redirects console logs to a specified file.
+* `--console_log_simple`: Enables a simpler log format.
+
+### 1.12. Hugging Face Hub Integration / Hugging Face Hub 連携
+
+* `--huggingface_repo_id`: The repository name on Hugging Face Hub to upload the model to (e.g., `your-username/your-model`).
+* `--huggingface_repo_type`: The type of repository on Hugging Face Hub. Usually `model`.
+* `--huggingface_path_in_repo`: The path within the repository to upload files to.
+* `--huggingface_token`: Your Hugging Face Hub authentication token.
+* `--huggingface_repo_visibility`: Sets the visibility of the repository (`public` or `private`).
+* `--resume_from_huggingface`: Resumes training from a state saved on Hugging Face Hub.
+* `--async_upload`: Enables asynchronous uploading of models to the Hub, preventing it from blocking the training process.
+* `--save_n_epoch_ratio`: Saves the model at a certain ratio of total epochs. For example, `5` will save at least 5 checkpoints throughout the training.
+
+### 1.13. Advanced Attention Settings / 高度なAttention設定
+
+* `--mem_eff_attn`: Use memory-efficient attention mechanism. This is an older implementation and `sdpa` or `xformers` are generally recommended.
+* `--xformers`: Use xformers library for memory-efficient attention. Requires `pip install xformers`.
+
+### 1.14. Advanced LR Scheduler Settings / 高度な学習率スケジューラ設定
+
+* `--lr_scheduler_type`: Specifies a custom scheduler module.
+* `--lr_scheduler_args`: Provides additional arguments to the custom scheduler (e.g., `"T_max=100"`).
+* `--lr_decay_steps`: Sets the number of steps for the learning rate to decay.
+* `--lr_scheduler_timescale`: The timescale for the inverse square root scheduler.
+* `--lr_scheduler_min_lr_ratio`: Sets the minimum learning rate as a ratio of the initial learning rate for certain schedulers.
+
+### 1.15. Differential Learning with LoRA / LoRAの差分学習
+
+This technique involves merging a pre-trained LoRA into the base model before starting a new training session. This is useful for fine-tuning an existing LoRA or for learning the 'difference' from it.
+
+* `--base_weights`: Path to one or more LoRA weight files to be merged into the base model before training begins.
+* `--base_weights_multiplier`: A multiplier for the weights of the LoRA specified by `--base_weights`. You can specify multiple values if you provide multiple weights.
+
+### 1.16. Other Miscellaneous Options / その他のオプション
+
+* `--tokenizer_cache_dir`: Specifies a directory to cache the tokenizer, which is useful for offline training.
+* `--scale_weight_norms`: Scales the weight norms of the LoRA modules. This can help prevent overfitting by controlling the magnitude of the weights. A value of `1.0` is a good starting point.
+* `--disable_mmap_load_safetensors`: Disables memory-mapped loading for `.safetensors` files. This can speed up model loading in some environments like WSL.
+
+## 2. Other Tips / その他のTips
+
+*   **VRAM Usage:** SDXL LoRA training requires a lot of VRAM. Even with 24GB VRAM, you might run out of memory depending on settings. Reduce VRAM usage with these settings:
+    *   `--mixed_precision=\"bf16\"` or `\"fp16\"` (essential)
+    *   `--gradient_checkpointing` (strongly recommended)
+    *   `--cache_latents` / `--cache_text_encoder_outputs` (highly effective, with limitations)
+    *   `--optimizer_type=\"AdamW8bit\"` or `\"Adafactor\"`
+    *   Increase `--gradient_accumulation_steps` (reduce batch size)
+    *   `--full_fp16` / `--full_bf16` (be mindful of stability)
+    *   `--fp8_base` / `--fp8_base_unet` (experimental)
+    *   `--fused_backward_pass` (Adafactor only, experimental)
+*   **Learning Rate:** Appropriate learning rates for SDXL LoRA depend on the dataset and `network_dim`/`alpha`. Starting around `1e-4` ~ `4e-5` (U-Net), `1e-5` ~ `2e-5` (Text Encoders) is common.
+*   **Training Time:** Training takes time due to high-resolution data and the size of the SDXL model. Using caching features and appropriate hardware is important.
+*   **Troubleshooting:**
+    *   **NaN Loss:** Learning rate might be too high, mixed precision settings incorrect (e.g., `--no_half_vae` not specified with `fp16`), or dataset issues.
+    *   **Out of Memory (OOM):** Try the VRAM reduction measures listed above.
+    *   **Training not progressing:** Learning rate might be too low, optimizer/scheduler settings incorrect, or dataset issues.
+
+## 3. Conclusion / おわりに
+
+`sdxl_train_network.py` offers many options to customize SDXL LoRA training. Refer to `--help`, other documents and the source code for further details.
+
+<details>
+<summary>日本語</summary>
+
+# 高度な設定: SDXL LoRA学習スクリプト `sdxl_train_network.py` 詳細ガイド
+
+このドキュメントでは、`sd-scripts` リポジトリに含まれる `sdxl_train_network.py` を使用した、SDXL (Stable Diffusion XL) モデルに対する LoRA (Low-Rank Adaptation) モデル学習の高度な設定オプションについて解説します。
+
+基本的な使い方については、以下のドキュメントを参照してください。
+
+*   [LoRA学習スクリプト `train_network.py` の使い方](train_network.md)
+*   [SDXL LoRA学習スクリプト `sdxl_train_network.py` の使い方](sdxl_train_network.md)
+
+このガイドは、基本的なLoRA学習の経験があり、より詳細な設定や高度な機能を試したい熟練した利用者を対象としています。
+
+**前提条件:**
+
+*   `sd-scripts` リポジトリのクローンと Python 環境のセットアップが完了していること。
+*   学習用データセットの準備と設定（`.toml`ファイル）が完了していること。（[データセット設定ガイド](link/to/dataset/config/doc)参照）
+*   基本的なLoRA学習のコマンドライン実行経験があること。
+
+## 1. コマンドライン引数 詳細解説
+
+`sdxl_train_network.py` は `train_network.py` の機能を継承しつつ、SDXL特有の機能を追加しています。ここでは、SDXL LoRA学習に関連する主要なコマンドライン引数について、機能別に分類して詳細に解説します。
+
+基本的な引数については、[LoRA学習スクリプト `train_network.py` の使い方](train_network.md#31-主要なコマンドライン引数) および [SDXL LoRA学習スクリプト `sdxl_train_network.py` の使い方](sdxl_train_network.md#31-主要なコマンドライン引数（差分）) を参照してください。
+
+### 1.1. モデル読み込み関連
+
+*   `--pretrained_model_name_or_path="<モデルパス>"` **[必須]**
+    *   学習のベースとなる **SDXLモデル** を指定します。Hugging Face HubのモデルID、ローカルのDiffusers形式モデルディレクトリ、または`.safetensors`ファイルを指定できます。
+    *   詳細は[基本ガイド](sdxl_train_network.md#モデル関連)を参照してください。
+*   `--vae="<VAEパス>"`
+    *   オプションで、学習に使用するVAEを指定します。SDXLモデルに含まれるVAE以外を使用する場合に指定します。`.ckpt`または`.safetensors`ファイルを指定できます。
+*   `--no_half_vae`
+    *   混合精度(`fp16`/`bf16`)使用時でもVAEを`float32`で動作させます。SDXLのVAEは`float16`で不安定になることがあるため、`fp16`指定時には有効にすることが推奨されます。`bf16`では通常不要です。
+*   `--fp8_base` / `--fp8_base_unet`
+    *   **実験的機能:** ベースモデル（U-Net, Text Encoder）またはU-NetのみをFP8で読み込み、VRAM使用量を削減します。PyTorch 2.1以上が必要です。詳細は TODO 後でドキュメントを追加 の関連セクションを参照してください (SD3の説明ですがSDXLにも適用されます)。
+
+### 1.2. データセット設定関連
+
+*   `--dataset_config="<設定ファイルのパス>"` 
+    *   データセットの設定を記述した`.toml`ファイルを指定します。SDXLでは高解像度データとバケツ機能（`.toml` で `enable_bucket = true` を指定）の利用が一般的です。
+    *   `.toml`ファイルの書き方の詳細は[データセット設定ガイド](link/to/dataset/config/doc)を参照してください。
+    *   アスペクト比バケツの解像度ステップ(`bucket_reso_steps`)は、SDXLでは32の倍数とする必要があります。
+
+### 1.3. 出力・保存関連
+
+基本的なオプションは `train_network.py` と共通です。
+
+*   `--output_dir="<出力先ディレクトリ>"` **[必須]**
+*   `--output_name="<出力ファイル名>"` **[必須]**
+*   `--save_model_as="safetensors"` (推奨), `ckpt`, `pt`, `diffusers`, `diffusers_safetensors`
+*   `--save_precision="fp16"`, `"bf16"`, `"float"`
+    *   モデルの保存精度を指定します。未指定時は学習時の精度(`fp16`, `bf16`等)で保存されます。
+*   `--save_every_n_epochs=N` / `--save_every_n_steps=N`
+    *   Nエポック/ステップごとにモデルを保存します。
+*   `--save_last_n_epochs=M` / `--save_last_n_steps=M`
+    *   エポック/ステップごとに保存する際、最新のM個のみを保持し、古いものは削除します。
+*   `--save_state` / `--save_state_on_train_end`
+    *   モデル保存時/学習終了時に、Optimizerの状態などを含む学習状態(`state`)を保存します。`--resume`オプションでの学習再開に必要です。
+*   `--save_last_n_epochs_state=M` / `--save_last_n_steps_state=M`
+    *   `state`の保存数をM個に制限します。`--save_last_n_epochs/steps`の指定を上書きします。
+*   `--no_metadata`
+    *   出力モデルにメタデータを保存しません。
+*   `--save_state_to_huggingface` / `--huggingface_repo_id` など
+    *   Hugging Face Hubへのモデルやstateのアップロード関連オプション。詳細は TODO ドキュメントを追加 を参照してください。
+
+### 1.4. ネットワークパラメータ (LoRA)
+
+基本的なオプションは `train_network.py` と共通です。
+
+*   `--network_module=networks.lora` **[必須]**
+*   `--network_dim=N` **[必須]**
+    *   LoRAのランク (次元数) を指定します。SDXLでは32や64などが試されることが多いですが、データセットや目的に応じて調整が必要です。
+*   `--network_alpha=M`
+    *   LoRAのアルファ値。`network_dim`の半分程度、または`network_dim`と同じ値などが一般的です。デフォルトは1。
+*   `--network_dropout=P`
+    *   LoRAモジュール内のドロップアウト率 (0.0~1.0)。過学習抑制の効果が期待できます。デフォルトはNone (ドロップアウトなし)。
+*   `--network_args ...`
+    *   ネットワークモジュールへの追加引数を `key=value` 形式で指定します。LoRAでは以下の高度な設定が可能です。
+        *   **階層別 (Block-wise) 次元数/アルファ:**
+            *   U-Netの各ブロックごとに異なる`dim`と`alpha`を指定できます。これにより、特定の層の影響を強めたり弱めたりする調整が可能です。
+            *   `block_dims`: U-NetのLinear層およびConv2d 1x1層に対するブロックごとのdimをカンマ区切りで指定します (SDXLでは23個の数値)。
+            *   `block_alphas`: 上記に対応するalpha値をカンマ区切りで指定します。
+            *   `conv_block_dims`: U-NetのConv2d 3x3層に対するブロックごとのdimをカンマ区切りで指定します。
+            *   `conv_block_alphas`: 上記に対応するalpha値をカンマ区切りで指定します。
+            *   指定しないブロックは `--network_dim`/`--network_alpha` または `--conv_dim`/`--conv_alpha` (存在する場合) の値が使用されます。
+            *   詳細は[LoRA の階層別学習率](train_network.md#lora-の階層別学習率) (train\_network.md内、SDXLでも同様に適用可能) や実装 ([lora.py](lora.py)) を参照してください。
+        *   **LoRA+:**
+            *   `loraplus_lr_ratio=R`: LoRAの上向き重み(UP)の学習率を、下向き重み(DOWN)の学習率のR倍にします。学習速度の向上が期待できます。論文推奨は16。
+            *   `loraplus_unet_lr_ratio=RU`: U-Net部分のLoRA+学習率比を個別に指定します。
+            *   `loraplus_text_encoder_lr_ratio=RT`: Text Encoder部分のLoRA+学習率比を個別に指定します。(`--text_encoder_lr1`, `--text_encoder_lr2`で指定した学習率に乗算されます)
+            *   詳細は[README](../README.md#jan-17-2025--2025-01-17-version-090)や実装 ([lora.py](lora.py)) を参照してください。
+*   `--network_train_unet_only`
+    *   U-NetのLoRAモジュールのみを学習します。Text Encoderの学習を行わない場合に指定します。`--cache_text_encoder_outputs` を使用する場合は必須です。
+*   `--network_train_text_encoder_only`
+    *   Text EncoderのLoRAモジュールのみを学習します。U-Netの学習を行わない場合に指定します。
+*   `--network_weights="<重みファイル>"`
+    *   学習済みのLoRA重みを読み込んで学習を開始します。ファインチューニングや学習再開に使用します。`--resume` との違いは、このオプションはLoRAモジュールの重みのみを読み込み、`--resume` はOptimizerの状態や学習ステップ数なども復元します。
+*   `--dim_from_weights`
+    *   `--network_weights` で指定した重みファイルからLoRAの次元数 (`dim`) を自動的に読み込みます。`--network_dim` の指定は不要になります。
+
+### 1.5. 学習パラメータ
+
+*   `--learning_rate=LR`
+    *   全体の学習率。各モジュール(`unet_lr`, `text_encoder_lr1`, `text_encoder_lr2`)のデフォルト値となります。`1e-3` や `1e-4` などが試されることが多いです。
+*   `--unet_lr=LR_U`
+    *   U-Net部分のLoRAモジュールの学習率。
+*   `--text_encoder_lr1=LR_TE1`
+    *   Text Encoder 1 (OpenCLIP ViT-G/14) のLoRAモジュールの学習率。通常、U-Netより小さい値 (例: `1e-5`, `2e-5`) が推奨されます。
+*   `--text_encoder_lr2=LR_TE2`
+    *   Text Encoder 2 (CLIP ViT-L/14) のLoRAモジュールの学習率。通常、U-Netより小さい値 (例: `1e-5`, `2e-5`) が推奨されます。
+*   `--optimizer_type="..."`
+    *   使用するOptimizerを指定します。`AdamW8bit` (省メモリ、一般的), `Adafactor` (さらに省メモリ、SDXLフルモデル学習で実績あり), `Lion`, `DAdaptation`, `Prodigy`などが選択可能です。各Optimizerには追加の引数が必要な場合があります (`--optimizer_args`参照)。
+    *   `AdamW8bit` や `PagedAdamW8bit` (要 `bitsandbytes`) が一般的です。
+    *   `Adafactor` はメモリ効率が良いですが、設定がやや複雑です (相対ステップ(`relative_step=True`)推奨、学習率スケジューラは`adafactor`推奨)。
+    *   `DAdaptation`, `Prodigy` は学習率の自動調整機能がありますが、LoRA+との併用はできません。学習率は`1.0`程度を指定します。
+    *   詳細は[train\_util.py](train_util.py)の`get_optimizer`関数を参照してください。
+*   `--optimizer_args ...`
+    *   Optimizerへの追加引数を `key=value` 形式で指定します (例: `"weight_decay=0.01"` `"betas=0.9,0.999"`).
+*   `--lr_scheduler="..."`
+    *   学習率スケジューラを指定します。`constant` (変化なし), `cosine` (コサインカーブ), `linear` (線形減衰), `constant_with_warmup` (ウォームアップ付き定数), `cosine_with_restarts` など。`constant` や `cosine` 、 `constant_with_warmup` がよく使われます。
+    *   スケジューラによっては追加の引数が必要です (`--lr_scheduler_args`参照)。
+    *   `DAdaptation` や `Prodigy` などの自己学習率調整機能付きOptimizerを使用する場合、スケジューラは不要です (`constant` を指定)。
+*   `--lr_warmup_steps=N`
+    *   学習率スケジューラのウォームアップステップ数。学習開始時に学習率を徐々に上げていく期間です。N < 1 の場合は全ステップ数に対する割合と解釈されます。
+*   `--lr_scheduler_num_cycles=N` / `--lr_scheduler_power=P`
+    *   特定のスケジューラ (`cosine_with_restarts`, `polynomial`) のためのパラメータ。
+*   `--max_train_steps=N` / `--max_train_epochs=N`
+    *   学習の総ステップ数またはエポック数を指定します。エポック指定が優先されます。
+*   `--mixed_precision="bf16"` / `"fp16"` / `"no"`
+    *   混合精度学習の設定。SDXLでは `bf16` (対応GPUの場合) または `fp16` の使用が強く推奨されます。VRAM使用量を削減し、学習速度を向上させます。
+*   `--full_fp16` / `--full_bf16`
+    *   勾配計算も含めて完全に半精度/bf16で行います。VRAM使用量をさらに削減できますが、学習の安定性に影響する可能性があります。VRAMがどうしても足りない場合に使用します。
+*   `--gradient_accumulation_steps=N`
+    *   勾配をNステップ分蓄積してからOptimizerを更新します。実質的なバッチサイズを `train_batch_size * N` に増やし、少ないVRAMで大きなバッチサイズ相当の効果を得られます。デフォルトは1。
+*   `--max_grad_norm=N`
+    *   勾配クリッピングの閾値。勾配のノルムがNを超える場合にクリッピングします。デフォルトは1.0。`0`で無効。
+*   `--gradient_checkpointing`
+    *   メモリ使用量を大幅に削減しますが、学習速度は若干低下します。SDXLではメモリ消費が大きいため、有効にすることが推奨されます。
+*   `--fused_backward_pass`
+    *   **実験的機能:** 勾配計算とOptimizerのステップを融合し、VRAM使用量を削減します。SDXLで利用可能です。現在 `Adafactor` Optimizerのみ対応。Gradient Accumulationとは併用できません。
+*   `--resume="<stateディレクトリ>"`
+    *   `--save_state`で保存された学習状態から学習を再開します。Optimizerの状態や学習ステップ数などが復元されます。
+
+### 1.6. キャッシュ機能関連
+
+SDXLは計算コストが高いため、キャッシュ機能が効果的です。
+
+*   `--cache_latents`
+    *   VAEの出力(Latent)をメモリにキャッシュします。VAEの計算を省略でき、VRAM使用量を削減し、学習を高速化します。**注意:** 画像に対するAugmentation (`color_aug`, `flip_aug`, `random_crop` 等) は無効になります。
+*   `--cache_latents_to_disk`
+    *   `--cache_latents` と併用し、キャッシュ先をディスクにします。大量のデータセットや複数回の学習で特に有効です。初回実行時にディスクにキャッシュが生成され、2回目以降はそれを読み込みます。
+*   `--cache_text_encoder_outputs`
+    *   Text Encoderの出力をメモリにキャッシュします。Text Encoderの計算を省略でき、VRAM使用量を削減し、学習を高速化します。**注意:** キャプションに対するAugmentation (`shuffle_caption`, `caption_dropout_rate` 等) は無効になります。**また、このオプションを使用する場合、Text EncoderのLoRAモジュールは学習できません (`--network_train_unet_only` の指定が必須です)。**
+*   `--cache_text_encoder_outputs_to_disk`
+    *   `--cache_text_encoder_outputs` と併用し、キャッシュ先をディスクにします。
+*   `--skip_cache_check`
+    *   キャッシュファイルの内容の検証をスキップします。ファイルの存在確認は行われ、存在しない場合はキャッシュが生成されます。デバッグ等で意図的に再キャッシュしたい場合を除き、通常は指定不要です。
+
+### 1.7. サンプル画像生成関連
+
+基本的なオプションは `train_network.py` と共通です。
+
+*   `--sample_every_n_steps=N` / `--sample_every_n_epochs=N`
+    *   Nステップ/エポックごとにサンプル画像を生成します。
+*   `--sample_at_first`
+    *   学習開始前にサンプル画像を生成します。
+*   `--sample_prompts="<プロンプトファイル>"`
+    *   サンプル画像生成に使用するプロンプトを記述したファイル (`.txt`, `.toml`, `.json`) を指定します。
+*   `--sample_sampler="..."`
+    *   サンプル画像生成時のサンプラー（スケジューラ）を指定します。`euler_a`, `dpm++_2m_karras` などが一般的です。選択肢は `--help` を参照してください。
+
+#### プロンプトファイルの書式
+プロンプトファイルは複数のプロンプトとオプションを含めることができます。例えば：
+
+```
+# prompt 1
+masterpiece, best quality, (1girl), in white shirts, upper body, looking at viewer, simple background --n low quality, worst quality, bad anatomy,bad composition, poor, low effort --w 768 --h 768 --d 1 --l 7.5 --s 28
+
+# prompt 2
+masterpiece, best quality, 1boy, in business suit, standing at street, looking back --n (low quality, worst quality), bad anatomy,bad composition, poor, low effort --w 576 --h 832 --d 2 --l 5.5 --s 40
+```
+
+`#`で始まる行はコメントです。生成画像のオプションはプロンプトの後に `--n` のように指定できます。以下のオプションが使用可能です。
+
+  * `--n` 次のオプションまでがネガティブプロンプトです。CFGスケールが `1.0` の場合は無視されます。
+  * `--w` 生成画像の幅を指定します。
+  * `--h` 生成画像の高さを指定します。
+  * `--d` 生成画像のシード値を指定します。
+  * `--l` 生成画像のCFGスケールを指定します。FLUX.1モデルでは、デフォルトは `1.0` でCFGなしを意味します。Chromaモデルでは、CFGを有効にするために `4.0` 程度に設定してください。
+  * `--g` 埋め込みガイダンス付きモデル（FLUX.1）の埋め込みガイダンススケールを指定、デフォルトは `3.5`。Chromaモデルでは `0.0` に設定してください。
+  * `--s` 生成時のステップ数を指定します。
+
+プロンプトの重み付け `( )` や `[ ]` はSD/SDXLモデルで動作し、FLUX.1など他のモデルでは動作しません。
+
+### 1.8. Logging & Tracking 関連
+
+*   `--logging_dir="<ログディレクトリ>"`
+    *   TensorBoardなどのログを出力するディレクトリを指定します。指定しない場合、ログは出力されません。
+*   `--log_with="tensorboard"` / `"wandb"` / `"all"`
+    *   使用するログツールを指定します。`wandb`を使用する場合、`pip install wandb`が必要です。
+*   `--log_prefix="<プレフィックス>"`
+    *   `logging_dir` 内に作成されるサブディレクトリ名の接頭辞を指定します。
+*   `--wandb_api_key="<APIキー>"` / `--wandb_run_name="<実行名>"`
+    *   Weights & Biases (wandb) 使用時のオプション。
+*   `--log_tracker_name` / `--log_tracker_config`
+    *   高度なトラッカー設定用オプション。通常は指定不要。
+*   `--log_config`
+    *   学習開始時に、使用された学習設定（一部の機密情報を除く）をログに出力します。再現性の確保に役立ちます。
+
+### 1.9. 正則化・高度な学習テクニック関連
+
+*   `--noise_offset=N`
+    *   ノイズオフセットを有効にし、その値を指定します。画像の明るさやコントラストの偏りを改善する効果が期待できます。SDXLのベースモデルはこの値で学習されているため、有効にすることが推奨されます (例: 0.0357)。元々の技術解説は[こちら](https://www.crosslabs.org/blog/diffusion-with-offset-noise)。
+*   `--noise_offset_random_strength`
+    *   ノイズオフセットの強度を0から指定値の間でランダムに変動させます。
+*   `--adaptive_noise_scale=N`
+    *   Latentの平均絶対値に応じてノイズオフセットを調整します。`--noise_offset`と併用します。
+*   `--multires_noise_iterations=N` / `--multires_noise_discount=D`
+    *   複数解像度ノイズを有効にします。異なる周波数成分のノイズを加えることで、ディテールの再現性を向上させる効果が期待できます。イテレーション回数N (6-10程度) と割引率D (0.3程度) を指定します。技術解説は[こちら](https://wandb.ai/johnowhitaker/multires_noise/reports/Multi-Resolution-Noise-for-Diffusion-Model-Training--VmlldzozNjYyOTU2)。
+*   `--ip_noise_gamma=G` / `--ip_noise_gamma_random_strength`
+    *   Input Perturbation Noiseを有効にします。入力(Latent)に微小なノイズを加えて正則化を行います。Gamma値 (0.1程度) を指定します。`random_strength`で強度をランダム化できます。
+*   `--min_snr_gamma=N`
+    *   Min-SNR Weighting Strategy を適用します。学習初期のノイズが大きいタイムステップでのLossの重みを調整し、学習を安定させます。`N=5` などが使用されます。
+*   `--scale_v_pred_loss_like_noise_pred`
+    *   v-predictionモデルにおいて、vの予測ロスをノイズ予測ロスと同様のスケールに調整します。SDXLはv-predictionではないため、**通常は使用しません**。
+*   `--v_pred_like_loss=N`
+    *   ノイズ予測モデルにv予測ライクなロスを追加します。`N`でその重みを指定します。SDXLでは**通常は使用しません**。
+*   `--debiased_estimation_loss`
+    *   Debiased EstimationによるLoss計算を行います。Min-SNRと類似の目的を持ちますが、異なるアプローチです。
+*   `--loss_type="l1"` / `"l2"` / `"huber"` / `"smooth_l1"`
+    *   損失関数を指定します。デフォルトは`l2` (MSE)。`huber`や`smooth_l1`は外れ値に頑健な損失関数です。
+*   `--huber_schedule="constant"` / `"exponential"` / `"snr"`
+    *   `huber`または`smooth_l1`損失使用時のスケジューリング方法。`snr`が推奨されています。
+*   `--huber_c=C` / `--huber_scale=S`
+    *   `huber`または`smooth_l1`損失のパラメータ。
+*   `--masked_loss`
+    *   マスク画像に基づいてLoss計算領域を限定します。データセット設定で`conditioning_data_dir`にマスク画像（白黒）を指定する必要があります。詳細は[マスクロスについて](masked_loss_README.md)を参照してください。
+
+### 1.10. 分散学習、その他学習関連
+
+*   `--seed=N`
+    *   乱数シードを指定します。学習の再現性を確保したい場合に設定します。
+*   `--max_token_length=N` (`75`, `150`, `225`)
+    *   Text Encoderが処理するトークンの最大長。SDXLでは通常`75` (デフォルト) または `150`, `225`。長くするとより複雑なプロンプトを扱えますが、VRAM使用量が増加します。
+*   `--clip_skip=N`
+    *   Text Encoderの最終層からN層スキップした層の出力を使用します。SDXLでは**通常使用しません**。
+*   `--lowram` / `--highvram`
+    *   メモリ使用量の最適化に関するオプション。`--lowram`はColabなどRAM < VRAM環境向け、`--highvram`はVRAM潤沢な環境向け。
+*   `--persistent_data_loader_workers` / `--max_data_loader_n_workers=N`
+    *   DataLoaderのワーカプロセスに関する設定。エポック間の待ち時間やメモリ使用量に影響します。
+*   `--config_file="<設定ファイル>"` / `--output_config`
+    *   コマンドライン引数の代わりに`.toml`ファイルを使用/出力するオプション。
+*   **Accelerate/DeepSpeed関連:** (`--ddp_timeout`, `--ddp_gradient_as_bucket_view`, `--ddp_static_graph`)
+    *   分散学習時の詳細設定。通常はAccelerateの設定 (`accelerate config`) で十分です。DeepSpeedを使用する場合は、別途設定が必要です。
+*   `--initial_epoch=<integer>` – 開始エポック番号を設定します。`1`で最初のエポック（未指定時と同じ）。注意：`initial_epoch`/`initial_step`はlr schedulerに影響しないため、`--resume`しない場合はlr schedulerは0から始まります。
+*   `--initial_step=<integer>` – 全エポックを含む開始ステップ番号を設定します。`0`で最初のステップ（未指定時と同じ）。`initial_epoch`を上書きします。
+*   `--skip_until_initial_step` – `initial_step`に到達するまで学習をスキップします。
+
+### 1.11. コンソールとログ
+
+* `--console_log_level`: コンソール出力のログレベルを設定します。`DEBUG`, `INFO`, `WARNING`, `ERROR`, `CRITICAL`から選択します。
+* `--console_log_file`: コンソールのログを指定されたファイルに出力します。
+* `--console_log_simple`: よりシンプルなログフォーマットを有効にします。
+
+### 1.12. Hugging Face Hub 連携
+
+* `--huggingface_repo_id`: モデルをアップロードするHugging Face Hubのリポジトリ名 (例: `your-username/your-model`)。
+* `--huggingface_repo_type`: Hugging Face Hubのリポジトリの種類。通常は`model`です。
+* `--huggingface_path_in_repo`: リポジトリ内でファイルをアップロードするパス。
+* `--huggingface_token`: Hugging Face Hubの認証トークン。
+* `--huggingface_repo_visibility`: リポジトリの公開設定 (`public`または`private`)。
+* `--resume_from_huggingface`: Hugging Face Hubに保存された状態から学習を再開します。
+* `--async_upload`: Hubへのモデルの非同期アップロードを有効にし、学習プロセスをブロックしないようにします。
+* `--save_n_epoch_ratio`: 総エポック数に対する特定の比率でモデルを保存します。例えば`5`を指定すると、学習全体で少なくとも5つのチェックポイントが保存されます。
+
+### 1.13. 高度なAttention設定
+
+* `--mem_eff_attn`: メモリ効率の良いAttentionメカニズムを使用します。これは古い実装であり、一般的には`sdpa`や`xformers`の使用が推奨されます。
+* `--xformers`: メモリ効率の良いAttentionのためにxformersライブラリを使用します。`pip install xformers`が必要です。
+
+### 1.14. 高度な学習率スケジューラ設定
+
+* `--lr_scheduler_type`: カスタムスケジューラモジュールを指定します。
+* `--lr_scheduler_args`: カスタムスケジューラに追加の引数を渡します (例: `"T_max=100"`)。
+* `--lr_decay_steps`: 学習率が減衰するステップ数を設定します。
+* `--lr_scheduler_timescale`: 逆平方根スケジューラのタイムスケール。
+* `--lr_scheduler_min_lr_ratio`: 特定のスケジューラについて、初期学習率に対する最小学習率の比率を設定します。
+
+### 1.15. LoRAの差分学習
+
+既存の学習済みLoRAをベースモデルにマージしてから、新たな学習を開始する手法です。既存LoRAのファインチューニングや、差分を学習させたい場合に有効です。
+
+* `--base_weights`: 学習開始前にベースモデルにマージするLoRAの重みファイルを1つ以上指定します。
+* `--base_weights_multiplier`: `--base_weights`で指定したLoRAの重みの倍率。複数指定も可能です。
+
+### 1.16. その他のオプション
+
+* `--tokenizer_cache_dir`: オフラインでの学習に便利なように、tokenizerをキャッシュするディレクトリを指定します。
+* `--scale_weight_norms`: LoRAモジュールの重みのノルムをスケーリングします。重みの大きさを制御することで過学習を防ぐ助けになります。`1.0`が良い出発点です。
+* `--disable_mmap_load_safetensors`: `.safetensors`ファイルのメモリマップドローディングを無効にします。WSLなどの一部環境でモデルの読み込みを高速化できます。
+
+## 2. その他のTips
+
+
+*   **VRAM使用量:** SDXL LoRA学習は多くのVRAMを必要とします。24GB VRAMでも設定によってはメモリ不足になることがあります。以下の設定でVRAM使用量を削減できます。
+    *   `--mixed_precision="bf16"` または `"fp16"` (必須級)
+    *   `--gradient_checkpointing` (強く推奨)
+    *   `--cache_latents` / `--cache_text_encoder_outputs` (効果大、制約あり)
+    *   `--optimizer_type="AdamW8bit"` または `"Adafactor"`
+    *   `--gradient_accumulation_steps` の値を増やす (バッチサイズを小さくする)
+    *   `--full_fp16` / `--full_bf16` (安定性に注意)
+    *   `--fp8_base` / `--fp8_base_unet` (実験的)
+    *   `--fused_backward_pass` (Adafactor限定、実験的)
+*   **学習率:** SDXL LoRAの適切な学習率はデータセットや`network_dim`/`alpha`に依存します。`1e-4` ~ `4e-5` (U-Net), `1e-5` ~ `2e-5` (Text Encoders) あたりから試すのが一般的です。
+*   **学習時間:** 高解像度データとSDXLモデルのサイズのため、学習には時間がかかります。キャッシュ機能や適切なハードウェアの利用が重要です。
+*   **トラブルシューティング:**
+    *   **NaN Loss:** 学習率が高すぎる、混合精度の設定が不適切 (`fp16`時の`--no_half_vae`未指定など)、データセットの問題などが考えられます。
+    *   **VRAM不足 (OOM):** 上記のVRAM削減策を試してください。
+    *   **学習が進まない:** 学習率が低すぎる、Optimizer/Schedulerの設定が不適切、データセットの問題などが考えられます。
+
+## 3. おわりに
+
+`sdxl_train_network.py` は非常に多くのオプションを提供しており、SDXL LoRA学習の様々な側面をカスタマイズできます。このドキュメントが、より高度な設定やチューニングを行う際の助けとなれば幸いです。
+
+不明な点や詳細については、各スクリプトの `--help` オプションや、リポジトリ内の他のドキュメント、実装コード自体を参照してください。
+
+</details>

docs/train_textual_inversion.md

@@ -0,0 +1,291 @@
+# How to use Textual Inversion training scripts / Textual Inversion学習スクリプトの使い方
+
+This document explains how to train Textual Inversion embeddings using the `train_textual_inversion.py` and `sdxl_train_textual_inversion.py` scripts included in the `sd-scripts` repository.
+
+<details>
+<summary>日本語</summary>
+このドキュメントでは、`sd-scripts` リポジトリに含まれる `train_textual_inversion.py` および `sdxl_train_textual_inversion.py` を使用してTextual Inversionの埋め込みを学習する方法について解説します。
+</details>
+
+## 1. Introduction / はじめに
+
+[Textual Inversion](https://textual-inversion.github.io/) is a technique that teaches Stable Diffusion new concepts by learning new token embeddings. Instead of fine-tuning the entire model, it only optimizes the text encoder's token embeddings, making it a lightweight approach to teaching the model specific characters, objects, or artistic styles.
+
+**Available Scripts:**
+- `train_textual_inversion.py`: For Stable Diffusion v1.x and v2.x models
+- `sdxl_train_textual_inversion.py`: For Stable Diffusion XL models
+
+**Prerequisites:**
+* The `sd-scripts` repository has been cloned and the Python environment has been set up.
+* The training dataset has been prepared. For dataset preparation, please refer to the [Dataset Configuration Guide](config_README-en.md).
+
+<details>
+<summary>日本語</summary>
+
+[Textual Inversion](https://textual-inversion.github.io/) は、新しいトークンの埋め込みを学習することで、Stable Diffusionに新しい概念を教える技術です。モデル全体をファインチューニングする代わりに、テキストエンコーダのトークン埋め込みのみを最適化するため、特定のキャラクター、オブジェクト、芸術的スタイルをモデルに教えるための軽量なアプローチです。
+
+**利用可能なスクリプト:**
+- `train_textual_inversion.py`: Stable Diffusion v1.xおよびv2.xモデル用
+- `sdxl_train_textual_inversion.py`: Stable Diffusion XLモデル用
+
+**前提条件:**
+* `sd-scripts` リポジトリのクローンとPython環境のセットアップが完了していること。
+* 学習用データセットの準備が完了していること。データセットの準備については[データセット設定ガイド](config_README-en.md)を参照してください。
+</details>
+
+## 2. Basic Usage / 基本的な使用方法
+
+### 2.1. For Stable Diffusion v1.x/v2.x Models / Stable Diffusion v1.x/v2.xモデル用
+
+```bash
+accelerate launch --num_cpu_threads_per_process 1 train_textual_inversion.py \
+  --pretrained_model_name_or_path="path/to/model.safetensors" \
+  --dataset_config="dataset_config.toml" \
+  --output_dir="output" \
+  --output_name="my_textual_inversion" \
+  --save_model_as="safetensors" \
+  --token_string="mychar" \
+  --init_word="girl" \
+  --num_vectors_per_token=4 \
+  --max_train_steps=1600 \
+  --learning_rate=1e-6 \
+  --optimizer_type="AdamW8bit" \
+  --mixed_precision="fp16" \
+  --cache_latents \
+  --sdpa
+```
+
+### 2.2. For SDXL Models / SDXLモデル用
+
+```bash
+accelerate launch --num_cpu_threads_per_process 1 sdxl_train_textual_inversion.py \
+  --pretrained_model_name_or_path="path/to/sdxl_model.safetensors" \
+  --dataset_config="dataset_config.toml" \
+  --output_dir="output" \
+  --output_name="my_sdxl_textual_inversion" \
+  --save_model_as="safetensors" \
+  --token_string="mychar" \
+  --init_word="girl" \
+  --num_vectors_per_token=4 \
+  --max_train_steps=1600 \
+  --learning_rate=1e-6 \
+  --optimizer_type="AdamW8bit" \
+  --mixed_precision="fp16" \
+  --cache_latents \
+  --sdpa
+```
+
+<details>
+<summary>日本語</summary>
+上記のコマンドは実際には1行で書く必要がありますが、見やすさのために改行しています（LinuxやMacでは行末に `\` を追加することで改行できます）。Windowsの場合は、改行せずに1行で書くか、`^` を行末に追加してください。
+</details>
+
+## 3. Key Command-Line Arguments / 主要なコマンドライン引数
+
+### 3.1. Textual Inversion Specific Arguments / Textual Inversion固有の引数
+
+#### Core Parameters / コアパラメータ
+
+* `--token_string="mychar"` **[Required]**
+  * Specifies the token string used in training. This must not exist in the tokenizer's vocabulary. In your training prompts, include this token string (e.g., if token_string is "mychar", use prompts like "mychar 1girl").
+  * 学習時に使用されるトークン文字列を指定します。tokenizerの語彙に存在しない文字である必要があります。学習時のプロンプトには、このトークン文字列を含める必要があります（例：token_stringが"mychar"なら、"mychar 1girl"のようなプロンプトを使用）。
+
+* `--init_word="girl"`
+  * Specifies the word to use for initializing the embedding vector. Choose a word that is conceptually close to what you want to teach. Must be a single token.
+  * 埋め込みベクトルの初期化に使用する単語を指定します。教えたい概念に近い単語を選ぶとよいでしょう。単一のトークンである必要があります。
+
+* `--num_vectors_per_token=4`
+  * Specifies how many embedding vectors to use for this token. More vectors provide greater expressiveness but consume more tokens from the 77-token limit.
+  * このトークンに使用する埋め込みベクトルの数を指定します。多いほど表現力が増しますが、77トークン制限からより多くのトークンを消費します。
+
+* `--weights="path/to/existing_embedding.safetensors"`
+  * Loads pre-trained embeddings to continue training from. Optional parameter for transfer learning.
+  * 既存の埋め込みを読み込んで、そこから追加で学習します。転移学習のオプションパラメータです。
+
+#### Template Options / テンプレートオプション
+
+* `--use_object_template`
+  * Ignores captions and uses predefined object templates (e.g., "a photo of a {}"). Same as the original implementation.
+  * キャプションを無視して、事前定義された物体用テンプレート（例："a photo of a {}"）を使用します。公式実装と同じです。
+
+* `--use_style_template`
+  * Ignores captions and uses predefined style templates (e.g., "a painting in the style of {}"). Same as the original implementation.
+  * キャプションを無視して、事前定義されたスタイル用テンプレート（例："a painting in the style of {}"）を使用します。公式実装と同じです。
+
+### 3.2. Model and Dataset Arguments / モデル・データセット引数
+
+For common model and dataset arguments, please refer to [LoRA Training Guide](train_network.md#31-main-command-line-arguments--主要なコマンドライン引数). The following arguments work the same way:
+
+* `--pretrained_model_name_or_path`
+* `--dataset_config`
+* `--v2`, `--v_parameterization`
+* `--resolution`
+* `--cache_latents`, `--vae_batch_size`
+* `--enable_bucket`, `--min_bucket_reso`, `--max_bucket_reso`
+
+<details>
+<summary>日本語</summary>
+一般的なモデル・データセット引数については、[LoRA学習ガイド](train_network.md#31-main-command-line-arguments--主要なコマンドライン引数)を参照してください。以下の引数は同様に動作します：
+
+* `--pretrained_model_name_or_path`
+* `--dataset_config`
+* `--v2`, `--v_parameterization`
+* `--resolution`
+* `--cache_latents`, `--vae_batch_size`
+* `--enable_bucket`, `--min_bucket_reso`, `--max_bucket_reso`
+</details>
+
+### 3.3. Training Parameters / 学習パラメータ
+
+For training parameters, please refer to [LoRA Training Guide](train_network.md#31-main-command-line-arguments--主要なコマンドライン引数). Textual Inversion typically uses these settings:
+
+* `--learning_rate=1e-6`: Lower learning rates are often used compared to LoRA training
+* `--max_train_steps=1600`: Fewer steps are usually sufficient
+* `--optimizer_type="AdamW8bit"`: Memory-efficient optimizer
+* `--mixed_precision="fp16"`: Reduces memory usage
+
+**Note:** Textual Inversion has lower memory requirements compared to full model fine-tuning, so you can often use larger batch sizes.
+
+<details>
+<summary>日本語</summary>
+学習パラメータについては、[LoRA学習ガイド](train_network.md#31-main-command-line-arguments--主要なコマンドライン引数)を参照してください。Textual Inversionでは通常以下の設定を使用します：
+
+* `--learning_rate=1e-6`: LoRA学習と比べて低い学習率がよく使用されます
+* `--max_train_steps=1600`: より少ないステップで十分な場合が多いです
+* `--optimizer_type="AdamW8bit"`: メモリ効率的なオプティマイザ
+* `--mixed_precision="fp16"`: メモリ使用量を削減
+
+**注意:** Textual Inversionはモデル全体のファインチューニングと比べてメモリ要件が低いため、多くの場合、より大きなバッチサイズを使用できます。
+</details>
+
+## 4. Dataset Preparation / データセット準備
+
+### 4.1. Dataset Configuration / データセット設定
+
+Create a TOML configuration file as described in the [Dataset Configuration Guide](config_README-en.md). Here's an example for Textual Inversion:
+
+```toml
+[general]
+shuffle_caption = false
+caption_extension = ".txt"
+keep_tokens = 1
+
+[[datasets]]
+resolution = 512                    # 1024 for SDXL
+batch_size = 4                      # Can use larger values than LoRA training
+enable_bucket = true
+
+  [[datasets.subsets]]
+  image_dir = "path/to/images"
+  caption_extension = ".txt"
+  num_repeats = 10
+```
+
+### 4.2. Caption Guidelines / キャプションガイドライン
+
+**Important:** Your captions must include the token string you specified. For example:
+
+* If `--token_string="mychar"`, captions should be like: "mychar, 1girl, blonde hair, blue eyes"
+* The token string can appear anywhere in the caption, but including it is essential
+
+You can verify that your token string is being recognized by using `--debug_dataset`, which will show token IDs. Look for tokens with IDs ≥ 49408 (these are the new custom tokens).
+
+<details>
+<summary>日本語</summary>
+
+**重要:** キャプションには指定したトークン文字列を含める必要があります。例：
+
+* `--token_string="mychar"` の場合、キャプションは "mychar, 1girl, blonde hair, blue eyes" のようにします
+* トークン文字列はキャプション内のどこに配置しても構いませんが、含めることが必須です
+
+`--debug_dataset` を使用してトークン文字列が認識されているかを確認できます。これによりトークンIDが表示されます。ID ≥ 49408 のトークン（これらは新しいカスタムトークン）を探してください。
+</details>
+
+## 5. Advanced Configuration / 高度な設定
+
+### 5.1. Multiple Token Vectors / 複数トークンベクトル
+
+When using `--num_vectors_per_token` > 1, the system creates additional token variations:
+- `--token_string="mychar"` with `--num_vectors_per_token=4` creates: "mychar", "mychar1", "mychar2", "mychar3"
+
+For generation, you can use either the base token or all tokens together.
+
+### 5.2. Memory Optimization / メモリ最適化
+
+* Use `--cache_latents` to cache VAE outputs and reduce VRAM usage
+* Use `--gradient_checkpointing` for additional memory savings
+* For SDXL, use `--cache_text_encoder_outputs` to cache text encoder outputs
+* Consider using `--mixed_precision="bf16"` on newer GPUs (RTX 30 series and later)
+
+### 5.3. Training Tips / 学習のコツ
+
+* **Learning Rate:** Start with 1e-6 and adjust based on results. Lower rates often work better than LoRA training.
+* **Steps:** 1000-2000 steps are usually sufficient, but this varies by dataset size and complexity.
+* **Batch Size:** Textual Inversion can handle larger batch sizes than full fine-tuning due to lower memory requirements.
+* **Templates:** Use `--use_object_template` for characters/objects, `--use_style_template` for artistic styles.
+
+<details>
+<summary>日本語</summary>
+
+* **学習率:** 1e-6から始めて、結果に基づいて調整してください。LoRA学習よりも低い率がよく機能します。
+* **ステップ数:** 通常1000-2000ステップで十分ですが、データセットのサイズと複雑さによって異なります。
+* **バッチサイズ:** メモリ要件が低いため、Textual Inversionは完全なファインチューニングよりも大きなバッチサイズを処理できます。
+* **テンプレート:** キャラクター/オブジェクトには `--use_object_template`、芸術的スタイルには `--use_style_template` を使用してください。
+</details>
+
+## 6. Usage After Training / 学習後の使用方法
+
+The trained Textual Inversion embeddings can be used in:
+
+* **Automatic1111 WebUI:** Place the `.safetensors` file in the `embeddings` folder
+* **ComfyUI:** Use the embedding file with appropriate nodes
+* **Other Diffusers-based applications:** Load using the embedding path
+
+In your prompts, simply use the token string you trained (e.g., "mychar") and the model will use the learned embedding.
+
+<details>
+<summary>日本語</summary>
+
+学習したTextual Inversionの埋め込みは以下で使用できます：
+
+* **Automatic1111 WebUI:** `.safetensors` ファイルを `embeddings` フォルダに配置
+* **ComfyUI:** 適切なノードで埋め込みファイルを使用
+* **その他のDiffusersベースアプリケーション:** 埋め込みパスを使用して読み込み
+
+プロンプトでは、学習したトークン文字列（例："mychar"）を単純に使用するだけで、モデルが学習した埋め込みを使用します。
+</details>
+
+## 7. Troubleshooting / トラブルシューティング
+
+### Common Issues / よくある問題
+
+1. **Token string already exists in tokenizer**
+   * Use a unique string that doesn't exist in the model's vocabulary
+   * Try adding numbers or special characters (e.g., "mychar123")
+
+2. **No improvement after training**
+   * Ensure your captions include the token string
+   * Try adjusting the learning rate (lower values like 5e-7)
+   * Increase the number of training steps
+
+   * Use `--cache_latents`
+
+<details>
+<summary>日本語</summary>
+
+1. **トークン文字列がtokenizerに既に存在する**
+   * モデルの語彙に存在しない固有の文字列を使用してください
+   * 数字や特殊文字を追加してみてください（例："mychar123"）
+
+2. **学習後に改善が見られない**
+   * キャプションにトークン文字列が含まれていることを確認してください
+   * 学習率を調整してみてください（5e-7のような低い値）
+   * 学習ステップ数を増やしてください
+
+3. **メモリ不足エラー**
+   * データセット設定でバッチサイズを減らしてください
+   * `--gradient_checkpointing` を使用してください
+   * `--cache_latents` を使用してください
+</details>
+
+For additional training options and advanced configurations, please refer to the [LoRA Training Guide](train_network.md) as many parameters are shared between training methods.

docs/validation.md

@@ -0,0 +1,261 @@
+# Validation Loss
+
+Validation loss is a crucial metric for monitoring the training process of a model. It helps you assess how well your model is generalizing to data it hasn't seen during training, which is essential for preventing overfitting. By periodically evaluating the model on a separate validation dataset, you can gain insights into its performance and make more informed decisions about when to stop training or adjust hyperparameters.
+
+This feature provides a stable and reliable validation loss metric by ensuring the validation process is deterministic.
+
+<details>
+<summary>日本語</summary>
+
+Validation loss（検証損失）は、モデルの学習過程を監視するための重要な指標です。モデルが学習中に見ていないデータに対してどの程度汎化できているかを評価するのに役立ち、過学習を防ぐために不可欠です。個別の検証データセットで定期的にモデルを評価することで、そのパフォーマンスに関する洞察を得て、学習をいつ停止するか、またはハイパーパラメータを調整するかについて、より多くの情報に基づいた決定を下すことができます。
+
+この機能は、検証プロセスが決定論的であることを保証することにより、安定して信頼性の高い検証損失指標を提供します。
+
+</details>
+
+## How It Works
+
+When validation is enabled, a portion of your dataset is set aside specifically for this purpose. The script then runs a validation step at regular intervals, calculating the loss on this validation data.
+
+To ensure that the validation loss is a reliable indicator of model performance, the process is deterministic. This means that for every validation run, the same random seed is used for noise generation and timestep selection. This consistency ensures that any fluctuations in the validation loss are due to changes in the model's weights, not random variations in the validation process itself.
+
+The average loss across all validation steps is then logged, providing a single, clear metric to track.
+
+For more technical details, please refer to the original pull request: [PR #1903](https://github.com/kohya-ss/sd-scripts/pull/1903).
+
+<details>
+<summary>日本語</summary>
+
+検証が有効になると、データセットの一部がこの目的のために特別に確保されます。スクリプトは定期的な間隔で検証ステップを実行し、この検証データに対する損失を計算します。
+
+検証損失がモデルのパフォーマンスの信頼できる指標であることを保証するために、プロセスは決定論的です。つまり、すべての検証実行で、ノイズ生成とタイムステップ選択に同じランダムシードが使用されます。この一貫性により、検証損失の変動が、検証プロセス自体のランダムな変動ではなく、モデルの重みの変化によるものであることが保証されます。
+
+すべての検証ステップにわたる平均損失がログに記録され、追跡するための単一の明確な指標が提供されます。
+
+より技術的な詳細については、元のプルリクエストを参照してください: [PR #1903](https://github.com/kohya-ss/sd-scripts/pull/1903).
+
+</details>
+
+## How to Use
+
+### Enabling Validation
+
+There are two primary ways to enable validation:
+
+1.  **Using a Dataset Config File (Recommended)**: You can specify a validation set directly within your dataset `.toml` file. This method offers the most control, allowing you to designate entire directories as validation sets or split a percentage of a specific subset for validation.
+
+    To use a whole directory for validation, add a subset and set `validation_split = 1.0`.
+
+    **Example: Separate Validation Set**
+    ```toml
+    [[datasets]]
+      # ... training subset ...
+      [[datasets.subsets]]
+        image_dir = "path/to/train_images"
+        # ... other settings ...
+
+      # Validation subset
+      [[datasets.subsets]]
+        image_dir = "path/to/validation_images"
+        validation_split = 1.0  # Use this entire subset for validation
+    ```
+
+    To use a fraction of a subset for validation, set `validation_split` to a value between 0.0 and 1.0.
+
+    **Example: Splitting a Subset**
+    ```toml
+    [[datasets]]
+      # ... dataset settings ...
+      [[datasets.subsets]]
+        image_dir = "path/to/images"
+        validation_split = 0.1  # Use 10% of this subset for validation
+    ```
+
+2.  **Using a Command-Line Argument**: For a simpler setup, you can use the `--validation_split` argument. This will take a random percentage of your *entire* training dataset for validation. This method is ignored if `validation_split` is defined in your dataset config file.
+
+    **Example Command:**
+    ```bash
+    accelerate launch train_network.py ... --validation_split 0.1
+    ```
+    This command will use 10% of the total training data for validation.
+
+<details>
+<summary>日本語</summary>
+
+### 検証を有効にする
+
+検証を有効にする主な方法は2つあります。
+
+1.  **データセット設定ファイルを使用する（推奨）**: データセットの`.toml`ファイル内で直接検証セットを指定できます。この方法は最も制御性が高く、ディレクトリ全体を検証セットとして指定したり、特定のサブセットのパーセンテージを検証用に分割したりすることができます。
+
+    ディレクトリ全体を検証に使用するには、サブセットを追加して`validation_split = 1.0`と設定します。
+
+    **例：個別の検証セット**
+    ```toml
+    [[datasets]]
+      # ... training subset ...
+      [[datasets.subsets]]
+        image_dir = "path/to/train_images"
+        # ... other settings ...
+
+      # Validation subset
+      [[datasets.subsets]]
+        image_dir = "path/to/validation_images"
+        validation_split = 1.0  # このサブセット全体を検証に使用します
+    ```
+
+    サブセットの一部を検証に使用するには、`validation_split`を0.0から1.0の間の値に設定します。
+
+    **例：サブセットの分割**
+    ```toml
+    [[datasets]]
+      # ... dataset settings ...
+      [[datasets.subsets]]
+        image_dir = "path/to/images"
+        validation_split = 0.1  # このサブセットの10%を検証に使用します
+    ```
+
+2.  **コマンドライン引数を使用する**: より簡単な設定のために、`--validation_split`引数を使用できます。これにより、*全*学習データセットのランダムなパーセンテージが検証に使用されます。この方法は、データセット設定ファイルで`validation_split`が定義されている場合は無視されます。
+
+    **コマンド例:**
+    ```bash
+    accelerate launch train_network.py ... --validation_split 0.1
+    ```
+    このコマンドは、全学習データの10%を検証に使用します。
+
+</details>
+
+### Configuration Options
+
+| Argument                    | TOML Option         | Description                                                                                                                            |
+| --------------------------- | ------------------- | -------------------------------------------------------------------------------------------------------------------------------------- |
+| `--validation_split`        | `validation_split`  | The fraction of the dataset to use for validation. The command-line argument applies globally, while the TOML option applies per-subset. The TOML setting takes precedence. |
+| `--validate_every_n_steps`  |                     | Run validation every N steps.                                                                                                          |
+| `--validate_every_n_epochs` |                     | Run validation every N epochs. If not specified, validation runs once per epoch by default.                                            |
+| `--max_validation_steps`    |                     | The maximum number of batches to use for a single validation run. If not set, the entire validation dataset is used.                     |
+| `--validation_seed`         | `validation_seed`   | A specific seed for the validation dataloader shuffling. If not set in the TOML file, the main training `--seed` is used.                 |
+
+<details>
+<summary>日本語</summary>
+
+### 設定オプション
+
+| 引数                        | TOMLオプション      | 説明                                                                                                                                   |
+| --------------------------- | ------------------- | -------------------------------------------------------------------------------------------------------------------------------------- |
+| `--validation_split`        | `validation_split`  | 検証に使用するデータセットの割合。コマンドライン引数は全体に適用され、TOMLオプションはサブセットごとに適用されます。TOML設定が優先されます。 |
+| `--validate_every_n_steps`  |                     | Nステップごとに検証を実行します。                                                                                                      |
+| `--validate_every_n_epochs` |                     | Nエポックごとに検証を実行します。指定しない場合、デフォルトでエポックごとに1回検証が実行されます。                                       |
+| `--max_validation_steps`    |                     | 1回の検証実行に使用するバッチの最大数。設定しない場合、検証データセット全体が使用されます。                                            |
+| `--validation_seed`         | `validation_seed`   | 検証データローダーのシャッフル用の特定のシード。TOMLファイルで設定されていない場合、メインの学習`--seed`が使用されます。                 |
+
+</details>
+
+### Viewing the Results
+
+The validation loss is logged to your tracking tool of choice (TensorBoard or Weights & Biases). Look for the metric `loss/validation` to monitor the performance.
+
+<details>
+<summary>日本語</summary>
+
+### 結果の表示
+
+検証損失は、選択した追跡ツール（TensorBoardまたはWeights & Biases）に記録されます。パフォーマンスを監視するには、`loss/validation`という指標を探してください。
+
+</details>
+
+### Practical Example
+
+Here is a complete example of how to run a LoRA training with validation enabled:
+
+**1. Prepare your `dataset_config.toml`:**
+
+```toml
+[general]
+shuffle_caption = true
+keep_tokens = 1
+
+[[datasets]]
+resolution = "1024,1024"
+batch_size = 2
+
+  [[datasets.subsets]]
+  image_dir = 'path/to/your_images'
+  caption_extension = '.txt'
+  num_repeats = 10
+
+  [[datasets.subsets]]
+  image_dir = 'path/to/your_validation_images'
+  caption_extension = '.txt'
+  validation_split = 1.0 # Use this entire subset for validation
+```
+
+**2. Run the training command:**
+
+```bash
+accelerate launch sdxl_train_network.py \
+  --pretrained_model_name_or_path="sd_xl_base_1.0.safetensors" \
+  --dataset_config="dataset_config.toml" \
+  --output_dir="output" \
+  --output_name="my_lora" \
+  --network_module=networks.lora \
+  --network_dim=32 \
+  --network_alpha=16 \
+  --save_every_n_epochs=1 \
+  --learning_rate=1e-4 \
+  --optimizer_type="AdamW8bit" \
+  --mixed_precision="bf16" \
+  --logging_dir=logs
+```
+
+The validation loss will be calculated once per epoch and saved to the `logs` directory, which you can view with TensorBoard.
+
+<details>
+<summary>日本語</summary>
+
+### 実践的な例
+
+検証を有効にしてLoRAの学習を実行する完全な例を次に示します。
+
+**1. `dataset_config.toml`を準備します:**
+
+```toml
+[general]
+shuffle_caption = true
+keep_tokens = 1
+
+[[datasets]]
+resolution = "1024,1024"
+batch_size = 2
+
+  [[datasets.subsets]]
+  image_dir = 'path/to/your_images'
+  caption_extension = '.txt'
+  num_repeats = 10
+
+  [[datasets.subsets]]
+  image_dir = 'path/to/your_validation_images'
+  caption_extension = '.txt'
+  validation_split = 1.0 # このサブセット全体を検証に使用します
+```
+
+**2. 学習コマンドを実行します:**
+
+```bash
+accelerate launch sdxl_train_network.py \
+  --pretrained_model_name_or_path="sd_xl_base_1.0.safetensors" \
+  --dataset_config="dataset_config.toml" \
+  --output_dir="output" \
+  --output_name="my_lora" \
+  --network_module=networks.lora \
+  --network_dim=32 \
+  --network_alpha=16 \
+  --save_every_n_epochs=1 \
+  --learning_rate=1e-4 \
+  --optimizer_type="AdamW8bit" \
+  --mixed_precision="bf16" \
+  --logging_dir=logs
+```
+
+検証損失はエポックごとに1回計算され、`logs`ディレクトリに保存されます。これはTensorBoardで表示できます。
+
+</details>

docs/wd14_tagger_README-en.md

@@ -0,0 +1,92 @@
+# Image Tagging using WD14Tagger
+
+This document is based on the information from this github page (https://github.com/toriato/stable-diffusion-webui-wd14-tagger#mrsmilingwolfs-model-aka-waifu-diffusion-14-tagger).
+
+Using onnx for inference is recommended. Please install onnx with the following command:
+
+```powershell
+pip install onnx onnxruntime-gpu
+```
+
+See [the official documentation](https://onnxruntime.ai/docs/install/#python-installs) for more details.
+
+The model weights will be automatically downloaded from Hugging Face.
+
+# Usage
+
+Run the script to perform tagging.
+
+```powershell
+python finetune/tag_images_by_wd14_tagger.py --onnx --repo_id <model repo id> --batch_size <batch size> <training data folder>
+```
+
+For example, if using the repository `SmilingWolf/wd-swinv2-tagger-v3` with a batch size of 4, and the training data is located in the parent folder `train_data`, it would be:
+
+```powershell
+python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagger-v3 --batch_size 4 ..\train_data
+```
+
+On the first run, the model files will be automatically downloaded to the `wd14_tagger_model` folder (the folder can be changed with an option). 
+
+Tag files will be created in the same directory as the training data images, with the same filename and a `.txt` extension.
+
+![Generated tag files](https://user-images.githubusercontent.com/52813779/208910534-ea514373-1185-4b7d-9ae3-61eb50bc294e.png)
+
+![Tags and image](https://user-images.githubusercontent.com/52813779/208910599-29070c15-7639-474f-b3e4-06bd5a3df29e.png)
+
+## Example
+
+To output in the Animagine XL 3.1 format, it would be as follows (enter on a single line in practice):
+
+```
+python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagger-v3 
+    --batch_size 4  --remove_underscore --undesired_tags "PUT,YOUR,UNDESIRED,TAGS" --recursive 
+    --use_rating_tags_as_last_tag --character_tags_first --character_tag_expand 
+    --always_first_tags "1girl,1boy"  ..\train_data
+```
+
+## Available Repository IDs
+
+[SmilingWolf's V2 and V3 models](https://huggingface.co/SmilingWolf) are available for use. Specify them in the format like `SmilingWolf/wd-vit-tagger-v3`. The default when omitted is `SmilingWolf/wd-v1-4-convnext-tagger-v2`.
+
+# Options 
+
+All options can be checked with `python tag_images_by_wd14_tagger.py --help`.
+
+## General Options
+
+- `--onnx`: Use ONNX for inference. If not specified, TensorFlow will be used. If using TensorFlow, please install TensorFlow separately. 
+- `--batch_size`: Number of images to process at once. Default is 1. Adjust according to VRAM capacity.
+- `--caption_extension`: File extension for caption files. Default is `.txt`.
+- `--max_data_loader_n_workers`: Maximum number of workers for DataLoader. Specifying a value of 1 or more will use DataLoader to speed up image loading. If unspecified, DataLoader will not be used.
+- `--thresh`: Confidence threshold for outputting tags. Default is 0.35. Lowering the value will assign more tags but accuracy will decrease. 
+- `--general_threshold`: Confidence threshold for general tags. If omitted, same as `--thresh`.
+- `--character_threshold`: Confidence threshold for character tags. If omitted, same as `--thresh`.
+- `--recursive`: If specified, subfolders within the specified folder will also be processed recursively.
+- `--append_tags`: Append tags to existing tag files.
+- `--frequency_tags`: Output tag frequencies.  
+- `--debug`: Debug mode. Outputs debug information if specified.
+
+## Model Download
+
+- `--model_dir`: Folder to save model files. Default is `wd14_tagger_model`.  
+- `--force_download`: Re-download model files if specified.
+
+## Tag Editing
+
+- `--remove_underscore`: Remove underscores from output tags.
+- `--undesired_tags`: Specify tags not to output. Multiple tags can be specified, separated by commas. For example, `black eyes,black hair`.
+- `--use_rating_tags`: Output rating tags at the beginning of the tags.
+- `--use_rating_tags_as_last_tag`: Add rating tags at the end of the tags.
+- `--character_tags_first`: Output character tags first.
+- `--character_tag_expand`: Expand character tag series names. For example, split the tag `chara_name_(series)` into `chara_name, series`.  
+- `--always_first_tags`: Specify tags to always output first when a certain tag appears in an image. Multiple tags can be specified, separated by commas. For example, `1girl,1boy`.
+- `--caption_separator`: Separate tags with this string in the output file. Default is `, `.
+- `--tag_replacement`: Perform tag replacement. Specify in the format `tag1,tag2;tag3,tag4`. If using `,` and `;`, escape them with `\`. \
+    For example, specify `aira tsubase,aira tsubase (uniform)` (when you want to train a specific costume), `aira tsubase,aira tsubase\, heir of shadows` (when the series name is not included in the tag).
+
+When using `tag_replacement`, it is applied after `character_tag_expand`.
+
+When specifying `remove_underscore`, specify `undesired_tags`, `always_first_tags`, and `tag_replacement` without including underscores.
+
+When specifying `caption_separator`, separate `undesired_tags` and `always_first_tags` with `caption_separator`. Always separate `tag_replacement` with `,`.

docs/wd14_tagger_README-ja.md

@@ -0,0 +1,92 @@
+# WD14Taggerによるタグ付け
+
+こちらのgithubページ（https://github.com/toriato/stable-diffusion-webui-wd14-tagger#mrsmilingwolfs-model-aka-waifu-diffusion-14-tagger ）の情報を参考にさせていただきました。
+
+onnx を用いた推論を推奨します。以下のコマンドで onnx をインストールしてください。
+
+```powershell
+pip install onnx onnxruntime-gpu
+```
+
+詳細は[公式ドキュメント](https://onnxruntime.ai/docs/install/#python-installs)をご覧ください。
+
+モデルの重みはHugging Faceから自動的にダウンロードしてきます。
+
+# 使い方
+
+スクリプトを実行してタグ付けを行います。
+```
+python fintune/tag_images_by_wd14_tagger.py --onnx --repo_id <モデルのrepo id> --batch_size <バッチサイズ> <教師データフォルダ>
+```
+
+レポジトリに `SmilingWolf/wd-swinv2-tagger-v3` を使用し、バッチサイズを4にして、教師データを親フォルダの `train_data`に置いた場合、以下のようになります。
+
+```
+python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagger-v3 --batch_size 4 ..\train_data
+```
+
+初回起動時にはモデルファイルが `wd14_tagger_model` フォルダに自動的にダウンロードされます（フォルダはオプションで変えられます）。
+
+タグファイルが教師データ画像と同じディレクトリに、同じファイル名、拡張子.txtで作成されます。
+
+![生成されたタグファイル](https://user-images.githubusercontent.com/52813779/208910534-ea514373-1185-4b7d-9ae3-61eb50bc294e.png)
+
+![タグと画像](https://user-images.githubusercontent.com/52813779/208910599-29070c15-7639-474f-b3e4-06bd5a3df29e.png)
+
+## 記述例
+
+Animagine XL 3.1 方式で出力する場合、以下のようになります（実際には 1 行で入力してください）。
+
+```
+python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagger-v3 
+    --batch_size 4  --remove_underscore --undesired_tags "PUT,YOUR,UNDESIRED,TAGS" --recursive 
+    --use_rating_tags_as_last_tag --character_tags_first --character_tag_expand 
+    --always_first_tags "1girl,1boy"  ..\train_data
+```
+
+## 使用可能なリポジトリID
+
+[SmilingWolf 氏の V2、V3 のモデル](https://huggingface.co/SmilingWolf)が使用可能です。`SmilingWolf/wd-vit-tagger-v3` のように指定してください。省略時のデフォルトは `SmilingWolf/wd-v1-4-convnext-tagger-v2` です。
+
+# オプション
+
+全てオプションは `python tag_images_by_wd14_tagger.py --help` で確認できます。
+
+## 一般オプション
+
+- `--onnx` : ONNX を使用して推論します。指定しない場合は TensorFlow を使用します。TensorFlow 使用時は別途 TensorFlow をインストールしてください。
+- `--batch_size` : 一度に処理する画像の数。デフォルトは1です。VRAMの容量に応じて増減してください。
+- `--caption_extension` : キャプションファイルの拡張子。デフォルトは `.txt` です。
+- `--max_data_loader_n_workers` : DataLoader の最大ワーカー数です。このオプションに 1 以上の数値を指定すると、DataLoader を用いて画像読み込みを高速化します。未指定時は DataLoader を用いません。
+- `--thresh` : 出力するタグの信頼度の閾値。デフォルトは0.35です。値を下げるとより多くのタグが付与されますが、精度は下がります。
+- `--general_threshold` : 一般タグの信頼度の閾値。省略時は `--thresh` と同じです。
+- `--character_threshold` : キャラクタータグの信頼度の閾値。省略時は `--thresh` と同じです。
+- `--recursive` : 指定すると、指定したフォルダ内のサブフォルダも再帰的に処理します。
+- `--append_tags` : 既存のタグファイルにタグを追加します。
+- `--frequency_tags` : タグの頻度を出力します。
+- `--debug` : デバッグモード。指定するとデバッグ情報を出力します。
+
+## モデルのダウンロード
+
+- `--model_dir` : モデルファイルの保存先フォルダ。デフォルトは `wd14_tagger_model` です。
+- `--force_download` : 指定するとモデルファイルを再ダウンロードします。
+
+## タグ編集関連
+
+- `--remove_underscore` : 出力するタグからアンダースコアを削除します。
+- `--undesired_tags` : 出力しないタグを指定します。カンマ区切りで複数指定できます。たとえば `black eyes,black hair` のように指定します。
+- `--use_rating_tags` : タグの最初にレーティングタグを出力します。
+- `--use_rating_tags_as_last_tag` : タグの最後にレーティングタグを追加します。
+- `--character_tags_first` : キャラクタータグを最初に出力します。
+- `--character_tag_expand` : キャラクタータグのシリーズ名を展開します。たとえば `chara_name_(series)` のタグを `chara_name, series` に分割します。
+- `--always_first_tags` : あるタグが画像に出力されたとき、そのタグを最初に出力するタグを指定します。カンマ区切りで複数指定できます。たとえば `1girl,1boy` のように指定します。
+- `--caption_separator` : 出力するファイルでタグをこの文字列で区切ります。デフォルトは `, ` です。
+- `--tag_replacement` : タグの置換を行います。`tag1,tag2;tag3,tag4` のように指定します。`,` および `;` を使う場合は `\` でエスケープしてください。\
+    たとえば `aira tsubase,aira tsubase (uniform)` （特定の衣装を学習させたいとき）、`aira tsubase,aira tsubase\, heir of shadows` （シリーズ名がタグに含まれないとき）のように指定します。
+
+`tag_replacement` は `character_tag_expand` の後に適用されます。
+
+`remove_underscore` 指定時は、`undesired_tags`、`always_first_tags`、`tag_replacement` はアンダースコアを含めずに指定してください。
+
+`caption_separator` 指定時は、`undesired_tags`、`always_first_tags` は `caption_separator`  で区切ってください。`tag_replacement` は必ず `,` で区切ってください。
+

fine_tune.py

@@ -2,29 +2,32 @@
 # XXX dropped option: hypernetwork training
 
 import argparse
-import gc
 import math
 import os
 from multiprocessing import Value
 import toml
 
 from tqdm import tqdm
+
 import torch
+from library import deepspeed_utils, strategy_base
+from library.device_utils import init_ipex, clean_memory_on_device
 
-try:
-    import intel_extension_for_pytorch as ipex
+init_ipex()
 
-    if torch.xpu.is_available():
-        from library.ipex import ipex_init
-
-        ipex_init()
-except Exception:
-    pass
 from accelerate.utils import set_seed
 from diffusers import DDPMScheduler
 
+from library.utils import setup_logging, add_logging_arguments
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
 import library.train_util as train_util
 import library.config_util as config_util
+import library.sai_model_spec as sai_model_spec
 from library.config_util import (
     ConfigSanitizer,
     BlueprintGenerator,
@@ -37,28 +40,39 @@ from library.custom_train_functions import (
     scale_v_prediction_loss_like_noise_prediction,
     apply_debiased_estimation,
 )
+import library.strategy_sd as strategy_sd
 
 
 def train(args):
     train_util.verify_training_args(args)
     train_util.prepare_dataset_args(args, True)
+    deepspeed_utils.prepare_deepspeed_args(args)
+    setup_logging(args, reset=True)
 
     cache_latents = args.cache_latents
 
     if args.seed is not None:
         set_seed(args.seed)  # 乱数系列を初期化する
 
-    tokenizer = train_util.load_tokenizer(args)
+    tokenize_strategy = strategy_sd.SdTokenizeStrategy(args.v2, args.max_token_length, args.tokenizer_cache_dir)
+    strategy_base.TokenizeStrategy.set_strategy(tokenize_strategy)
+
+    # prepare caching strategy: this must be set before preparing dataset. because dataset may use this strategy for initialization.
+    if cache_latents:
+        latents_caching_strategy = strategy_sd.SdSdxlLatentsCachingStrategy(
+            False, args.cache_latents_to_disk, args.vae_batch_size, args.skip_cache_check
+        )
+        strategy_base.LatentsCachingStrategy.set_strategy(latents_caching_strategy)
 
     # データセットを準備する
     if args.dataset_class is None:
         blueprint_generator = BlueprintGenerator(ConfigSanitizer(False, True, False, True))
         if args.dataset_config is not None:
-            print(f"Load dataset config from {args.dataset_config}")
+            logger.info(f"Load dataset config from {args.dataset_config}")
             user_config = config_util.load_user_config(args.dataset_config)
             ignored = ["train_data_dir", "in_json"]
             if any(getattr(args, attr) is not None for attr in ignored):
-                print(
+                logger.warning(
                     "ignore following options because config file is found: {0} / 設定ファイルが利用されるため以下のオプションは無視されます: {0}".format(
                         ", ".join(ignored)
                     )
@@ -77,21 +91,24 @@ def train(args):
                 ]
             }
 
-        blueprint = blueprint_generator.generate(user_config, args, tokenizer=tokenizer)
-        train_dataset_group = config_util.generate_dataset_group_by_blueprint(blueprint.dataset_group)
+        blueprint = blueprint_generator.generate(user_config, args)
+        train_dataset_group, val_dataset_group = config_util.generate_dataset_group_by_blueprint(blueprint.dataset_group)
     else:
-        train_dataset_group = train_util.load_arbitrary_dataset(args, tokenizer)
+        train_dataset_group = train_util.load_arbitrary_dataset(args)
+        val_dataset_group = None
 
     current_epoch = Value("i", 0)
     current_step = Value("i", 0)
     ds_for_collator = train_dataset_group if args.max_data_loader_n_workers == 0 else None
     collator = train_util.collator_class(current_epoch, current_step, ds_for_collator)
 
+    train_dataset_group.verify_bucket_reso_steps(64)
+
     if args.debug_dataset:
         train_util.debug_dataset(train_dataset_group)
         return
     if len(train_dataset_group) == 0:
-        print(
+        logger.error(
             "No data found. Please verify the metadata file and train_data_dir option. / 画像がありません。メタデータおよびtrain_data_dirオプションを確認してください。"
         )
         return
@@ -102,11 +119,12 @@ def train(args):
         ), "when caching latents, either color_aug or random_crop cannot be used / latentをキャッシュするときはcolor_augとrandom_cropは使えません"
 
     # acceleratorを準備する
-    print("prepare accelerator")
+    logger.info("prepare accelerator")
     accelerator = train_util.prepare_accelerator(args)
 
     # mixed precisionに対応した型を用意しておき適宜castする
     weight_dtype, save_dtype = train_util.prepare_dtype(args)
+    vae_dtype = torch.float32 if args.no_half_vae else weight_dtype
 
     # モデルを読み込む
     text_encoder, vae, unet, load_stable_diffusion_format = train_util.load_target_model(args, weight_dtype, accelerator)
@@ -157,15 +175,14 @@ def train(args):
 
     # 学習を準備する
     if cache_latents:
-        vae.to(accelerator.device, dtype=weight_dtype)
+        vae.to(accelerator.device, dtype=vae_dtype)
         vae.requires_grad_(False)
         vae.eval()
-        with torch.no_grad():
-            train_dataset_group.cache_latents(vae, args.vae_batch_size, args.cache_latents_to_disk, accelerator.is_main_process)
+
+        train_dataset_group.new_cache_latents(vae, accelerator)
+
         vae.to("cpu")
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
-        gc.collect()
+        clean_memory_on_device(accelerator.device)
 
         accelerator.wait_for_everyone()
 
@@ -189,10 +206,13 @@ def train(args):
         else:
             text_encoder.eval()
 
+    text_encoding_strategy = strategy_sd.SdTextEncodingStrategy(args.clip_skip)
+    strategy_base.TextEncodingStrategy.set_strategy(text_encoding_strategy)
+
     if not cache_latents:
         vae.requires_grad_(False)
         vae.eval()
-        vae.to(accelerator.device, dtype=weight_dtype)
+        vae.to(accelerator.device, dtype=vae_dtype)
 
     for m in training_models:
         m.requires_grad_(True)
@@ -211,9 +231,13 @@ def train(args):
     accelerator.print("prepare optimizer, data loader etc.")
     _, _, optimizer = train_util.get_optimizer(args, trainable_params=trainable_params)
 
-    # dataloaderを準備する
-    # DataLoaderのプロセス数：0はメインプロセスになる
-    n_workers = min(args.max_data_loader_n_workers, os.cpu_count() - 1)  # cpu_count-1 ただし最大で指定された数まで
+    # prepare dataloader
+    # strategies are set here because they cannot be referenced in another process. Copy them with the dataset
+    # some strategies can be None
+    train_dataset_group.set_current_strategies()
+
+    # DataLoaderのプロセス数：0 は persistent_workers が使えないので注意
+    n_workers = min(args.max_data_loader_n_workers, os.cpu_count())  # cpu_count or max_data_loader_n_workers
     train_dataloader = torch.utils.data.DataLoader(
         train_dataset_group,
         batch_size=1,
@@ -228,7 +252,9 @@ def train(args):
         args.max_train_steps = args.max_train_epochs * math.ceil(
             len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
         )
-        accelerator.print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}")
+        accelerator.print(
+            f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
+        )
 
     # データセット側にも学習ステップを送信
     train_dataset_group.set_max_train_steps(args.max_train_steps)
@@ -245,13 +271,23 @@ def train(args):
         unet.to(weight_dtype)
         text_encoder.to(weight_dtype)
 
-    # acceleratorがなんかよろしくやってくれるらしい
-    if args.train_text_encoder:
-        unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-            unet, text_encoder, optimizer, train_dataloader, lr_scheduler
+    if args.deepspeed:
+        if args.train_text_encoder:
+            ds_model = deepspeed_utils.prepare_deepspeed_model(args, unet=unet, text_encoder=text_encoder)
+        else:
+            ds_model = deepspeed_utils.prepare_deepspeed_model(args, unet=unet)
+        ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+            ds_model, optimizer, train_dataloader, lr_scheduler
         )
+        training_models = [ds_model]
     else:
-        unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(unet, optimizer, train_dataloader, lr_scheduler)
+        # acceleratorがなんかよろしくやってくれるらしい
+        if args.train_text_encoder:
+            unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+                unet, text_encoder, optimizer, train_dataloader, lr_scheduler
+            )
+        else:
+            unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(unet, optimizer, train_dataloader, lr_scheduler)
 
     # 実験的機能：勾配も含めたfp16学習を行う　PyTorchにパッチを当ててfp16でのgrad scaleを有効にする
     if args.full_fp16:
@@ -292,13 +328,22 @@ def train(args):
     if accelerator.is_main_process:
         init_kwargs = {}
         if args.wandb_run_name:
-            init_kwargs['wandb'] = {'name': args.wandb_run_name}
+            init_kwargs["wandb"] = {"name": args.wandb_run_name}
         if args.log_tracker_config is not None:
             init_kwargs = toml.load(args.log_tracker_config)
-        accelerator.init_trackers("finetuning" if args.log_tracker_name is None else args.log_tracker_name, init_kwargs=init_kwargs)
+        accelerator.init_trackers(
+            "finetuning" if args.log_tracker_name is None else args.log_tracker_name,
+            config=train_util.get_sanitized_config_or_none(args),
+            init_kwargs=init_kwargs,
+        )
 
     # For --sample_at_first
-    train_util.sample_images(accelerator, args, 0, global_step, accelerator.device, vae, tokenizer, text_encoder, unet)
+    train_util.sample_images(
+        accelerator, args, 0, global_step, accelerator.device, vae, tokenize_strategy.tokenizer, text_encoder, unet
+    )
+    if len(accelerator.trackers) > 0:
+        # log empty object to commit the sample images to wandb
+        accelerator.log({}, step=0)
 
     loss_recorder = train_util.LossRecorder()
     for epoch in range(num_train_epochs):
@@ -310,32 +355,30 @@ def train(args):
 
         for step, batch in enumerate(train_dataloader):
             current_step.value = global_step
-            with accelerator.accumulate(training_models[0]):  # 複数モデルに対応していない模様だがとりあえずこうしておく
+            with accelerator.accumulate(*training_models):
                 with torch.no_grad():
                     if "latents" in batch and batch["latents"] is not None:
-                        latents = batch["latents"].to(accelerator.device)  # .to(dtype=weight_dtype)
+                        latents = batch["latents"].to(accelerator.device).to(dtype=weight_dtype)
                     else:
                         # latentに変換
-                        latents = vae.encode(batch["images"].to(dtype=weight_dtype)).latent_dist.sample()
+                        latents = vae.encode(batch["images"].to(dtype=vae_dtype)).latent_dist.sample().to(weight_dtype)
                     latents = latents * 0.18215
                 b_size = latents.shape[0]
 
                 with torch.set_grad_enabled(args.train_text_encoder):
                     # Get the text embedding for conditioning
                     if args.weighted_captions:
-                        encoder_hidden_states = get_weighted_text_embeddings(
-                            tokenizer,
-                            text_encoder,
-                            batch["captions"],
-                            accelerator.device,
-                            args.max_token_length // 75 if args.max_token_length else 1,
-                            clip_skip=args.clip_skip,
-                        )
+                        input_ids_list, weights_list = tokenize_strategy.tokenize_with_weights(batch["captions"])
+                        encoder_hidden_states = text_encoding_strategy.encode_tokens_with_weights(
+                            tokenize_strategy, [text_encoder], input_ids_list, weights_list
+                        )[0]
                     else:
-                        input_ids = batch["input_ids"].to(accelerator.device)
-                        encoder_hidden_states = train_util.get_hidden_states(
-                            args, input_ids, tokenizer, text_encoder, None if not args.full_fp16 else weight_dtype
-                        )
+                        input_ids = batch["input_ids_list"][0].to(accelerator.device)
+                        encoder_hidden_states = text_encoding_strategy.encode_tokens(
+                            tokenize_strategy, [text_encoder], [input_ids]
+                        )[0]
+                    if args.full_fp16:
+                        encoder_hidden_states = encoder_hidden_states.to(weight_dtype)
 
                 # Sample noise, sample a random timestep for each image, and add noise to the latents,
                 # with noise offset and/or multires noise if specified
@@ -351,9 +394,10 @@ def train(args):
                 else:
                     target = noise
 
+                huber_c = train_util.get_huber_threshold_if_needed(args, timesteps, noise_scheduler)
                 if args.min_snr_gamma or args.scale_v_pred_loss_like_noise_pred or args.debiased_estimation_loss:
                     # do not mean over batch dimension for snr weight or scale v-pred loss
-                    loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
+                    loss = train_util.conditional_loss(noise_pred.float(), target.float(), args.loss_type, "none", huber_c)
                     loss = loss.mean([1, 2, 3])
 
                     if args.min_snr_gamma:
@@ -361,11 +405,11 @@ def train(args):
                     if args.scale_v_pred_loss_like_noise_pred:
                         loss = scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler)
                     if args.debiased_estimation_loss:
-                        loss = apply_debiased_estimation(loss, timesteps, noise_scheduler)
+                        loss = apply_debiased_estimation(loss, timesteps, noise_scheduler, args.v_parameterization)
 
                     loss = loss.mean()  # mean over batch dimension
                 else:
-                    loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="mean")
+                    loss = train_util.conditional_loss(noise_pred.float(), target.float(), args.loss_type, "mean", huber_c)
 
                 accelerator.backward(loss)
                 if accelerator.sync_gradients and args.max_grad_norm != 0.0:
@@ -384,7 +428,7 @@ def train(args):
                 global_step += 1
 
                 train_util.sample_images(
-                    accelerator, args, None, global_step, accelerator.device, vae, tokenizer, text_encoder, unet
+                    accelerator, args, None, global_step, accelerator.device, vae, tokenize_strategy.tokenizer, text_encoder, unet
                 )
 
                 # 指定ステップごとにモデルを保存
@@ -409,7 +453,7 @@ def train(args):
                         )
 
             current_loss = loss.detach().item()  # 平均なのでbatch sizeは関係ないはず
-            if args.logging_dir is not None:
+            if len(accelerator.trackers) > 0:
                 logs = {"loss": current_loss}
                 train_util.append_lr_to_logs(logs, lr_scheduler, args.optimizer_type, including_unet=True)
                 accelerator.log(logs, step=global_step)
@@ -422,7 +466,7 @@ def train(args):
             if global_step >= args.max_train_steps:
                 break
 
-        if args.logging_dir is not None:
+        if len(accelerator.trackers) > 0:
             logs = {"loss/epoch": loss_recorder.moving_average}
             accelerator.log(logs, step=epoch + 1)
 
@@ -447,7 +491,9 @@ def train(args):
                     vae,
                 )
 
-        train_util.sample_images(accelerator, args, epoch + 1, global_step, accelerator.device, vae, tokenizer, text_encoder, unet)
+        train_util.sample_images(
+            accelerator, args, epoch + 1, global_step, accelerator.device, vae, tokenize_strategy.tokenizer, text_encoder, unet
+        )
 
     is_main_process = accelerator.is_main_process
     if is_main_process:
@@ -456,7 +502,7 @@ def train(args):
 
     accelerator.end_training()
 
-    if args.save_state and is_main_process:
+    if is_main_process and (args.save_state or args.save_state_on_train_end):
         train_util.save_state_on_train_end(args, accelerator)
 
     del accelerator  # この後メモリを使うのでこれは消す
@@ -466,21 +512,26 @@ def train(args):
         train_util.save_sd_model_on_train_end(
             args, src_path, save_stable_diffusion_format, use_safetensors, save_dtype, epoch, global_step, text_encoder, unet, vae
         )
-        print("model saved.")
+        logger.info("model saved.")
 
 
 def setup_parser() -> argparse.ArgumentParser:
     parser = argparse.ArgumentParser()
 
+    add_logging_arguments(parser)
     train_util.add_sd_models_arguments(parser)
+    sai_model_spec.add_model_spec_arguments(parser)
     train_util.add_dataset_arguments(parser, False, True, True)
     train_util.add_training_arguments(parser, False)
+    deepspeed_utils.add_deepspeed_arguments(parser)
     train_util.add_sd_saving_arguments(parser)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
     custom_train_functions.add_custom_train_arguments(parser)
 
-    parser.add_argument("--diffusers_xformers", action="store_true", help="use xformers by diffusers / Diffusersでxformersを使用する")
+    parser.add_argument(
+        "--diffusers_xformers", action="store_true", help="use xformers by diffusers / Diffusersでxformersを使用する"
+    )
     parser.add_argument("--train_text_encoder", action="store_true", help="train text encoder / text encoderも学習する")
     parser.add_argument(
         "--learning_rate_te",
@@ -488,6 +539,11 @@ def setup_parser() -> argparse.ArgumentParser:
         default=None,
         help="learning rate for text encoder, default is same as unet / Text Encoderの学習率、デフォルトはunetと同じ",
     )
+    parser.add_argument(
+        "--no_half_vae",
+        action="store_true",
+        help="do not use fp16/bf16 VAE in mixed precision (use float VAE) / mixed precisionでも fp16/bf16 VAEを使わずfloat VAEを使う",
+    )
 
     return parser
 
@@ -496,6 +552,7 @@ if __name__ == "__main__":
     parser = setup_parser()
 
     args = parser.parse_args()
+    train_util.verify_command_line_training_args(args)
     args = train_util.read_config_from_file(args, parser)
 
     train(args)

finetune/blip/blip.py

@@ -21,6 +21,10 @@ import torch.nn.functional as F
 import os
 from urllib.parse import urlparse
 from timm.models.hub import download_cached_file
+from library.utils import setup_logging
+setup_logging()
+import logging
+logger = logging.getLogger(__name__)
 
 class BLIP_Base(nn.Module):
     def __init__(self,                 
@@ -130,8 +134,9 @@ class BLIP_Decoder(nn.Module):
     def generate(self, image, sample=False, num_beams=3, max_length=30, min_length=10, top_p=0.9, repetition_penalty=1.0):
         image_embeds = self.visual_encoder(image)
 
-        if not sample:
-            image_embeds = image_embeds.repeat_interleave(num_beams,dim=0)
+        # recent version of transformers seems to do repeat_interleave automatically
+        # if not sample:
+        #     image_embeds = image_embeds.repeat_interleave(num_beams,dim=0)
             
         image_atts = torch.ones(image_embeds.size()[:-1],dtype=torch.long).to(image.device)
         model_kwargs = {"encoder_hidden_states": image_embeds, "encoder_attention_mask":image_atts}
@@ -235,6 +240,6 @@ def load_checkpoint(model,url_or_filename):
                 del state_dict[key]
     
     msg = model.load_state_dict(state_dict,strict=False)
-    print('load checkpoint from %s'%url_or_filename)  
+    logger.info('load checkpoint from %s'%url_or_filename)  
     return model,msg
     

finetune/clean_captions_and_tags.py

@@ -8,6 +8,10 @@ import json
 import re
 
 from tqdm import tqdm
+from library.utils import setup_logging
+setup_logging()
+import logging
+logger = logging.getLogger(__name__)
 
 PATTERN_HAIR_LENGTH = re.compile(r', (long|short|medium) hair, ')
 PATTERN_HAIR_CUT = re.compile(r', (bob|hime) cut, ')
@@ -36,13 +40,13 @@ def clean_tags(image_key, tags):
   tokens = tags.split(", rating")
   if len(tokens) == 1:
     # WD14 taggerのときはこちらになるのでメッセージは出さない
-    # print("no rating:")
-    # print(f"{image_key} {tags}")
+    # logger.info("no rating:")
+    # logger.info(f"{image_key} {tags}")
     pass
   else:
     if len(tokens) > 2:
-      print("multiple ratings:")
-      print(f"{image_key} {tags}")
+      logger.info("multiple ratings:")
+      logger.info(f"{image_key} {tags}")
     tags = tokens[0]
 
   tags = ", " + tags.replace(", ", ", , ") + ", "     # カンマ付きで検索をするための身も蓋もない対策
@@ -124,43 +128,43 @@ def clean_caption(caption):
 
 def main(args):
   if os.path.exists(args.in_json):
-    print(f"loading existing metadata: {args.in_json}")
+    logger.info(f"loading existing metadata: {args.in_json}")
     with open(args.in_json, "rt", encoding='utf-8') as f:
       metadata = json.load(f)
   else:
-    print("no metadata / メタデータファイルがありません")
+    logger.error("no metadata / メタデータファイルがありません")
     return
 
-  print("cleaning captions and tags.")
+  logger.info("cleaning captions and tags.")
   image_keys = list(metadata.keys())
   for image_key in tqdm(image_keys):
     tags = metadata[image_key].get('tags')
     if tags is None:
-      print(f"image does not have tags / メタデータにタグがありません: {image_key}")
+      logger.error(f"image does not have tags / メタデータにタグがありません: {image_key}")
     else:
       org = tags
       tags = clean_tags(image_key, tags)
       metadata[image_key]['tags'] = tags
       if args.debug and org != tags:
-        print("FROM: " + org)
-        print("TO:   " + tags)
+        logger.info("FROM: " + org)
+        logger.info("TO:   " + tags)
 
     caption = metadata[image_key].get('caption')
     if caption is None:
-      print(f"image does not have caption / メタデータにキャプションがありません: {image_key}")
+      logger.error(f"image does not have caption / メタデータにキャプションがありません: {image_key}")
     else:
       org = caption
       caption = clean_caption(caption)
       metadata[image_key]['caption'] = caption
       if args.debug and org != caption:
-        print("FROM: " + org)
-        print("TO:   " + caption)
+        logger.info("FROM: " + org)
+        logger.info("TO:   " + caption)
 
   # metadataを書き出して終わり
-  print(f"writing metadata: {args.out_json}")
+  logger.info(f"writing metadata: {args.out_json}")
   with open(args.out_json, "wt", encoding='utf-8') as f:
     json.dump(metadata, f, indent=2)
-  print("done!")
+  logger.info("done!")
 
 
 def setup_parser() -> argparse.ArgumentParser:
@@ -178,10 +182,10 @@ if __name__ == '__main__':
 
   args, unknown = parser.parse_known_args()
   if len(unknown) == 1:
-    print("WARNING: train_data_dir argument is removed. This script will not work with three arguments in future. Please specify two arguments: in_json and out_json.")
-    print("All captions and tags in the metadata are processed.")
-    print("警告: train_data_dir引数は不要になりました。将来的には三つの引数を指定すると動かなくなる予定です。読み込み元のメタデータと書き出し先の二つの引数だけ指定してください。")
-    print("メタデータ内のすべてのキャプションとタグが処理されます。")
+    logger.warning("WARNING: train_data_dir argument is removed. This script will not work with three arguments in future. Please specify two arguments: in_json and out_json.")
+    logger.warning("All captions and tags in the metadata are processed.")
+    logger.warning("警告: train_data_dir引数は不要になりました。将来的には三つの引数を指定すると動かなくなる予定です。読み込み元のメタデータと書き出し先の二つの引数だけ指定してください。")
+    logger.warning("メタデータ内のすべてのキャプションとタグが処理されます。")
     args.in_json = args.out_json
     args.out_json = unknown[0]
   elif len(unknown) > 0:

finetune/make_captions.py

@@ -9,14 +9,22 @@ from pathlib import Path
 from PIL import Image
 from tqdm import tqdm
 import numpy as np
+
 import torch
+from library.device_utils import init_ipex, get_preferred_device
+init_ipex()
+
 from torchvision import transforms
 from torchvision.transforms.functional import InterpolationMode
 sys.path.append(os.path.dirname(__file__))
 from blip.blip import blip_decoder, is_url
 import library.train_util as train_util
+from library.utils import setup_logging
+setup_logging()
+import logging
+logger = logging.getLogger(__name__)
 
-DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+DEVICE = get_preferred_device()
 
 
 IMAGE_SIZE = 384
@@ -47,7 +55,7 @@ class ImageLoadingTransformDataset(torch.utils.data.Dataset):
             # convert to tensor temporarily so dataloader will accept it
             tensor = IMAGE_TRANSFORM(image)
         except Exception as e:
-            print(f"Could not load image path / 画像を読み込めません: {img_path}, error: {e}")
+            logger.error(f"Could not load image path / 画像を読み込めません: {img_path}, error: {e}")
             return None
 
         return (tensor, img_path)
@@ -74,21 +82,21 @@ def main(args):
         args.train_data_dir = os.path.abspath(args.train_data_dir)  # convert to absolute path
 
         cwd = os.getcwd()
-        print("Current Working Directory is: ", cwd)
+        logger.info(f"Current Working Directory is: {cwd}")
         os.chdir("finetune")
         if not is_url(args.caption_weights) and not os.path.isfile(args.caption_weights):
             args.caption_weights = os.path.join("..", args.caption_weights)
 
-    print(f"load images from {args.train_data_dir}")
+    logger.info(f"load images from {args.train_data_dir}")
     train_data_dir_path = Path(args.train_data_dir)
     image_paths = train_util.glob_images_pathlib(train_data_dir_path, args.recursive)
-    print(f"found {len(image_paths)} images.")
+    logger.info(f"found {len(image_paths)} images.")
 
-    print(f"loading BLIP caption: {args.caption_weights}")
+    logger.info(f"loading BLIP caption: {args.caption_weights}")
     model = blip_decoder(pretrained=args.caption_weights, image_size=IMAGE_SIZE, vit="large", med_config="./blip/med_config.json")
     model.eval()
     model = model.to(DEVICE)
-    print("BLIP loaded")
+    logger.info("BLIP loaded")
 
     # captioningする
     def run_batch(path_imgs):
@@ -108,7 +116,7 @@ def main(args):
             with open(os.path.splitext(image_path)[0] + args.caption_extension, "wt", encoding="utf-8") as f:
                 f.write(caption + "\n")
                 if args.debug:
-                    print(image_path, caption)
+                    logger.info(f'{image_path} {caption}')
 
     # 読み込みの高速化のためにDataLoaderを使うオプション
     if args.max_data_loader_n_workers is not None:
@@ -138,7 +146,7 @@ def main(args):
                         raw_image = raw_image.convert("RGB")
                     img_tensor = IMAGE_TRANSFORM(raw_image)
                 except Exception as e:
-                    print(f"Could not load image path / 画像を読み込めません: {image_path}, error: {e}")
+                    logger.error(f"Could not load image path / 画像を読み込めません: {image_path}, error: {e}")
                     continue
 
             b_imgs.append((image_path, img_tensor))
@@ -148,7 +156,7 @@ def main(args):
     if len(b_imgs) > 0:
         run_batch(b_imgs)
 
-    print("done!")
+    logger.info("done!")
 
 
 def setup_parser() -> argparse.ArgumentParser:

finetune/make_captions_by_git.py

@@ -5,12 +5,19 @@ import re
 from pathlib import Path
 from PIL import Image
 from tqdm import tqdm
+
 import torch
+from library.device_utils import init_ipex, get_preferred_device
+init_ipex()
+
 from transformers import AutoProcessor, AutoModelForCausalLM
 from transformers.generation.utils import GenerationMixin
 
 import library.train_util as train_util
-
+from library.utils import setup_logging
+setup_logging()
+import logging
+logger = logging.getLogger(__name__)
 
 DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
@@ -35,8 +42,8 @@ def remove_words(captions, debug):
         for pat in PATTERN_REPLACE:
             cap = pat.sub("", cap)
         if debug and cap != caption:
-            print(caption)
-            print(cap)
+            logger.info(caption)
+            logger.info(cap)
         removed_caps.append(cap)
     return removed_caps
 
@@ -70,16 +77,16 @@ def main(args):
     GenerationMixin._prepare_input_ids_for_generation = _prepare_input_ids_for_generation_patch
     """
 
-    print(f"load images from {args.train_data_dir}")
+    logger.info(f"load images from {args.train_data_dir}")
     train_data_dir_path = Path(args.train_data_dir)
     image_paths = train_util.glob_images_pathlib(train_data_dir_path, args.recursive)
-    print(f"found {len(image_paths)} images.")
+    logger.info(f"found {len(image_paths)} images.")
 
     # できればcacheに依存せず明示的にダウンロードしたい
-    print(f"loading GIT: {args.model_id}")
+    logger.info(f"loading GIT: {args.model_id}")
     git_processor = AutoProcessor.from_pretrained(args.model_id)
     git_model = AutoModelForCausalLM.from_pretrained(args.model_id).to(DEVICE)
-    print("GIT loaded")
+    logger.info("GIT loaded")
 
     # captioningする
     def run_batch(path_imgs):
@@ -97,7 +104,7 @@ def main(args):
             with open(os.path.splitext(image_path)[0] + args.caption_extension, "wt", encoding="utf-8") as f:
                 f.write(caption + "\n")
                 if args.debug:
-                    print(image_path, caption)
+                    logger.info(f"{image_path} {caption}")
 
     # 読み込みの高速化のためにDataLoaderを使うオプション
     if args.max_data_loader_n_workers is not None:
@@ -126,7 +133,7 @@ def main(args):
                     if image.mode != "RGB":
                         image = image.convert("RGB")
                 except Exception as e:
-                    print(f"Could not load image path / 画像を読み込めません: {image_path}, error: {e}")
+                    logger.error(f"Could not load image path / 画像を読み込めません: {image_path}, error: {e}")
                     continue
 
             b_imgs.append((image_path, image))
@@ -137,7 +144,7 @@ def main(args):
     if len(b_imgs) > 0:
         run_batch(b_imgs)
 
-    print("done!")
+    logger.info("done!")
 
 
 def setup_parser() -> argparse.ArgumentParser:

finetune/merge_captions_to_metadata.py

@@ -5,72 +5,96 @@ from typing import List
 from tqdm import tqdm
 import library.train_util as train_util
 import os
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
 
 def main(args):
-  assert not args.recursive or (args.recursive and args.full_path), "recursive requires full_path / recursiveはfull_pathと同時に指定してください"
+    assert not args.recursive or (
+        args.recursive and args.full_path
+    ), "recursive requires full_path / recursiveはfull_pathと同時に指定してください"
 
-  train_data_dir_path = Path(args.train_data_dir)
-  image_paths: List[Path] = train_util.glob_images_pathlib(train_data_dir_path, args.recursive)
-  print(f"found {len(image_paths)} images.")
+    train_data_dir_path = Path(args.train_data_dir)
+    image_paths: List[Path] = train_util.glob_images_pathlib(train_data_dir_path, args.recursive)
+    logger.info(f"found {len(image_paths)} images.")
 
-  if args.in_json is None and Path(args.out_json).is_file():
-    args.in_json = args.out_json
+    if args.in_json is None and Path(args.out_json).is_file():
+        args.in_json = args.out_json
 
-  if args.in_json is not None:
-    print(f"loading existing metadata: {args.in_json}")
-    metadata = json.loads(Path(args.in_json).read_text(encoding='utf-8'))
-    print("captions for existing images will be overwritten / 既存の画像のキャプションは上書きされます")
-  else:
-    print("new metadata will be created / 新しいメタデータファイルが作成されます")
-    metadata = {}
+    if args.in_json is not None:
+        logger.info(f"loading existing metadata: {args.in_json}")
+        metadata = json.loads(Path(args.in_json).read_text(encoding="utf-8"))
+        logger.warning("captions for existing images will be overwritten / 既存の画像のキャプションは上書きされます")
+    else:
+        logger.info("new metadata will be created / 新しいメタデータファイルが作成されます")
+        metadata = {}
 
-  print("merge caption texts to metadata json.")
-  for image_path in tqdm(image_paths):
-    caption_path = image_path.with_suffix(args.caption_extension)
-    caption = caption_path.read_text(encoding='utf-8').strip()
+    logger.info("merge caption texts to metadata json.")
+    for image_path in tqdm(image_paths):
+        caption_path = image_path.with_suffix(args.caption_extension)
+        caption = caption_path.read_text(encoding="utf-8").strip()
 
-    if not os.path.exists(caption_path):
-      caption_path = os.path.join(image_path, args.caption_extension)
+        if not os.path.exists(caption_path):
+            caption_path = os.path.join(image_path, args.caption_extension)
 
-    image_key = str(image_path) if args.full_path else image_path.stem
-    if image_key not in metadata:
-      metadata[image_key] = {}
+        image_key = str(image_path) if args.full_path else image_path.stem
+        if image_key not in metadata:
+            metadata[image_key] = {}
 
-    metadata[image_key]['caption'] = caption
-    if args.debug:
-      print(image_key, caption)
+        metadata[image_key]["caption"] = caption
+        if args.debug:
+            logger.info(f"{image_key} {caption}")
 
-  # metadataを書き出して終わり
-  print(f"writing metadata: {args.out_json}")
-  Path(args.out_json).write_text(json.dumps(metadata, indent=2), encoding='utf-8')
-  print("done!")
+    # metadataを書き出して終わり
+    logger.info(f"writing metadata: {args.out_json}")
+    Path(args.out_json).write_text(json.dumps(metadata, indent=2), encoding="utf-8")
+    logger.info("done!")
 
 
 def setup_parser() -> argparse.ArgumentParser:
-  parser = argparse.ArgumentParser()
-  parser.add_argument("train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ")
-  parser.add_argument("out_json", type=str, help="metadata file to output / メタデータファイル書き出し先")
-  parser.add_argument("--in_json", type=str,
-                      help="metadata file to input (if omitted and out_json exists, existing out_json is read) / 読み込むメタデータファイル（省略時、out_jsonが存在すればそれを読み込む）")
-  parser.add_argument("--caption_extention", type=str, default=None,
-                      help="extension of caption file (for backward compatibility) / 読み込むキャプションファイルの拡張子（スペルミスしていたのを残してあります）")
-  parser.add_argument("--caption_extension", type=str, default=".caption", help="extension of caption file / 読み込むキャプションファイルの拡張子")
-  parser.add_argument("--full_path", action="store_true",
-                      help="use full path as image-key in metadata (supports multiple directories) / メタデータで画像キーをフルパスにする（複数の学習画像ディレクトリに対応）")
-  parser.add_argument("--recursive", action="store_true",
-                      help="recursively look for training tags in all child folders of train_data_dir / train_data_dirのすべての子フォルダにある学習タグを再帰的に探す")
-  parser.add_argument("--debug", action="store_true", help="debug mode")
+    parser = argparse.ArgumentParser()
+    parser.add_argument("train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ")
+    parser.add_argument("out_json", type=str, help="metadata file to output / メタデータファイル書き出し先")
+    parser.add_argument(
+        "--in_json",
+        type=str,
+        help="metadata file to input (if omitted and out_json exists, existing out_json is read) / 読み込むメタデータファイル（省略時、out_jsonが存在すればそれを読み込む）",
+    )
+    parser.add_argument(
+        "--caption_extention",
+        type=str,
+        default=None,
+        help="extension of caption file (for backward compatibility) / 読み込むキャプションファイルの拡張子（スペルミスしていたのを残してあります）",
+    )
+    parser.add_argument(
+        "--caption_extension", type=str, default=".caption", help="extension of caption file / 読み込むキャプションファイルの拡張子"
+    )
+    parser.add_argument(
+        "--full_path",
+        action="store_true",
+        help="use full path as image-key in metadata (supports multiple directories) / メタデータで画像キーをフルパスにする（複数の学習画像ディレクトリに対応）",
+    )
+    parser.add_argument(
+        "--recursive",
+        action="store_true",
+        help="recursively look for training tags in all child folders of train_data_dir / train_data_dirのすべての子フォルダにある学習タグを再帰的に探す",
+    )
+    parser.add_argument("--debug", action="store_true", help="debug mode")
 
-  return parser
+    return parser
 
 
-if __name__ == '__main__':
-  parser = setup_parser()
+if __name__ == "__main__":
+    parser = setup_parser()
 
-  args = parser.parse_args()
+    args = parser.parse_args()
 
-  # スペルミスしていたオプションを復元する
-  if args.caption_extention is not None:
-    args.caption_extension = args.caption_extention
+    # スペルミスしていたオプションを復元する
+    if args.caption_extention is not None:
+        args.caption_extension = args.caption_extention
 
-  main(args)
+    main(args)

finetune/merge_dd_tags_to_metadata.py

@@ -5,67 +5,89 @@ from typing import List
 from tqdm import tqdm
 import library.train_util as train_util
 import os
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
 
 def main(args):
-  assert not args.recursive or (args.recursive and args.full_path), "recursive requires full_path / recursiveはfull_pathと同時に指定してください"
+    assert not args.recursive or (
+        args.recursive and args.full_path
+    ), "recursive requires full_path / recursiveはfull_pathと同時に指定してください"
 
-  train_data_dir_path = Path(args.train_data_dir)
-  image_paths: List[Path] = train_util.glob_images_pathlib(train_data_dir_path, args.recursive)
-  print(f"found {len(image_paths)} images.")
+    train_data_dir_path = Path(args.train_data_dir)
+    image_paths: List[Path] = train_util.glob_images_pathlib(train_data_dir_path, args.recursive)
+    logger.info(f"found {len(image_paths)} images.")
 
-  if args.in_json is None and Path(args.out_json).is_file():
-    args.in_json = args.out_json
+    if args.in_json is None and Path(args.out_json).is_file():
+        args.in_json = args.out_json
 
-  if args.in_json is not None:
-    print(f"loading existing metadata: {args.in_json}")
-    metadata = json.loads(Path(args.in_json).read_text(encoding='utf-8'))
-    print("tags data for existing images will be overwritten / 既存の画像のタグは上書きされます")
-  else:
-    print("new metadata will be created / 新しいメタデータファイルが作成されます")
-    metadata = {}
+    if args.in_json is not None:
+        logger.info(f"loading existing metadata: {args.in_json}")
+        metadata = json.loads(Path(args.in_json).read_text(encoding="utf-8"))
+        logger.warning("tags data for existing images will be overwritten / 既存の画像のタグは上書きされます")
+    else:
+        logger.info("new metadata will be created / 新しいメタデータファイルが作成されます")
+        metadata = {}
 
-  print("merge tags to metadata json.")
-  for image_path in tqdm(image_paths):
-    tags_path = image_path.with_suffix(args.caption_extension)
-    tags = tags_path.read_text(encoding='utf-8').strip()
+    logger.info("merge tags to metadata json.")
+    for image_path in tqdm(image_paths):
+        tags_path = image_path.with_suffix(args.caption_extension)
+        tags = tags_path.read_text(encoding="utf-8").strip()
 
-    if not os.path.exists(tags_path):
-      tags_path = os.path.join(image_path, args.caption_extension)
+        if not os.path.exists(tags_path):
+            tags_path = os.path.join(image_path, args.caption_extension)
 
-    image_key = str(image_path) if args.full_path else image_path.stem
-    if image_key not in metadata:
-      metadata[image_key] = {}
+        image_key = str(image_path) if args.full_path else image_path.stem
+        if image_key not in metadata:
+            metadata[image_key] = {}
 
-    metadata[image_key]['tags'] = tags
-    if args.debug:
-      print(image_key, tags)
+        metadata[image_key]["tags"] = tags
+        if args.debug:
+            logger.info(f"{image_key} {tags}")
 
-  # metadataを書き出して終わり
-  print(f"writing metadata: {args.out_json}")
-  Path(args.out_json).write_text(json.dumps(metadata, indent=2), encoding='utf-8')
+    # metadataを書き出して終わり
+    logger.info(f"writing metadata: {args.out_json}")
+    Path(args.out_json).write_text(json.dumps(metadata, indent=2), encoding="utf-8")
 
-  print("done!")
+    logger.info("done!")
 
 
 def setup_parser() -> argparse.ArgumentParser:
-  parser = argparse.ArgumentParser()
-  parser.add_argument("train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ")
-  parser.add_argument("out_json", type=str, help="metadata file to output / メタデータファイル書き出し先")
-  parser.add_argument("--in_json", type=str,
-                      help="metadata file to input (if omitted and out_json exists, existing out_json is read) / 読み込むメタデータファイル（省略時、out_jsonが存在すればそれを読み込む）")
-  parser.add_argument("--full_path", action="store_true",
-                      help="use full path as image-key in metadata (supports multiple directories) / メタデータで画像キーをフルパスにする（複数の学習画像ディレクトリに対応）")
-  parser.add_argument("--recursive", action="store_true",
-                      help="recursively look for training tags in all child folders of train_data_dir / train_data_dirのすべての子フォルダにある学習タグを再帰的に探す")
-  parser.add_argument("--caption_extension", type=str, default=".txt",
-                      help="extension of caption (tag) file / 読み込むキャプション（タグ）ファイルの拡張子")
-  parser.add_argument("--debug", action="store_true", help="debug mode, print tags")
+    parser = argparse.ArgumentParser()
+    parser.add_argument("train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ")
+    parser.add_argument("out_json", type=str, help="metadata file to output / メタデータファイル書き出し先")
+    parser.add_argument(
+        "--in_json",
+        type=str,
+        help="metadata file to input (if omitted and out_json exists, existing out_json is read) / 読み込むメタデータファイル（省略時、out_jsonが存在すればそれを読み込む）",
+    )
+    parser.add_argument(
+        "--full_path",
+        action="store_true",
+        help="use full path as image-key in metadata (supports multiple directories) / メタデータで画像キーをフルパスにする（複数の学習画像ディレクトリに対応）",
+    )
+    parser.add_argument(
+        "--recursive",
+        action="store_true",
+        help="recursively look for training tags in all child folders of train_data_dir / train_data_dirのすべての子フォルダにある学習タグを再帰的に探す",
+    )
+    parser.add_argument(
+        "--caption_extension",
+        type=str,
+        default=".txt",
+        help="extension of caption (tag) file / 読み込むキャプション（タグ）ファイルの拡張子",
+    )
+    parser.add_argument("--debug", action="store_true", help="debug mode, print tags")
 
-  return parser
+    return parser
 
 
-if __name__ == '__main__':
-  parser = setup_parser()
+if __name__ == "__main__":
+    parser = setup_parser()
 
-  args = parser.parse_args()
-  main(args)
+    args = parser.parse_args()
+    main(args)

finetune/prepare_buckets_latents.py

@@ -8,13 +8,24 @@ from tqdm import tqdm
 import numpy as np
 from PIL import Image
 import cv2
+
 import torch
+from library.device_utils import init_ipex, get_preferred_device
+
+init_ipex()
+
 from torchvision import transforms
 
 import library.model_util as model_util
 import library.train_util as train_util
+from library.utils import setup_logging
 
-DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+DEVICE = get_preferred_device()
 
 IMAGE_TRANSFORMS = transforms.Compose(
     [
@@ -51,22 +62,22 @@ def get_npz_filename(data_dir, image_key, is_full_path, recursive):
 def main(args):
     # assert args.bucket_reso_steps % 8 == 0, f"bucket_reso_steps must be divisible by 8 / bucket_reso_stepは8で割り切れる必要があります"
     if args.bucket_reso_steps % 8 > 0:
-        print(f"resolution of buckets in training time is a multiple of 8 / 学習時の各bucketの解像度は8単位になります")
+        logger.warning(f"resolution of buckets in training time is a multiple of 8 / 学習時の各bucketの解像度は8単位になります")
     if args.bucket_reso_steps % 32 > 0:
-        print(
+        logger.warning(
             f"WARNING: bucket_reso_steps is not divisible by 32. It is not working with SDXL / bucket_reso_stepsが32で割り切れません。SDXLでは動作しません"
         )
 
     train_data_dir_path = Path(args.train_data_dir)
     image_paths: List[str] = [str(p) for p in train_util.glob_images_pathlib(train_data_dir_path, args.recursive)]
-    print(f"found {len(image_paths)} images.")
+    logger.info(f"found {len(image_paths)} images.")
 
     if os.path.exists(args.in_json):
-        print(f"loading existing metadata: {args.in_json}")
+        logger.info(f"loading existing metadata: {args.in_json}")
         with open(args.in_json, "rt", encoding="utf-8") as f:
             metadata = json.load(f)
     else:
-        print(f"no metadata / メタデータファイルがありません: {args.in_json}")
+        logger.error(f"no metadata / メタデータファイルがありません: {args.in_json}")
         return
 
     weight_dtype = torch.float32
@@ -81,7 +92,9 @@ def main(args):
 
     # bucketのサイズを計算する
     max_reso = tuple([int(t) for t in args.max_resolution.split(",")])
-    assert len(max_reso) == 2, f"illegal resolution (not 'width,height') / 画像サイズに誤りがあります。'幅,高さ'で指定してください: {args.max_resolution}"
+    assert (
+        len(max_reso) == 2
+    ), f"illegal resolution (not 'width,height') / 画像サイズに誤りがあります。'幅,高さ'で指定してください: {args.max_resolution}"
 
     bucket_manager = train_util.BucketManager(
         args.bucket_no_upscale, max_reso, args.min_bucket_reso, args.max_bucket_reso, args.bucket_reso_steps
@@ -89,7 +102,7 @@ def main(args):
     if not args.bucket_no_upscale:
         bucket_manager.make_buckets()
     else:
-        print(
+        logger.warning(
             "min_bucket_reso and max_bucket_reso are ignored if bucket_no_upscale is set, because bucket reso is defined by image size automatically / bucket_no_upscaleが指定された場合は、bucketの解像度は画像サイズから自動計算されるため、min_bucket_resoとmax_bucket_resoは無視されます"
         )
 
@@ -99,7 +112,7 @@ def main(args):
     def process_batch(is_last):
         for bucket in bucket_manager.buckets:
             if (is_last and len(bucket) > 0) or len(bucket) >= args.batch_size:
-                train_util.cache_batch_latents(vae, True, bucket, args.flip_aug, False)
+                train_util.cache_batch_latents(vae, True, bucket, args.flip_aug, args.alpha_mask, False)
                 bucket.clear()
 
     # 読み込みの高速化のためにDataLoaderを使うオプション
@@ -130,7 +143,7 @@ def main(args):
                 if image.mode != "RGB":
                     image = image.convert("RGB")
             except Exception as e:
-                print(f"Could not load image path / 画像を読み込めません: {image_path}, error: {e}")
+                logger.error(f"Could not load image path / 画像を読み込めません: {image_path}, error: {e}")
                 continue
 
         image_key = image_path if args.full_path else os.path.splitext(os.path.basename(image_path))[0]
@@ -183,15 +196,15 @@ def main(args):
     for i, reso in enumerate(bucket_manager.resos):
         count = bucket_counts.get(reso, 0)
         if count > 0:
-            print(f"bucket {i} {reso}: {count}")
+            logger.info(f"bucket {i} {reso}: {count}")
     img_ar_errors = np.array(img_ar_errors)
-    print(f"mean ar error: {np.mean(img_ar_errors)}")
+    logger.info(f"mean ar error: {np.mean(img_ar_errors)}")
 
     # metadataを書き出して終わり
-    print(f"writing metadata: {args.out_json}")
+    logger.info(f"writing metadata: {args.out_json}")
     with open(args.out_json, "wt", encoding="utf-8") as f:
         json.dump(metadata, f, indent=2)
-    print("done!")
+    logger.info("done!")
 
 
 def setup_parser() -> argparse.ArgumentParser:
@@ -200,7 +213,9 @@ def setup_parser() -> argparse.ArgumentParser:
     parser.add_argument("in_json", type=str, help="metadata file to input / 読み込むメタデータファイル")
     parser.add_argument("out_json", type=str, help="metadata file to output / メタデータファイル書き出し先")
     parser.add_argument("model_name_or_path", type=str, help="model name or path to encode latents / latentを取得するためのモデル")
-    parser.add_argument("--v2", action="store_true", help="not used (for backward compatibility) / 使用されません（互換性のため残してあります）")
+    parser.add_argument(
+        "--v2", action="store_true", help="not used (for backward compatibility) / 使用されません（互換性のため残してあります）"
+    )
     parser.add_argument("--batch_size", type=int, default=1, help="batch size in inference / 推論時のバッチサイズ")
     parser.add_argument(
         "--max_data_loader_n_workers",
@@ -223,10 +238,16 @@ def setup_parser() -> argparse.ArgumentParser:
         help="steps of resolution for buckets, divisible by 8 is recommended / bucketの解像度の単位、8で割り切れる値を推奨します",
     )
     parser.add_argument(
-        "--bucket_no_upscale", action="store_true", help="make bucket for each image without upscaling / 画像を拡大せずbucketを作成します"
+        "--bucket_no_upscale",
+        action="store_true",
+        help="make bucket for each image without upscaling / 画像を拡大せずbucketを作成します",
     )
     parser.add_argument(
-        "--mixed_precision", type=str, default="no", choices=["no", "fp16", "bf16"], help="use mixed precision / 混合精度を使う場合、その精度"
+        "--mixed_precision",
+        type=str,
+        default="no",
+        choices=["no", "fp16", "bf16"],
+        help="use mixed precision / 混合精度を使う場合、その精度",
     )
     parser.add_argument(
         "--full_path",
@@ -234,7 +255,15 @@ def setup_parser() -> argparse.ArgumentParser:
         help="use full path as image-key in metadata (supports multiple directories) / メタデータで画像キーをフルパスにする（複数の学習画像ディレクトリに対応）",
     )
     parser.add_argument(
-        "--flip_aug", action="store_true", help="flip augmentation, save latents for flipped images / 左右反転した画像もlatentを取得、保存する"
+        "--flip_aug",
+        action="store_true",
+        help="flip augmentation, save latents for flipped images / 左右反転した画像もlatentを取得、保存する",
+    )
+    parser.add_argument(
+        "--alpha_mask",
+        type=str,
+        default="",
+        help="save alpha mask for images for loss calculation / 損失計算用に画像のアルファマスクを保存する",
     )
     parser.add_argument(
         "--skip_existing",

finetune/tag_images_by_wd14_tagger.py

@@ -1,9 +1,11 @@
 import argparse
 import csv
+import json
+import math
 import os
 from pathlib import Path
+from typing import Optional
 
-import cv2
 import numpy as np
 import torch
 from huggingface_hub import hf_hub_download
@@ -11,6 +13,12 @@ from PIL import Image
 from tqdm import tqdm
 
 import library.train_util as train_util
+from library.utils import setup_logging, resize_image
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
 
 # from wd14 tagger
 IMAGE_SIZE = 448
@@ -23,8 +31,22 @@ SUB_DIR = "variables"
 SUB_DIR_FILES = ["variables.data-00000-of-00001", "variables.index"]
 CSV_FILE = FILES[-1]
 
+TAG_JSON_FILE = "tag_mapping.json"
+
+
+def preprocess_image(image: Image.Image) -> np.ndarray:
+    # If image has transparency, convert to RGBA. If not, convert to RGB
+    if image.mode in ("RGBA", "LA") or "transparency" in image.info:
+        image = image.convert("RGBA")
+    elif image.mode != "RGB":
+        image = image.convert("RGB")
+
+    # If image is RGBA, combine with white background
+    if image.mode == "RGBA":
+        background = Image.new("RGB", image.size, (255, 255, 255))
+        background.paste(image, mask=image.split()[3])  # Use alpha channel as mask
+        image = background
 
-def preprocess_image(image):
     image = np.array(image)
     image = image[:, :, ::-1]  # RGB->BGR
 
@@ -36,75 +58,121 @@ def preprocess_image(image):
     pad_t = pad_y // 2
     image = np.pad(image, ((pad_t, pad_y - pad_t), (pad_l, pad_x - pad_l), (0, 0)), mode="constant", constant_values=255)
 
-    interp = cv2.INTER_AREA if size > IMAGE_SIZE else cv2.INTER_LANCZOS4
-    image = cv2.resize(image, (IMAGE_SIZE, IMAGE_SIZE), interpolation=interp)
+    image = resize_image(image, image.shape[0], image.shape[1], IMAGE_SIZE, IMAGE_SIZE)
 
     image = image.astype(np.float32)
     return image
 
 
 class ImageLoadingPrepDataset(torch.utils.data.Dataset):
-    def __init__(self, image_paths):
-        self.images = image_paths
+    def __init__(self, image_paths: list[str], batch_size: int):
+        self.image_paths = image_paths
+        self.batch_size = batch_size
 
     def __len__(self):
-        return len(self.images)
+        return math.ceil(len(self.image_paths) / self.batch_size)
 
-    def __getitem__(self, idx):
-        img_path = str(self.images[idx])
+    def __getitem__(self, batch_index: int) -> tuple[str, np.ndarray, tuple[int, int]]:
+        image_index_start = batch_index * self.batch_size
+        image_index_end = min((batch_index + 1) * self.batch_size, len(self.image_paths))
 
-        try:
-            image = Image.open(img_path).convert("RGB")
-            image = preprocess_image(image)
-            tensor = torch.tensor(image)
-        except Exception as e:
-            print(f"Could not load image path / 画像を読み込めません: {img_path}, error: {e}")
-            return None
+        batch_image_paths = []
+        images = []
+        image_sizes = []
+        for idx in range(image_index_start, image_index_end):
+            img_path = str(self.image_paths[idx])
 
-        return (tensor, img_path)
+            try:
+                image = Image.open(img_path)
+                image_size = image.size
+                image = preprocess_image(image)
+
+                batch_image_paths.append(img_path)
+                images.append(image)
+                image_sizes.append(image_size)
+            except Exception as e:
+                logger.error(f"Could not load image path / 画像を読み込めません: {img_path}, error: {e}")
+
+        images = np.stack(images) if len(images) > 0 else np.zeros((0, IMAGE_SIZE, IMAGE_SIZE, 3))
+        return batch_image_paths, images, image_sizes
 
 
-def collate_fn_remove_corrupted(batch):
-    """Collate function that allows to remove corrupted examples in the
-    dataloader. It expects that the dataloader returns 'None' when that occurs.
-    The 'None's in the batch are removed.
-    """
-    # Filter out all the Nones (corrupted examples)
-    batch = list(filter(lambda x: x is not None, batch))
+def collate_fn_no_op(batch):
+    """Collate function that does nothing and returns the batch as is."""
     return batch
 
 
 def main(args):
-    # hf_hub_downloadをそのまま使うとsymlink関係で問題があるらしいので、キャッシュディレクトリとforce_filenameを指定してなんとかする
-    # depreacatedの警告が出るけどなくなったらその時
-    # https://github.com/toriato/stable-diffusion-webui-wd14-tagger/issues/22
-    if not os.path.exists(args.model_dir) or args.force_download:
-        print(f"downloading wd14 tagger model from hf_hub. id: {args.repo_id}")
-        files = FILES
-        if args.onnx:
-            files += FILES_ONNX
-        for file in files:
-            hf_hub_download(args.repo_id, file, cache_dir=args.model_dir, force_download=True, force_filename=file)
-        for file in SUB_DIR_FILES:
-            hf_hub_download(
-                args.repo_id,
-                file,
-                subfolder=SUB_DIR,
-                cache_dir=os.path.join(args.model_dir, SUB_DIR),
-                force_download=True,
-                force_filename=file,
-            )
-    else:
-        print("using existing wd14 tagger model")
+    # model location is model_dir + repo_id
+    # given repo_id may be "namespace/repo_name" or "namespace/repo_name/subdir"
+    # so we split it to "namespace/reponame" and "subdir"
+    tokens = args.repo_id.split("/")
 
-    # 画像を読み込む
+    if len(tokens) > 2:
+        repo_id = "/".join(tokens[:2])
+        subdir = "/".join(tokens[2:])
+        model_location = os.path.join(args.model_dir, repo_id.replace("/", "_"), subdir)
+        onnx_model_name = "model_optimized.onnx"
+        default_format = False
+    else:
+        repo_id = args.repo_id
+        subdir = None
+        model_location = os.path.join(args.model_dir, repo_id.replace("/", "_"))
+        onnx_model_name = "model.onnx"
+        default_format = True
+
+    # https://github.com/toriato/stable-diffusion-webui-wd14-tagger/issues/22
+
+    if not os.path.exists(model_location) or args.force_download:
+        os.makedirs(args.model_dir, exist_ok=True)
+        logger.info(f"downloading wd14 tagger model from hf_hub. id: {args.repo_id}")
+
+        if subdir is None:
+            # SmilingWolf structure
+            files = FILES
+            if args.onnx:
+                files = ["selected_tags.csv"]
+                files += FILES_ONNX
+            else:
+                for file in SUB_DIR_FILES:
+                    hf_hub_download(
+                        repo_id=args.repo_id,
+                        filename=file,
+                        subfolder=SUB_DIR,
+                        local_dir=os.path.join(model_location, SUB_DIR),
+                        force_download=True,
+                    )
+
+            for file in files:
+                hf_hub_download(
+                    repo_id=args.repo_id,
+                    filename=file,
+                    local_dir=model_location,
+                    force_download=True,
+                )
+        else:
+            # another structure
+            files = [onnx_model_name, "tag_mapping.json"]
+
+            for file in files:
+                hf_hub_download(
+                    repo_id=repo_id,
+                    filename=file,
+                    subfolder=subdir,
+                    local_dir=os.path.join(args.model_dir, repo_id.replace("/", "_")),  # because subdir is specified
+                    force_download=True,
+                )
+    else:
+        logger.info("using existing wd14 tagger model")
+
+    # モデルを読み込む
     if args.onnx:
         import onnx
         import onnxruntime as ort
 
-        onnx_path = f"{args.model_dir}/model.onnx"
-        print("Running wd14 tagger with onnx")
-        print(f"loading onnx model: {onnx_path}")
+        onnx_path = os.path.join(model_location, onnx_model_name)
+        logger.info("Running wd14 tagger with onnx")
+        logger.info(f"loading onnx model: {onnx_path}")
 
         if not os.path.exists(onnx_path):
             raise Exception(
@@ -116,103 +184,317 @@ def main(args):
         input_name = model.graph.input[0].name
         try:
             batch_size = model.graph.input[0].type.tensor_type.shape.dim[0].dim_value
-        except:
+        except Exception:
             batch_size = model.graph.input[0].type.tensor_type.shape.dim[0].dim_param
 
-        if args.batch_size != batch_size and type(batch_size) != str:
+        if args.batch_size != batch_size and not isinstance(batch_size, str) and batch_size > 0:
             # some rebatch model may use 'N' as dynamic axes
-            print(
+            logger.warning(
                 f"Batch size {args.batch_size} doesn't match onnx model batch size {batch_size}, use model batch size {batch_size}"
             )
             args.batch_size = batch_size
 
         del model
 
-        ort_sess = ort.InferenceSession(
-            onnx_path,
-            providers=["CUDAExecutionProvider"]
-            if "CUDAExecutionProvider" in ort.get_available_providers()
-            else ["CPUExecutionProvider"],
-        )
+        if "OpenVINOExecutionProvider" in ort.get_available_providers():
+            # requires provider options for gpu support
+            # fp16 causes nonsense outputs
+            ort_sess = ort.InferenceSession(
+                onnx_path,
+                providers=(["OpenVINOExecutionProvider"]),
+                provider_options=[{"device_type": "GPU", "precision": "FP32"}],
+            )
+        else:
+            providers = (
+                ["CUDAExecutionProvider"]
+                if "CUDAExecutionProvider" in ort.get_available_providers()
+                else (
+                    ["ROCMExecutionProvider"]
+                    if "ROCMExecutionProvider" in ort.get_available_providers()
+                    else ["CPUExecutionProvider"]
+                )
+            )
+            logger.info(f"Using onnxruntime providers: {providers}")
+            ort_sess = ort.InferenceSession(onnx_path, providers=providers)
     else:
         from tensorflow.keras.models import load_model
 
-        model = load_model(f"{args.model_dir}")
+        model = load_model(f"{model_location}")
 
+    # We read the CSV file manually to avoid adding dependencies.
     # label_names = pd.read_csv("2022_0000_0899_6549/selected_tags.csv")
-    # 依存ライブラリを増やしたくないので自力で読むよ
 
-    with open(os.path.join(args.model_dir, CSV_FILE), "r", encoding="utf-8") as f:
-        reader = csv.reader(f)
-        l = [row for row in reader]
-        header = l[0]  # tag_id,name,category,count
-        rows = l[1:]
-    assert header[0] == "tag_id" and header[1] == "name" and header[2] == "category", f"unexpected csv format: {header}"
+    def expand_character_tags(char_tags):
+        for i, tag in enumerate(char_tags):
+            if tag.endswith(")"):
+                # chara_name_(series) -> chara_name, series
+                # chara_name_(costume)_(series) -> chara_name_(costume), series
+                tags = tag.split("(")
+                character_tag = "(".join(tags[:-1])
+                if character_tag.endswith("_"):
+                    character_tag = character_tag[:-1]
+                series_tag = tags[-1].replace(")", "")
+                char_tags[i] = character_tag + args.caption_separator + series_tag
 
-    general_tags = [row[1] for row in rows[1:] if row[2] == "0"]
-    character_tags = [row[1] for row in rows[1:] if row[2] == "4"]
+    def remove_underscore(tags):
+        return [tag.replace("_", " ") if len(tag) > 3 else tag for tag in tags]
+
+    def process_tag_replacement(tags: list[str], tag_replacements_arg: str) -> list[str]:
+        # escape , and ; in tag_replacement: wd14 tag names may contain , and ;,
+        # so user must be specified them like `aa\,bb,AA\,BB;cc\;dd,CC\;DD` which means
+        # `aa,bb` is replaced with `AA,BB` and `cc;dd` is replaced with `CC;DD`
+        escaped_tag_replacements = tag_replacements_arg.replace("\\,", "@@@@").replace("\\;", "####")
+        tag_replacements = escaped_tag_replacements.split(";")
+
+        for tag_replacements_arg in tag_replacements:
+            tags = tag_replacements_arg.split(",")  # source, target
+            assert (
+                len(tags) == 2
+            ), f"tag replacement must be in the format of `source,target` / タグの置換は `置換元,置換先` の形式で指定してください: {args.tag_replacement}"
+
+            source, target = [tag.replace("@@@@", ",").replace("####", ";") for tag in tags]
+            logger.info(f"replacing tag: {source} -> {target}")
+
+            if source in tags:
+                tags[tags.index(source)] = target
+
+        return tags
+
+    if default_format:
+        with open(os.path.join(model_location, CSV_FILE), "r", encoding="utf-8") as f:
+            reader = csv.reader(f)
+            line = [row for row in reader]
+            header = line[0]  # tag_id,name,category,count
+            rows = line[1:]
+        assert header[0] == "tag_id" and header[1] == "name" and header[2] == "category", f"unexpected csv format: {header}"
+
+        rating_tags = [row[1] for row in rows[0:] if row[2] == "9"]
+        general_tags = [row[1] for row in rows[0:] if row[2] == "0"]
+        character_tags = [row[1] for row in rows[0:] if row[2] == "4"]
+
+        if args.character_tag_expand:
+            expand_character_tags(character_tags)
+        if args.remove_underscore:
+            rating_tags = remove_underscore(rating_tags)
+            character_tags = remove_underscore(character_tags)
+            general_tags = remove_underscore(general_tags)
+        if args.tag_replacement is not None:
+            process_tag_replacement(rating_tags, args.tag_replacement)
+            process_tag_replacement(general_tags, args.tag_replacement)
+            process_tag_replacement(character_tags, args.tag_replacement)
+    else:
+        with open(os.path.join(model_location, TAG_JSON_FILE), "r", encoding="utf-8") as f:
+            tag_mapping = json.load(f)
+
+        rating_tags = []
+        general_tags = []
+        character_tags = []
+
+        tag_id_to_tag_mapping = {}
+        tag_id_to_category_mapping = {}
+        for tag_id, tag_info in tag_mapping.items():
+            tag = tag_info["tag"]
+            category = tag_info["category"]
+            assert category in [
+                "Rating",
+                "General",
+                "Character",
+                "Copyright",
+                "Meta",
+                "Model",
+                "Quality",
+                "Artist",
+            ], f"unexpected category: {category}"
+
+            if args.remove_underscore:
+                tag = remove_underscore([tag])[0]
+            if args.tag_replacement is not None:
+                tag = process_tag_replacement([tag], args.tag_replacement)[0]
+            if category == "Character" and args.character_tag_expand:
+                tag_list = [tag]
+                expand_character_tags(tag_list)
+                tag = tag_list[0]
+
+            tag_id_to_tag_mapping[int(tag_id)] = tag
+            tag_id_to_category_mapping[int(tag_id)] = category
 
     # 画像を読み込む
-
     train_data_dir_path = Path(args.train_data_dir)
     image_paths = train_util.glob_images_pathlib(train_data_dir_path, args.recursive)
-    print(f"found {len(image_paths)} images.")
+    logger.info(f"found {len(image_paths)} images.")
+    image_paths = [str(ip) for ip in image_paths]
 
     tag_freq = {}
 
     caption_separator = args.caption_separator
     stripped_caption_separator = caption_separator.strip()
-    undesired_tags = set(args.undesired_tags.split(stripped_caption_separator))
+    undesired_tags = args.undesired_tags.split(stripped_caption_separator)
+    undesired_tags = set([tag.strip() for tag in undesired_tags if tag.strip() != ""])
 
-    def run_batch(path_imgs):
-        imgs = np.array([im for _, im in path_imgs])
+    always_first_tags = None
+    if args.always_first_tags is not None:
+        always_first_tags = [tag for tag in args.always_first_tags.split(stripped_caption_separator) if tag.strip() != ""]
+
+    def run_batch(path_imgs: tuple[list[str], np.ndarray, list[tuple[int, int]]]) -> Optional[dict[str, dict]]:
+        nonlocal args, default_format, model, ort_sess, input_name, tag_freq
+
+        imgs = path_imgs[1]
+        result = {}
 
         if args.onnx:
-            if len(imgs) < args.batch_size:
-                imgs = np.concatenate([imgs, np.zeros((args.batch_size - len(imgs), IMAGE_SIZE, IMAGE_SIZE, 3))], axis=0)
+            # if len(imgs) < args.batch_size:
+            #     imgs = np.concatenate([imgs, np.zeros((args.batch_size - len(imgs), IMAGE_SIZE, IMAGE_SIZE, 3))], axis=0)
+            if not default_format:
+                imgs = imgs.transpose(0, 3, 1, 2)  # to NCHW
+                imgs = imgs / 127.5 - 1.0
             probs = ort_sess.run(None, {input_name: imgs})[0]  # onnx output numpy
-            probs = probs[: len(path_imgs)]
+            probs = probs[: len(imgs)]  # remove padding
         else:
             probs = model(imgs, training=False)
             probs = probs.numpy()
 
-        for (image_path, _), prob in zip(path_imgs, probs):
-            # 最初の4つはratingなので無視する
-            # # First 4 labels are actually ratings: pick one with argmax
-            # ratings_names = label_names[:4]
-            # rating_index = ratings_names["probs"].argmax()
-            # found_rating = ratings_names[rating_index: rating_index + 1][["name", "probs"]]
-
-            # それ以降はタグなのでconfidenceがthresholdより高いものを追加する
-            # Everything else is tags: pick any where prediction confidence > threshold
+        for image_path, image_size, prob in zip(path_imgs[0], path_imgs[2], probs):
             combined_tags = []
-            general_tag_text = ""
+            rating_tag_text = ""
             character_tag_text = ""
-            for i, p in enumerate(prob[4:]):
-                if i < len(general_tags) and p >= args.general_threshold:
-                    tag_name = general_tags[i]
-                    if args.remove_underscore and len(tag_name) > 3:  # ignore emoji tags like >_< and ^_^
-                        tag_name = tag_name.replace("_", " ")
+            general_tag_text = ""
+            other_tag_text = ""
 
-                    if tag_name not in undesired_tags:
+            if default_format:
+                # 最初の4つ以降はタグなのでconfidenceがthreshold以上のものを追加する
+                # First 4 labels are ratings, the rest are tags: pick any where prediction confidence >= threshold
+                for i, p in enumerate(prob[4:]):
+                    if i < len(general_tags) and p >= args.general_threshold:
+                        tag_name = general_tags[i]
+
+                        if tag_name not in undesired_tags:
+                            tag_freq[tag_name] = tag_freq.get(tag_name, 0) + 1
+                            general_tag_text += caption_separator + tag_name
+                            combined_tags.append(tag_name)
+                    elif i >= len(general_tags) and p >= args.character_threshold:
+                        tag_name = character_tags[i - len(general_tags)]
+
+                        if tag_name not in undesired_tags:
+                            tag_freq[tag_name] = tag_freq.get(tag_name, 0) + 1
+                            character_tag_text += caption_separator + tag_name
+                            if args.character_tags_first:  # insert to the beginning
+                                combined_tags.insert(0, tag_name)
+                            else:
+                                combined_tags.append(tag_name)
+
+                # 最初の4つはratingなのでargmaxで選ぶ
+                # First 4 labels are actually ratings: pick one with argmax
+                if args.use_rating_tags or args.use_rating_tags_as_last_tag:
+                    ratings_probs = prob[:4]
+                    rating_index = ratings_probs.argmax()
+                    found_rating = rating_tags[rating_index]
+
+                    if found_rating not in undesired_tags:
+                        tag_freq[found_rating] = tag_freq.get(found_rating, 0) + 1
+                        rating_tag_text = found_rating
+                        if args.use_rating_tags:
+                            combined_tags.insert(0, found_rating)  # insert to the beginning
+                        else:
+                            combined_tags.append(found_rating)
+            else:
+                # apply sigmoid to probabilities
+                prob = 1 / (1 + np.exp(-prob))
+
+                rating_max_prob = -1
+                rating_tag = None
+                quality_max_prob = -1
+                quality_tag = None
+                character_tags = []
+
+                min_thres = min(
+                    args.thresh,
+                    args.general_threshold,
+                    args.character_threshold,
+                    args.copyright_threshold,
+                    args.meta_threshold,
+                    args.model_threshold,
+                    args.artist_threshold,
+                )
+                prob_indices = np.where(prob >= min_thres)[0]
+                # for i, p in enumerate(prob):
+                for i in prob_indices:
+                    if i not in tag_id_to_tag_mapping:
+                        continue
+                    p = prob[i]
+
+                    tag_name = tag_id_to_tag_mapping[i]
+                    category = tag_id_to_category_mapping[i]
+                    if tag_name in undesired_tags:
+                        continue
+
+                    if category == "Rating":
+                        if p > rating_max_prob:
+                            rating_max_prob = p
+                            rating_tag = tag_name
+                            rating_tag_text = tag_name
+                        continue
+                    elif category == "Quality":
+                        if p > quality_max_prob:
+                            quality_max_prob = p
+                            quality_tag = tag_name
+                            if args.use_quality_tags or args.use_quality_tags_as_last_tag:
+                                other_tag_text += caption_separator + tag_name
+                        continue
+
+                    if category == "General" and p >= args.general_threshold:
                         tag_freq[tag_name] = tag_freq.get(tag_name, 0) + 1
                         general_tag_text += caption_separator + tag_name
-                        combined_tags.append(tag_name)
-                elif i >= len(general_tags) and p >= args.character_threshold:
-                    tag_name = character_tags[i - len(general_tags)]
-                    if args.remove_underscore and len(tag_name) > 3:
-                        tag_name = tag_name.replace("_", " ")
-
-                    if tag_name not in undesired_tags:
+                        combined_tags.append((tag_name, p))
+                    elif category == "Character" and p >= args.character_threshold:
                         tag_freq[tag_name] = tag_freq.get(tag_name, 0) + 1
                         character_tag_text += caption_separator + tag_name
-                        combined_tags.append(tag_name)
+                        if args.character_tags_first:  # we separate character tags
+                            character_tags.append((tag_name, p))
+                        else:
+                            combined_tags.append((tag_name, p))
+                    elif (
+                        (category == "Copyright" and p >= args.copyright_threshold)
+                        or (category == "Meta" and p >= args.meta_threshold)
+                        or (category == "Model" and p >= args.model_threshold)
+                        or (category == "Artist" and p >= args.artist_threshold)
+                    ):
+                        tag_freq[tag_name] = tag_freq.get(tag_name, 0) + 1
+                        other_tag_text += f"{caption_separator}{tag_name} ({category})"
+                        combined_tags.append((tag_name, p))
+
+                # sort by probability
+                combined_tags.sort(key=lambda x: x[1], reverse=True)
+                if character_tags:
+                    character_tags.sort(key=lambda x: x[1], reverse=True)
+                    combined_tags = character_tags + combined_tags
+                combined_tags = [t[0] for t in combined_tags]  # remove probability
+
+                if quality_tag is not None:
+                    if args.use_quality_tags_as_last_tag:
+                        combined_tags.append(quality_tag)
+                    elif args.use_quality_tags:
+                        combined_tags.insert(0, quality_tag)
+                if rating_tag is not None:
+                    if args.use_rating_tags_as_last_tag:
+                        combined_tags.append(rating_tag)
+                    elif args.use_rating_tags:
+                        combined_tags.insert(0, rating_tag)
+
+            # 一番最初に置くタグを指定する
+            # Always put some tags at the beginning
+            if always_first_tags is not None:
+                for tag in always_first_tags:
+                    if tag in combined_tags:
+                        combined_tags.remove(tag)
+                        combined_tags.insert(0, tag)
 
             # 先頭のカンマを取る
             if len(general_tag_text) > 0:
                 general_tag_text = general_tag_text[len(caption_separator) :]
             if len(character_tag_text) > 0:
                 character_tag_text = character_tag_text[len(caption_separator) :]
+            if len(other_tag_text) > 0:
+                other_tag_text = other_tag_text[len(caption_separator) :]
 
             caption_file = os.path.splitext(image_path)[0] + args.caption_extension
 
@@ -234,61 +516,87 @@ def main(args):
                     # Create new tag_text
                     tag_text = caption_separator.join(existing_tags + new_tags)
 
-            with open(caption_file, "wt", encoding="utf-8") as f:
-                f.write(tag_text + "\n")
-                if args.debug:
-                    print(f"\n{image_path}:\n  Character tags: {character_tag_text}\n  General tags: {general_tag_text}")
+            if not args.output_path:
+                with open(caption_file, "wt", encoding="utf-8") as f:
+                    f.write(tag_text + "\n")
+            else:
+                entry = {"tags": tag_text, "image_size": list(image_size)}
+                result[image_path] = entry
+
+            if args.debug:
+                logger.info("")
+                logger.info(f"{image_path}:")
+                logger.info(f"\tRating tags: {rating_tag_text}")
+                logger.info(f"\tCharacter tags: {character_tag_text}")
+                logger.info(f"\tGeneral tags: {general_tag_text}")
+                if other_tag_text:
+                    logger.info(f"\tOther tags: {other_tag_text}")
+
+        return result
 
     # 読み込みの高速化のためにDataLoaderを使うオプション
     if args.max_data_loader_n_workers is not None:
-        dataset = ImageLoadingPrepDataset(image_paths)
+        dataset = ImageLoadingPrepDataset(image_paths, args.batch_size)
         data = torch.utils.data.DataLoader(
             dataset,
-            batch_size=args.batch_size,
+            batch_size=1,
             shuffle=False,
             num_workers=args.max_data_loader_n_workers,
-            collate_fn=collate_fn_remove_corrupted,
+            collate_fn=collate_fn_no_op,
             drop_last=False,
         )
     else:
-        data = [[(None, ip)] for ip in image_paths]
+        # data = [[(ip, None, None)] for ip in image_paths]
+        data = [[]]
+        for ip in image_paths:
+            if len(data[-1]) >= args.batch_size:
+                data.append([])
+            data[-1].append((ip, None, None))
 
-    b_imgs = []
+    results = {}
     for data_entry in tqdm(data, smoothing=0.0):
-        for data in data_entry:
-            if data is None:
-                continue
+        if data_entry is None or len(data_entry) == 0:
+            continue
 
-            image, image_path = data
-            if image is not None:
-                image = image.detach().numpy()
-            else:
-                try:
-                    image = Image.open(image_path)
-                    if image.mode != "RGB":
-                        image = image.convert("RGB")
-                    image = preprocess_image(image)
-                except Exception as e:
-                    print(f"Could not load image path / 画像を読み込めません: {image_path}, error: {e}")
-                    continue
-            b_imgs.append((image_path, image))
+        if data_entry[0][1] is None:
+            # No preloaded image, need to load
+            images = []
+            image_sizes = []
+            for image_path, _, _ in data_entry:
+                image = Image.open(image_path)
+                image_size = image.size
+                image = preprocess_image(image)
+                images.append(image)
+                image_sizes.append(image_size)
+            b_imgs = ([ip for ip, _, _ in data_entry], np.stack(images), image_sizes)
+        else:
+            b_imgs = data_entry[0]
 
-            if len(b_imgs) >= args.batch_size:
-                b_imgs = [(str(image_path), image) for image_path, image in b_imgs]  # Convert image_path to string
-                run_batch(b_imgs)
-                b_imgs.clear()
+        r = run_batch(b_imgs)
+        if args.output_path and r is not None:
+            results.update(r)
 
-    if len(b_imgs) > 0:
-        b_imgs = [(str(image_path), image) for image_path, image in b_imgs]  # Convert image_path to string
-        run_batch(b_imgs)
+    if args.output_path:
+        if args.output_path.endswith(".jsonl"):
+            # optional JSONL metadata
+            with open(args.output_path, "wt", encoding="utf-8") as f:
+                for image_path, entry in results.items():
+                    f.write(
+                        json.dumps({"image_path": image_path, "caption": entry["tags"], "image_size": entry["image_size"]}) + "\n"
+                    )
+        else:
+            # standard JSON metadata
+            with open(args.output_path, "wt", encoding="utf-8") as f:
+                json.dump(results, f, ensure_ascii=False, indent=4)
+            logger.info(f"captions saved to {args.output_path}")
 
     if args.frequency_tags:
         sorted_tags = sorted(tag_freq.items(), key=lambda x: x[1], reverse=True)
-        print("\nTag frequencies:")
+        print("Tag frequencies:")
         for tag, freq in sorted_tags:
             print(f"{tag}: {freq}")
 
-    print("done!")
+    logger.info("done!")
 
 
 def setup_parser() -> argparse.ArgumentParser:
@@ -307,7 +615,9 @@ def setup_parser() -> argparse.ArgumentParser:
         help="directory to store wd14 tagger model / wd14 taggerのモデルを格納するディレクトリ",
     )
     parser.add_argument(
-        "--force_download", action="store_true", help="force downloading wd14 tagger models / wd14 taggerのモデルを再ダウンロードします"
+        "--force_download",
+        action="store_true",
+        help="force downloading wd14 tagger models / wd14 taggerのモデルを再ダウンロードします",
     )
     parser.add_argument("--batch_size", type=int, default=1, help="batch size in inference / 推論時のバッチサイズ")
     parser.add_argument(
@@ -316,14 +626,24 @@ def setup_parser() -> argparse.ArgumentParser:
         default=None,
         help="enable image reading by DataLoader with this number of workers (faster) / DataLoaderによる画像読み込みを有効にしてこのワーカー数を適用する（読み込みを高速化）",
     )
+    parser.add_argument(
+        "--output_path",
+        type=str,
+        default=None,
+        help="path for output captions (json format). if this is set, captions will be saved to this file / 出力キャプションのパス（json形式）。このオプションが設定されている場合、キャプションはこのファイルに保存されます",
+    )
     parser.add_argument(
         "--caption_extention",
         type=str,
         default=None,
         help="extension of caption file (for backward compatibility) / 出力されるキャプションファイルの拡張子（スペルミスしていたのを残してあります）",
     )
-    parser.add_argument("--caption_extension", type=str, default=".txt", help="extension of caption file / 出力されるキャプションファイルの拡張子")
-    parser.add_argument("--thresh", type=float, default=0.35, help="threshold of confidence to add a tag / タグを追加するか判定する閾値")
+    parser.add_argument(
+        "--caption_extension", type=str, default=".txt", help="extension of caption file / 出力されるキャプションファイルの拡張子"
+    )
+    parser.add_argument(
+        "--thresh", type=float, default=0.35, help="threshold of confidence to add a tag / タグを追加するか判定する閾値"
+    )
     parser.add_argument(
         "--general_threshold",
         type=float,
@@ -334,9 +654,40 @@ def setup_parser() -> argparse.ArgumentParser:
         "--character_threshold",
         type=float,
         default=None,
-        help="threshold of confidence to add a tag for character category, same as --thres if omitted / characterカテゴリのタグを追加するための確信度の閾値、省略時は --thresh と同じ",
+        help="threshold of confidence to add a tag for character category, same as --thres if omitted. set above 1 to disable character tags"
+        " / characterカテゴリのタグを追加するための確信度の閾値、省略時は --thresh と同じ。1以上にするとcharacterタグを無効化できる",
+    )
+    parser.add_argument(
+        "--meta_threshold",
+        type=float,
+        default=None,
+        help="threshold of confidence to add a tag for meta category, same as --thresh if omitted. set above 1 to disable meta tags"
+        " / metaカテゴリのタグを追加するための確信度の閾値、省略時は --thresh と同じ。1以上にするとmetaタグを無効化できる",
+    )
+    parser.add_argument(
+        "--model_threshold",
+        type=float,
+        default=None,
+        help="threshold of confidence to add a tag for model category, same as --thresh if omitted. set above 1 to disable model tags"
+        " / modelカテゴリのタグを追加するための確信度の閾値、省略時は --thresh と同じ。1以上にするとmodelタグを無効化できる",
+    )
+    parser.add_argument(
+        "--copyright_threshold",
+        type=float,
+        default=None,
+        help="threshold of confidence to add a tag for copyright category, same as --thresh if omitted. set above 1 to disable copyright tags"
+        " / copyrightカテゴリのタグを追加するための確信度の閾値、省略時は --thresh と同じ。1以上にするとcopyrightタグを無効化できる",
+    )
+    parser.add_argument(
+        "--artist_threshold",
+        type=float,
+        default=None,
+        help="threshold of confidence to add a tag for artist category, same as --thresh if omitted. set above 1 to disable artist tags"
+        " / artistカテゴリのタグを追加するための確信度の閾値、省略時は --thresh と同じ。1以上にするとartistタグを無効化できる",
+    )
+    parser.add_argument(
+        "--recursive", action="store_true", help="search for images in subfolders recursively / サブフォルダを再帰的に検索する"
     )
-    parser.add_argument("--recursive", action="store_true", help="search for images in subfolders recursively / サブフォルダを再帰的に検索する")
     parser.add_argument(
         "--remove_underscore",
         action="store_true",
@@ -349,15 +700,64 @@ def setup_parser() -> argparse.ArgumentParser:
         default="",
         help="comma-separated list of undesired tags to remove from the output / 出力から除外したいタグのカンマ区切りのリスト",
     )
-    parser.add_argument("--frequency_tags", action="store_true", help="Show frequency of tags for images / 画像ごとのタグの出現頻度を表示する")
+    parser.add_argument(
+        "--frequency_tags", action="store_true", help="Show frequency of tags for images / タグの出現頻度を表示する"
+    )
     parser.add_argument("--onnx", action="store_true", help="use onnx model for inference / onnxモデルを推論に使用する")
-    parser.add_argument("--append_tags", action="store_true", help="Append captions instead of overwriting / 上書きではなくキャプションを追記する")
+    parser.add_argument(
+        "--append_tags", action="store_true", help="Append captions instead of overwriting / 上書きではなくキャプションを追記する"
+    )
+    parser.add_argument(
+        "--use_rating_tags",
+        action="store_true",
+        help="Adds rating tags as the first tag / レーティングタグを最初のタグとして追加する",
+    )
+    parser.add_argument(
+        "--use_rating_tags_as_last_tag",
+        action="store_true",
+        help="Adds rating tags as the last tag / レーティングタグを最後のタグとして追加する",
+    )
+    parser.add_argument(
+        "--use_quality_tags",
+        action="store_true",
+        help="Adds quality tags as the first tag / クオリティタグを最初のタグとして追加する",
+    )
+    parser.add_argument(
+        "--use_quality_tags_as_last_tag",
+        action="store_true",
+        help="Adds quality tags as the last tag / クオリティタグを最後のタグとして追加する",
+    )
+    parser.add_argument(
+        "--character_tags_first",
+        action="store_true",
+        help="Always inserts character tags before the general tags / characterタグを常にgeneralタグの前に出力する",
+    )
+    parser.add_argument(
+        "--always_first_tags",
+        type=str,
+        default=None,
+        help="comma-separated list of tags to always put at the beginning, e.g. `1girl,1boy`"
+        + " / 必ず先頭に置くタグのカンマ区切りリスト、例 : `1girl,1boy`",
+    )
     parser.add_argument(
         "--caption_separator",
         type=str,
         default=", ",
         help="Separator for captions, include space if needed / キャプションの区切り文字、必要ならスペースを含めてください",
     )
+    parser.add_argument(
+        "--tag_replacement",
+        type=str,
+        default=None,
+        help="tag replacement in the format of `source1,target1;source2,target2; ...`. Escape `,` and `;` with `\`. e.g. `tag1,tag2;tag3,tag4`"
+        + " / タグの置換を `置換元1,置換先1;置換元2,置換先2; ...`で指定する。`\` で `,` と `;` をエスケープできる。例: `tag1,tag2;tag3,tag4`",
+    )
+    parser.add_argument(
+        "--character_tag_expand",
+        action="store_true",
+        help="expand tag tail parenthesis to another tag for character tags. `chara_name_(series)` becomes `chara_name, series`"
+        + " / キャラクタタグの末尾の括弧を別のタグに展開する。`chara_name_(series)` は `chara_name, series` になる",
+    )
 
     return parser
 
@@ -375,5 +775,13 @@ if __name__ == "__main__":
         args.general_threshold = args.thresh
     if args.character_threshold is None:
         args.character_threshold = args.thresh
+    if args.meta_threshold is None:
+        args.meta_threshold = args.thresh
+    if args.model_threshold is None:
+        args.model_threshold = args.thresh
+    if args.copyright_threshold is None:
+        args.copyright_threshold = args.thresh
+    if args.artist_threshold is None:
+        args.artist_threshold = args.thresh
 
     main(args)

flux_minimal_inference.py

@@ -0,0 +1,596 @@
+# Minimum Inference Code for FLUX
+
+import argparse
+import datetime
+import math
+import os
+import random
+from typing import Callable, List, Optional
+import einops
+import numpy as np
+
+import torch
+from tqdm import tqdm
+from PIL import Image
+import accelerate
+from transformers import CLIPTextModel
+from safetensors.torch import load_file
+
+from library import device_utils
+from library.device_utils import init_ipex, get_preferred_device
+from networks import oft_flux
+
+init_ipex()
+
+
+from library.utils import setup_logging, str_to_dtype
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+import networks.lora_flux as lora_flux
+from library import flux_models, flux_utils, sd3_utils, strategy_flux
+
+
+def time_shift(mu: float, sigma: float, t: torch.Tensor):
+    return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
+
+
+def get_lin_function(x1: float = 256, y1: float = 0.5, x2: float = 4096, y2: float = 1.15) -> Callable[[float], float]:
+    m = (y2 - y1) / (x2 - x1)
+    b = y1 - m * x1
+    return lambda x: m * x + b
+
+
+def get_schedule(
+    num_steps: int,
+    image_seq_len: int,
+    base_shift: float = 0.5,
+    max_shift: float = 1.15,
+    shift: bool = True,
+) -> list[float]:
+    # extra step for zero
+    timesteps = torch.linspace(1, 0, num_steps + 1)
+
+    # shifting the schedule to favor high timesteps for higher signal images
+    if shift:
+        # eastimate mu based on linear estimation between two points
+        mu = get_lin_function(y1=base_shift, y2=max_shift)(image_seq_len)
+        timesteps = time_shift(mu, 1.0, timesteps)
+
+    return timesteps.tolist()
+
+
+def denoise(
+    model: flux_models.Flux,
+    img: torch.Tensor,
+    img_ids: torch.Tensor,
+    txt: torch.Tensor,
+    txt_ids: torch.Tensor,
+    vec: torch.Tensor,
+    timesteps: list[float],
+    guidance: float = 4.0,
+    t5_attn_mask: Optional[torch.Tensor] = None,
+    neg_txt: Optional[torch.Tensor] = None,
+    neg_vec: Optional[torch.Tensor] = None,
+    neg_t5_attn_mask: Optional[torch.Tensor] = None,
+    cfg_scale: Optional[float] = None,
+):
+    # prepare classifier free guidance
+    logger.info(f"guidance: {guidance}, cfg_scale: {cfg_scale}")
+    do_cfg = neg_txt is not None and (cfg_scale is not None and cfg_scale != 1.0)
+
+    # this is ignored for schnell
+    guidance_vec = torch.full((img.shape[0] * (2 if do_cfg else 1),), guidance, device=img.device, dtype=img.dtype)
+
+    if do_cfg:
+        print("Using classifier free guidance")
+        b_img_ids = torch.cat([img_ids, img_ids], dim=0)
+        b_txt_ids = torch.cat([txt_ids, txt_ids], dim=0)
+        b_txt = torch.cat([neg_txt, txt], dim=0)
+        b_vec = torch.cat([neg_vec, vec], dim=0) if neg_vec is not None else None
+        if t5_attn_mask is not None and neg_t5_attn_mask is not None:
+            b_t5_attn_mask = torch.cat([neg_t5_attn_mask, t5_attn_mask], dim=0)
+        else:
+            b_t5_attn_mask = None
+    else:
+        b_img_ids = img_ids
+        b_txt_ids = txt_ids
+        b_txt = txt
+        b_vec = vec
+        b_t5_attn_mask = t5_attn_mask
+
+    for t_curr, t_prev in zip(tqdm(timesteps[:-1]), timesteps[1:]):
+        t_vec = torch.full((b_img_ids.shape[0],), t_curr, dtype=img.dtype, device=img.device)
+
+        # classifier free guidance
+        if do_cfg:
+            b_img = torch.cat([img, img], dim=0)
+        else:
+            b_img = img
+
+        y_input = b_vec
+
+        mod_vectors = model.get_mod_vectors(timesteps=t_vec, guidance=guidance_vec, batch_size=b_img.shape[0])
+
+        pred = model(
+            img=b_img,
+            img_ids=b_img_ids,
+            txt=b_txt,
+            txt_ids=b_txt_ids,
+            y=y_input,
+            timesteps=t_vec,
+            guidance=guidance_vec,
+            txt_attention_mask=b_t5_attn_mask,
+            mod_vectors=mod_vectors,
+        )
+
+        # classifier free guidance
+        if do_cfg:
+            pred_uncond, pred = torch.chunk(pred, 2, dim=0)
+            pred = pred_uncond + cfg_scale * (pred - pred_uncond)
+
+        img = img + (t_prev - t_curr) * pred
+
+    return img
+
+
+def do_sample(
+    accelerator: Optional[accelerate.Accelerator],
+    model: flux_models.Flux,
+    img: torch.Tensor,
+    img_ids: torch.Tensor,
+    l_pooled: Optional[torch.Tensor],
+    t5_out: torch.Tensor,
+    txt_ids: torch.Tensor,
+    num_steps: int,
+    guidance: float,
+    t5_attn_mask: Optional[torch.Tensor],
+    is_schnell: bool,
+    device: torch.device,
+    flux_dtype: torch.dtype,
+    neg_l_pooled: Optional[torch.Tensor] = None,
+    neg_t5_out: Optional[torch.Tensor] = None,
+    neg_t5_attn_mask: Optional[torch.Tensor] = None,
+    cfg_scale: Optional[float] = None,
+):
+    logger.info(f"num_steps: {num_steps}")
+    timesteps = get_schedule(num_steps, img.shape[1], shift=not is_schnell)
+
+    # denoise initial noise
+    if accelerator:
+        with accelerator.autocast(), torch.no_grad():
+            x = denoise(
+                model,
+                img,
+                img_ids,
+                t5_out,
+                txt_ids,
+                l_pooled,
+                timesteps,
+                guidance,
+                t5_attn_mask,
+                neg_t5_out,
+                neg_l_pooled,
+                neg_t5_attn_mask,
+                cfg_scale,
+            )
+    else:
+        with torch.autocast(device_type=device.type, dtype=flux_dtype), torch.no_grad():
+            x = denoise(
+                model,
+                img,
+                img_ids,
+                t5_out,
+                txt_ids,
+                l_pooled,
+                timesteps,
+                guidance,
+                t5_attn_mask,
+                neg_t5_out,
+                neg_l_pooled,
+                neg_t5_attn_mask,
+                cfg_scale,
+            )
+
+    return x
+
+
+def generate_image(
+    model,
+    clip_l: Optional[CLIPTextModel],
+    t5xxl,
+    ae,
+    prompt: str,
+    seed: Optional[int],
+    image_width: int,
+    image_height: int,
+    steps: Optional[int],
+    guidance: float,
+    negative_prompt: Optional[str],
+    cfg_scale: float,
+):
+    seed = seed if seed is not None else random.randint(0, 2**32 - 1)
+    logger.info(f"Seed: {seed}")
+
+    # make first noise with packed shape
+    # original: b,16,2*h//16,2*w//16, packed: b,h//16*w//16,16*2*2
+    packed_latent_height, packed_latent_width = math.ceil(image_height / 16), math.ceil(image_width / 16)
+    noise_dtype = torch.float32 if is_fp8(dtype) else dtype
+    noise = torch.randn(
+        1,
+        packed_latent_height * packed_latent_width,
+        16 * 2 * 2,
+        device=device,
+        dtype=noise_dtype,
+        generator=torch.Generator(device=device).manual_seed(seed),
+    )
+
+    # prepare img and img ids
+
+    # this is needed only for img2img
+    # img = rearrange(img, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
+    # if img.shape[0] == 1 and bs > 1:
+    #     img = repeat(img, "1 ... -> bs ...", bs=bs)
+
+    # txt2img only needs img_ids
+    img_ids = flux_utils.prepare_img_ids(1, packed_latent_height, packed_latent_width)
+
+    # prepare fp8 models
+    if clip_l is not None and is_fp8(clip_l_dtype) and (not hasattr(clip_l, "fp8_prepared") or not clip_l.fp8_prepared):
+        logger.info(f"prepare CLIP-L for fp8: set to {clip_l_dtype}, set embeddings to {torch.bfloat16}")
+        clip_l.to(clip_l_dtype)  # fp8
+        clip_l.text_model.embeddings.to(dtype=torch.bfloat16)
+        clip_l.fp8_prepared = True
+
+    if is_fp8(t5xxl_dtype) and (not hasattr(t5xxl, "fp8_prepared") or not t5xxl.fp8_prepared):
+        logger.info(f"prepare T5xxl for fp8: set to {t5xxl_dtype}")
+
+        def prepare_fp8(text_encoder, target_dtype):
+            def forward_hook(module):
+                def forward(hidden_states):
+                    hidden_gelu = module.act(module.wi_0(hidden_states))
+                    hidden_linear = module.wi_1(hidden_states)
+                    hidden_states = hidden_gelu * hidden_linear
+                    hidden_states = module.dropout(hidden_states)
+
+                    hidden_states = module.wo(hidden_states)
+                    return hidden_states
+
+                return forward
+
+            for module in text_encoder.modules():
+                if module.__class__.__name__ in ["T5LayerNorm", "Embedding"]:
+                    # print("set", module.__class__.__name__, "to", target_dtype)
+                    module.to(target_dtype)
+                if module.__class__.__name__ in ["T5DenseGatedActDense"]:
+                    # print("set", module.__class__.__name__, "hooks")
+                    module.forward = forward_hook(module)
+
+        t5xxl.to(t5xxl_dtype)
+        prepare_fp8(t5xxl.encoder, torch.bfloat16)
+        t5xxl.fp8_prepared = True
+
+    # prepare embeddings
+    logger.info("Encoding prompts...")
+    if clip_l is not None:
+        clip_l = clip_l.to(device)
+    t5xxl = t5xxl.to(device)
+
+    def encode(prpt: str):
+        tokens_and_masks = tokenize_strategy.tokenize(prpt)
+        with torch.no_grad():
+            if clip_l is not None:
+                if is_fp8(clip_l_dtype):
+                    with accelerator.autocast():
+                        l_pooled, _, _, _ = encoding_strategy.encode_tokens(tokenize_strategy, [clip_l, None], tokens_and_masks)
+                else:
+                    with torch.autocast(device_type=device.type, dtype=clip_l_dtype):
+                        l_pooled, _, _, _ = encoding_strategy.encode_tokens(tokenize_strategy, [clip_l, None], tokens_and_masks)
+            else:
+                l_pooled = None
+
+            if is_fp8(t5xxl_dtype):
+                with accelerator.autocast():
+                    _, t5_out, txt_ids, t5_attn_mask = encoding_strategy.encode_tokens(
+                        tokenize_strategy, [clip_l, t5xxl], tokens_and_masks, args.apply_t5_attn_mask
+                    )
+            else:
+                with torch.autocast(device_type=device.type, dtype=t5xxl_dtype):
+                    _, t5_out, txt_ids, t5_attn_mask = encoding_strategy.encode_tokens(
+                        tokenize_strategy, [clip_l, t5xxl], tokens_and_masks, args.apply_t5_attn_mask
+                    )
+        return l_pooled, t5_out, txt_ids, t5_attn_mask
+
+    l_pooled, t5_out, txt_ids, t5_attn_mask = encode(prompt)
+    if negative_prompt:
+        neg_l_pooled, neg_t5_out, _, neg_t5_attn_mask = encode(negative_prompt)
+    else:
+        neg_l_pooled, neg_t5_out, neg_t5_attn_mask = None, None, None
+
+    # NaN check
+    if l_pooled is not None and torch.isnan(l_pooled).any():
+        raise ValueError("NaN in l_pooled")
+    if torch.isnan(t5_out).any():
+        raise ValueError("NaN in t5_out")
+
+    if args.offload:
+        if clip_l is not None:
+            clip_l = clip_l.cpu()
+        t5xxl = t5xxl.cpu()
+    # del clip_l, t5xxl
+    device_utils.clean_memory()
+
+    # generate image
+    logger.info("Generating image...")
+    model = model.to(device)
+    if steps is None:
+        steps = 4 if is_schnell else 50
+
+    img_ids = img_ids.to(device)
+    t5_attn_mask = t5_attn_mask.to(device) if args.apply_t5_attn_mask else None
+    neg_t5_attn_mask = neg_t5_attn_mask.to(device) if neg_t5_attn_mask is not None and args.apply_t5_attn_mask else None
+
+    x = do_sample(
+        accelerator,
+        model,
+        noise,
+        img_ids,
+        l_pooled,
+        t5_out,
+        txt_ids,
+        steps,
+        guidance,
+        t5_attn_mask,
+        is_schnell,
+        device,
+        flux_dtype,
+        neg_l_pooled,
+        neg_t5_out,
+        neg_t5_attn_mask,
+        cfg_scale,
+    )
+    if args.offload:
+        model = model.cpu()
+    # del model
+    device_utils.clean_memory()
+
+    # unpack
+    x = x.float()
+    x = einops.rearrange(x, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=packed_latent_height, w=packed_latent_width, ph=2, pw=2)
+
+    # decode
+    logger.info("Decoding image...")
+    ae = ae.to(device)
+    with torch.no_grad():
+        if is_fp8(ae_dtype):
+            with accelerator.autocast():
+                x = ae.decode(x)
+        else:
+            with torch.autocast(device_type=device.type, dtype=ae_dtype):
+                x = ae.decode(x)
+    if args.offload:
+        ae = ae.cpu()
+
+    x = x.clamp(-1, 1)
+    x = x.permute(0, 2, 3, 1)
+    img = Image.fromarray((127.5 * (x + 1.0)).float().cpu().numpy().astype(np.uint8)[0])
+
+    # save image
+    output_dir = args.output_dir
+    os.makedirs(output_dir, exist_ok=True)
+    output_path = os.path.join(output_dir, f"{datetime.datetime.now().strftime('%Y%m%d_%H%M%S')}.png")
+    img.save(output_path)
+
+    logger.info(f"Saved image to {output_path}")
+
+
+if __name__ == "__main__":
+    target_height = 768  # 1024
+    target_width = 1360  # 1024
+
+    # steps = 50  # 28  # 50
+    # guidance_scale = 5
+    # seed = 1  # None  # 1
+
+    device = get_preferred_device()
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--ckpt_path", type=str, required=True)
+    parser.add_argument("--model_type", type=str, choices=["flux", "chroma"], default="flux", help="Model type to use")
+    parser.add_argument("--clip_l", type=str, required=False)
+    parser.add_argument("--t5xxl", type=str, required=False)
+    parser.add_argument("--ae", type=str, required=False)
+    parser.add_argument("--apply_t5_attn_mask", action="store_true")
+    parser.add_argument("--prompt", type=str, default="A photo of a cat")
+    parser.add_argument("--output_dir", type=str, default=".")
+    parser.add_argument("--dtype", type=str, default="bfloat16", help="base dtype")
+    parser.add_argument("--clip_l_dtype", type=str, default=None, help="dtype for clip_l")
+    parser.add_argument("--ae_dtype", type=str, default=None, help="dtype for ae")
+    parser.add_argument("--t5xxl_dtype", type=str, default=None, help="dtype for t5xxl")
+    parser.add_argument("--flux_dtype", type=str, default=None, help="dtype for flux")
+    parser.add_argument("--seed", type=int, default=None)
+    parser.add_argument("--steps", type=int, default=None, help="Number of steps. Default is 4 for schnell, 50 for dev")
+    parser.add_argument("--guidance", type=float, default=3.5)
+    parser.add_argument("--negative_prompt", type=str, default=None)
+    parser.add_argument("--cfg_scale", type=float, default=1.0)
+    parser.add_argument("--offload", action="store_true", help="Offload to CPU")
+    parser.add_argument(
+        "--lora_weights",
+        type=str,
+        nargs="*",
+        default=[],
+        help="LoRA weights, only supports networks.lora_flux and lora_oft, each argument is a `path;multiplier` (semi-colon separated)",
+    )
+    parser.add_argument("--merge_lora_weights", action="store_true", help="Merge LoRA weights to model")
+    parser.add_argument("--width", type=int, default=target_width)
+    parser.add_argument("--height", type=int, default=target_height)
+    parser.add_argument("--interactive", action="store_true")
+    args = parser.parse_args()
+
+    seed = args.seed
+    steps = args.steps
+    guidance_scale = args.guidance
+
+    def is_fp8(dt):
+        return dt in [torch.float8_e4m3fn, torch.float8_e4m3fnuz, torch.float8_e5m2, torch.float8_e5m2fnuz]
+
+    dtype = str_to_dtype(args.dtype)
+    clip_l_dtype = str_to_dtype(args.clip_l_dtype, dtype)
+    t5xxl_dtype = str_to_dtype(args.t5xxl_dtype, dtype)
+    ae_dtype = str_to_dtype(args.ae_dtype, dtype)
+    flux_dtype = str_to_dtype(args.flux_dtype, dtype)
+
+    logger.info(f"Dtypes for clip_l, t5xxl, ae, flux: {clip_l_dtype}, {t5xxl_dtype}, {ae_dtype}, {flux_dtype}")
+
+    loading_device = "cpu" if args.offload else device
+
+    use_fp8 = [is_fp8(d) for d in [dtype, clip_l_dtype, t5xxl_dtype, ae_dtype, flux_dtype]]
+    if any(use_fp8):
+        accelerator = accelerate.Accelerator(mixed_precision="bf16")
+    else:
+        accelerator = None
+
+    # load clip_l (skip for chroma model)
+    if args.model_type == "flux":
+        logger.info(f"Loading clip_l from {args.clip_l}...")
+        clip_l = flux_utils.load_clip_l(args.clip_l, clip_l_dtype, loading_device, disable_mmap=True)
+        clip_l.eval()
+    else:
+        clip_l = None
+
+    logger.info(f"Loading t5xxl from {args.t5xxl}...")
+    t5xxl = flux_utils.load_t5xxl(args.t5xxl, t5xxl_dtype, loading_device, disable_mmap=True)
+    t5xxl.eval()
+
+    # if is_fp8(clip_l_dtype):
+    #     clip_l = accelerator.prepare(clip_l)
+    # if is_fp8(t5xxl_dtype):
+    #     t5xxl = accelerator.prepare(t5xxl)
+
+    # DiT
+    is_schnell, model = flux_utils.load_flow_model(
+        args.ckpt_path, None, loading_device, disable_mmap=True, model_type=args.model_type
+    )
+    model.eval()
+    logger.info(f"Casting model to {flux_dtype}")
+    model.to(flux_dtype)  # make sure model is dtype
+    # if is_fp8(flux_dtype):
+    #     model = accelerator.prepare(model)
+    #     if args.offload:
+    #         model = model.to("cpu")
+
+    t5xxl_max_length = 256 if is_schnell else 512
+    tokenize_strategy = strategy_flux.FluxTokenizeStrategy(t5xxl_max_length)
+    encoding_strategy = strategy_flux.FluxTextEncodingStrategy()
+
+    # AE
+    ae = flux_utils.load_ae(args.ae, ae_dtype, loading_device)
+    ae.eval()
+    # if is_fp8(ae_dtype):
+    #     ae = accelerator.prepare(ae)
+
+    # LoRA
+    lora_models: List[lora_flux.LoRANetwork] = []
+    for weights_file in args.lora_weights:
+        if ";" in weights_file:
+            weights_file, multiplier = weights_file.split(";")
+            multiplier = float(multiplier)
+        else:
+            multiplier = 1.0
+
+        weights_sd = load_file(weights_file)
+        is_lora = is_oft = False
+        for key in weights_sd.keys():
+            if key.startswith("lora"):
+                is_lora = True
+            if key.startswith("oft"):
+                is_oft = True
+            if is_lora or is_oft:
+                break
+
+        module = lora_flux if is_lora else oft_flux
+        lora_model, _ = module.create_network_from_weights(multiplier, None, ae, [clip_l, t5xxl], model, weights_sd, True)
+
+        if args.merge_lora_weights:
+            lora_model.merge_to([clip_l, t5xxl], model, weights_sd)
+        else:
+            lora_model.apply_to([clip_l, t5xxl], model)
+            info = lora_model.load_state_dict(weights_sd, strict=True)
+            logger.info(f"Loaded LoRA weights from {weights_file}: {info}")
+            lora_model.eval()
+            lora_model.to(device)
+
+        lora_models.append(lora_model)
+
+    if not args.interactive:
+        generate_image(
+            model,
+            clip_l,
+            t5xxl,
+            ae,
+            args.prompt,
+            args.seed,
+            args.width,
+            args.height,
+            args.steps,
+            args.guidance,
+            args.negative_prompt,
+            args.cfg_scale,
+        )
+    else:
+        # loop for interactive
+        width = target_width
+        height = target_height
+        steps = None
+        guidance = args.guidance
+        cfg_scale = args.cfg_scale
+
+        while True:
+            print(
+                "Enter prompt (empty to exit). Options: --w <width> --h <height> --s <steps> --d <seed> --g <guidance> --m <multipliers for LoRA>"
+                " --n <negative prompt>, `-` for empty negative prompt --c <cfg_scale>"
+            )
+            prompt = input()
+            if prompt == "":
+                break
+
+            # parse options
+            options = prompt.split("--")
+            prompt = options[0].strip()
+            seed = None
+            negative_prompt = None
+            for opt in options[1:]:
+                try:
+                    opt = opt.strip()
+                    if opt.startswith("w"):
+                        width = int(opt[1:].strip())
+                    elif opt.startswith("h"):
+                        height = int(opt[1:].strip())
+                    elif opt.startswith("s"):
+                        steps = int(opt[1:].strip())
+                    elif opt.startswith("d"):
+                        seed = int(opt[1:].strip())
+                    elif opt.startswith("g"):
+                        guidance = float(opt[1:].strip())
+                    elif opt.startswith("m"):
+                        mutipliers = opt[1:].strip().split(",")
+                        if len(mutipliers) != len(lora_models):
+                            logger.error(f"Invalid number of multipliers, expected {len(lora_models)}")
+                            continue
+                        for i, lora_model in enumerate(lora_models):
+                            lora_model.set_multiplier(float(mutipliers[i]))
+                    elif opt.startswith("n"):
+                        negative_prompt = opt[1:].strip()
+                        if negative_prompt == "-":
+                            negative_prompt = ""
+                    elif opt.startswith("c"):
+                        cfg_scale = float(opt[1:].strip())
+                except ValueError as e:
+                    logger.error(f"Invalid option: {opt}, {e}")
+
+            generate_image(model, clip_l, t5xxl, ae, prompt, seed, width, height, steps, guidance, negative_prompt, cfg_scale)
+
+    logger.info("Done!")

flux_train.py

@@ -0,0 +1,851 @@
+# training with captions
+
+# Swap blocks between CPU and GPU:
+# This implementation is inspired by and based on the work of 2kpr.
+# Many thanks to 2kpr for the original concept and implementation of memory-efficient offloading.
+# The original idea has been adapted and extended to fit the current project's needs.
+
+# Key features:
+# - CPU offloading during forward and backward passes
+# - Use of fused optimizer and grad_hook for efficient gradient processing
+# - Per-block fused optimizer instances
+
+import argparse
+from concurrent.futures import ThreadPoolExecutor
+import copy
+import math
+import os
+from multiprocessing import Value
+import time
+from typing import List, Optional, Tuple, Union
+import toml
+
+from tqdm import tqdm
+
+import torch
+import torch.nn as nn
+from library import utils
+from library.device_utils import init_ipex, clean_memory_on_device
+
+init_ipex()
+
+from accelerate.utils import set_seed
+from library import deepspeed_utils, flux_train_utils, flux_utils, strategy_base, strategy_flux, sai_model_spec
+from library.sd3_train_utils import FlowMatchEulerDiscreteScheduler
+
+import library.train_util as train_util
+
+from library.utils import setup_logging, add_logging_arguments
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+import library.config_util as config_util
+
+# import library.sdxl_train_util as sdxl_train_util
+from library.config_util import (
+    ConfigSanitizer,
+    BlueprintGenerator,
+)
+from library.custom_train_functions import apply_masked_loss, add_custom_train_arguments
+
+
+def train(args):
+    train_util.verify_training_args(args)
+    train_util.prepare_dataset_args(args, True)
+    # sdxl_train_util.verify_sdxl_training_args(args)
+    deepspeed_utils.prepare_deepspeed_args(args)
+    setup_logging(args, reset=True)
+
+    # temporary: backward compatibility for deprecated options. remove in the future
+    if not args.skip_cache_check:
+        args.skip_cache_check = args.skip_latents_validity_check
+
+    # assert (
+    #     not args.weighted_captions
+    # ), "weighted_captions is not supported currently / weighted_captionsは現在サポートされていません"
+    if args.cache_text_encoder_outputs_to_disk and not args.cache_text_encoder_outputs:
+        logger.warning(
+            "cache_text_encoder_outputs_to_disk is enabled, so cache_text_encoder_outputs is also enabled / cache_text_encoder_outputs_to_diskが有効になっているため、cache_text_encoder_outputsも有効になります"
+        )
+        args.cache_text_encoder_outputs = True
+
+    if args.cpu_offload_checkpointing and not args.gradient_checkpointing:
+        logger.warning(
+            "cpu_offload_checkpointing is enabled, so gradient_checkpointing is also enabled / cpu_offload_checkpointingが有効になっているため、gradient_checkpointingも有効になります"
+        )
+        args.gradient_checkpointing = True
+
+    assert (
+        args.blocks_to_swap is None or args.blocks_to_swap == 0
+    ) or not args.cpu_offload_checkpointing, (
+        "blocks_to_swap is not supported with cpu_offload_checkpointing / blocks_to_swapはcpu_offload_checkpointingと併用できません"
+    )
+
+    cache_latents = args.cache_latents
+    use_dreambooth_method = args.in_json is None
+
+    if args.seed is not None:
+        set_seed(args.seed)  # 乱数系列を初期化する
+
+    # prepare caching strategy: this must be set before preparing dataset. because dataset may use this strategy for initialization.
+    if args.cache_latents:
+        latents_caching_strategy = strategy_flux.FluxLatentsCachingStrategy(
+            args.cache_latents_to_disk, args.vae_batch_size, args.skip_cache_check
+        )
+        strategy_base.LatentsCachingStrategy.set_strategy(latents_caching_strategy)
+
+    # データセットを準備する
+    if args.dataset_class is None:
+        blueprint_generator = BlueprintGenerator(ConfigSanitizer(True, True, args.masked_loss, True))
+        if args.dataset_config is not None:
+            logger.info(f"Load dataset config from {args.dataset_config}")
+            user_config = config_util.load_user_config(args.dataset_config)
+            ignored = ["train_data_dir", "in_json"]
+            if any(getattr(args, attr) is not None for attr in ignored):
+                logger.warning(
+                    "ignore following options because config file is found: {0} / 設定ファイルが利用されるため以下のオプションは無視されます: {0}".format(
+                        ", ".join(ignored)
+                    )
+                )
+        else:
+            if use_dreambooth_method:
+                logger.info("Using DreamBooth method.")
+                user_config = {
+                    "datasets": [
+                        {
+                            "subsets": config_util.generate_dreambooth_subsets_config_by_subdirs(
+                                args.train_data_dir, args.reg_data_dir
+                            )
+                        }
+                    ]
+                }
+            else:
+                logger.info("Training with captions.")
+                user_config = {
+                    "datasets": [
+                        {
+                            "subsets": [
+                                {
+                                    "image_dir": args.train_data_dir,
+                                    "metadata_file": args.in_json,
+                                }
+                            ]
+                        }
+                    ]
+                }
+
+        blueprint = blueprint_generator.generate(user_config, args)
+        train_dataset_group, val_dataset_group = config_util.generate_dataset_group_by_blueprint(blueprint.dataset_group)
+    else:
+        train_dataset_group = train_util.load_arbitrary_dataset(args)
+        val_dataset_group = None
+
+    current_epoch = Value("i", 0)
+    current_step = Value("i", 0)
+    ds_for_collator = train_dataset_group if args.max_data_loader_n_workers == 0 else None
+    collator = train_util.collator_class(current_epoch, current_step, ds_for_collator)
+
+    train_dataset_group.verify_bucket_reso_steps(16)  # TODO これでいいか確認
+
+    _, is_schnell, _, _ = flux_utils.analyze_checkpoint_state(args.pretrained_model_name_or_path)
+    if args.debug_dataset:
+        if args.cache_text_encoder_outputs:
+            strategy_base.TextEncoderOutputsCachingStrategy.set_strategy(
+                strategy_flux.FluxTextEncoderOutputsCachingStrategy(
+                    args.cache_text_encoder_outputs_to_disk, args.text_encoder_batch_size, args.skip_cache_check, False
+                )
+            )
+        t5xxl_max_token_length = (
+            args.t5xxl_max_token_length if args.t5xxl_max_token_length is not None else (256 if is_schnell else 512)
+        )
+        strategy_base.TokenizeStrategy.set_strategy(strategy_flux.FluxTokenizeStrategy(t5xxl_max_token_length))
+
+        train_dataset_group.set_current_strategies()
+        train_util.debug_dataset(train_dataset_group, True)
+        return
+    if len(train_dataset_group) == 0:
+        logger.error(
+            "No data found. Please verify the metadata file and train_data_dir option. / 画像がありません。メタデータおよびtrain_data_dirオプションを確認してください。"
+        )
+        return
+
+    if cache_latents:
+        assert (
+            train_dataset_group.is_latent_cacheable()
+        ), "when caching latents, either color_aug or random_crop cannot be used / latentをキャッシュするときはcolor_augとrandom_cropは使えません"
+
+    if args.cache_text_encoder_outputs:
+        assert (
+            train_dataset_group.is_text_encoder_output_cacheable()
+        ), "when caching text encoder output, either caption_dropout_rate, shuffle_caption, token_warmup_step or caption_tag_dropout_rate cannot be used / text encoderの出力をキャッシュするときはcaption_dropout_rate, shuffle_caption, token_warmup_step, caption_tag_dropout_rateは使えません"
+
+    # acceleratorを準備する
+    logger.info("prepare accelerator")
+    accelerator = train_util.prepare_accelerator(args)
+
+    # mixed precisionに対応した型を用意しておき適宜castする
+    weight_dtype, save_dtype = train_util.prepare_dtype(args)
+
+    # モデルを読み込む
+
+    # load VAE for caching latents
+    ae = None
+    if cache_latents:
+        ae = flux_utils.load_ae(args.ae, weight_dtype, "cpu", args.disable_mmap_load_safetensors)
+        ae.to(accelerator.device, dtype=weight_dtype)
+        ae.requires_grad_(False)
+        ae.eval()
+
+        train_dataset_group.new_cache_latents(ae, accelerator)
+
+        ae.to("cpu")  # if no sampling, vae can be deleted
+        clean_memory_on_device(accelerator.device)
+
+        accelerator.wait_for_everyone()
+
+    # prepare tokenize strategy
+    if args.t5xxl_max_token_length is None:
+        if is_schnell:
+            t5xxl_max_token_length = 256
+        else:
+            t5xxl_max_token_length = 512
+    else:
+        t5xxl_max_token_length = args.t5xxl_max_token_length
+
+    flux_tokenize_strategy = strategy_flux.FluxTokenizeStrategy(t5xxl_max_token_length)
+    strategy_base.TokenizeStrategy.set_strategy(flux_tokenize_strategy)
+
+    # load clip_l, t5xxl for caching text encoder outputs
+    clip_l = flux_utils.load_clip_l(args.clip_l, weight_dtype, "cpu", args.disable_mmap_load_safetensors)
+    t5xxl = flux_utils.load_t5xxl(args.t5xxl, weight_dtype, "cpu", args.disable_mmap_load_safetensors)
+    clip_l.eval()
+    t5xxl.eval()
+    clip_l.requires_grad_(False)
+    t5xxl.requires_grad_(False)
+
+    text_encoding_strategy = strategy_flux.FluxTextEncodingStrategy(args.apply_t5_attn_mask)
+    strategy_base.TextEncodingStrategy.set_strategy(text_encoding_strategy)
+
+    # cache text encoder outputs
+    sample_prompts_te_outputs = None
+    if args.cache_text_encoder_outputs:
+        # Text Encodes are eval and no grad here
+        clip_l.to(accelerator.device)
+        t5xxl.to(accelerator.device)
+
+        text_encoder_caching_strategy = strategy_flux.FluxTextEncoderOutputsCachingStrategy(
+            args.cache_text_encoder_outputs_to_disk, args.text_encoder_batch_size, False, False, args.apply_t5_attn_mask
+        )
+        strategy_base.TextEncoderOutputsCachingStrategy.set_strategy(text_encoder_caching_strategy)
+
+        with accelerator.autocast():
+            train_dataset_group.new_cache_text_encoder_outputs([clip_l, t5xxl], accelerator)
+
+        # cache sample prompt's embeddings to free text encoder's memory
+        if args.sample_prompts is not None:
+            logger.info(f"cache Text Encoder outputs for sample prompt: {args.sample_prompts}")
+
+            text_encoding_strategy: strategy_flux.FluxTextEncodingStrategy = strategy_base.TextEncodingStrategy.get_strategy()
+
+            prompts = train_util.load_prompts(args.sample_prompts)
+            sample_prompts_te_outputs = {}  # key: prompt, value: text encoder outputs
+            with accelerator.autocast(), torch.no_grad():
+                for prompt_dict in prompts:
+                    for p in [prompt_dict.get("prompt", ""), prompt_dict.get("negative_prompt", "")]:
+                        if p not in sample_prompts_te_outputs:
+                            logger.info(f"cache Text Encoder outputs for prompt: {p}")
+                            tokens_and_masks = flux_tokenize_strategy.tokenize(p)
+                            sample_prompts_te_outputs[p] = text_encoding_strategy.encode_tokens(
+                                flux_tokenize_strategy, [clip_l, t5xxl], tokens_and_masks, args.apply_t5_attn_mask
+                            )
+
+        accelerator.wait_for_everyone()
+
+        # now we can delete Text Encoders to free memory
+        clip_l = None
+        t5xxl = None
+        clean_memory_on_device(accelerator.device)
+
+    # load FLUX
+    _, flux = flux_utils.load_flow_model(
+        args.pretrained_model_name_or_path, weight_dtype, "cpu", args.disable_mmap_load_safetensors, model_type="flux"
+    )
+
+    if args.gradient_checkpointing:
+        flux.enable_gradient_checkpointing(cpu_offload=args.cpu_offload_checkpointing)
+
+    flux.requires_grad_(True)
+
+    # block swap
+
+    # backward compatibility
+    if args.blocks_to_swap is None:
+        blocks_to_swap = args.double_blocks_to_swap or 0
+        if args.single_blocks_to_swap is not None:
+            blocks_to_swap += args.single_blocks_to_swap // 2
+        if blocks_to_swap > 0:
+            logger.warning(
+                "double_blocks_to_swap and single_blocks_to_swap are deprecated. Use blocks_to_swap instead."
+                " / double_blocks_to_swapとsingle_blocks_to_swapは非推奨です。blocks_to_swapを使ってください。"
+            )
+            logger.info(
+                f"double_blocks_to_swap={args.double_blocks_to_swap} and single_blocks_to_swap={args.single_blocks_to_swap} are converted to blocks_to_swap={blocks_to_swap}."
+            )
+            args.blocks_to_swap = blocks_to_swap
+        del blocks_to_swap
+
+    is_swapping_blocks = args.blocks_to_swap is not None and args.blocks_to_swap > 0
+    if is_swapping_blocks:
+        # Swap blocks between CPU and GPU to reduce memory usage, in forward and backward passes.
+        # This idea is based on 2kpr's great work. Thank you!
+        logger.info(f"enable block swap: blocks_to_swap={args.blocks_to_swap}")
+        flux.enable_block_swap(args.blocks_to_swap, accelerator.device)
+
+    if not cache_latents:
+        # load VAE here if not cached
+        ae = flux_utils.load_ae(args.ae, weight_dtype, "cpu")
+        ae.requires_grad_(False)
+        ae.eval()
+        ae.to(accelerator.device, dtype=weight_dtype)
+
+    training_models = []
+    params_to_optimize = []
+    training_models.append(flux)
+    name_and_params = list(flux.named_parameters())
+    # single param group for now
+    params_to_optimize.append({"params": [p for _, p in name_and_params], "lr": args.learning_rate})
+    param_names = [[n for n, _ in name_and_params]]
+
+    # calculate number of trainable parameters
+    n_params = 0
+    for group in params_to_optimize:
+        for p in group["params"]:
+            n_params += p.numel()
+
+    accelerator.print(f"number of trainable parameters: {n_params}")
+
+    # 学習に必要なクラスを準備する
+    accelerator.print("prepare optimizer, data loader etc.")
+
+    if args.blockwise_fused_optimizers:
+        # fused backward pass: https://pytorch.org/tutorials/intermediate/optimizer_step_in_backward_tutorial.html
+        # Instead of creating an optimizer for all parameters as in the tutorial, we create an optimizer for each block of parameters.
+        # This balances memory usage and management complexity.
+
+        # split params into groups. currently different learning rates are not supported
+        grouped_params = []
+        param_group = {}
+        for group in params_to_optimize:
+            named_parameters = list(flux.named_parameters())
+            assert len(named_parameters) == len(group["params"]), "number of parameters does not match"
+            for p, np in zip(group["params"], named_parameters):
+                # determine target layer and block index for each parameter
+                block_type = "other"  # double, single or other
+                if np[0].startswith("double_blocks"):
+                    block_index = int(np[0].split(".")[1])
+                    block_type = "double"
+                elif np[0].startswith("single_blocks"):
+                    block_index = int(np[0].split(".")[1])
+                    block_type = "single"
+                else:
+                    block_index = -1
+
+                param_group_key = (block_type, block_index)
+                if param_group_key not in param_group:
+                    param_group[param_group_key] = []
+                param_group[param_group_key].append(p)
+
+        block_types_and_indices = []
+        for param_group_key, param_group in param_group.items():
+            block_types_and_indices.append(param_group_key)
+            grouped_params.append({"params": param_group, "lr": args.learning_rate})
+
+            num_params = 0
+            for p in param_group:
+                num_params += p.numel()
+            accelerator.print(f"block {param_group_key}: {num_params} parameters")
+
+        # prepare optimizers for each group
+        optimizers = []
+        for group in grouped_params:
+            _, _, optimizer = train_util.get_optimizer(args, trainable_params=[group])
+            optimizers.append(optimizer)
+        optimizer = optimizers[0]  # avoid error in the following code
+
+        logger.info(f"using {len(optimizers)} optimizers for blockwise fused optimizers")
+
+        if train_util.is_schedulefree_optimizer(optimizers[0], args):
+            raise ValueError("Schedule-free optimizer is not supported with blockwise fused optimizers")
+        optimizer_train_fn = lambda: None  # dummy function
+        optimizer_eval_fn = lambda: None  # dummy function
+    else:
+        _, _, optimizer = train_util.get_optimizer(args, trainable_params=params_to_optimize)
+        optimizer_train_fn, optimizer_eval_fn = train_util.get_optimizer_train_eval_fn(optimizer, args)
+
+    # prepare dataloader
+    # strategies are set here because they cannot be referenced in another process. Copy them with the dataset
+    # some strategies can be None
+    train_dataset_group.set_current_strategies()
+
+    # DataLoaderのプロセス数：0 は persistent_workers が使えないので注意
+    n_workers = min(args.max_data_loader_n_workers, os.cpu_count())  # cpu_count or max_data_loader_n_workers
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset_group,
+        batch_size=1,
+        shuffle=True,
+        collate_fn=collator,
+        num_workers=n_workers,
+        persistent_workers=args.persistent_data_loader_workers,
+    )
+
+    # 学習ステップ数を計算する
+    if args.max_train_epochs is not None:
+        args.max_train_steps = args.max_train_epochs * math.ceil(
+            len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
+        )
+        accelerator.print(
+            f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
+        )
+
+    # データセット側にも学習ステップを送信
+    train_dataset_group.set_max_train_steps(args.max_train_steps)
+
+    # lr schedulerを用意する
+    if args.blockwise_fused_optimizers:
+        # prepare lr schedulers for each optimizer
+        lr_schedulers = [train_util.get_scheduler_fix(args, optimizer, accelerator.num_processes) for optimizer in optimizers]
+        lr_scheduler = lr_schedulers[0]  # avoid error in the following code
+    else:
+        lr_scheduler = train_util.get_scheduler_fix(args, optimizer, accelerator.num_processes)
+
+    # 実験的機能：勾配も含めたfp16/bf16学習を行う　モデル全体をfp16/bf16にする
+    if args.full_fp16:
+        assert (
+            args.mixed_precision == "fp16"
+        ), "full_fp16 requires mixed precision='fp16' / full_fp16を使う場合はmixed_precision='fp16'を指定してください。"
+        accelerator.print("enable full fp16 training.")
+        flux.to(weight_dtype)
+        if clip_l is not None:
+            clip_l.to(weight_dtype)
+            t5xxl.to(weight_dtype)  # TODO check works with fp16 or not
+    elif args.full_bf16:
+        assert (
+            args.mixed_precision == "bf16"
+        ), "full_bf16 requires mixed precision='bf16' / full_bf16を使う場合はmixed_precision='bf16'を指定してください。"
+        accelerator.print("enable full bf16 training.")
+        flux.to(weight_dtype)
+        if clip_l is not None:
+            clip_l.to(weight_dtype)
+            t5xxl.to(weight_dtype)
+
+    # if we don't cache text encoder outputs, move them to device
+    if not args.cache_text_encoder_outputs:
+        clip_l.to(accelerator.device)
+        t5xxl.to(accelerator.device)
+
+    clean_memory_on_device(accelerator.device)
+
+    if args.deepspeed:
+        ds_model = deepspeed_utils.prepare_deepspeed_model(args, mmdit=flux)
+        # most of ZeRO stage uses optimizer partitioning, so we have to prepare optimizer and ds_model at the same time. # pull/1139#issuecomment-1986790007
+        ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+            ds_model, optimizer, train_dataloader, lr_scheduler
+        )
+        training_models = [ds_model]
+
+    else:
+        # accelerator does some magic
+        # if we doesn't swap blocks, we can move the model to device
+        flux = accelerator.prepare(flux, device_placement=[not is_swapping_blocks])
+        if is_swapping_blocks:
+            accelerator.unwrap_model(flux).move_to_device_except_swap_blocks(accelerator.device)  # reduce peak memory usage
+        optimizer, train_dataloader, lr_scheduler = accelerator.prepare(optimizer, train_dataloader, lr_scheduler)
+
+    # 実験的機能：勾配も含めたfp16学習を行う　PyTorchにパッチを当ててfp16でのgrad scaleを有効にする
+    if args.full_fp16:
+        # During deepseed training, accelerate not handles fp16/bf16|mixed precision directly via scaler. Let deepspeed engine do.
+        # -> But we think it's ok to patch accelerator even if deepspeed is enabled.
+        train_util.patch_accelerator_for_fp16_training(accelerator)
+
+    # resumeする
+    train_util.resume_from_local_or_hf_if_specified(accelerator, args)
+
+    if args.fused_backward_pass:
+        # use fused optimizer for backward pass: other optimizers will be supported in the future
+        import library.adafactor_fused
+
+        library.adafactor_fused.patch_adafactor_fused(optimizer)
+
+        for param_group, param_name_group in zip(optimizer.param_groups, param_names):
+            for parameter, param_name in zip(param_group["params"], param_name_group):
+                if parameter.requires_grad:
+
+                    def create_grad_hook(p_name, p_group):
+                        def grad_hook(tensor: torch.Tensor):
+                            if accelerator.sync_gradients and args.max_grad_norm != 0.0:
+                                accelerator.clip_grad_norm_(tensor, args.max_grad_norm)
+                            optimizer.step_param(tensor, p_group)
+                            tensor.grad = None
+
+                        return grad_hook
+
+                    parameter.register_post_accumulate_grad_hook(create_grad_hook(param_name, param_group))
+
+    elif args.blockwise_fused_optimizers:
+        # prepare for additional optimizers and lr schedulers
+        for i in range(1, len(optimizers)):
+            optimizers[i] = accelerator.prepare(optimizers[i])
+            lr_schedulers[i] = accelerator.prepare(lr_schedulers[i])
+
+        # counters are used to determine when to step the optimizer
+        global optimizer_hooked_count
+        global num_parameters_per_group
+        global parameter_optimizer_map
+
+        optimizer_hooked_count = {}
+        num_parameters_per_group = [0] * len(optimizers)
+        parameter_optimizer_map = {}
+
+        for opt_idx, optimizer in enumerate(optimizers):
+            for param_group in optimizer.param_groups:
+                for parameter in param_group["params"]:
+                    if parameter.requires_grad:
+
+                        def grad_hook(parameter: torch.Tensor):
+                            if accelerator.sync_gradients and args.max_grad_norm != 0.0:
+                                accelerator.clip_grad_norm_(parameter, args.max_grad_norm)
+
+                            i = parameter_optimizer_map[parameter]
+                            optimizer_hooked_count[i] += 1
+                            if optimizer_hooked_count[i] == num_parameters_per_group[i]:
+                                optimizers[i].step()
+                                optimizers[i].zero_grad(set_to_none=True)
+
+                        parameter.register_post_accumulate_grad_hook(grad_hook)
+                        parameter_optimizer_map[parameter] = opt_idx
+                        num_parameters_per_group[opt_idx] += 1
+
+    # epoch数を計算する
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
+    if (args.save_n_epoch_ratio is not None) and (args.save_n_epoch_ratio > 0):
+        args.save_every_n_epochs = math.floor(num_train_epochs / args.save_n_epoch_ratio) or 1
+
+    # 学習する
+    # total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
+    accelerator.print("running training / 学習開始")
+    accelerator.print(f"  num examples / サンプル数: {train_dataset_group.num_train_images}")
+    accelerator.print(f"  num batches per epoch / 1epochのバッチ数: {len(train_dataloader)}")
+    accelerator.print(f"  num epochs / epoch数: {num_train_epochs}")
+    accelerator.print(
+        f"  batch size per device / バッチサイズ: {', '.join([str(d.batch_size) for d in train_dataset_group.datasets])}"
+    )
+    # accelerator.print(
+    #     f"  total train batch size (with parallel & distributed & accumulation) / 総バッチサイズ（並列学習、勾配合計含む）: {total_batch_size}"
+    # )
+    accelerator.print(f"  gradient accumulation steps / 勾配を合計するステップ数 = {args.gradient_accumulation_steps}")
+    accelerator.print(f"  total optimization steps / 学習ステップ数: {args.max_train_steps}")
+
+    progress_bar = tqdm(range(args.max_train_steps), smoothing=0, disable=not accelerator.is_local_main_process, desc="steps")
+    global_step = 0
+
+    noise_scheduler = FlowMatchEulerDiscreteScheduler(num_train_timesteps=1000, shift=args.discrete_flow_shift)
+    noise_scheduler_copy = copy.deepcopy(noise_scheduler)
+
+    if accelerator.is_main_process:
+        init_kwargs = {}
+        if args.wandb_run_name:
+            init_kwargs["wandb"] = {"name": args.wandb_run_name}
+        if args.log_tracker_config is not None:
+            init_kwargs = toml.load(args.log_tracker_config)
+        accelerator.init_trackers(
+            "finetuning" if args.log_tracker_name is None else args.log_tracker_name,
+            config=train_util.get_sanitized_config_or_none(args),
+            init_kwargs=init_kwargs,
+        )
+
+    if is_swapping_blocks:
+        accelerator.unwrap_model(flux).prepare_block_swap_before_forward()
+
+    # For --sample_at_first
+    optimizer_eval_fn()
+    flux_train_utils.sample_images(accelerator, args, 0, global_step, flux, ae, [clip_l, t5xxl], sample_prompts_te_outputs)
+    optimizer_train_fn()
+    if len(accelerator.trackers) > 0:
+        # log empty object to commit the sample images to wandb
+        accelerator.log({}, step=0)
+
+    loss_recorder = train_util.LossRecorder()
+    epoch = 0  # avoid error when max_train_steps is 0
+    for epoch in range(num_train_epochs):
+        accelerator.print(f"\nepoch {epoch+1}/{num_train_epochs}")
+        current_epoch.value = epoch + 1
+
+        for m in training_models:
+            m.train()
+
+        for step, batch in enumerate(train_dataloader):
+            current_step.value = global_step
+
+            if args.blockwise_fused_optimizers:
+                optimizer_hooked_count = {i: 0 for i in range(len(optimizers))}  # reset counter for each step
+
+            with accelerator.accumulate(*training_models):
+                if "latents" in batch and batch["latents"] is not None:
+                    latents = batch["latents"].to(accelerator.device, dtype=weight_dtype)
+                else:
+                    with torch.no_grad():
+                        # encode images to latents. images are [-1, 1]
+                        latents = ae.encode(batch["images"].to(ae.dtype)).to(accelerator.device, dtype=weight_dtype)
+
+                    # NaNが含まれていれば警告を表示し0に置き換える
+                    if torch.any(torch.isnan(latents)):
+                        accelerator.print("NaN found in latents, replacing with zeros")
+                        latents = torch.nan_to_num(latents, 0, out=latents)
+
+                text_encoder_outputs_list = batch.get("text_encoder_outputs_list", None)
+                if text_encoder_outputs_list is not None:
+                    text_encoder_conds = text_encoder_outputs_list
+                else:
+                    # not cached or training, so get from text encoders
+                    tokens_and_masks = batch["input_ids_list"]
+                    with torch.no_grad():
+                        input_ids = [ids.to(accelerator.device) for ids in batch["input_ids_list"]]
+                        text_encoder_conds = text_encoding_strategy.encode_tokens(
+                            flux_tokenize_strategy, [clip_l, t5xxl], input_ids, args.apply_t5_attn_mask
+                        )
+                        if args.full_fp16:
+                            text_encoder_conds = [c.to(weight_dtype) for c in text_encoder_conds]
+
+                # TODO support some features for noise implemented in get_noise_noisy_latents_and_timesteps
+
+                # Sample noise that we'll add to the latents
+                noise = torch.randn_like(latents)
+                bsz = latents.shape[0]
+
+                # get noisy model input and timesteps
+                noisy_model_input, timesteps, sigmas = flux_train_utils.get_noisy_model_input_and_timesteps(
+                    args, noise_scheduler_copy, latents, noise, accelerator.device, weight_dtype
+                )
+
+                # pack latents and get img_ids
+                packed_noisy_model_input = flux_utils.pack_latents(noisy_model_input)  # b, c, h*2, w*2 -> b, h*w, c*4
+                packed_latent_height, packed_latent_width = noisy_model_input.shape[2] // 2, noisy_model_input.shape[3] // 2
+                img_ids = flux_utils.prepare_img_ids(bsz, packed_latent_height, packed_latent_width).to(device=accelerator.device)
+
+                # get guidance: ensure args.guidance_scale is float
+                guidance_vec = torch.full((bsz,), float(args.guidance_scale), device=accelerator.device)
+
+                # call model
+                l_pooled, t5_out, txt_ids, t5_attn_mask = text_encoder_conds
+                if not args.apply_t5_attn_mask:
+                    t5_attn_mask = None
+
+                with accelerator.autocast():
+                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transformer model (we should not keep it but I want to keep the inputs same for the model for testing)
+                    model_pred = flux(
+                        img=packed_noisy_model_input,
+                        img_ids=img_ids,
+                        txt=t5_out,
+                        txt_ids=txt_ids,
+                        y=l_pooled,
+                        timesteps=timesteps / 1000,
+                        guidance=guidance_vec,
+                        txt_attention_mask=t5_attn_mask,
+                    )
+
+                # unpack latents
+                model_pred = flux_utils.unpack_latents(model_pred, packed_latent_height, packed_latent_width)
+
+                # apply model prediction type
+                model_pred, weighting = flux_train_utils.apply_model_prediction_type(args, model_pred, noisy_model_input, sigmas)
+
+                # flow matching loss: this is different from SD3
+                target = noise - latents
+
+                # calculate loss
+                huber_c = train_util.get_huber_threshold_if_needed(args, timesteps, noise_scheduler)
+                loss = train_util.conditional_loss(model_pred.float(), target.float(), args.loss_type, "none", huber_c)
+                if weighting is not None:
+                    loss = loss * weighting
+                if args.masked_loss or ("alpha_masks" in batch and batch["alpha_masks"] is not None):
+                    loss = apply_masked_loss(loss, batch)
+                loss = loss.mean([1, 2, 3])
+
+                loss_weights = batch["loss_weights"]  # 各sampleごとのweight
+                loss = loss * loss_weights
+                loss = loss.mean()
+
+                # backward
+                accelerator.backward(loss)
+
+                if not (args.fused_backward_pass or args.blockwise_fused_optimizers):
+                    if accelerator.sync_gradients and args.max_grad_norm != 0.0:
+                        params_to_clip = []
+                        for m in training_models:
+                            params_to_clip.extend(m.parameters())
+                        accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
+
+                    optimizer.step()
+                    lr_scheduler.step()
+                    optimizer.zero_grad(set_to_none=True)
+                else:
+                    # optimizer.step() and optimizer.zero_grad() are called in the optimizer hook
+                    lr_scheduler.step()
+                    if args.blockwise_fused_optimizers:
+                        for i in range(1, len(optimizers)):
+                            lr_schedulers[i].step()
+
+            # Checks if the accelerator has performed an optimization step behind the scenes
+            if accelerator.sync_gradients:
+                progress_bar.update(1)
+                global_step += 1
+
+                optimizer_eval_fn()
+                flux_train_utils.sample_images(
+                    accelerator, args, None, global_step, flux, ae, [clip_l, t5xxl], sample_prompts_te_outputs
+                )
+
+                # 指定ステップごとにモデルを保存
+                if args.save_every_n_steps is not None and global_step % args.save_every_n_steps == 0:
+                    accelerator.wait_for_everyone()
+                    if accelerator.is_main_process:
+                        flux_train_utils.save_flux_model_on_epoch_end_or_stepwise(
+                            args,
+                            False,
+                            accelerator,
+                            save_dtype,
+                            epoch,
+                            num_train_epochs,
+                            global_step,
+                            accelerator.unwrap_model(flux),
+                        )
+                optimizer_train_fn()
+
+            current_loss = loss.detach().item()  # 平均なのでbatch sizeは関係ないはず
+            if len(accelerator.trackers) > 0:
+                logs = {"loss": current_loss}
+                train_util.append_lr_to_logs(logs, lr_scheduler, args.optimizer_type, including_unet=True)
+
+                accelerator.log(logs, step=global_step)
+
+            loss_recorder.add(epoch=epoch, step=step, loss=current_loss)
+            avr_loss: float = loss_recorder.moving_average
+            logs = {"avr_loss": avr_loss}  # , "lr": lr_scheduler.get_last_lr()[0]}
+            progress_bar.set_postfix(**logs)
+
+            if global_step >= args.max_train_steps:
+                break
+
+        if len(accelerator.trackers) > 0:
+            logs = {"loss/epoch": loss_recorder.moving_average}
+            accelerator.log(logs, step=epoch + 1)
+
+        accelerator.wait_for_everyone()
+
+        optimizer_eval_fn()
+        if args.save_every_n_epochs is not None:
+            if accelerator.is_main_process:
+                flux_train_utils.save_flux_model_on_epoch_end_or_stepwise(
+                    args,
+                    True,
+                    accelerator,
+                    save_dtype,
+                    epoch,
+                    num_train_epochs,
+                    global_step,
+                    accelerator.unwrap_model(flux),
+                )
+
+        flux_train_utils.sample_images(
+            accelerator, args, epoch + 1, global_step, flux, ae, [clip_l, t5xxl], sample_prompts_te_outputs
+        )
+        optimizer_train_fn()
+
+    is_main_process = accelerator.is_main_process
+    # if is_main_process:
+    flux = accelerator.unwrap_model(flux)
+
+    accelerator.end_training()
+    optimizer_eval_fn()
+
+    if args.save_state or args.save_state_on_train_end:
+        train_util.save_state_on_train_end(args, accelerator)
+
+    del accelerator  # この後メモリを使うのでこれは消す
+
+    if is_main_process:
+        flux_train_utils.save_flux_model_on_train_end(args, save_dtype, epoch, global_step, flux)
+        logger.info("model saved.")
+
+
+def setup_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser()
+
+    add_logging_arguments(parser)
+    train_util.add_sd_models_arguments(parser)  # TODO split this
+    sai_model_spec.add_model_spec_arguments(parser)
+    train_util.add_dataset_arguments(parser, True, True, True)
+    train_util.add_training_arguments(parser, False)
+    train_util.add_masked_loss_arguments(parser)
+    deepspeed_utils.add_deepspeed_arguments(parser)
+    train_util.add_sd_saving_arguments(parser)
+    train_util.add_optimizer_arguments(parser)
+    config_util.add_config_arguments(parser)
+    add_custom_train_arguments(parser)  # TODO remove this from here
+    train_util.add_dit_training_arguments(parser)
+    flux_train_utils.add_flux_train_arguments(parser)
+
+    parser.add_argument(
+        "--mem_eff_save",
+        action="store_true",
+        help="[EXPERIMENTAL] use memory efficient custom model saving method / メモリ効率の良い独自のモデル保存方法を使う",
+    )
+
+    parser.add_argument(
+        "--fused_optimizer_groups",
+        type=int,
+        default=None,
+        help="**this option is not working** will be removed in the future / このオプションは動作しません。将来削除されます",
+    )
+    parser.add_argument(
+        "--blockwise_fused_optimizers",
+        action="store_true",
+        help="enable blockwise optimizers for fused backward pass and optimizer step / fused backward passとoptimizer step のためブロック単位のoptimizerを有効にする",
+    )
+    parser.add_argument(
+        "--skip_latents_validity_check",
+        action="store_true",
+        help="[Deprecated] use 'skip_cache_check' instead / 代わりに 'skip_cache_check' を使用してください",
+    )
+    parser.add_argument(
+        "--double_blocks_to_swap",
+        type=int,
+        default=None,
+        help="[Deprecated] use 'blocks_to_swap' instead / 代わりに 'blocks_to_swap' を使用してください",
+    )
+    parser.add_argument(
+        "--single_blocks_to_swap",
+        type=int,
+        default=None,
+        help="[Deprecated] use 'blocks_to_swap' instead / 代わりに 'blocks_to_swap' を使用してください",
+    )
+    parser.add_argument(
+        "--cpu_offload_checkpointing",
+        action="store_true",
+        help="[EXPERIMENTAL] enable offloading of tensors to CPU during checkpointing / チェックポイント時にテンソルをCPUにオフロードする",
+    )
+    return parser
+
+
+if __name__ == "__main__":
+    parser = setup_parser()
+
+    args = parser.parse_args()
+    train_util.verify_command_line_training_args(args)
+    args = train_util.read_config_from_file(args, parser)
+
+    train(args)

flux_train_control_net.py

@@ -0,0 +1,885 @@
+# training with captions
+
+# Swap blocks between CPU and GPU:
+# This implementation is inspired by and based on the work of 2kpr.
+# Many thanks to 2kpr for the original concept and implementation of memory-efficient offloading.
+# The original idea has been adapted and extended to fit the current project's needs.
+
+# Key features:
+# - CPU offloading during forward and backward passes
+# - Use of fused optimizer and grad_hook for efficient gradient processing
+# - Per-block fused optimizer instances
+
+import argparse
+import copy
+import math
+import os
+import time
+from concurrent.futures import ThreadPoolExecutor
+from multiprocessing import Value
+from typing import List, Optional, Tuple, Union
+
+import toml
+import torch
+import torch.nn as nn
+from tqdm import tqdm
+
+from library import utils
+from library.device_utils import clean_memory_on_device, init_ipex
+
+init_ipex()
+
+from accelerate.utils import set_seed
+
+import library.train_util as train_util
+import library.sai_model_spec as sai_model_spec
+from library import (
+    deepspeed_utils,
+    flux_train_utils,
+    flux_utils,
+    strategy_base,
+    strategy_flux,
+)
+from library.sd3_train_utils import FlowMatchEulerDiscreteScheduler
+from library.utils import add_logging_arguments, setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+import library.config_util as config_util
+
+# import library.sdxl_train_util as sdxl_train_util
+from library.config_util import (
+    BlueprintGenerator,
+    ConfigSanitizer,
+)
+from library.custom_train_functions import add_custom_train_arguments, apply_masked_loss
+
+
+def train(args):
+    train_util.verify_training_args(args)
+    train_util.prepare_dataset_args(args, True)
+    # sdxl_train_util.verify_sdxl_training_args(args)
+    deepspeed_utils.prepare_deepspeed_args(args)
+    setup_logging(args, reset=True)
+
+    # temporary: backward compatibility for deprecated options. remove in the future
+    if not args.skip_cache_check:
+        args.skip_cache_check = args.skip_latents_validity_check
+
+    if args.model_type != "flux":
+        raise ValueError(
+            f"FLUX.1 ControlNet training requires model_type='flux'. / FLUX.1 ControlNetの学習にはmodel_type='flux'を指定してください。"
+        )
+
+    # assert (
+    #     not args.weighted_captions
+    # ), "weighted_captions is not supported currently / weighted_captionsは現在サポートされていません"
+    if args.cache_text_encoder_outputs_to_disk and not args.cache_text_encoder_outputs:
+        logger.warning(
+            "cache_text_encoder_outputs_to_disk is enabled, so cache_text_encoder_outputs is also enabled / cache_text_encoder_outputs_to_diskが有効になっているため、cache_text_encoder_outputsも有効になります"
+        )
+        args.cache_text_encoder_outputs = True
+
+    if args.cpu_offload_checkpointing and not args.gradient_checkpointing:
+        logger.warning(
+            "cpu_offload_checkpointing is enabled, so gradient_checkpointing is also enabled / cpu_offload_checkpointingが有効になっているため、gradient_checkpointingも有効になります"
+        )
+        args.gradient_checkpointing = True
+
+    assert (
+        args.blocks_to_swap is None or args.blocks_to_swap == 0
+    ) or not args.cpu_offload_checkpointing, (
+        "blocks_to_swap is not supported with cpu_offload_checkpointing / blocks_to_swapはcpu_offload_checkpointingと併用できません"
+    )
+
+    cache_latents = args.cache_latents
+
+    if args.seed is not None:
+        set_seed(args.seed)  # 乱数系列を初期化する
+
+    # prepare caching strategy: this must be set before preparing dataset. because dataset may use this strategy for initialization.
+    if args.cache_latents:
+        latents_caching_strategy = strategy_flux.FluxLatentsCachingStrategy(
+            args.cache_latents_to_disk, args.vae_batch_size, args.skip_cache_check
+        )
+        strategy_base.LatentsCachingStrategy.set_strategy(latents_caching_strategy)
+
+    # データセットを準備する
+    if args.dataset_class is None:
+        blueprint_generator = BlueprintGenerator(ConfigSanitizer(False, False, True, True))
+        if args.dataset_config is not None:
+            logger.info(f"Load dataset config from {args.dataset_config}")
+            user_config = config_util.load_user_config(args.dataset_config)
+            ignored = ["train_data_dir", "conditioning_data_dir"]
+            if any(getattr(args, attr) is not None for attr in ignored):
+                logger.warning(
+                    "ignore following options because config file is found: {0} / 設定ファイルが利用されるため以下のオプションは無視されます: {0}".format(
+                        ", ".join(ignored)
+                    )
+                )
+        else:
+            user_config = {
+                "datasets": [
+                    {
+                        "subsets": config_util.generate_controlnet_subsets_config_by_subdirs(
+                            args.train_data_dir, args.conditioning_data_dir, args.caption_extension
+                        )
+                    }
+                ]
+            }
+
+        blueprint = blueprint_generator.generate(user_config, args)
+        train_dataset_group, val_dataset_group = config_util.generate_dataset_group_by_blueprint(blueprint.dataset_group)
+    else:
+        train_dataset_group = train_util.load_arbitrary_dataset(args)
+        val_dataset_group = None
+
+    current_epoch = Value("i", 0)
+    current_step = Value("i", 0)
+    ds_for_collator = train_dataset_group if args.max_data_loader_n_workers == 0 else None
+    collator = train_util.collator_class(current_epoch, current_step, ds_for_collator)
+
+    train_dataset_group.verify_bucket_reso_steps(16)  # TODO これでいいか確認
+
+    _, is_schnell, _, _ = flux_utils.analyze_checkpoint_state(args.pretrained_model_name_or_path)
+    if args.debug_dataset:
+        if args.cache_text_encoder_outputs:
+            strategy_base.TextEncoderOutputsCachingStrategy.set_strategy(
+                strategy_flux.FluxTextEncoderOutputsCachingStrategy(
+                    args.cache_text_encoder_outputs_to_disk, args.text_encoder_batch_size, args.skip_cache_check, False
+                )
+            )
+        t5xxl_max_token_length = (
+            args.t5xxl_max_token_length if args.t5xxl_max_token_length is not None else (256 if is_schnell else 512)
+        )
+        strategy_base.TokenizeStrategy.set_strategy(strategy_flux.FluxTokenizeStrategy(t5xxl_max_token_length))
+
+        train_dataset_group.set_current_strategies()
+        train_util.debug_dataset(train_dataset_group, True)
+        return
+    if len(train_dataset_group) == 0:
+        logger.error(
+            "No data found. Please verify the metadata file and train_data_dir option. / 画像がありません。メタデータおよびtrain_data_dirオプションを確認してください。"
+        )
+        return
+
+    if cache_latents:
+        assert (
+            train_dataset_group.is_latent_cacheable()
+        ), "when caching latents, either color_aug or random_crop cannot be used / latentをキャッシュするときはcolor_augとrandom_cropは使えません"
+
+    if args.cache_text_encoder_outputs:
+        assert (
+            train_dataset_group.is_text_encoder_output_cacheable()
+        ), "when caching text encoder output, either caption_dropout_rate, shuffle_caption, token_warmup_step or caption_tag_dropout_rate cannot be used / text encoderの出力をキャッシュするときはcaption_dropout_rate, shuffle_caption, token_warmup_step, caption_tag_dropout_rateは使えません"
+
+    # acceleratorを準備する
+    logger.info("prepare accelerator")
+    accelerator = train_util.prepare_accelerator(args)
+
+    # mixed precisionに対応した型を用意しておき適宜castする
+    weight_dtype, save_dtype = train_util.prepare_dtype(args)
+
+    # モデルを読み込む
+
+    # load VAE for caching latents
+    ae = None
+    if cache_latents:
+        ae = flux_utils.load_ae(args.ae, weight_dtype, "cpu", args.disable_mmap_load_safetensors)
+        ae.to(accelerator.device, dtype=weight_dtype)
+        ae.requires_grad_(False)
+        ae.eval()
+
+        train_dataset_group.new_cache_latents(ae, accelerator)
+
+        ae.to("cpu")  # if no sampling, vae can be deleted
+        clean_memory_on_device(accelerator.device)
+
+        accelerator.wait_for_everyone()
+
+    # prepare tokenize strategy
+    if args.t5xxl_max_token_length is None:
+        if is_schnell:
+            t5xxl_max_token_length = 256
+        else:
+            t5xxl_max_token_length = 512
+    else:
+        t5xxl_max_token_length = args.t5xxl_max_token_length
+
+    flux_tokenize_strategy = strategy_flux.FluxTokenizeStrategy(t5xxl_max_token_length)
+    strategy_base.TokenizeStrategy.set_strategy(flux_tokenize_strategy)
+
+    # load clip_l, t5xxl for caching text encoder outputs
+    clip_l = flux_utils.load_clip_l(args.clip_l, weight_dtype, "cpu", args.disable_mmap_load_safetensors)
+    t5xxl = flux_utils.load_t5xxl(args.t5xxl, weight_dtype, "cpu", args.disable_mmap_load_safetensors)
+    clip_l.eval()
+    t5xxl.eval()
+    clip_l.requires_grad_(False)
+    t5xxl.requires_grad_(False)
+
+    text_encoding_strategy = strategy_flux.FluxTextEncodingStrategy(args.apply_t5_attn_mask)
+    strategy_base.TextEncodingStrategy.set_strategy(text_encoding_strategy)
+
+    # cache text encoder outputs
+    sample_prompts_te_outputs = None
+    if args.cache_text_encoder_outputs:
+        # Text Encodes are eval and no grad here
+        clip_l.to(accelerator.device)
+        t5xxl.to(accelerator.device)
+
+        text_encoder_caching_strategy = strategy_flux.FluxTextEncoderOutputsCachingStrategy(
+            args.cache_text_encoder_outputs_to_disk, args.text_encoder_batch_size, False, False, args.apply_t5_attn_mask
+        )
+        strategy_base.TextEncoderOutputsCachingStrategy.set_strategy(text_encoder_caching_strategy)
+
+        with accelerator.autocast():
+            train_dataset_group.new_cache_text_encoder_outputs([clip_l, t5xxl], accelerator)
+
+        # cache sample prompt's embeddings to free text encoder's memory
+        if args.sample_prompts is not None:
+            logger.info(f"cache Text Encoder outputs for sample prompt: {args.sample_prompts}")
+
+            text_encoding_strategy: strategy_flux.FluxTextEncodingStrategy = strategy_base.TextEncodingStrategy.get_strategy()
+
+            prompts = train_util.load_prompts(args.sample_prompts)
+            sample_prompts_te_outputs = {}  # key: prompt, value: text encoder outputs
+            with accelerator.autocast(), torch.no_grad():
+                for prompt_dict in prompts:
+                    for p in [prompt_dict.get("prompt", ""), prompt_dict.get("negative_prompt", "")]:
+                        if p not in sample_prompts_te_outputs:
+                            logger.info(f"cache Text Encoder outputs for prompt: {p}")
+                            tokens_and_masks = flux_tokenize_strategy.tokenize(p)
+                            sample_prompts_te_outputs[p] = text_encoding_strategy.encode_tokens(
+                                flux_tokenize_strategy, [clip_l, t5xxl], tokens_and_masks, args.apply_t5_attn_mask
+                            )
+
+        accelerator.wait_for_everyone()
+
+        # now we can delete Text Encoders to free memory
+        clip_l = None
+        t5xxl = None
+        clean_memory_on_device(accelerator.device)
+
+    # load FLUX
+    is_schnell, flux = flux_utils.load_flow_model(
+        args.pretrained_model_name_or_path, weight_dtype, "cpu", args.disable_mmap_load_safetensors, model_type="flux"
+    )
+    flux.requires_grad_(False)
+
+    # load controlnet
+    controlnet_dtype = torch.float32 if args.deepspeed else weight_dtype
+    controlnet = flux_utils.load_controlnet(
+        args.controlnet_model_name_or_path, is_schnell, controlnet_dtype, accelerator.device, args.disable_mmap_load_safetensors
+    )
+    controlnet.train()
+
+    if args.gradient_checkpointing:
+        if not args.deepspeed:
+            flux.enable_gradient_checkpointing(cpu_offload=args.cpu_offload_checkpointing)
+        controlnet.enable_gradient_checkpointing(cpu_offload=args.cpu_offload_checkpointing)
+
+    # block swap
+
+    # backward compatibility
+    if args.blocks_to_swap is None:
+        blocks_to_swap = args.double_blocks_to_swap or 0
+        if args.single_blocks_to_swap is not None:
+            blocks_to_swap += args.single_blocks_to_swap // 2
+        if blocks_to_swap > 0:
+            logger.warning(
+                "double_blocks_to_swap and single_blocks_to_swap are deprecated. Use blocks_to_swap instead."
+                " / double_blocks_to_swapとsingle_blocks_to_swapは非推奨です。blocks_to_swapを使ってください。"
+            )
+            logger.info(
+                f"double_blocks_to_swap={args.double_blocks_to_swap} and single_blocks_to_swap={args.single_blocks_to_swap} are converted to blocks_to_swap={blocks_to_swap}."
+            )
+            args.blocks_to_swap = blocks_to_swap
+        del blocks_to_swap
+
+    is_swapping_blocks = args.blocks_to_swap is not None and args.blocks_to_swap > 0
+    if is_swapping_blocks:
+        # Swap blocks between CPU and GPU to reduce memory usage, in forward and backward passes.
+        # This idea is based on 2kpr's great work. Thank you!
+        logger.info(f"enable block swap: blocks_to_swap={args.blocks_to_swap}")
+        flux.enable_block_swap(args.blocks_to_swap, accelerator.device)
+        flux.move_to_device_except_swap_blocks(accelerator.device)  # reduce peak memory usage
+        # ControlNet only has two blocks, so we can keep it on GPU
+        # controlnet.enable_block_swap(args.blocks_to_swap, accelerator.device)
+    else:
+        flux.to(accelerator.device)
+
+    if not cache_latents:
+        # load VAE here if not cached
+        ae = flux_utils.load_ae(args.ae, weight_dtype, "cpu")
+        ae.requires_grad_(False)
+        ae.eval()
+        ae.to(accelerator.device, dtype=weight_dtype)
+
+    training_models = []
+    params_to_optimize = []
+    training_models.append(controlnet)
+    name_and_params = list(controlnet.named_parameters())
+    # single param group for now
+    params_to_optimize.append({"params": [p for _, p in name_and_params], "lr": args.learning_rate})
+    param_names = [[n for n, _ in name_and_params]]
+
+    # calculate number of trainable parameters
+    n_params = 0
+    for group in params_to_optimize:
+        for p in group["params"]:
+            n_params += p.numel()
+
+    accelerator.print(f"number of trainable parameters: {n_params}")
+
+    # 学習に必要なクラスを準備する
+    accelerator.print("prepare optimizer, data loader etc.")
+
+    if args.blockwise_fused_optimizers:
+        # fused backward pass: https://pytorch.org/tutorials/intermediate/optimizer_step_in_backward_tutorial.html
+        # Instead of creating an optimizer for all parameters as in the tutorial, we create an optimizer for each block of parameters.
+        # This balances memory usage and management complexity.
+
+        # split params into groups. currently different learning rates are not supported
+        grouped_params = []
+        param_group = {}
+        for group in params_to_optimize:
+            named_parameters = list(controlnet.named_parameters())
+            assert len(named_parameters) == len(group["params"]), "number of parameters does not match"
+            for p, np in zip(group["params"], named_parameters):
+                # determine target layer and block index for each parameter
+                block_type = "other"  # double, single or other
+                if np[0].startswith("double_blocks"):
+                    block_index = int(np[0].split(".")[1])
+                    block_type = "double"
+                elif np[0].startswith("single_blocks"):
+                    block_index = int(np[0].split(".")[1])
+                    block_type = "single"
+                else:
+                    block_index = -1
+
+                param_group_key = (block_type, block_index)
+                if param_group_key not in param_group:
+                    param_group[param_group_key] = []
+                param_group[param_group_key].append(p)
+
+        block_types_and_indices = []
+        for param_group_key, param_group in param_group.items():
+            block_types_and_indices.append(param_group_key)
+            grouped_params.append({"params": param_group, "lr": args.learning_rate})
+
+            num_params = 0
+            for p in param_group:
+                num_params += p.numel()
+            accelerator.print(f"block {param_group_key}: {num_params} parameters")
+
+        # prepare optimizers for each group
+        optimizers = []
+        for group in grouped_params:
+            _, _, optimizer = train_util.get_optimizer(args, trainable_params=[group])
+            optimizers.append(optimizer)
+        optimizer = optimizers[0]  # avoid error in the following code
+
+        logger.info(f"using {len(optimizers)} optimizers for blockwise fused optimizers")
+
+        if train_util.is_schedulefree_optimizer(optimizers[0], args):
+            raise ValueError("Schedule-free optimizer is not supported with blockwise fused optimizers")
+        optimizer_train_fn = lambda: None  # dummy function
+        optimizer_eval_fn = lambda: None  # dummy function
+    else:
+        _, _, optimizer = train_util.get_optimizer(args, trainable_params=params_to_optimize)
+        optimizer_train_fn, optimizer_eval_fn = train_util.get_optimizer_train_eval_fn(optimizer, args)
+
+    # prepare dataloader
+    # strategies are set here because they cannot be referenced in another process. Copy them with the dataset
+    # some strategies can be None
+    train_dataset_group.set_current_strategies()
+
+    # DataLoaderのプロセス数：0 は persistent_workers が使えないので注意
+    n_workers = min(args.max_data_loader_n_workers, os.cpu_count())  # cpu_count or max_data_loader_n_workers
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset_group,
+        batch_size=1,
+        shuffle=True,
+        collate_fn=collator,
+        num_workers=n_workers,
+        persistent_workers=args.persistent_data_loader_workers,
+    )
+
+    # 学習ステップ数を計算する
+    if args.max_train_epochs is not None:
+        args.max_train_steps = args.max_train_epochs * math.ceil(
+            len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
+        )
+        accelerator.print(
+            f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
+        )
+
+    # データセット側にも学習ステップを送信
+    train_dataset_group.set_max_train_steps(args.max_train_steps)
+
+    # lr schedulerを用意する
+    if args.blockwise_fused_optimizers:
+        # prepare lr schedulers for each optimizer
+        lr_schedulers = [train_util.get_scheduler_fix(args, optimizer, accelerator.num_processes) for optimizer in optimizers]
+        lr_scheduler = lr_schedulers[0]  # avoid error in the following code
+    else:
+        lr_scheduler = train_util.get_scheduler_fix(args, optimizer, accelerator.num_processes)
+
+    # 実験的機能：勾配も含めたfp16/bf16学習を行う　モデル全体をfp16/bf16にする
+    if args.full_fp16:
+        assert (
+            args.mixed_precision == "fp16"
+        ), "full_fp16 requires mixed precision='fp16' / full_fp16を使う場合はmixed_precision='fp16'を指定してください。"
+        accelerator.print("enable full fp16 training.")
+        flux.to(weight_dtype)
+        controlnet.to(weight_dtype)
+        if clip_l is not None:
+            clip_l.to(weight_dtype)
+            t5xxl.to(weight_dtype)  # TODO check works with fp16 or not
+    elif args.full_bf16:
+        assert (
+            args.mixed_precision == "bf16"
+        ), "full_bf16 requires mixed precision='bf16' / full_bf16を使う場合はmixed_precision='bf16'を指定してください。"
+        accelerator.print("enable full bf16 training.")
+        flux.to(weight_dtype)
+        controlnet.to(weight_dtype)
+        if clip_l is not None:
+            clip_l.to(weight_dtype)
+            t5xxl.to(weight_dtype)
+
+    # if we don't cache text encoder outputs, move them to device
+    if not args.cache_text_encoder_outputs:
+        clip_l.to(accelerator.device)
+        t5xxl.to(accelerator.device)
+
+    clean_memory_on_device(accelerator.device)
+
+    if args.deepspeed:
+        ds_model = deepspeed_utils.prepare_deepspeed_model(args, mmdit=controlnet)
+        # most of ZeRO stage uses optimizer partitioning, so we have to prepare optimizer and ds_model at the same time. # pull/1139#issuecomment-1986790007
+        ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+            ds_model, optimizer, train_dataloader, lr_scheduler
+        )
+        training_models = [ds_model]
+
+    else:
+        # accelerator does some magic
+        # if we doesn't swap blocks, we can move the model to device
+        controlnet = accelerator.prepare(controlnet)  # , device_placement=[not is_swapping_blocks])
+        optimizer, train_dataloader, lr_scheduler = accelerator.prepare(optimizer, train_dataloader, lr_scheduler)
+
+    # 実験的機能：勾配も含めたfp16学習を行う　PyTorchにパッチを当ててfp16でのgrad scaleを有効にする
+    if args.full_fp16:
+        # During deepseed training, accelerate not handles fp16/bf16|mixed precision directly via scaler. Let deepspeed engine do.
+        # -> But we think it's ok to patch accelerator even if deepspeed is enabled.
+        train_util.patch_accelerator_for_fp16_training(accelerator)
+
+    # resumeする
+    train_util.resume_from_local_or_hf_if_specified(accelerator, args)
+
+    if args.fused_backward_pass:
+        # use fused optimizer for backward pass: other optimizers will be supported in the future
+        import library.adafactor_fused
+
+        library.adafactor_fused.patch_adafactor_fused(optimizer)
+
+        for param_group, param_name_group in zip(optimizer.param_groups, param_names):
+            for parameter, param_name in zip(param_group["params"], param_name_group):
+                if parameter.requires_grad:
+
+                    def create_grad_hook(p_name, p_group):
+                        def grad_hook(tensor: torch.Tensor):
+                            if accelerator.sync_gradients and args.max_grad_norm != 0.0:
+                                accelerator.clip_grad_norm_(tensor, args.max_grad_norm)
+                            optimizer.step_param(tensor, p_group)
+                            tensor.grad = None
+
+                        return grad_hook
+
+                    parameter.register_post_accumulate_grad_hook(create_grad_hook(param_name, param_group))
+
+    elif args.blockwise_fused_optimizers:
+        # prepare for additional optimizers and lr schedulers
+        for i in range(1, len(optimizers)):
+            optimizers[i] = accelerator.prepare(optimizers[i])
+            lr_schedulers[i] = accelerator.prepare(lr_schedulers[i])
+
+        # counters are used to determine when to step the optimizer
+        global optimizer_hooked_count
+        global num_parameters_per_group
+        global parameter_optimizer_map
+
+        optimizer_hooked_count = {}
+        num_parameters_per_group = [0] * len(optimizers)
+        parameter_optimizer_map = {}
+
+        for opt_idx, optimizer in enumerate(optimizers):
+            for param_group in optimizer.param_groups:
+                for parameter in param_group["params"]:
+                    if parameter.requires_grad:
+
+                        def grad_hook(parameter: torch.Tensor):
+                            if accelerator.sync_gradients and args.max_grad_norm != 0.0:
+                                accelerator.clip_grad_norm_(parameter, args.max_grad_norm)
+
+                            i = parameter_optimizer_map[parameter]
+                            optimizer_hooked_count[i] += 1
+                            if optimizer_hooked_count[i] == num_parameters_per_group[i]:
+                                optimizers[i].step()
+                                optimizers[i].zero_grad(set_to_none=True)
+
+                        parameter.register_post_accumulate_grad_hook(grad_hook)
+                        parameter_optimizer_map[parameter] = opt_idx
+                        num_parameters_per_group[opt_idx] += 1
+
+    # epoch数を計算する
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
+    if (args.save_n_epoch_ratio is not None) and (args.save_n_epoch_ratio > 0):
+        args.save_every_n_epochs = math.floor(num_train_epochs / args.save_n_epoch_ratio) or 1
+
+    # 学習する
+    # total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
+    accelerator.print("running training / 学習開始")
+    accelerator.print(f"  num examples / サンプル数: {train_dataset_group.num_train_images}")
+    accelerator.print(f"  num batches per epoch / 1epochのバッチ数: {len(train_dataloader)}")
+    accelerator.print(f"  num epochs / epoch数: {num_train_epochs}")
+    accelerator.print(
+        f"  batch size per device / バッチサイズ: {', '.join([str(d.batch_size) for d in train_dataset_group.datasets])}"
+    )
+    # accelerator.print(
+    #     f"  total train batch size (with parallel & distributed & accumulation) / 総バッチサイズ（並列学習、勾配合計含む）: {total_batch_size}"
+    # )
+    accelerator.print(f"  gradient accumulation steps / 勾配を合計するステップ数 = {args.gradient_accumulation_steps}")
+    accelerator.print(f"  total optimization steps / 学習ステップ数: {args.max_train_steps}")
+
+    progress_bar = tqdm(range(args.max_train_steps), smoothing=0, disable=not accelerator.is_local_main_process, desc="steps")
+    global_step = 0
+
+    noise_scheduler = FlowMatchEulerDiscreteScheduler(num_train_timesteps=1000, shift=args.discrete_flow_shift)
+    noise_scheduler_copy = copy.deepcopy(noise_scheduler)
+
+    if accelerator.is_main_process:
+        init_kwargs = {}
+        if args.wandb_run_name:
+            init_kwargs["wandb"] = {"name": args.wandb_run_name}
+        if args.log_tracker_config is not None:
+            init_kwargs = toml.load(args.log_tracker_config)
+        accelerator.init_trackers(
+            "finetuning" if args.log_tracker_name is None else args.log_tracker_name,
+            config=train_util.get_sanitized_config_or_none(args),
+            init_kwargs=init_kwargs,
+        )
+
+    if is_swapping_blocks:
+        flux.prepare_block_swap_before_forward()
+
+    # For --sample_at_first
+    optimizer_eval_fn()
+    flux_train_utils.sample_images(
+        accelerator, args, 0, global_step, flux, ae, [clip_l, t5xxl], sample_prompts_te_outputs, controlnet=controlnet
+    )
+    optimizer_train_fn()
+    if len(accelerator.trackers) > 0:
+        # log empty object to commit the sample images to wandb
+        accelerator.log({}, step=0)
+
+    loss_recorder = train_util.LossRecorder()
+    epoch = 0  # avoid error when max_train_steps is 0
+    for epoch in range(num_train_epochs):
+        accelerator.print(f"\nepoch {epoch+1}/{num_train_epochs}")
+        current_epoch.value = epoch + 1
+
+        for m in training_models:
+            m.train()
+
+        for step, batch in enumerate(train_dataloader):
+            current_step.value = global_step
+
+            if args.blockwise_fused_optimizers:
+                optimizer_hooked_count = {i: 0 for i in range(len(optimizers))}  # reset counter for each step
+
+            with accelerator.accumulate(*training_models):
+                if "latents" in batch and batch["latents"] is not None:
+                    latents = batch["latents"].to(accelerator.device, dtype=weight_dtype)
+                else:
+                    with torch.no_grad():
+                        # encode images to latents. images are [-1, 1]
+                        latents = ae.encode(batch["images"].to(ae.dtype)).to(accelerator.device, dtype=weight_dtype)
+
+                    # NaNが含まれていれば警告を表示し0に置き換える
+                    if torch.any(torch.isnan(latents)):
+                        accelerator.print("NaN found in latents, replacing with zeros")
+                        latents = torch.nan_to_num(latents, 0, out=latents)
+
+                text_encoder_outputs_list = batch.get("text_encoder_outputs_list", None)
+                if text_encoder_outputs_list is not None:
+                    text_encoder_conds = text_encoder_outputs_list
+                else:
+                    # not cached or training, so get from text encoders
+                    tokens_and_masks = batch["input_ids_list"]
+                    with torch.no_grad():
+                        input_ids = [ids.to(accelerator.device) for ids in batch["input_ids_list"]]
+                        text_encoder_conds = text_encoding_strategy.encode_tokens(
+                            flux_tokenize_strategy, [clip_l, t5xxl], input_ids, args.apply_t5_attn_mask
+                        )
+                text_encoder_conds = [c.to(weight_dtype) for c in text_encoder_conds]
+
+                # TODO support some features for noise implemented in get_noise_noisy_latents_and_timesteps
+
+                # Sample noise that we'll add to the latents
+                noise = torch.randn_like(latents)
+                bsz = latents.shape[0]
+
+                # get noisy model input and timesteps
+                noisy_model_input, timesteps, sigmas = flux_train_utils.get_noisy_model_input_and_timesteps(
+                    args, noise_scheduler_copy, latents, noise, accelerator.device, weight_dtype
+                )
+
+                # pack latents and get img_ids
+                packed_noisy_model_input = flux_utils.pack_latents(noisy_model_input)  # b, c, h*2, w*2 -> b, h*w, c*4
+                packed_latent_height, packed_latent_width = noisy_model_input.shape[2] // 2, noisy_model_input.shape[3] // 2
+                img_ids = (
+                    flux_utils.prepare_img_ids(bsz, packed_latent_height, packed_latent_width)
+                    .to(device=accelerator.device)
+                    .to(weight_dtype)
+                )
+
+                # get guidance: ensure args.guidance_scale is float
+                guidance_vec = torch.full((bsz,), float(args.guidance_scale), device=accelerator.device, dtype=weight_dtype)
+
+                # call model
+                l_pooled, t5_out, txt_ids, t5_attn_mask = text_encoder_conds
+                if not args.apply_t5_attn_mask:
+                    t5_attn_mask = None
+
+                with accelerator.autocast():
+                    block_samples, block_single_samples = controlnet(
+                        img=packed_noisy_model_input,
+                        img_ids=img_ids,
+                        controlnet_cond=batch["conditioning_images"].to(accelerator.device).to(weight_dtype),
+                        txt=t5_out,
+                        txt_ids=txt_ids,
+                        y=l_pooled,
+                        timesteps=timesteps / 1000,
+                        guidance=guidance_vec,
+                        txt_attention_mask=t5_attn_mask,
+                    )
+                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transformer model (we should not keep it but I want to keep the inputs same for the model for testing)
+                    model_pred = flux(
+                        img=packed_noisy_model_input,
+                        img_ids=img_ids,
+                        txt=t5_out,
+                        txt_ids=txt_ids,
+                        y=l_pooled,
+                        block_controlnet_hidden_states=block_samples,
+                        block_controlnet_single_hidden_states=block_single_samples,
+                        timesteps=timesteps / 1000,
+                        guidance=guidance_vec,
+                        txt_attention_mask=t5_attn_mask,
+                    )
+
+                # unpack latents
+                model_pred = flux_utils.unpack_latents(model_pred, packed_latent_height, packed_latent_width)
+
+                # apply model prediction type
+                model_pred, weighting = flux_train_utils.apply_model_prediction_type(args, model_pred, noisy_model_input, sigmas)
+
+                # flow matching loss: this is different from SD3
+                target = noise - latents
+
+                # calculate loss
+                loss = train_util.conditional_loss(
+                    model_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=None
+                )
+                if weighting is not None:
+                    loss = loss * weighting
+                if args.masked_loss or ("alpha_masks" in batch and batch["alpha_masks"] is not None):
+                    loss = apply_masked_loss(loss, batch)
+                loss = loss.mean([1, 2, 3])
+
+                loss_weights = batch["loss_weights"]  # 各sampleごとのweight
+                loss = loss * loss_weights
+                loss = loss.mean()
+
+                # backward
+                accelerator.backward(loss)
+
+                if not (args.fused_backward_pass or args.blockwise_fused_optimizers):
+                    if accelerator.sync_gradients and args.max_grad_norm != 0.0:
+                        params_to_clip = []
+                        for m in training_models:
+                            params_to_clip.extend(m.parameters())
+                        accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
+
+                    optimizer.step()
+                    lr_scheduler.step()
+                    optimizer.zero_grad(set_to_none=True)
+                else:
+                    # optimizer.step() and optimizer.zero_grad() are called in the optimizer hook
+                    lr_scheduler.step()
+                    if args.blockwise_fused_optimizers:
+                        for i in range(1, len(optimizers)):
+                            lr_schedulers[i].step()
+
+            # Checks if the accelerator has performed an optimization step behind the scenes
+            if accelerator.sync_gradients:
+                progress_bar.update(1)
+                global_step += 1
+
+                optimizer_eval_fn()
+                flux_train_utils.sample_images(
+                    accelerator,
+                    args,
+                    None,
+                    global_step,
+                    flux,
+                    ae,
+                    [clip_l, t5xxl],
+                    sample_prompts_te_outputs,
+                    controlnet=controlnet,
+                )
+
+                # 指定ステップごとにモデルを保存
+                if args.save_every_n_steps is not None and global_step % args.save_every_n_steps == 0:
+                    accelerator.wait_for_everyone()
+                    if accelerator.is_main_process:
+                        flux_train_utils.save_flux_model_on_epoch_end_or_stepwise(
+                            args,
+                            False,
+                            accelerator,
+                            save_dtype,
+                            epoch,
+                            num_train_epochs,
+                            global_step,
+                            accelerator.unwrap_model(controlnet),
+                        )
+                optimizer_train_fn()
+
+            current_loss = loss.detach().item()  # 平均なのでbatch sizeは関係ないはず
+            if len(accelerator.trackers) > 0:
+                logs = {"loss": current_loss}
+                train_util.append_lr_to_logs(logs, lr_scheduler, args.optimizer_type, including_unet=True)
+
+                accelerator.log(logs, step=global_step)
+
+            loss_recorder.add(epoch=epoch, step=step, loss=current_loss)
+            avr_loss: float = loss_recorder.moving_average
+            logs = {"avr_loss": avr_loss}  # , "lr": lr_scheduler.get_last_lr()[0]}
+            progress_bar.set_postfix(**logs)
+
+            if global_step >= args.max_train_steps:
+                break
+
+        if len(accelerator.trackers) > 0:
+            logs = {"loss/epoch": loss_recorder.moving_average}
+            accelerator.log(logs, step=epoch + 1)
+
+        accelerator.wait_for_everyone()
+
+        optimizer_eval_fn()
+        if args.save_every_n_epochs is not None:
+            if accelerator.is_main_process:
+                flux_train_utils.save_flux_model_on_epoch_end_or_stepwise(
+                    args,
+                    True,
+                    accelerator,
+                    save_dtype,
+                    epoch,
+                    num_train_epochs,
+                    global_step,
+                    accelerator.unwrap_model(controlnet),
+                )
+
+        flux_train_utils.sample_images(
+            accelerator, args, epoch + 1, global_step, flux, ae, [clip_l, t5xxl], sample_prompts_te_outputs, controlnet=controlnet
+        )
+        optimizer_train_fn()
+
+    is_main_process = accelerator.is_main_process
+    # if is_main_process:
+    controlnet = accelerator.unwrap_model(controlnet)
+
+    accelerator.end_training()
+    optimizer_eval_fn()
+
+    if args.save_state or args.save_state_on_train_end:
+        train_util.save_state_on_train_end(args, accelerator)
+
+    del accelerator  # この後メモリを使うのでこれは消す
+
+    if is_main_process:
+        flux_train_utils.save_flux_model_on_train_end(args, save_dtype, epoch, global_step, controlnet)
+        logger.info("model saved.")
+
+
+def setup_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser()
+
+    add_logging_arguments(parser)
+    train_util.add_sd_models_arguments(parser)  # TODO split this
+    sai_model_spec.add_model_spec_arguments(parser)
+    train_util.add_dataset_arguments(parser, False, True, True)
+    train_util.add_training_arguments(parser, False)
+    train_util.add_masked_loss_arguments(parser)
+    deepspeed_utils.add_deepspeed_arguments(parser)
+    train_util.add_sd_saving_arguments(parser)
+    train_util.add_optimizer_arguments(parser)
+    config_util.add_config_arguments(parser)
+    add_custom_train_arguments(parser)  # TODO remove this from here
+    train_util.add_dit_training_arguments(parser)
+    flux_train_utils.add_flux_train_arguments(parser)
+
+    parser.add_argument(
+        "--mem_eff_save",
+        action="store_true",
+        help="[EXPERIMENTAL] use memory efficient custom model saving method / メモリ効率の良い独自のモデル保存方法を使う",
+    )
+
+    parser.add_argument(
+        "--fused_optimizer_groups",
+        type=int,
+        default=None,
+        help="**this option is not working** will be removed in the future / このオプションは動作しません。将来削除されます",
+    )
+    parser.add_argument(
+        "--blockwise_fused_optimizers",
+        action="store_true",
+        help="enable blockwise optimizers for fused backward pass and optimizer step / fused backward passとoptimizer step のためブロック単位のoptimizerを有効にする",
+    )
+    parser.add_argument(
+        "--skip_latents_validity_check",
+        action="store_true",
+        help="[Deprecated] use 'skip_cache_check' instead / 代わりに 'skip_cache_check' を使用してください",
+    )
+    parser.add_argument(
+        "--double_blocks_to_swap",
+        type=int,
+        default=None,
+        help="[Deprecated] use 'blocks_to_swap' instead / 代わりに 'blocks_to_swap' を使用してください",
+    )
+    parser.add_argument(
+        "--single_blocks_to_swap",
+        type=int,
+        default=None,
+        help="[Deprecated] use 'blocks_to_swap' instead / 代わりに 'blocks_to_swap' を使用してください",
+    )
+    parser.add_argument(
+        "--cpu_offload_checkpointing",
+        action="store_true",
+        help="[EXPERIMENTAL] enable offloading of tensors to CPU during checkpointing / チェックポイント時にテンソルをCPUにオフロードする",
+    )
+    return parser
+
+
+if __name__ == "__main__":
+    parser = setup_parser()
+
+    args = parser.parse_args()
+    train_util.verify_command_line_training_args(args)
+    args = train_util.read_config_from_file(args, parser)
+
+    train(args)

flux_train_network.py

@@ -0,0 +1,547 @@
+import argparse
+import copy
+import math
+import random
+from typing import Any, Optional, Union
+
+import torch
+from accelerate import Accelerator
+
+from library.device_utils import clean_memory_on_device, init_ipex
+
+init_ipex()
+
+import train_network
+from library import (
+    flux_models,
+    flux_train_utils,
+    flux_utils,
+    sd3_train_utils,
+    strategy_base,
+    strategy_flux,
+    train_util,
+)
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class FluxNetworkTrainer(train_network.NetworkTrainer):
+    def __init__(self):
+        super().__init__()
+        self.sample_prompts_te_outputs = None
+        self.is_schnell: Optional[bool] = None
+        self.is_swapping_blocks: bool = False
+        self.model_type: Optional[str] = None
+
+    def assert_extra_args(
+        self,
+        args,
+        train_dataset_group: Union[train_util.DatasetGroup, train_util.MinimalDataset],
+        val_dataset_group: Optional[train_util.DatasetGroup],
+    ):
+        super().assert_extra_args(args, train_dataset_group, val_dataset_group)
+        # sdxl_train_util.verify_sdxl_training_args(args)
+
+        self.model_type = args.model_type  # "flux" or "chroma"
+        if self.model_type != "chroma":
+            self.use_clip_l = True
+        else:
+            self.use_clip_l = False  # Chroma does not use CLIP-L
+            assert args.apply_t5_attn_mask, "apply_t5_attn_mask must be True for Chroma / Chromaではapply_t5_attn_maskを指定する必要があります"
+
+        if args.fp8_base_unet:
+            args.fp8_base = True  # if fp8_base_unet is enabled, fp8_base is also enabled for FLUX.1
+
+        if args.cache_text_encoder_outputs_to_disk and not args.cache_text_encoder_outputs:
+            logger.warning(
+                "cache_text_encoder_outputs_to_disk is enabled, so cache_text_encoder_outputs is also enabled / cache_text_encoder_outputs_to_diskが有効になっているため、cache_text_encoder_outputsも有効になります"
+            )
+            args.cache_text_encoder_outputs = True
+
+        if args.cache_text_encoder_outputs:
+            assert (
+                train_dataset_group.is_text_encoder_output_cacheable()
+            ), "when caching Text Encoder output, either caption_dropout_rate, shuffle_caption, token_warmup_step or caption_tag_dropout_rate cannot be used / Text Encoderの出力をキャッシュするときはcaption_dropout_rate, shuffle_caption, token_warmup_step, caption_tag_dropout_rateは使えません"
+
+        # prepare CLIP-L/T5XXL training flags
+        self.train_clip_l = not args.network_train_unet_only and self.use_clip_l
+        self.train_t5xxl = False  # default is False even if args.network_train_unet_only is False
+
+        if args.max_token_length is not None:
+            logger.warning("max_token_length is not used in Flux training / max_token_lengthはFluxのトレーニングでは使用されません")
+
+        assert (
+            args.blocks_to_swap is None or args.blocks_to_swap == 0
+        ) or not args.cpu_offload_checkpointing, "blocks_to_swap is not supported with cpu_offload_checkpointing / blocks_to_swapはcpu_offload_checkpointingと併用できません"
+
+        # deprecated split_mode option
+        if args.split_mode:
+            if args.blocks_to_swap is not None:
+                logger.warning(
+                    "split_mode is deprecated. Because `--blocks_to_swap` is set, `--split_mode` is ignored."
+                    " / split_modeは非推奨です。`--blocks_to_swap`が設定されているため、`--split_mode`は無視されます。"
+                )
+            else:
+                logger.warning(
+                    "split_mode is deprecated. Please use `--blocks_to_swap` instead. `--blocks_to_swap 18` is automatically set."
+                    " / split_modeは非推奨です。代わりに`--blocks_to_swap`を使用してください。`--blocks_to_swap 18`が自動的に設定されました。"
+                )
+                args.blocks_to_swap = 18  # 18 is safe for most cases
+
+        train_dataset_group.verify_bucket_reso_steps(32)  # TODO check this
+        if val_dataset_group is not None:
+            val_dataset_group.verify_bucket_reso_steps(32)  # TODO check this
+
+    def load_target_model(self, args, weight_dtype, accelerator):
+        # currently offload to cpu for some models
+
+        # if the file is fp8 and we are using fp8_base, we can load it as is (fp8)
+        loading_dtype = None if args.fp8_base else weight_dtype
+
+        # if we load to cpu, flux.to(fp8) takes a long time, so we should load to gpu in future
+        _, model = flux_utils.load_flow_model(
+            args.pretrained_model_name_or_path,
+            loading_dtype,
+            "cpu",
+            disable_mmap=args.disable_mmap_load_safetensors,
+            model_type=self.model_type,
+        )
+        if args.fp8_base:
+            # check dtype of model
+            if model.dtype == torch.float8_e4m3fnuz or model.dtype == torch.float8_e5m2 or model.dtype == torch.float8_e5m2fnuz:
+                raise ValueError(f"Unsupported fp8 model dtype: {model.dtype}")
+            elif model.dtype == torch.float8_e4m3fn:
+                logger.info("Loaded fp8 FLUX model")
+            else:
+                logger.info(
+                    "Cast FLUX model to fp8. This may take a while. You can reduce the time by using fp8 checkpoint."
+                    " / FLUXモデルをfp8に変換しています。これには時間がかかる場合があります。fp8チェックポイントを使用することで時間を短縮できます。"
+                )
+                model.to(torch.float8_e4m3fn)
+
+        # if args.split_mode:
+        #     model = self.prepare_split_model(model, weight_dtype, accelerator)
+
+        self.is_swapping_blocks = args.blocks_to_swap is not None and args.blocks_to_swap > 0
+        if self.is_swapping_blocks:
+            # Swap blocks between CPU and GPU to reduce memory usage, in forward and backward passes.
+            logger.info(f"enable block swap: blocks_to_swap={args.blocks_to_swap}")
+            model.enable_block_swap(args.blocks_to_swap, accelerator.device)
+
+        if self.use_clip_l:
+            clip_l = flux_utils.load_clip_l(args.clip_l, weight_dtype, "cpu", disable_mmap=args.disable_mmap_load_safetensors)
+        else:
+            clip_l = flux_utils.dummy_clip_l()  # dummy CLIP-L for Chroma, which does not use CLIP-L
+        clip_l.eval()
+
+        # if the file is fp8 and we are using fp8_base (not unet), we can load it as is (fp8)
+        if args.fp8_base and not args.fp8_base_unet:
+            loading_dtype = None  # as is
+        else:
+            loading_dtype = weight_dtype
+
+        # loading t5xxl to cpu takes a long time, so we should load to gpu in future
+        t5xxl = flux_utils.load_t5xxl(args.t5xxl, loading_dtype, "cpu", disable_mmap=args.disable_mmap_load_safetensors)
+        t5xxl.eval()
+        if args.fp8_base and not args.fp8_base_unet:
+            # check dtype of model
+            if t5xxl.dtype == torch.float8_e4m3fnuz or t5xxl.dtype == torch.float8_e5m2 or t5xxl.dtype == torch.float8_e5m2fnuz:
+                raise ValueError(f"Unsupported fp8 model dtype: {t5xxl.dtype}")
+            elif t5xxl.dtype == torch.float8_e4m3fn:
+                logger.info("Loaded fp8 T5XXL model")
+
+        ae = flux_utils.load_ae(args.ae, weight_dtype, "cpu", disable_mmap=args.disable_mmap_load_safetensors)
+
+        model_version = flux_utils.MODEL_VERSION_FLUX_V1 if self.model_type != "chroma" else flux_utils.MODEL_VERSION_CHROMA
+        return model_version, [clip_l, t5xxl], ae, model
+
+    def get_tokenize_strategy(self, args):
+        # This method is called before `assert_extra_args`, so we cannot use `self.is_schnell` here.
+        # Instead, we analyze the checkpoint state to determine if it is schnell.
+        if args.model_type != "chroma":
+            _, is_schnell, _, _ = flux_utils.analyze_checkpoint_state(args.pretrained_model_name_or_path)
+        else:
+            is_schnell = False
+        self.is_schnell = is_schnell
+
+        if args.t5xxl_max_token_length is None:
+            if self.is_schnell:
+                t5xxl_max_token_length = 256
+            else:
+                t5xxl_max_token_length = 512
+        else:
+            t5xxl_max_token_length = args.t5xxl_max_token_length
+
+        logger.info(f"t5xxl_max_token_length: {t5xxl_max_token_length}")
+        return strategy_flux.FluxTokenizeStrategy(t5xxl_max_token_length, args.tokenizer_cache_dir)
+
+    def get_tokenizers(self, tokenize_strategy: strategy_flux.FluxTokenizeStrategy):
+        return [tokenize_strategy.clip_l, tokenize_strategy.t5xxl]
+
+    def get_latents_caching_strategy(self, args):
+        latents_caching_strategy = strategy_flux.FluxLatentsCachingStrategy(args.cache_latents_to_disk, args.vae_batch_size, False)
+        return latents_caching_strategy
+
+    def get_text_encoding_strategy(self, args):
+        return strategy_flux.FluxTextEncodingStrategy(apply_t5_attn_mask=args.apply_t5_attn_mask)
+
+    def post_process_network(self, args, accelerator, network, text_encoders, unet):
+        # check t5xxl is trained or not
+        self.train_t5xxl = network.train_t5xxl
+
+        if self.train_t5xxl and args.cache_text_encoder_outputs:
+            raise ValueError(
+                "T5XXL is trained, so cache_text_encoder_outputs cannot be used / T5XXL学習時はcache_text_encoder_outputsは使用できません"
+            )
+
+    def get_models_for_text_encoding(self, args, accelerator, text_encoders):
+        if args.cache_text_encoder_outputs:
+            if self.train_clip_l and not self.train_t5xxl:
+                return text_encoders[0:1]  # only CLIP-L is needed for encoding because T5XXL is cached
+            else:
+                return None  # no text encoders are needed for encoding because both are cached
+        else:
+            return text_encoders  # both CLIP-L and T5XXL are needed for encoding
+
+    def get_text_encoders_train_flags(self, args, text_encoders):
+        return [self.train_clip_l, self.train_t5xxl]
+
+    def get_text_encoder_outputs_caching_strategy(self, args):
+        if args.cache_text_encoder_outputs:
+            # if the text encoders is trained, we need tokenization, so is_partial is True
+            return strategy_flux.FluxTextEncoderOutputsCachingStrategy(
+                args.cache_text_encoder_outputs_to_disk,
+                args.text_encoder_batch_size,
+                args.skip_cache_check,
+                is_partial=self.train_clip_l or self.train_t5xxl,
+                apply_t5_attn_mask=args.apply_t5_attn_mask,
+            )
+        else:
+            return None
+
+    def cache_text_encoder_outputs_if_needed(
+        self, args, accelerator: Accelerator, unet, vae, text_encoders, dataset: train_util.DatasetGroup, weight_dtype
+    ):
+        if args.cache_text_encoder_outputs:
+            if not args.lowram:
+                # メモリ消費を減らす
+                logger.info("move vae and unet to cpu to save memory")
+                org_vae_device = vae.device
+                org_unet_device = unet.device
+                vae.to("cpu")
+                unet.to("cpu")
+                clean_memory_on_device(accelerator.device)
+
+            # When TE is not be trained, it will not be prepared so we need to use explicit autocast
+            logger.info("move text encoders to gpu")
+            text_encoders[0].to(accelerator.device, dtype=weight_dtype)  # always not fp8
+            text_encoders[1].to(accelerator.device)
+
+            if text_encoders[1].dtype == torch.float8_e4m3fn:
+                # if we load fp8 weights, the model is already fp8, so we use it as is
+                self.prepare_text_encoder_fp8(1, text_encoders[1], text_encoders[1].dtype, weight_dtype)
+            else:
+                # otherwise, we need to convert it to target dtype
+                text_encoders[1].to(weight_dtype)
+
+            with accelerator.autocast():
+                dataset.new_cache_text_encoder_outputs(text_encoders, accelerator)
+
+            # cache sample prompts
+            if args.sample_prompts is not None:
+                logger.info(f"cache Text Encoder outputs for sample prompt: {args.sample_prompts}")
+
+                tokenize_strategy: strategy_flux.FluxTokenizeStrategy = strategy_base.TokenizeStrategy.get_strategy()
+                text_encoding_strategy: strategy_flux.FluxTextEncodingStrategy = strategy_base.TextEncodingStrategy.get_strategy()
+
+                prompts = train_util.load_prompts(args.sample_prompts)
+                sample_prompts_te_outputs = {}  # key: prompt, value: text encoder outputs
+                with accelerator.autocast(), torch.no_grad():
+                    for prompt_dict in prompts:
+                        for p in [prompt_dict.get("prompt", ""), prompt_dict.get("negative_prompt", "")]:
+                            if p not in sample_prompts_te_outputs:
+                                logger.info(f"cache Text Encoder outputs for prompt: {p}")
+                                tokens_and_masks = tokenize_strategy.tokenize(p)
+                                sample_prompts_te_outputs[p] = text_encoding_strategy.encode_tokens(
+                                    tokenize_strategy, text_encoders, tokens_and_masks, args.apply_t5_attn_mask
+                                )
+                self.sample_prompts_te_outputs = sample_prompts_te_outputs
+
+            accelerator.wait_for_everyone()
+
+            # move back to cpu
+            if not self.is_train_text_encoder(args):
+                logger.info("move CLIP-L back to cpu")
+                text_encoders[0].to("cpu")
+            logger.info("move t5XXL back to cpu")
+            text_encoders[1].to("cpu")
+            clean_memory_on_device(accelerator.device)
+
+            if not args.lowram:
+                logger.info("move vae and unet back to original device")
+                vae.to(org_vae_device)
+                unet.to(org_unet_device)
+        else:
+            # Text Encoderから毎回出力を取得するので、GPUに乗せておく
+            text_encoders[0].to(accelerator.device, dtype=weight_dtype)
+            text_encoders[1].to(accelerator.device)
+
+    def sample_images(self, accelerator, args, epoch, global_step, device, ae, tokenizer, text_encoder, flux):
+        text_encoders = text_encoder  # for compatibility
+        text_encoders = self.get_models_for_text_encoding(args, accelerator, text_encoders)
+
+        flux_train_utils.sample_images(
+            accelerator, args, epoch, global_step, flux, ae, text_encoders, self.sample_prompts_te_outputs
+        )
+
+    def get_noise_scheduler(self, args: argparse.Namespace, device: torch.device) -> Any:
+        noise_scheduler = sd3_train_utils.FlowMatchEulerDiscreteScheduler(num_train_timesteps=1000, shift=args.discrete_flow_shift)
+        self.noise_scheduler_copy = copy.deepcopy(noise_scheduler)
+        return noise_scheduler
+
+    def encode_images_to_latents(self, args, vae, images):
+        return vae.encode(images)
+
+    def shift_scale_latents(self, args, latents):
+        return latents
+
+    def get_noise_pred_and_target(
+        self,
+        args,
+        accelerator,
+        noise_scheduler,
+        latents,
+        batch,
+        text_encoder_conds,
+        unet: flux_models.Flux,
+        network,
+        weight_dtype,
+        train_unet,
+        is_train=True,
+    ):
+        # Sample noise that we'll add to the latents
+        noise = torch.randn_like(latents)
+        bsz = latents.shape[0]
+
+        # get noisy model input and timesteps
+        noisy_model_input, timesteps, sigmas = flux_train_utils.get_noisy_model_input_and_timesteps(
+            args, noise_scheduler, latents, noise, accelerator.device, weight_dtype
+        )
+
+        # pack latents and get img_ids
+        packed_noisy_model_input = flux_utils.pack_latents(noisy_model_input)  # b, c, h*2, w*2 -> b, h*w, c*4
+        packed_latent_height, packed_latent_width = noisy_model_input.shape[2] // 2, noisy_model_input.shape[3] // 2
+        img_ids = flux_utils.prepare_img_ids(bsz, packed_latent_height, packed_latent_width).to(device=accelerator.device)
+
+        # get guidance
+        # ensure guidance_scale in args is float
+        guidance_vec = torch.full((bsz,), float(args.guidance_scale), device=accelerator.device)
+
+        # get modulation vectors for Chroma
+        with accelerator.autocast(), torch.no_grad():
+            mod_vectors = unet.get_mod_vectors(timesteps=timesteps / 1000, guidance=guidance_vec, batch_size=bsz)
+
+        if args.gradient_checkpointing:
+            noisy_model_input.requires_grad_(True)
+            for t in text_encoder_conds:
+                if t is not None and t.dtype.is_floating_point:
+                    t.requires_grad_(True)
+            img_ids.requires_grad_(True)
+            guidance_vec.requires_grad_(True)
+            if mod_vectors is not None:
+                mod_vectors.requires_grad_(True)
+
+        # Predict the noise residual
+        l_pooled, t5_out, txt_ids, t5_attn_mask = text_encoder_conds
+        if not args.apply_t5_attn_mask:
+            t5_attn_mask = None
+
+        def call_dit(img, img_ids, t5_out, txt_ids, l_pooled, timesteps, guidance_vec, t5_attn_mask, mod_vectors):
+            # grad is enabled even if unet is not in train mode, because Text Encoder is in train mode
+            with torch.set_grad_enabled(is_train), accelerator.autocast():
+                # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transformer model (we should not keep it but I want to keep the inputs same for the model for testing)
+                model_pred = unet(
+                    img=img,
+                    img_ids=img_ids,
+                    txt=t5_out,
+                    txt_ids=txt_ids,
+                    y=l_pooled,
+                    timesteps=timesteps / 1000,
+                    guidance=guidance_vec,
+                    txt_attention_mask=t5_attn_mask,
+                    mod_vectors=mod_vectors,
+                )
+            return model_pred
+
+        model_pred = call_dit(
+            img=packed_noisy_model_input,
+            img_ids=img_ids,
+            t5_out=t5_out,
+            txt_ids=txt_ids,
+            l_pooled=l_pooled,
+            timesteps=timesteps,
+            guidance_vec=guidance_vec,
+            t5_attn_mask=t5_attn_mask,
+            mod_vectors=mod_vectors,
+        )
+
+        # unpack latents
+        model_pred = flux_utils.unpack_latents(model_pred, packed_latent_height, packed_latent_width)
+
+        # apply model prediction type
+        model_pred, weighting = flux_train_utils.apply_model_prediction_type(args, model_pred, noisy_model_input, sigmas)
+
+        # flow matching loss: this is different from SD3
+        target = noise - latents
+
+        # differential output preservation
+        if "custom_attributes" in batch:
+            diff_output_pr_indices = []
+            for i, custom_attributes in enumerate(batch["custom_attributes"]):
+                if "diff_output_preservation" in custom_attributes and custom_attributes["diff_output_preservation"]:
+                    diff_output_pr_indices.append(i)
+
+            if len(diff_output_pr_indices) > 0:
+                network.set_multiplier(0.0)
+                unet.prepare_block_swap_before_forward()
+                with torch.no_grad():
+                    model_pred_prior = call_dit(
+                        img=packed_noisy_model_input[diff_output_pr_indices],
+                        img_ids=img_ids[diff_output_pr_indices],
+                        t5_out=t5_out[diff_output_pr_indices],
+                        txt_ids=txt_ids[diff_output_pr_indices],
+                        l_pooled=l_pooled[diff_output_pr_indices],
+                        timesteps=timesteps[diff_output_pr_indices],
+                        guidance_vec=guidance_vec[diff_output_pr_indices] if guidance_vec is not None else None,
+                        t5_attn_mask=t5_attn_mask[diff_output_pr_indices] if t5_attn_mask is not None else None,
+                        mod_vectors=mod_vectors[diff_output_pr_indices] if mod_vectors is not None else None,
+                    )
+                network.set_multiplier(1.0)  # may be overwritten by "network_multipliers" in the next step
+
+                model_pred_prior = flux_utils.unpack_latents(model_pred_prior, packed_latent_height, packed_latent_width)
+                model_pred_prior, _ = flux_train_utils.apply_model_prediction_type(
+                    args,
+                    model_pred_prior,
+                    noisy_model_input[diff_output_pr_indices],
+                    sigmas[diff_output_pr_indices] if sigmas is not None else None,
+                )
+                target[diff_output_pr_indices] = model_pred_prior.to(target.dtype)
+
+        return model_pred, target, timesteps, weighting
+
+    def post_process_loss(self, loss, args, timesteps, noise_scheduler):
+        return loss
+
+    def get_sai_model_spec(self, args):
+        if self.model_type != "chroma":
+            model_description = "schnell" if self.is_schnell else "dev"
+        else:
+            model_description = "chroma"
+        return train_util.get_sai_model_spec(None, args, False, True, False, flux=model_description)
+
+    def update_metadata(self, metadata, args):
+        metadata["ss_model_type"] = args.model_type
+        metadata["ss_apply_t5_attn_mask"] = args.apply_t5_attn_mask
+        metadata["ss_weighting_scheme"] = args.weighting_scheme
+        metadata["ss_logit_mean"] = args.logit_mean
+        metadata["ss_logit_std"] = args.logit_std
+        metadata["ss_mode_scale"] = args.mode_scale
+        metadata["ss_guidance_scale"] = args.guidance_scale
+        metadata["ss_timestep_sampling"] = args.timestep_sampling
+        metadata["ss_sigmoid_scale"] = args.sigmoid_scale
+        metadata["ss_model_prediction_type"] = args.model_prediction_type
+        metadata["ss_discrete_flow_shift"] = args.discrete_flow_shift
+
+    def is_text_encoder_not_needed_for_training(self, args):
+        return args.cache_text_encoder_outputs and not self.is_train_text_encoder(args)
+
+    def prepare_text_encoder_grad_ckpt_workaround(self, index, text_encoder):
+        if index == 0:  # CLIP-L
+            return super().prepare_text_encoder_grad_ckpt_workaround(index, text_encoder)
+        else:  # T5XXL
+            text_encoder.encoder.embed_tokens.requires_grad_(True)
+
+    def prepare_text_encoder_fp8(self, index, text_encoder, te_weight_dtype, weight_dtype):
+        if index == 0:  # CLIP-L
+            logger.info(f"prepare CLIP-L for fp8: set to {te_weight_dtype}, set embeddings to {weight_dtype}")
+            text_encoder.to(te_weight_dtype)  # fp8
+            text_encoder.text_model.embeddings.to(dtype=weight_dtype)
+        else:  # T5XXL
+
+            def prepare_fp8(text_encoder, target_dtype):
+                def forward_hook(module):
+                    def forward(hidden_states):
+                        hidden_gelu = module.act(module.wi_0(hidden_states))
+                        hidden_linear = module.wi_1(hidden_states)
+                        hidden_states = hidden_gelu * hidden_linear
+                        hidden_states = module.dropout(hidden_states)
+
+                        hidden_states = module.wo(hidden_states)
+                        return hidden_states
+
+                    return forward
+
+                for module in text_encoder.modules():
+                    if module.__class__.__name__ in ["T5LayerNorm", "Embedding"]:
+                        # print("set", module.__class__.__name__, "to", target_dtype)
+                        module.to(target_dtype)
+                    if module.__class__.__name__ in ["T5DenseGatedActDense"]:
+                        # print("set", module.__class__.__name__, "hooks")
+                        module.forward = forward_hook(module)
+
+            if flux_utils.get_t5xxl_actual_dtype(text_encoder) == torch.float8_e4m3fn and text_encoder.dtype == weight_dtype:
+                logger.info(f"T5XXL already prepared for fp8")
+            else:
+                logger.info(f"prepare T5XXL for fp8: set to {te_weight_dtype}, set embeddings to {weight_dtype}, add hooks")
+                text_encoder.to(te_weight_dtype)  # fp8
+                prepare_fp8(text_encoder, weight_dtype)
+
+    def on_validation_step_end(self, args, accelerator, network, text_encoders, unet, batch, weight_dtype):
+        if self.is_swapping_blocks:
+            # prepare for next forward: because backward pass is not called, we need to prepare it here
+            accelerator.unwrap_model(unet).prepare_block_swap_before_forward()
+
+    def prepare_unet_with_accelerator(
+        self, args: argparse.Namespace, accelerator: Accelerator, unet: torch.nn.Module
+    ) -> torch.nn.Module:
+        if not self.is_swapping_blocks:
+            return super().prepare_unet_with_accelerator(args, accelerator, unet)
+
+        # if we doesn't swap blocks, we can move the model to device
+        flux: flux_models.Flux = unet
+        flux = accelerator.prepare(flux, device_placement=[not self.is_swapping_blocks])
+        accelerator.unwrap_model(flux).move_to_device_except_swap_blocks(accelerator.device)  # reduce peak memory usage
+        accelerator.unwrap_model(flux).prepare_block_swap_before_forward()
+
+        return flux
+
+
+def setup_parser() -> argparse.ArgumentParser:
+    parser = train_network.setup_parser()
+    train_util.add_dit_training_arguments(parser)
+    flux_train_utils.add_flux_train_arguments(parser)
+
+    parser.add_argument(
+        "--split_mode",
+        action="store_true",
+        # help="[EXPERIMENTAL] use split mode for Flux model, network arg `train_blocks=single` is required"
+        # + "/[実験的] Fluxモデルの分割モードを使用する。ネットワーク引数`train_blocks=single`が必要",
+        help="[Deprecated] This option is deprecated. Please use `--blocks_to_swap` instead."
+        " / このオプションは非推奨です。代わりに`--blocks_to_swap`を使用してください。",
+    )
+    return parser
+
+
+if __name__ == "__main__":
+    parser = setup_parser()
+
+    args = parser.parse_args()
+    train_util.verify_command_line_training_args(args)
+    args = train_util.read_config_from_file(args, parser)
+
+    trainer = FluxNetworkTrainer()
+    trainer.train(args)

hunyuan_image_train_network.py

@@ -0,0 +1,717 @@
+import argparse
+import copy
+import gc
+from typing import Any, Optional, Union, cast
+import os
+import time
+from types import SimpleNamespace
+
+import numpy as np
+import torch
+import torch.nn as nn
+from PIL import Image
+from accelerate import Accelerator, PartialState
+
+from library import flux_utils, hunyuan_image_models, hunyuan_image_vae, strategy_base, train_util
+from library.device_utils import clean_memory_on_device, init_ipex
+
+init_ipex()
+
+import train_network
+from library import (
+    flux_train_utils,
+    hunyuan_image_models,
+    hunyuan_image_text_encoder,
+    hunyuan_image_utils,
+    hunyuan_image_vae,
+    sd3_train_utils,
+    strategy_base,
+    strategy_hunyuan_image,
+    train_util,
+)
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+# region sampling
+
+
+# TODO commonize with flux_utils
+def sample_images(
+    accelerator: Accelerator,
+    args: argparse.Namespace,
+    epoch,
+    steps,
+    dit: hunyuan_image_models.HYImageDiffusionTransformer,
+    vae,
+    text_encoders,
+    sample_prompts_te_outputs,
+    prompt_replacement=None,
+):
+    if steps == 0:
+        if not args.sample_at_first:
+            return
+    else:
+        if args.sample_every_n_steps is None and args.sample_every_n_epochs is None:
+            return
+        if args.sample_every_n_epochs is not None:
+            # sample_every_n_steps は無視する
+            if epoch is None or epoch % args.sample_every_n_epochs != 0:
+                return
+        else:
+            if steps % args.sample_every_n_steps != 0 or epoch is not None:  # steps is not divisible or end of epoch
+                return
+
+    logger.info("")
+    logger.info(f"generating sample images at step / サンプル画像生成 ステップ: {steps}")
+    if not os.path.isfile(args.sample_prompts) and sample_prompts_te_outputs is None:
+        logger.error(f"No prompt file / プロンプトファイルがありません: {args.sample_prompts}")
+        return
+
+    distributed_state = PartialState()  # for multi gpu distributed inference. this is a singleton, so it's safe to use it here
+
+    # unwrap unet and text_encoder(s)
+    dit = accelerator.unwrap_model(dit)
+    dit = cast(hunyuan_image_models.HYImageDiffusionTransformer, dit)
+    dit.switch_block_swap_for_inference()
+    if text_encoders is not None:
+        text_encoders = [(accelerator.unwrap_model(te) if te is not None else None) for te in text_encoders]
+    # print([(te.parameters().__next__().device if te is not None else None) for te in text_encoders])
+
+    prompts = train_util.load_prompts(args.sample_prompts)
+
+    save_dir = args.output_dir + "/sample"
+    os.makedirs(save_dir, exist_ok=True)
+
+    # save random state to restore later
+    rng_state = torch.get_rng_state()
+    cuda_rng_state = None
+    try:
+        cuda_rng_state = torch.cuda.get_rng_state() if torch.cuda.is_available() else None
+    except Exception:
+        pass
+
+    if distributed_state.num_processes <= 1:
+        # If only one device is available, just use the original prompt list. We don't need to care about the distribution of prompts.
+        with torch.no_grad(), accelerator.autocast():
+            for prompt_dict in prompts:
+                sample_image_inference(
+                    accelerator,
+                    args,
+                    dit,
+                    text_encoders,
+                    vae,
+                    save_dir,
+                    prompt_dict,
+                    epoch,
+                    steps,
+                    sample_prompts_te_outputs,
+                    prompt_replacement,
+                )
+    else:
+        # Creating list with N elements, where each element is a list of prompt_dicts, and N is the number of processes available (number of devices available)
+        # prompt_dicts are assigned to lists based on order of processes, to attempt to time the image creation time to match enum order. Probably only works when steps and sampler are identical.
+        per_process_prompts = []  # list of lists
+        for i in range(distributed_state.num_processes):
+            per_process_prompts.append(prompts[i :: distributed_state.num_processes])
+
+        with torch.no_grad():
+            with distributed_state.split_between_processes(per_process_prompts) as prompt_dict_lists:
+                for prompt_dict in prompt_dict_lists[0]:
+                    sample_image_inference(
+                        accelerator,
+                        args,
+                        dit,
+                        text_encoders,
+                        vae,
+                        save_dir,
+                        prompt_dict,
+                        epoch,
+                        steps,
+                        sample_prompts_te_outputs,
+                        prompt_replacement,
+                    )
+
+    torch.set_rng_state(rng_state)
+    if cuda_rng_state is not None:
+        torch.cuda.set_rng_state(cuda_rng_state)
+
+    dit.switch_block_swap_for_training()
+    clean_memory_on_device(accelerator.device)
+
+
+def sample_image_inference(
+    accelerator: Accelerator,
+    args: argparse.Namespace,
+    dit: hunyuan_image_models.HYImageDiffusionTransformer,
+    text_encoders: Optional[list[nn.Module]],
+    vae: hunyuan_image_vae.HunyuanVAE2D,
+    save_dir,
+    prompt_dict,
+    epoch,
+    steps,
+    sample_prompts_te_outputs,
+    prompt_replacement,
+):
+    assert isinstance(prompt_dict, dict)
+    negative_prompt = prompt_dict.get("negative_prompt")
+    sample_steps = prompt_dict.get("sample_steps", 20)
+    width = prompt_dict.get("width", 512)
+    height = prompt_dict.get("height", 512)
+    cfg_scale = prompt_dict.get("scale", 3.5)
+    seed = prompt_dict.get("seed")
+    prompt: str = prompt_dict.get("prompt", "")
+    flow_shift: float = prompt_dict.get("flow_shift", 5.0)
+    # sampler_name: str = prompt_dict.get("sample_sampler", args.sample_sampler)
+
+    if prompt_replacement is not None:
+        prompt = prompt.replace(prompt_replacement[0], prompt_replacement[1])
+        if negative_prompt is not None:
+            negative_prompt = negative_prompt.replace(prompt_replacement[0], prompt_replacement[1])
+
+    if seed is not None:
+        torch.manual_seed(seed)
+        torch.cuda.manual_seed(seed)
+    else:
+        # True random sample image generation
+        torch.seed()
+        torch.cuda.seed()
+
+    if negative_prompt is None:
+        negative_prompt = ""
+    height = max(64, height - height % 16)  # round to divisible by 16
+    width = max(64, width - width % 16)  # round to divisible by 16
+    logger.info(f"prompt: {prompt}")
+    if cfg_scale != 1.0:
+        logger.info(f"negative_prompt: {negative_prompt}")
+    elif negative_prompt != "":
+        logger.info(f"negative prompt is ignored because scale is 1.0")
+    logger.info(f"height: {height}")
+    logger.info(f"width: {width}")
+    logger.info(f"sample_steps: {sample_steps}")
+    if cfg_scale != 1.0:
+        logger.info(f"CFG scale: {cfg_scale}")
+    logger.info(f"flow_shift: {flow_shift}")
+    # logger.info(f"sample_sampler: {sampler_name}")
+    if seed is not None:
+        logger.info(f"seed: {seed}")
+
+    # encode prompts
+    tokenize_strategy = strategy_base.TokenizeStrategy.get_strategy()
+    encoding_strategy = strategy_base.TextEncodingStrategy.get_strategy()
+
+    def encode_prompt(prpt):
+        text_encoder_conds = []
+        if sample_prompts_te_outputs and prpt in sample_prompts_te_outputs:
+            text_encoder_conds = sample_prompts_te_outputs[prpt]
+            # print(f"Using cached text encoder outputs for prompt: {prpt}")
+        if text_encoders is not None:
+            # print(f"Encoding prompt: {prpt}")
+            tokens_and_masks = tokenize_strategy.tokenize(prpt)
+            encoded_text_encoder_conds = encoding_strategy.encode_tokens(tokenize_strategy, text_encoders, tokens_and_masks)
+
+            # if text_encoder_conds is not cached, use encoded_text_encoder_conds
+            if len(text_encoder_conds) == 0:
+                text_encoder_conds = encoded_text_encoder_conds
+            else:
+                # if encoded_text_encoder_conds is not None, update cached text_encoder_conds
+                for i in range(len(encoded_text_encoder_conds)):
+                    if encoded_text_encoder_conds[i] is not None:
+                        text_encoder_conds[i] = encoded_text_encoder_conds[i]
+        return text_encoder_conds
+
+    vl_embed, vl_mask, byt5_embed, byt5_mask, ocr_mask = encode_prompt(prompt)
+    arg_c = {
+        "embed": vl_embed,
+        "mask": vl_mask,
+        "embed_byt5": byt5_embed,
+        "mask_byt5": byt5_mask,
+        "ocr_mask": ocr_mask,
+        "prompt": prompt,
+    }
+
+    # encode negative prompts
+    if cfg_scale != 1.0:
+        neg_vl_embed, neg_vl_mask, neg_byt5_embed, neg_byt5_mask, neg_ocr_mask = encode_prompt(negative_prompt)
+        arg_c_null = {
+            "embed": neg_vl_embed,
+            "mask": neg_vl_mask,
+            "embed_byt5": neg_byt5_embed,
+            "mask_byt5": neg_byt5_mask,
+            "ocr_mask": neg_ocr_mask,
+            "prompt": negative_prompt,
+        }
+    else:
+        arg_c_null = None
+
+    gen_args = SimpleNamespace(
+        image_size=(height, width),
+        infer_steps=sample_steps,
+        flow_shift=flow_shift,
+        guidance_scale=cfg_scale,
+        fp8=args.fp8_scaled,
+        apg_start_step_ocr=38,
+        apg_start_step_general=5,
+        guidance_rescale=0.0,
+        guidance_rescale_apg=0.0,
+    )
+
+    from hunyuan_image_minimal_inference import generate_body  # import here to avoid circular import
+
+    dit_is_training = dit.training
+    dit.eval()
+    x = generate_body(gen_args, dit, arg_c, arg_c_null, accelerator.device, seed)
+    if dit_is_training:
+        dit.train()
+    clean_memory_on_device(accelerator.device)
+
+    # latent to image
+    org_vae_device = vae.device  # will be on cpu
+    vae.to(accelerator.device)  # distributed_state.device is same as accelerator.device
+    with torch.no_grad():
+        x = x / vae.scaling_factor
+        x = vae.decode(x.to(vae.device, dtype=vae.dtype))
+    vae.to(org_vae_device)
+
+    clean_memory_on_device(accelerator.device)
+
+    x = x.clamp(-1, 1)
+    x = x.permute(0, 2, 3, 1)
+    image = Image.fromarray((127.5 * (x + 1.0)).float().cpu().numpy().astype(np.uint8)[0])
+
+    # adding accelerator.wait_for_everyone() here should sync up and ensure that sample images are saved in the same order as the original prompt list
+    # but adding 'enum' to the filename should be enough
+
+    ts_str = time.strftime("%Y%m%d%H%M%S", time.localtime())
+    num_suffix = f"e{epoch:06d}" if epoch is not None else f"{steps:06d}"
+    seed_suffix = "" if seed is None else f"_{seed}"
+    i: int = prompt_dict["enum"]
+    img_filename = f"{'' if args.output_name is None else args.output_name + '_'}{num_suffix}_{i:02d}_{ts_str}{seed_suffix}.png"
+    image.save(os.path.join(save_dir, img_filename))
+
+    # send images to wandb if enabled
+    if "wandb" in [tracker.name for tracker in accelerator.trackers]:
+        wandb_tracker = accelerator.get_tracker("wandb")
+
+        import wandb
+
+        # not to commit images to avoid inconsistency between training and logging steps
+        wandb_tracker.log({f"sample_{i}": wandb.Image(image, caption=prompt)}, commit=False)  # positive prompt as a caption
+
+
+# endregion
+
+
+class HunyuanImageNetworkTrainer(train_network.NetworkTrainer):
+    def __init__(self):
+        super().__init__()
+        self.sample_prompts_te_outputs = None
+        self.is_swapping_blocks: bool = False
+        self.rotary_pos_emb_cache = {}
+
+    def assert_extra_args(
+        self,
+        args,
+        train_dataset_group: Union[train_util.DatasetGroup, train_util.MinimalDataset],
+        val_dataset_group: Optional[train_util.DatasetGroup],
+    ):
+        super().assert_extra_args(args, train_dataset_group, val_dataset_group)
+        # sdxl_train_util.verify_sdxl_training_args(args)
+
+        if args.mixed_precision == "fp16":
+            logger.warning(
+                "mixed_precision bf16 is recommended for HunyuanImage-2.1 / HunyuanImage-2.1ではmixed_precision bf16が推奨されます"
+            )
+
+        if (args.fp8_base or args.fp8_base_unet) and not args.fp8_scaled:
+            logger.warning(
+                "fp8_base and fp8_base_unet are not supported. Use fp8_scaled instead / fp8_baseとfp8_base_unetはサポートされていません。代わりにfp8_scaledを使用してください"
+            )
+        if args.fp8_scaled and (args.fp8_base or args.fp8_base_unet):
+            logger.info(
+                "fp8_scaled is used, so fp8_base and fp8_base_unet are ignored / fp8_scaledが使われているので、fp8_baseとfp8_base_unetは無視されます"
+            )
+            args.fp8_base = False
+            args.fp8_base_unet = False
+
+        if args.cache_text_encoder_outputs_to_disk and not args.cache_text_encoder_outputs:
+            logger.warning(
+                "cache_text_encoder_outputs_to_disk is enabled, so cache_text_encoder_outputs is also enabled / cache_text_encoder_outputs_to_diskが有効になっているため、cache_text_encoder_outputsも有効になります"
+            )
+            args.cache_text_encoder_outputs = True
+
+        if args.cache_text_encoder_outputs:
+            assert (
+                train_dataset_group.is_text_encoder_output_cacheable()
+            ), "when caching Text Encoder output, either caption_dropout_rate, shuffle_caption, token_warmup_step or caption_tag_dropout_rate cannot be used / Text Encoderの出力をキャッシュするときはcaption_dropout_rate, shuffle_caption, token_warmup_step, caption_tag_dropout_rateは使えません"
+
+        train_dataset_group.verify_bucket_reso_steps(32)
+        if val_dataset_group is not None:
+            val_dataset_group.verify_bucket_reso_steps(32)
+
+    def load_target_model(self, args, weight_dtype, accelerator):
+        self.is_swapping_blocks = args.blocks_to_swap is not None and args.blocks_to_swap > 0
+
+        vl_dtype = torch.float8_e4m3fn if args.fp8_vl else torch.bfloat16
+        vl_device = "cpu"  # loading to cpu and move to gpu later in cache_text_encoder_outputs_if_needed
+        _, text_encoder_vlm = hunyuan_image_text_encoder.load_qwen2_5_vl(
+            args.text_encoder, dtype=vl_dtype, device=vl_device, disable_mmap=args.disable_mmap_load_safetensors
+        )
+        _, text_encoder_byt5 = hunyuan_image_text_encoder.load_byt5(
+            args.byt5, dtype=torch.float16, device=vl_device, disable_mmap=args.disable_mmap_load_safetensors
+        )
+
+        vae = hunyuan_image_vae.load_vae(
+            args.vae, "cpu", disable_mmap=args.disable_mmap_load_safetensors, chunk_size=args.vae_chunk_size
+        )
+        vae.to(dtype=torch.float16)  # VAE is always fp16
+        vae.eval()
+
+        model_version = hunyuan_image_utils.MODEL_VERSION_2_1
+        return model_version, [text_encoder_vlm, text_encoder_byt5], vae, None  # unet will be loaded later
+
+    def load_unet_lazily(self, args, weight_dtype, accelerator, text_encoders) -> tuple[nn.Module, list[nn.Module]]:
+        if args.cache_text_encoder_outputs:
+            logger.info("Replace text encoders with dummy models to save memory")
+
+            # This doesn't free memory, so we move text encoders to meta device in cache_text_encoder_outputs_if_needed
+            text_encoders = [flux_utils.dummy_clip_l() for _ in text_encoders]
+            clean_memory_on_device(accelerator.device)
+            gc.collect()
+
+        loading_dtype = None if args.fp8_scaled else weight_dtype
+        loading_device = "cpu" if self.is_swapping_blocks else accelerator.device
+
+        attn_mode = "torch"
+        if args.xformers:
+            attn_mode = "xformers"
+        if args.attn_mode is not None:
+            attn_mode = args.attn_mode
+
+        logger.info(f"Loading DiT model with attn_mode: {attn_mode}, split_attn: {args.split_attn}, fp8_scaled: {args.fp8_scaled}")
+        model = hunyuan_image_models.load_hunyuan_image_model(
+            accelerator.device,
+            args.pretrained_model_name_or_path,
+            attn_mode,
+            args.split_attn,
+            loading_device,
+            loading_dtype,
+            args.fp8_scaled,
+        )
+
+        if self.is_swapping_blocks:
+            # Swap blocks between CPU and GPU to reduce memory usage, in forward and backward passes.
+            logger.info(f"enable block swap: blocks_to_swap={args.blocks_to_swap}")
+            model.enable_block_swap(args.blocks_to_swap, accelerator.device, supports_backward=True)
+
+        return model, text_encoders
+
+    def get_tokenize_strategy(self, args):
+        return strategy_hunyuan_image.HunyuanImageTokenizeStrategy(args.tokenizer_cache_dir)
+
+    def get_tokenizers(self, tokenize_strategy: strategy_hunyuan_image.HunyuanImageTokenizeStrategy):
+        return [tokenize_strategy.vlm_tokenizer, tokenize_strategy.byt5_tokenizer]
+
+    def get_latents_caching_strategy(self, args):
+        return strategy_hunyuan_image.HunyuanImageLatentsCachingStrategy(args.cache_latents_to_disk, args.vae_batch_size, False)
+
+    def get_text_encoding_strategy(self, args):
+        return strategy_hunyuan_image.HunyuanImageTextEncodingStrategy()
+
+    def post_process_network(self, args, accelerator, network, text_encoders, unet):
+        pass
+
+    def get_models_for_text_encoding(self, args, accelerator, text_encoders):
+        if args.cache_text_encoder_outputs:
+            return None  # no text encoders are needed for encoding because both are cached
+        else:
+            return text_encoders
+
+    def get_text_encoders_train_flags(self, args, text_encoders):
+        # HunyuanImage-2.1 does not support training VLM or byT5
+        return [False, False]
+
+    def get_text_encoder_outputs_caching_strategy(self, args):
+        if args.cache_text_encoder_outputs:
+            return strategy_hunyuan_image.HunyuanImageTextEncoderOutputsCachingStrategy(
+                args.cache_text_encoder_outputs_to_disk, args.text_encoder_batch_size, args.skip_cache_check, False
+            )
+        else:
+            return None
+
+    def cache_text_encoder_outputs_if_needed(
+        self, args, accelerator: Accelerator, unet, vae, text_encoders, dataset: train_util.DatasetGroup, weight_dtype
+    ):
+        vlm_device = "cpu" if args.text_encoder_cpu else accelerator.device
+        if args.cache_text_encoder_outputs:
+            if not args.lowram:
+                # メモリ消費を減らす
+                logger.info("move vae to cpu to save memory")
+                org_vae_device = vae.device
+                vae.to("cpu")
+                clean_memory_on_device(accelerator.device)
+
+            logger.info(f"move text encoders to {vlm_device} to encode and cache text encoder outputs")
+            text_encoders[0].to(vlm_device)
+            text_encoders[1].to(vlm_device)
+
+            # VLM (bf16) and byT5 (fp16) are used for encoding, so we cannot use autocast here
+            dataset.new_cache_text_encoder_outputs(text_encoders, accelerator)
+
+            # cache sample prompts
+            if args.sample_prompts is not None:
+                logger.info(f"cache Text Encoder outputs for sample prompt: {args.sample_prompts}")
+
+                tokenize_strategy: strategy_hunyuan_image.HunyuanImageTokenizeStrategy = (
+                    strategy_base.TokenizeStrategy.get_strategy()
+                )
+                text_encoding_strategy: strategy_hunyuan_image.HunyuanImageTextEncodingStrategy = (
+                    strategy_base.TextEncodingStrategy.get_strategy()
+                )
+
+                prompts = train_util.load_prompts(args.sample_prompts)
+                sample_prompts_te_outputs = {}  # key: prompt, value: text encoder outputs
+                with accelerator.autocast(), torch.no_grad():
+                    for prompt_dict in prompts:
+                        for p in [prompt_dict.get("prompt", ""), prompt_dict.get("negative_prompt", "")]:
+                            if p not in sample_prompts_te_outputs:
+                                logger.info(f"cache Text Encoder outputs for prompt: {p}")
+                                tokens_and_masks = tokenize_strategy.tokenize(p)
+                                sample_prompts_te_outputs[p] = text_encoding_strategy.encode_tokens(
+                                    tokenize_strategy, text_encoders, tokens_and_masks
+                                )
+                self.sample_prompts_te_outputs = sample_prompts_te_outputs
+
+            accelerator.wait_for_everyone()
+
+            # text encoders are not needed for training, so we move to meta device
+            logger.info("move text encoders to meta device to save memory")
+            text_encoders = [te.to("meta") for te in text_encoders]
+            clean_memory_on_device(accelerator.device)
+
+            if not args.lowram:
+                logger.info("move vae back to original device")
+                vae.to(org_vae_device)
+        else:
+            # Text Encoderから毎回出力を取得するので、GPUに乗せておく
+            text_encoders[0].to(vlm_device)
+            text_encoders[1].to(vlm_device)
+
+    def sample_images(self, accelerator, args, epoch, global_step, device, ae, tokenizer, text_encoder, flux):
+        text_encoders = text_encoder  # for compatibility
+        text_encoders = self.get_models_for_text_encoding(args, accelerator, text_encoders)
+
+        sample_images(accelerator, args, epoch, global_step, flux, ae, text_encoders, self.sample_prompts_te_outputs)
+
+    def get_noise_scheduler(self, args: argparse.Namespace, device: torch.device) -> Any:
+        noise_scheduler = sd3_train_utils.FlowMatchEulerDiscreteScheduler(num_train_timesteps=1000, shift=args.discrete_flow_shift)
+        self.noise_scheduler_copy = copy.deepcopy(noise_scheduler)
+        return noise_scheduler
+
+    def encode_images_to_latents(self, args, vae: hunyuan_image_vae.HunyuanVAE2D, images):
+        return vae.encode(images).sample()
+
+    def shift_scale_latents(self, args, latents):
+        # for encoding, we need to scale the latents
+        return latents * hunyuan_image_vae.LATENT_SCALING_FACTOR
+
+    def get_noise_pred_and_target(
+        self,
+        args,
+        accelerator,
+        noise_scheduler,
+        latents,
+        batch,
+        text_encoder_conds,
+        unet: hunyuan_image_models.HYImageDiffusionTransformer,
+        network,
+        weight_dtype,
+        train_unet,
+        is_train=True,
+    ):
+        # Sample noise that we'll add to the latents
+        noise = torch.randn_like(latents)
+
+        # get noisy model input and timesteps
+        noisy_model_input, _, sigmas = flux_train_utils.get_noisy_model_input_and_timesteps(
+            args, noise_scheduler, latents, noise, accelerator.device, weight_dtype
+        )
+        # bfloat16 is too low precision for 0-1000 TODO fix get_noisy_model_input_and_timesteps
+        timesteps = (sigmas[:, 0, 0, 0] * 1000).to(torch.int64)
+        # print(
+        #     f"timestep: {timesteps}, noisy_model_input shape: {noisy_model_input.shape}, mean: {noisy_model_input.mean()}, std: {noisy_model_input.std()}"
+        # )
+
+        if args.gradient_checkpointing:
+            noisy_model_input.requires_grad_(True)
+            for t in text_encoder_conds:
+                if t is not None and t.dtype.is_floating_point:
+                    t.requires_grad_(True)
+
+        # Predict the noise residual
+        # ocr_mask is for inference only, so it is not used here
+        vlm_embed, vlm_mask, byt5_embed, byt5_mask, ocr_mask = text_encoder_conds
+
+        # print(f"embed shape: {vlm_embed.shape}, mean: {vlm_embed.mean()}, std: {vlm_embed.std()}")
+        # print(f"embed_byt5 shape: {byt5_embed.shape}, mean: {byt5_embed.mean()}, std: {byt5_embed.std()}")
+        # print(f"latents shape: {latents.shape}, mean: {latents.mean()}, std: {latents.std()}")
+        # print(f"mask shape: {vlm_mask.shape}, sum: {vlm_mask.sum()}")
+        # print(f"mask_byt5 shape: {byt5_mask.shape}, sum: {byt5_mask.sum()}")
+        with torch.set_grad_enabled(is_train), accelerator.autocast():
+            model_pred = unet(
+                noisy_model_input, timesteps, vlm_embed, vlm_mask, byt5_embed, byt5_mask  # , self.rotary_pos_emb_cache
+            )
+
+        # apply model prediction type
+        model_pred, weighting = flux_train_utils.apply_model_prediction_type(args, model_pred, noisy_model_input, sigmas)
+
+        # flow matching loss
+        target = noise - latents
+
+        # differential output preservation is not used for HunyuanImage-2.1 currently
+
+        return model_pred, target, timesteps, weighting
+
+    def post_process_loss(self, loss, args, timesteps, noise_scheduler):
+        return loss
+
+    def get_sai_model_spec(self, args):
+        return train_util.get_sai_model_spec_dataclass(None, args, False, True, False, hunyuan_image="2.1").to_metadata_dict()
+
+    def update_metadata(self, metadata, args):
+        metadata["ss_logit_mean"] = args.logit_mean
+        metadata["ss_logit_std"] = args.logit_std
+        metadata["ss_mode_scale"] = args.mode_scale
+        metadata["ss_timestep_sampling"] = args.timestep_sampling
+        metadata["ss_sigmoid_scale"] = args.sigmoid_scale
+        metadata["ss_model_prediction_type"] = args.model_prediction_type
+        metadata["ss_discrete_flow_shift"] = args.discrete_flow_shift
+
+    def is_text_encoder_not_needed_for_training(self, args):
+        return args.cache_text_encoder_outputs and not self.is_train_text_encoder(args)
+
+    def prepare_text_encoder_grad_ckpt_workaround(self, index, text_encoder):
+        # do not support text encoder training for HunyuanImage-2.1
+        pass
+
+    def cast_text_encoder(self, args):
+        return False  # VLM is bf16, byT5 is fp16, so do not cast to other dtype
+
+    def cast_vae(self, args):
+        return False  # VAE is fp16, so do not cast to other dtype
+
+    def cast_unet(self, args):
+        return not args.fp8_scaled  # if fp8_scaled is used, do not cast to other dtype
+
+    def prepare_text_encoder_fp8(self, index, text_encoder, te_weight_dtype, weight_dtype):
+        # fp8 text encoder for HunyuanImage-2.1 is not supported currently
+        pass
+
+    def on_validation_step_end(self, args, accelerator, network, text_encoders, unet, batch, weight_dtype):
+        if self.is_swapping_blocks:
+            # prepare for next forward: because backward pass is not called, we need to prepare it here
+            accelerator.unwrap_model(unet).prepare_block_swap_before_forward()
+
+    def prepare_unet_with_accelerator(
+        self, args: argparse.Namespace, accelerator: Accelerator, unet: torch.nn.Module
+    ) -> torch.nn.Module:
+        if not self.is_swapping_blocks:
+            return super().prepare_unet_with_accelerator(args, accelerator, unet)
+
+        # if we doesn't swap blocks, we can move the model to device
+        model: hunyuan_image_models.HYImageDiffusionTransformer = unet
+        model = accelerator.prepare(model, device_placement=[not self.is_swapping_blocks])
+        accelerator.unwrap_model(model).move_to_device_except_swap_blocks(accelerator.device)  # reduce peak memory usage
+        accelerator.unwrap_model(model).prepare_block_swap_before_forward()
+
+        return model
+
+
+def setup_parser() -> argparse.ArgumentParser:
+    parser = train_network.setup_parser()
+    train_util.add_dit_training_arguments(parser)
+
+    parser.add_argument(
+        "--text_encoder",
+        type=str,
+        help="path to Qwen2.5-VL (*.sft or *.safetensors), should be bfloat16 / Qwen2.5-VLのパス（*.sftまたは*.safetensors）、bfloat16が前提",
+    )
+    parser.add_argument(
+        "--byt5",
+        type=str,
+        help="path to byt5 (*.sft or *.safetensors), should be float16 / byt5のパス（*.sftまたは*.safetensors）、float16が前提",
+    )
+
+    parser.add_argument(
+        "--timestep_sampling",
+        choices=["sigma", "uniform", "sigmoid", "shift", "flux_shift"],
+        default="sigma",
+        help="Method to sample timesteps: sigma-based, uniform random, sigmoid of random normal, shift of sigmoid and FLUX.1 shifting."
+        " / タイムステップをサンプリングする方法：sigma、random uniform、random normalのsigmoid、sigmoidのシフト、FLUX.1のシフト。",
+    )
+    parser.add_argument(
+        "--sigmoid_scale",
+        type=float,
+        default=1.0,
+        help='Scale factor for sigmoid timestep sampling (only used when timestep-sampling is "sigmoid"). / sigmoidタイムステップサンプリングの倍率（timestep-samplingが"sigmoid"の場合のみ有効）。',
+    )
+    parser.add_argument(
+        "--model_prediction_type",
+        choices=["raw", "additive", "sigma_scaled"],
+        default="raw",
+        help="How to interpret and process the model prediction: "
+        "raw (use as is), additive (add to noisy input), sigma_scaled (apply sigma scaling). Default is raw unlike FLUX.1."
+        " / モデル予測の解釈と処理方法："
+        "raw（そのまま使用）、additive（ノイズ入力に加算）、sigma_scaled（シグマスケーリングを適用）。デフォルトはFLUX.1とは異なりrawです。",
+    )
+    parser.add_argument(
+        "--discrete_flow_shift",
+        type=float,
+        default=5.0,
+        help="Discrete flow shift for the Euler Discrete Scheduler, default is 5.0. / Euler Discrete Schedulerの離散フローシフト、デフォルトは5.0。",
+    )
+    parser.add_argument("--fp8_scaled", action="store_true", help="Use scaled fp8 for DiT / DiTにスケーリングされたfp8を使う")
+    parser.add_argument("--fp8_vl", action="store_true", help="Use fp8 for VLM text encoder / VLMテキストエンコーダにfp8を使用する")
+    parser.add_argument(
+        "--text_encoder_cpu", action="store_true", help="Inference on CPU for Text Encoders / テキストエンコーダをCPUで推論する"
+    )
+    parser.add_argument(
+        "--vae_chunk_size",
+        type=int,
+        default=None,  # default is None (no chunking)
+        help="Chunk size for VAE decoding to reduce memory usage. Default is None (no chunking). 16 is recommended if enabled"
+        " / メモリ使用量を減らすためのVAEデコードのチャンクサイズ。デフォルトはNone（チャンクなし）。有効にする場合は16程度を推奨。",
+    )
+
+    parser.add_argument(
+        "--attn_mode",
+        choices=["torch", "xformers", "flash", "sageattn", "sdpa"],  # "sdpa" is for backward compatibility
+        default=None,
+        help="Attention implementation to use. Default is None (torch). xformers requires --split_attn. sageattn does not support training (inference only). This option overrides --xformers or --sdpa."
+        " / 使用するAttentionの実装。デフォルトはNone（torch）です。xformersは--split_attnの指定が必要です。sageattnはトレーニングをサポートしていません（推論のみ）。このオプションは--xformersまたは--sdpaを上書きします。",
+    )
+    parser.add_argument(
+        "--split_attn",
+        action="store_true",
+        help="split attention computation to reduce memory usage / メモリ使用量を減らすためにattention時にバッチを分割する",
+    )
+
+    return parser
+
+
+if __name__ == "__main__":
+    parser = setup_parser()
+
+    args = parser.parse_args()
+    train_util.verify_command_line_training_args(args)
+    args = train_util.read_config_from_file(args, parser)
+
+    if args.attn_mode == "sdpa":
+        args.attn_mode = "torch"  # backward compatibility
+
+    trainer = HunyuanImageNetworkTrainer()
+    trainer.train(args)

library/adafactor_fused.py

@@ -0,0 +1,138 @@
+import math
+import torch
+from transformers import Adafactor
+
+# stochastic rounding for bfloat16
+# The implementation was provided by 2kpr. Thank you very much!
+
+def copy_stochastic_(target: torch.Tensor, source: torch.Tensor):
+    """
+    copies source into target using stochastic rounding
+
+    Args:
+        target: the target tensor with dtype=bfloat16
+        source: the target tensor with dtype=float32
+    """
+    # create a random 16 bit integer
+    result = torch.randint_like(source, dtype=torch.int32, low=0, high=(1 << 16))
+
+    # add the random number to the lower 16 bit of the mantissa
+    result.add_(source.view(dtype=torch.int32))
+
+    # mask off the lower 16 bit of the mantissa
+    result.bitwise_and_(-65536)  # -65536 = FFFF0000 as a signed int32
+
+    # copy the higher 16 bit into the target tensor
+    target.copy_(result.view(dtype=torch.float32))
+
+    del result
+
+
+@torch.no_grad()
+def adafactor_step_param(self, p, group):
+    if p.grad is None:
+        return
+    grad = p.grad
+    if grad.dtype in {torch.float16, torch.bfloat16}:
+        grad = grad.float()
+    if grad.is_sparse:
+        raise RuntimeError("Adafactor does not support sparse gradients.")
+
+    state = self.state[p]
+    grad_shape = grad.shape
+
+    factored, use_first_moment = Adafactor._get_options(group, grad_shape)
+    # State Initialization
+    if len(state) == 0:
+        state["step"] = 0
+
+        if use_first_moment:
+            # Exponential moving average of gradient values
+            state["exp_avg"] = torch.zeros_like(grad)
+        if factored:
+            state["exp_avg_sq_row"] = torch.zeros(grad_shape[:-1]).to(grad)
+            state["exp_avg_sq_col"] = torch.zeros(grad_shape[:-2] + grad_shape[-1:]).to(grad)
+        else:
+            state["exp_avg_sq"] = torch.zeros_like(grad)
+
+        state["RMS"] = 0
+    else:
+        if use_first_moment:
+            state["exp_avg"] = state["exp_avg"].to(grad)
+        if factored:
+            state["exp_avg_sq_row"] = state["exp_avg_sq_row"].to(grad)
+            state["exp_avg_sq_col"] = state["exp_avg_sq_col"].to(grad)
+        else:
+            state["exp_avg_sq"] = state["exp_avg_sq"].to(grad)
+
+    p_data_fp32 = p
+    if p.dtype in {torch.float16, torch.bfloat16}:
+        p_data_fp32 = p_data_fp32.float()
+
+    state["step"] += 1
+    state["RMS"] = Adafactor._rms(p_data_fp32)
+    lr = Adafactor._get_lr(group, state)
+
+    beta2t = 1.0 - math.pow(state["step"], group["decay_rate"])
+    update = (grad**2) + group["eps"][0]
+    if factored:
+        exp_avg_sq_row = state["exp_avg_sq_row"]
+        exp_avg_sq_col = state["exp_avg_sq_col"]
+
+        exp_avg_sq_row.mul_(beta2t).add_(update.mean(dim=-1), alpha=(1.0 - beta2t))
+        exp_avg_sq_col.mul_(beta2t).add_(update.mean(dim=-2), alpha=(1.0 - beta2t))
+
+        # Approximation of exponential moving average of square of gradient
+        update = Adafactor._approx_sq_grad(exp_avg_sq_row, exp_avg_sq_col)
+        update.mul_(grad)
+    else:
+        exp_avg_sq = state["exp_avg_sq"]
+
+        exp_avg_sq.mul_(beta2t).add_(update, alpha=(1.0 - beta2t))
+        update = exp_avg_sq.rsqrt().mul_(grad)
+
+    update.div_((Adafactor._rms(update) / group["clip_threshold"]).clamp_(min=1.0))
+    update.mul_(lr)
+
+    if use_first_moment:
+        exp_avg = state["exp_avg"]
+        exp_avg.mul_(group["beta1"]).add_(update, alpha=(1 - group["beta1"]))
+        update = exp_avg
+
+    if group["weight_decay"] != 0:
+        p_data_fp32.add_(p_data_fp32, alpha=(-group["weight_decay"] * lr))
+
+    p_data_fp32.add_(-update)
+
+    # if p.dtype in {torch.float16, torch.bfloat16}:
+    #    p.copy_(p_data_fp32)
+
+    if p.dtype == torch.bfloat16:
+        copy_stochastic_(p, p_data_fp32)
+    elif p.dtype == torch.float16:
+        p.copy_(p_data_fp32)
+
+
+@torch.no_grad()
+def adafactor_step(self, closure=None):
+    """
+    Performs a single optimization step
+
+    Arguments:
+        closure (callable, optional): A closure that reevaluates the model
+            and returns the loss.
+    """
+    loss = None
+    if closure is not None:
+        loss = closure()
+
+    for group in self.param_groups:
+        for p in group["params"]:
+            adafactor_step_param(self, p, group)
+
+    return loss
+
+
+def patch_adafactor_fused(optimizer: Adafactor):
+    optimizer.step_param = adafactor_step_param.__get__(optimizer)
+    optimizer.step = adafactor_step.__get__(optimizer)

library/attention.py

@@ -0,0 +1,260 @@
+# Unified attention function supporting various implementations
+
+from dataclasses import dataclass
+import torch
+from typing import Optional, Union
+
+try:
+    import flash_attn
+    from flash_attn.flash_attn_interface import _flash_attn_forward
+    from flash_attn.flash_attn_interface import flash_attn_varlen_func
+    from flash_attn.flash_attn_interface import flash_attn_func
+except ImportError:
+    flash_attn = None
+    flash_attn_varlen_func = None
+    _flash_attn_forward = None
+    flash_attn_func = None
+
+try:
+    from sageattention import sageattn_varlen, sageattn
+except ImportError:
+    sageattn_varlen = None
+    sageattn = None
+
+try:
+    import xformers.ops as xops
+except ImportError:
+    xops = None
+
+
+@dataclass
+class AttentionParams:
+    attn_mode: Optional[str] = None
+    split_attn: bool = False
+    img_len: Optional[int] = None
+    attention_mask: Optional[torch.Tensor] = None
+    seqlens: Optional[torch.Tensor] = None
+    cu_seqlens: Optional[torch.Tensor] = None
+    max_seqlen: Optional[int] = None
+
+    @staticmethod
+    def create_attention_params(attn_mode: Optional[str], split_attn: bool) -> "AttentionParams":
+        return AttentionParams(attn_mode, split_attn)
+
+    @staticmethod
+    def create_attention_params_from_mask(
+        attn_mode: Optional[str], split_attn: bool, img_len: Optional[int], attention_mask: Optional[torch.Tensor]
+    ) -> "AttentionParams":
+        if attention_mask is None:
+            # No attention mask provided: assume all tokens are valid
+            return AttentionParams(attn_mode, split_attn, None, None, None, None, None)
+        else:
+            # Note: attention_mask is only for text tokens, not including image tokens
+            seqlens = attention_mask.sum(dim=1).to(torch.int32) + img_len  # [B]
+            max_seqlen = attention_mask.shape[1] + img_len
+
+            if split_attn:
+                # cu_seqlens is not needed for split attention
+                return AttentionParams(attn_mode, split_attn, img_len, attention_mask, seqlens, None, max_seqlen)
+
+            # Convert attention mask to cumulative sequence lengths for flash attention
+            batch_size = attention_mask.shape[0]
+            cu_seqlens = torch.zeros([2 * batch_size + 1], dtype=torch.int32, device=attention_mask.device)
+            for i in range(batch_size):
+                cu_seqlens[2 * i + 1] = i * max_seqlen + seqlens[i]  # end of valid tokens for query
+                cu_seqlens[2 * i + 2] = (i + 1) * max_seqlen  # end of all tokens for query
+
+            # Expand attention mask to include image tokens
+            attention_mask = torch.nn.functional.pad(attention_mask, (img_len, 0), value=1)  # [B, img_len + L]
+
+            if attn_mode == "xformers":
+                seqlens_list = seqlens.cpu().tolist()
+                attention_mask = xops.fmha.attn_bias.BlockDiagonalMask.from_seqlens(
+                    seqlens_list, seqlens_list, device=attention_mask.device
+                )
+            elif attn_mode == "torch":
+                attention_mask = attention_mask[:, None, None, :].to(torch.bool)  # [B, 1, 1, img_len + L]
+
+            return AttentionParams(attn_mode, split_attn, img_len, attention_mask, seqlens, cu_seqlens, max_seqlen)
+
+
+def attention(
+    qkv_or_q: Union[torch.Tensor, list],
+    k: Optional[torch.Tensor] = None,
+    v: Optional[torch.Tensor] = None,
+    attn_params: Optional[AttentionParams] = None,
+    drop_rate: float = 0.0,
+) -> torch.Tensor:
+    """
+    Compute scaled dot-product attention with variable sequence lengths.
+
+    Handles batches with different sequence lengths by splitting and
+    processing each sequence individually.
+
+    Args:
+        qkv_or_q: Query tensor [B, L, H, D]. or list of such tensors.
+        k: Key tensor [B, L, H, D].
+        v: Value tensor [B, L, H, D].
+        attn_param: Attention parameters including mask and sequence lengths.
+        drop_rate: Attention dropout rate.
+
+    Returns:
+        Attention output tensor [B, L, H*D].
+    """
+    if isinstance(qkv_or_q, list):
+        q, k, v = qkv_or_q
+        q: torch.Tensor = q
+        qkv_or_q.clear()
+        del qkv_or_q
+    else:
+        q: torch.Tensor = qkv_or_q
+        del qkv_or_q
+        assert k is not None and v is not None, "k and v must be provided if qkv_or_q is a tensor"
+    if attn_params is None:
+        attn_params = AttentionParams.create_attention_params("torch", False)
+
+    # If split attn is False, attention mask is provided and all sequence lengths are same, we can trim the sequence
+    seqlen_trimmed = False
+    if not attn_params.split_attn and attn_params.attention_mask is not None and attn_params.seqlens is not None:
+        if torch.all(attn_params.seqlens == attn_params.seqlens[0]):
+            seqlen = attn_params.seqlens[0].item()
+            q = q[:, :seqlen]
+            k = k[:, :seqlen]
+            v = v[:, :seqlen]
+            max_seqlen = attn_params.max_seqlen
+            attn_params = AttentionParams.create_attention_params(attn_params.attn_mode, False)  # do not in-place modify
+            attn_params.max_seqlen = max_seqlen  # keep max_seqlen for padding
+            seqlen_trimmed = True
+
+    # Determine tensor layout based on attention implementation
+    if attn_params.attn_mode == "torch" or (
+        attn_params.attn_mode == "sageattn" and (attn_params.split_attn or attn_params.cu_seqlens is None)
+    ):
+        transpose_fn = lambda x: x.transpose(1, 2)  # [B, H, L, D] for SDPA and sageattn with fixed length
+        # pad on sequence length dimension
+        pad_fn = lambda x, pad_to: torch.nn.functional.pad(x, (0, 0, 0, pad_to - x.shape[-2]), value=0)
+    else:
+        transpose_fn = lambda x: x  # [B, L, H, D] for other implementations
+        # pad on sequence length dimension
+        pad_fn = lambda x, pad_to: torch.nn.functional.pad(x, (0, 0, 0, 0, 0, pad_to - x.shape[-3]), value=0)
+
+    # Process each batch element with its valid sequence lengths
+    if attn_params.split_attn:
+        if attn_params.seqlens is None:
+            # If no seqlens provided, assume all tokens are valid
+            attn_params = AttentionParams.create_attention_params(attn_params.attn_mode, True)  # do not in-place modify
+            attn_params.seqlens = torch.tensor([q.shape[1]] * q.shape[0], device=q.device)
+            attn_params.max_seqlen = q.shape[1]
+        q = [transpose_fn(q[i : i + 1, : attn_params.seqlens[i]]) for i in range(len(q))]
+        k = [transpose_fn(k[i : i + 1, : attn_params.seqlens[i]]) for i in range(len(k))]
+        v = [transpose_fn(v[i : i + 1, : attn_params.seqlens[i]]) for i in range(len(v))]
+    else:
+        q = transpose_fn(q)
+        k = transpose_fn(k)
+        v = transpose_fn(v)
+
+    if attn_params.attn_mode == "torch":
+        if attn_params.split_attn:
+            x = []
+            for i in range(len(q)):
+                x_i = torch.nn.functional.scaled_dot_product_attention(q[i], k[i], v[i], dropout_p=drop_rate)
+                q[i] = None
+                k[i] = None
+                v[i] = None
+                x.append(pad_fn(x_i, attn_params.max_seqlen))  # B, H, L, D
+            x = torch.cat(x, dim=0)
+            del q, k, v
+
+        else:
+            x = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=attn_params.attention_mask, dropout_p=drop_rate)
+            del q, k, v
+
+    elif attn_params.attn_mode == "xformers":
+        if attn_params.split_attn:
+            x = []
+            for i in range(len(q)):
+                x_i = xops.memory_efficient_attention(q[i], k[i], v[i], p=drop_rate)
+                q[i] = None
+                k[i] = None
+                v[i] = None
+                x.append(pad_fn(x_i, attn_params.max_seqlen))  # B, L, H, D
+            x = torch.cat(x, dim=0)
+            del q, k, v
+
+        else:
+            x = xops.memory_efficient_attention(q, k, v, attn_bias=attn_params.attention_mask, p=drop_rate)
+            del q, k, v
+
+    elif attn_params.attn_mode == "sageattn":
+        if attn_params.split_attn:
+            x = []
+            for i in range(len(q)):
+                # HND seems to cause an error
+                x_i = sageattn(q[i], k[i], v[i])  # B, H, L, D. No dropout support
+                q[i] = None
+                k[i] = None
+                v[i] = None
+                x.append(pad_fn(x_i, attn_params.max_seqlen))  # B, H, L, D
+            x = torch.cat(x, dim=0)
+            del q, k, v
+        elif attn_params.cu_seqlens is None:  # all tokens are valid
+            x = sageattn(q, k, v)  # B, L, H, D. No dropout support
+            del q, k, v
+        else:
+            # Reshape to [(bxs), a, d]
+            batch_size, seqlen = q.shape[0], q.shape[1]
+            q = q.view(q.shape[0] * q.shape[1], *q.shape[2:])  # [B*L, H, D]
+            k = k.view(k.shape[0] * k.shape[1], *k.shape[2:])  # [B*L, H, D]
+            v = v.view(v.shape[0] * v.shape[1], *v.shape[2:])  # [B*L, H, D]
+
+            # Assume cu_seqlens_q == cu_seqlens_kv and max_seqlen_q == max_seqlen_kv. No dropout support
+            x = sageattn_varlen(
+                q, k, v, attn_params.cu_seqlens, attn_params.cu_seqlens, attn_params.max_seqlen, attn_params.max_seqlen
+            )
+            del q, k, v
+
+            # Reshape x with shape [(bxs), a, d] to [b, s, a, d]
+            x = x.view(batch_size, seqlen, x.shape[-2], x.shape[-1])  # B, L, H, D
+
+    elif attn_params.attn_mode == "flash":
+        if attn_params.split_attn:
+            x = []
+            for i in range(len(q)):
+                # HND seems to cause an error
+                x_i = flash_attn_func(q[i], k[i], v[i], drop_rate)  # B, L, H, D
+                q[i] = None
+                k[i] = None
+                v[i] = None
+                x.append(pad_fn(x_i, attn_params.max_seqlen))  # B, L, H, D
+            x = torch.cat(x, dim=0)
+            del q, k, v
+        elif attn_params.cu_seqlens is None:  # all tokens are valid
+            x = flash_attn_func(q, k, v, drop_rate)  # B, L, H, D
+            del q, k, v
+        else:
+            # Reshape to [(bxs), a, d]
+            batch_size, seqlen = q.shape[0], q.shape[1]
+            q = q.view(q.shape[0] * q.shape[1], *q.shape[2:])  # [B*L, H, D]
+            k = k.view(k.shape[0] * k.shape[1], *k.shape[2:])  # [B*L, H, D]
+            v = v.view(v.shape[0] * v.shape[1], *v.shape[2:])  # [B*L, H, D]
+
+            # Assume cu_seqlens_q == cu_seqlens_kv and max_seqlen_q == max_seqlen_kv
+            x = flash_attn_varlen_func(
+                q, k, v, attn_params.cu_seqlens, attn_params.cu_seqlens, attn_params.max_seqlen, attn_params.max_seqlen, drop_rate
+            )
+            del q, k, v
+
+            # Reshape x with shape [(bxs), a, d] to [b, s, a, d]
+            x = x.view(batch_size, seqlen, x.shape[-2], x.shape[-1])  # B, L, H, D
+
+    else:
+        # Currently only PyTorch SDPA and xformers are implemented
+        raise ValueError(f"Unsupported attention mode: {attn_params.attn_mode}")
+
+    x = transpose_fn(x)  # [B, L, H, D]
+    x = x.reshape(x.shape[0], x.shape[1], -1)  # [B, L, H*D]
+
+    if seqlen_trimmed:
+        x = torch.nn.functional.pad(x, (0, 0, 0, attn_params.max_seqlen - x.shape[1]), value=0)  # pad back to max_seqlen
+
+    return x

library/chroma_models.py

@@ -0,0 +1,744 @@
+# copy from the official repo: https://github.com/lodestone-rock/flow/blob/master/src/models/chroma/model.py
+# and modified
+# licensed under Apache License 2.0
+
+import math
+from dataclasses import dataclass
+
+import torch
+from einops import rearrange
+from torch import Tensor, nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as ckpt
+
+from .flux_models import attention, rope, apply_rope, EmbedND, timestep_embedding, MLPEmbedder, RMSNorm, QKNorm, SelfAttention, Flux
+from . import custom_offloading_utils
+
+
+def distribute_modulations(tensor: torch.Tensor, depth_single_blocks, depth_double_blocks):
+    """
+    Distributes slices of the tensor into the block_dict as ModulationOut objects.
+
+    Args:
+        tensor (torch.Tensor): Input tensor with shape [batch_size, vectors, dim].
+    """
+    batch_size, vectors, dim = tensor.shape
+
+    block_dict = {}
+
+    # HARD CODED VALUES! lookup table for the generated vectors
+    # TODO: move this into chroma config!
+    # Add 38 single mod blocks
+    for i in range(depth_single_blocks):
+        key = f"single_blocks.{i}.modulation.lin"
+        block_dict[key] = None
+
+    # Add 19 image double blocks
+    for i in range(depth_double_blocks):
+        key = f"double_blocks.{i}.img_mod.lin"
+        block_dict[key] = None
+
+    # Add 19 text double blocks
+    for i in range(depth_double_blocks):
+        key = f"double_blocks.{i}.txt_mod.lin"
+        block_dict[key] = None
+
+    # Add the final layer
+    block_dict["final_layer.adaLN_modulation.1"] = None
+    # 6.2b version
+    # block_dict["lite_double_blocks.4.img_mod.lin"] = None
+    # block_dict["lite_double_blocks.4.txt_mod.lin"] = None
+
+    idx = 0  # Index to keep track of the vector slices
+
+    for key in block_dict.keys():
+        if "single_blocks" in key:
+            # Single block: 1 ModulationOut
+            block_dict[key] = ModulationOut(
+                shift=tensor[:, idx : idx + 1, :],
+                scale=tensor[:, idx + 1 : idx + 2, :],
+                gate=tensor[:, idx + 2 : idx + 3, :],
+            )
+            idx += 3  # Advance by 3 vectors
+
+        elif "img_mod" in key:
+            # Double block: List of 2 ModulationOut
+            double_block = []
+            for _ in range(2):  # Create 2 ModulationOut objects
+                double_block.append(
+                    ModulationOut(
+                        shift=tensor[:, idx : idx + 1, :],
+                        scale=tensor[:, idx + 1 : idx + 2, :],
+                        gate=tensor[:, idx + 2 : idx + 3, :],
+                    )
+                )
+                idx += 3  # Advance by 3 vectors per ModulationOut
+            block_dict[key] = double_block
+
+        elif "txt_mod" in key:
+            # Double block: List of 2 ModulationOut
+            double_block = []
+            for _ in range(2):  # Create 2 ModulationOut objects
+                double_block.append(
+                    ModulationOut(
+                        shift=tensor[:, idx : idx + 1, :],
+                        scale=tensor[:, idx + 1 : idx + 2, :],
+                        gate=tensor[:, idx + 2 : idx + 3, :],
+                    )
+                )
+                idx += 3  # Advance by 3 vectors per ModulationOut
+            block_dict[key] = double_block
+
+        elif "final_layer" in key:
+            # Final layer: 1 ModulationOut
+            block_dict[key] = [
+                tensor[:, idx : idx + 1, :],
+                tensor[:, idx + 1 : idx + 2, :],
+            ]
+            idx += 2  # Advance by 3 vectors
+
+    return block_dict
+
+
+class Approximator(nn.Module):
+    def __init__(self, in_dim: int, out_dim: int, hidden_dim: int, n_layers=4):
+        super().__init__()
+        self.in_proj = nn.Linear(in_dim, hidden_dim, bias=True)
+        self.layers = nn.ModuleList([MLPEmbedder(hidden_dim, hidden_dim) for x in range(n_layers)])
+        self.norms = nn.ModuleList([RMSNorm(hidden_dim) for x in range(n_layers)])
+        self.out_proj = nn.Linear(hidden_dim, out_dim)
+
+    @property
+    def device(self):
+        # Get the device of the module (assumes all parameters are on the same device)
+        return next(self.parameters()).device
+
+    def enable_gradient_checkpointing(self):
+        for layer in self.layers:
+            layer.enable_gradient_checkpointing()
+
+    def disable_gradient_checkpointing(self):
+        for layer in self.layers:
+            layer.disable_gradient_checkpointing()
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = self.in_proj(x)
+
+        for layer, norms in zip(self.layers, self.norms):
+            x = x + layer(norms(x))
+
+        x = self.out_proj(x)
+
+        return x
+
+
+@dataclass
+class ModulationOut:
+    shift: Tensor
+    scale: Tensor
+    gate: Tensor
+
+
+def _modulation_shift_scale_fn(x, scale, shift):
+    return (1 + scale) * x + shift
+
+
+def _modulation_gate_fn(x, gate, gate_params):
+    return x + gate * gate_params
+
+
+class DoubleStreamBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        mlp_ratio: float,
+        qkv_bias: bool = False,
+    ):
+        super().__init__()
+
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.img_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.img_attn = SelfAttention(
+            dim=hidden_size,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+        )
+
+        self.img_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.img_mlp = nn.Sequential(
+            nn.Linear(hidden_size, mlp_hidden_dim, bias=True),
+            nn.GELU(approximate="tanh"),
+            nn.Linear(mlp_hidden_dim, hidden_size, bias=True),
+        )
+
+        self.txt_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.txt_attn = SelfAttention(
+            dim=hidden_size,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+        )
+
+        self.txt_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.txt_mlp = nn.Sequential(
+            nn.Linear(hidden_size, mlp_hidden_dim, bias=True),
+            nn.GELU(approximate="tanh"),
+            nn.Linear(mlp_hidden_dim, hidden_size, bias=True),
+        )
+
+        self.gradient_checkpointing = False
+
+    @property
+    def device(self):
+        # Get the device of the module (assumes all parameters are on the same device)
+        return next(self.parameters()).device
+
+    def modulation_shift_scale_fn(self, x, scale, shift):
+        return _modulation_shift_scale_fn(x, scale, shift)
+
+    def modulation_gate_fn(self, x, gate, gate_params):
+        return _modulation_gate_fn(x, gate, gate_params)
+
+    def enable_gradient_checkpointing(self):
+        self.gradient_checkpointing = True
+
+    def disable_gradient_checkpointing(self):
+        self.gradient_checkpointing = False
+
+    def _forward(
+        self,
+        img: Tensor,
+        txt: Tensor,
+        pe: list[Tensor],
+        distill_vec: list[ModulationOut],
+        txt_seq_len: Tensor,
+    ) -> tuple[Tensor, Tensor]:
+        (img_mod1, img_mod2), (txt_mod1, txt_mod2) = distill_vec
+
+        # prepare image for attention
+        img_modulated = self.img_norm1(img)
+        # replaced with compiled fn
+        # img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
+        img_modulated = self.modulation_shift_scale_fn(img_modulated, img_mod1.scale, img_mod1.shift)
+        img_qkv = self.img_attn.qkv(img_modulated)
+        del img_modulated
+
+        img_q, img_k, img_v = rearrange(img_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        del img_qkv
+        img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
+
+        # prepare txt for attention
+        txt_modulated = self.txt_norm1(txt)
+        # replaced with compiled fn
+        # txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
+        txt_modulated = self.modulation_shift_scale_fn(txt_modulated, txt_mod1.scale, txt_mod1.shift)
+        txt_qkv = self.txt_attn.qkv(txt_modulated)
+        del txt_modulated
+
+        txt_q, txt_k, txt_v = rearrange(txt_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        del txt_qkv
+        txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
+
+        # run actual attention: we split the batch into each element
+        max_txt_len = torch.max(txt_seq_len).item()
+        img_len = img_q.shape[-2]  # max 64
+        txt_q = list(torch.chunk(txt_q, txt_q.shape[0], dim=0))  # list of [B, H, L, D] tensors
+        txt_k = list(torch.chunk(txt_k, txt_k.shape[0], dim=0))
+        txt_v = list(torch.chunk(txt_v, txt_v.shape[0], dim=0))
+        img_q = list(torch.chunk(img_q, img_q.shape[0], dim=0))
+        img_k = list(torch.chunk(img_k, img_k.shape[0], dim=0))
+        img_v = list(torch.chunk(img_v, img_v.shape[0], dim=0))
+        txt_attn = []
+        img_attn = []
+        for i in range(txt.shape[0]):
+            txt_q[i] = txt_q[i][:, :, : txt_seq_len[i]]
+            q = torch.cat((img_q[i], txt_q[i]), dim=2)
+            txt_q[i] = None
+            img_q[i] = None
+
+            txt_k[i] = txt_k[i][:, :, : txt_seq_len[i]]
+            k = torch.cat((img_k[i], txt_k[i]), dim=2)
+            txt_k[i] = None
+            img_k[i] = None
+
+            txt_v[i] = txt_v[i][:, :, : txt_seq_len[i]]
+            v = torch.cat((img_v[i], txt_v[i]), dim=2)
+            txt_v[i] = None
+            img_v[i] = None
+
+            attn = attention(q, k, v, pe=pe[i : i + 1, :, : q.shape[2]], attn_mask=None)  # attn = (1, L, D)
+            del q, k, v
+            img_attn_i = attn[:, :img_len, :]
+            txt_attn_i = torch.zeros((1, max_txt_len, attn.shape[-1]), dtype=attn.dtype, device=self.device)
+            txt_attn_i[:, : txt_seq_len[i], :] = attn[:, img_len:, :]
+            del attn
+            txt_attn.append(txt_attn_i)
+            img_attn.append(img_attn_i)
+
+        txt_attn = torch.cat(txt_attn, dim=0)
+        img_attn = torch.cat(img_attn, dim=0)
+
+        # q = torch.cat((txt_q, img_q), dim=2)
+        # k = torch.cat((txt_k, img_k), dim=2)
+        # v = torch.cat((txt_v, img_v), dim=2)
+
+        # attn = attention(q, k, v, pe=pe, attn_mask=mask)
+        # txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
+
+        # calculate the img blocks
+        # replaced with compiled fn
+        # img = img + img_mod1.gate * self.img_attn.proj(img_attn)
+        # img = img + img_mod2.gate * self.img_mlp((1 + img_mod2.scale) * self.img_norm2(img) + img_mod2.shift)
+        img = self.modulation_gate_fn(img, img_mod1.gate, self.img_attn.proj(img_attn))
+        del img_attn, img_mod1
+        img = self.modulation_gate_fn(
+            img,
+            img_mod2.gate,
+            self.img_mlp(self.modulation_shift_scale_fn(self.img_norm2(img), img_mod2.scale, img_mod2.shift)),
+        )
+        del img_mod2
+
+        # calculate the txt blocks
+        # replaced with compiled fn
+        # txt = txt + txt_mod1.gate * self.txt_attn.proj(txt_attn)
+        # txt = txt + txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift)
+        txt = self.modulation_gate_fn(txt, txt_mod1.gate, self.txt_attn.proj(txt_attn))
+        del txt_attn, txt_mod1
+        txt = self.modulation_gate_fn(
+            txt,
+            txt_mod2.gate,
+            self.txt_mlp(self.modulation_shift_scale_fn(self.txt_norm2(txt), txt_mod2.scale, txt_mod2.shift)),
+        )
+        del txt_mod2
+
+        return img, txt
+
+    def forward(
+        self,
+        img: Tensor,
+        txt: Tensor,
+        pe: Tensor,
+        distill_vec: list[ModulationOut],
+        txt_seq_len: Tensor,
+    ) -> tuple[Tensor, Tensor]:
+        if self.training and self.gradient_checkpointing:
+            return ckpt.checkpoint(self._forward, img, txt, pe, distill_vec, txt_seq_len, use_reentrant=False)
+        else:
+            return self._forward(img, txt, pe, distill_vec, txt_seq_len)
+
+
+class SingleStreamBlock(nn.Module):
+    """
+    A DiT block with parallel linear layers as described in
+    https://arxiv.org/abs/2302.05442 and adapted modulation interface.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qk_scale: float | None = None,
+    ):
+        super().__init__()
+        self.hidden_dim = hidden_size
+        self.num_heads = num_heads
+        head_dim = hidden_size // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        # qkv and mlp_in
+        self.linear1 = nn.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim)
+        # proj and mlp_out
+        self.linear2 = nn.Linear(hidden_size + self.mlp_hidden_dim, hidden_size)
+
+        self.norm = QKNorm(head_dim)
+
+        self.hidden_size = hidden_size
+        self.pre_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+
+        self.mlp_act = nn.GELU(approximate="tanh")
+
+        self.gradient_checkpointing = False
+
+    @property
+    def device(self):
+        # Get the device of the module (assumes all parameters are on the same device)
+        return next(self.parameters()).device
+
+    def modulation_shift_scale_fn(self, x, scale, shift):
+        return _modulation_shift_scale_fn(x, scale, shift)
+
+    def modulation_gate_fn(self, x, gate, gate_params):
+        return _modulation_gate_fn(x, gate, gate_params)
+
+    def enable_gradient_checkpointing(self):
+        self.gradient_checkpointing = True
+
+    def disable_gradient_checkpointing(self):
+        self.gradient_checkpointing = False
+
+    def _forward(self, x: Tensor, pe: list[Tensor], distill_vec: list[ModulationOut], txt_seq_len: Tensor) -> Tensor:
+        mod = distill_vec
+        # replaced with compiled fn
+        # x_mod = (1 + mod.scale) * self.pre_norm(x) + mod.shift
+        x_mod = self.modulation_shift_scale_fn(self.pre_norm(x), mod.scale, mod.shift)
+        qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
+        del x_mod
+
+        q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        del qkv
+        q, k = self.norm(q, k, v)
+
+        # # compute attention
+        # attn = attention(q, k, v, pe=pe, attn_mask=mask)
+
+        # compute attention: we split the batch into each element
+        max_txt_len = torch.max(txt_seq_len).item()
+        img_len = q.shape[-2] - max_txt_len
+        q = list(torch.chunk(q, q.shape[0], dim=0))
+        k = list(torch.chunk(k, k.shape[0], dim=0))
+        v = list(torch.chunk(v, v.shape[0], dim=0))
+        attn = []
+        for i in range(x.size(0)):
+            q[i] = q[i][:, :, : img_len + txt_seq_len[i]]
+            k[i] = k[i][:, :, : img_len + txt_seq_len[i]]
+            v[i] = v[i][:, :, : img_len + txt_seq_len[i]]
+            attn_trimmed = attention(q[i], k[i], v[i], pe=pe[i : i + 1, :, : img_len + txt_seq_len[i]], attn_mask=None)
+            q[i] = None
+            k[i] = None
+            v[i] = None
+
+            attn_i = torch.zeros((1, x.shape[1], attn_trimmed.shape[-1]), dtype=attn_trimmed.dtype, device=self.device)
+            attn_i[:, : img_len + txt_seq_len[i], :] = attn_trimmed
+            del attn_trimmed
+            attn.append(attn_i)
+
+        attn = torch.cat(attn, dim=0)
+
+        # compute activation in mlp stream, cat again and run second linear layer
+        mlp = self.mlp_act(mlp)
+        output = self.linear2(torch.cat((attn, mlp), 2))
+        del attn, mlp
+        # replaced with compiled fn
+        # return x + mod.gate * output
+        return self.modulation_gate_fn(x, mod.gate, output)
+
+    def forward(self, x: Tensor, pe: Tensor, distill_vec: list[ModulationOut], txt_seq_len: Tensor) -> Tensor:
+        if self.training and self.gradient_checkpointing:
+            return ckpt.checkpoint(self._forward, x, pe, distill_vec, txt_seq_len, use_reentrant=False)
+        else:
+            return self._forward(x, pe, distill_vec, txt_seq_len)
+
+
+class LastLayer(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        patch_size: int,
+        out_channels: int,
+    ):
+        super().__init__()
+        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
+
+    @property
+    def device(self):
+        # Get the device of the module (assumes all parameters are on the same device)
+        return next(self.parameters()).device
+
+    def modulation_shift_scale_fn(self, x, scale, shift):
+        return _modulation_shift_scale_fn(x, scale, shift)
+
+    def forward(self, x: Tensor, distill_vec: list[Tensor]) -> Tensor:
+        shift, scale = distill_vec
+        shift = shift.squeeze(1)
+        scale = scale.squeeze(1)
+        # replaced with compiled fn
+        # x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
+        x = self.modulation_shift_scale_fn(self.norm_final(x), scale[:, None, :], shift[:, None, :])
+        x = self.linear(x)
+        return x
+
+
+@dataclass
+class ChromaParams:
+    in_channels: int
+    context_in_dim: int
+    hidden_size: int
+    mlp_ratio: float
+    num_heads: int
+    depth: int
+    depth_single_blocks: int
+    axes_dim: list[int]
+    theta: int
+    qkv_bias: bool
+    guidance_embed: bool
+    approximator_in_dim: int
+    approximator_depth: int
+    approximator_hidden_size: int
+    _use_compiled: bool
+
+
+chroma_params = ChromaParams(
+    in_channels=64,
+    context_in_dim=4096,
+    hidden_size=3072,
+    mlp_ratio=4.0,
+    num_heads=24,
+    depth=19,
+    depth_single_blocks=38,
+    axes_dim=[16, 56, 56],
+    theta=10_000,
+    qkv_bias=True,
+    guidance_embed=True,
+    approximator_in_dim=64,
+    approximator_depth=5,
+    approximator_hidden_size=5120,
+    _use_compiled=False,
+)
+
+
+def modify_mask_to_attend_padding(mask, max_seq_length, num_extra_padding=8):
+    """
+    Modifies attention mask to allow attention to a few extra padding tokens.
+
+    Args:
+        mask: Original attention mask (1 for tokens to attend to, 0 for masked tokens)
+        max_seq_length: Maximum sequence length of the model
+        num_extra_padding: Number of padding tokens to unmask
+
+    Returns:
+        Modified mask
+    """
+    # Get the actual sequence length from the mask
+    seq_length = mask.sum(dim=-1)
+    batch_size = mask.shape[0]
+
+    modified_mask = mask.clone()
+
+    for i in range(batch_size):
+        current_seq_len = int(seq_length[i].item())
+
+        # Only add extra padding tokens if there's room
+        if current_seq_len < max_seq_length:
+            # Calculate how many padding tokens we can unmask
+            available_padding = max_seq_length - current_seq_len
+            tokens_to_unmask = min(num_extra_padding, available_padding)
+
+            # Unmask the specified number of padding tokens right after the sequence
+            modified_mask[i, current_seq_len : current_seq_len + tokens_to_unmask] = 1
+
+    return modified_mask
+
+
+class Chroma(Flux):
+    """
+    Transformer model for flow matching on sequences.
+    """
+
+    def __init__(self, params: ChromaParams):
+        nn.Module.__init__(self)
+        self.params = params
+        self.in_channels = params.in_channels
+        self.out_channels = self.in_channels
+        if params.hidden_size % params.num_heads != 0:
+            raise ValueError(f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}")
+        pe_dim = params.hidden_size // params.num_heads
+        if sum(params.axes_dim) != pe_dim:
+            raise ValueError(f"Got {params.axes_dim} but expected positional dim {pe_dim}")
+        self.hidden_size = params.hidden_size
+        self.num_heads = params.num_heads
+        self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)
+        self.img_in = nn.Linear(self.in_channels, self.hidden_size, bias=True)
+
+        # TODO: need proper mapping for this approximator output!
+        # currently the mapping is hardcoded in distribute_modulations function
+        self.distilled_guidance_layer = Approximator(
+            params.approximator_in_dim,
+            self.hidden_size,
+            params.approximator_hidden_size,
+            params.approximator_depth,
+        )
+        self.txt_in = nn.Linear(params.context_in_dim, self.hidden_size)
+
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    self.hidden_size,
+                    self.num_heads,
+                    mlp_ratio=params.mlp_ratio,
+                    qkv_bias=params.qkv_bias,
+                )
+                for _ in range(params.depth)
+            ]
+        )
+
+        self.single_blocks = nn.ModuleList(
+            [
+                SingleStreamBlock(
+                    self.hidden_size,
+                    self.num_heads,
+                    mlp_ratio=params.mlp_ratio,
+                )
+                for _ in range(params.depth_single_blocks)
+            ]
+        )
+
+        self.final_layer = LastLayer(
+            self.hidden_size,
+            1,
+            self.out_channels,
+        )
+
+        # TODO: move this hardcoded value to config
+        # single layer has 3 modulation vectors
+        # double layer has 6 modulation vectors for each expert
+        # final layer has 2 modulation vectors
+        self.mod_index_length = 3 * params.depth_single_blocks + 2 * 6 * params.depth + 2
+        self.depth_single_blocks = params.depth_single_blocks
+        self.depth_double_blocks = params.depth
+        # self.mod_index = torch.tensor(list(range(self.mod_index_length)), device=0)
+        self.register_buffer(
+            "mod_index",
+            torch.tensor(list(range(self.mod_index_length)), device="cpu"),
+            persistent=False,
+        )
+        self.approximator_in_dim = params.approximator_in_dim
+
+        self.blocks_to_swap = None
+        self.offloader_double = None
+        self.offloader_single = None
+        self.num_double_blocks = len(self.double_blocks)
+        self.num_single_blocks = len(self.single_blocks)
+
+        # Initialize properties required by Flux parent class
+        self.gradient_checkpointing = False
+        self.cpu_offload_checkpointing = False
+
+    def get_model_type(self) -> str:
+        return "chroma"
+
+    def enable_gradient_checkpointing(self, cpu_offload: bool = False):
+        self.gradient_checkpointing = True
+        self.cpu_offload_checkpointing = cpu_offload
+
+        self.distilled_guidance_layer.enable_gradient_checkpointing()
+        for block in self.double_blocks + self.single_blocks:
+            block.enable_gradient_checkpointing()
+
+        print(f"Chroma: Gradient checkpointing enabled.")
+
+    def disable_gradient_checkpointing(self):
+        self.gradient_checkpointing = False
+        self.cpu_offload_checkpointing = False
+
+        self.distilled_guidance_layer.disable_gradient_checkpointing()
+        for block in self.double_blocks + self.single_blocks:
+            block.disable_gradient_checkpointing()
+
+        print("Chroma: Gradient checkpointing disabled.")
+
+    def get_mod_vectors(self, timesteps: Tensor, guidance: Tensor | None = None, batch_size: int | None = None) -> Tensor:
+        # We extract this logic from forward to clarify the propagation of the gradients
+        # original comment: https://github.com/lodestone-rock/flow/blob/c76f63058980d0488826936025889e256a2e0458/src/models/chroma/model.py#L195
+
+        # print(f"Chroma get_input_vec: timesteps {timesteps}, guidance: {guidance}, batch_size: {batch_size}")
+        distill_timestep = timestep_embedding(timesteps, self.approximator_in_dim // 4)
+        # TODO: need to add toggle to omit this from schnell but that's not a priority
+        distil_guidance = timestep_embedding(guidance, self.approximator_in_dim // 4)
+        # get all modulation index
+        modulation_index = timestep_embedding(self.mod_index, self.approximator_in_dim // 2)
+        # we need to broadcast the modulation index here so each batch has all of the index
+        modulation_index = modulation_index.unsqueeze(0).repeat(batch_size, 1, 1)
+        # and we need to broadcast timestep and guidance along too
+        timestep_guidance = torch.cat([distill_timestep, distil_guidance], dim=1).unsqueeze(1).repeat(1, self.mod_index_length, 1)
+        # then and only then we could concatenate it together
+        input_vec = torch.cat([timestep_guidance, modulation_index], dim=-1)
+
+        mod_vectors = self.distilled_guidance_layer(input_vec)
+        return mod_vectors
+
+    def forward(
+        self,
+        img: Tensor,
+        img_ids: Tensor,
+        txt: Tensor,
+        txt_ids: Tensor,
+        timesteps: Tensor,
+        y: Tensor,
+        block_controlnet_hidden_states=None,
+        block_controlnet_single_hidden_states=None,
+        guidance: Tensor | None = None,
+        txt_attention_mask: Tensor | None = None,
+        attn_padding: int = 1,
+        mod_vectors: Tensor | None = None,
+    ) -> Tensor:
+        # print(
+        #     f"Chroma forward: img shape {img.shape}, txt shape {txt.shape}, img_ids shape {img_ids.shape}, txt_ids shape {txt_ids.shape}"
+        # )
+        # print(f"input_vec shape: {input_vec.shape if input_vec is not None else 'None'}")
+        # print(f"timesteps: {timesteps}, guidance: {guidance}")
+
+        if img.ndim != 3 or txt.ndim != 3:
+            raise ValueError("Input img and txt tensors must have 3 dimensions.")
+
+        # running on sequences img
+        img = self.img_in(img)
+        txt = self.txt_in(txt)
+
+        if mod_vectors is None:  # fallback to the original logic
+            with torch.no_grad():
+                mod_vectors = self.get_mod_vectors(timesteps, guidance, img.shape[0])
+        mod_vectors_dict = distribute_modulations(mod_vectors, self.depth_single_blocks, self.depth_double_blocks)
+
+        # calculate text length for each batch instead of masking
+        txt_emb_len = txt.shape[1]
+        txt_seq_len = txt_attention_mask[:, :txt_emb_len].sum(dim=-1).to(torch.int64)  # (batch_size, )
+        txt_seq_len = torch.clip(txt_seq_len + attn_padding, 0, txt_emb_len)
+        max_txt_len = torch.max(txt_seq_len).item()  # max text length in the batch
+        # print(f"max_txt_len: {max_txt_len}, txt_seq_len: {txt_seq_len}")
+
+        # trim txt embedding to the text length
+        txt = txt[:, :max_txt_len, :]
+
+        # create positional encoding for the text and image
+        ids = torch.cat((img_ids, txt_ids[:, :max_txt_len]), dim=1)  # reverse order of ids for faster attention
+        pe = self.pe_embedder(ids)  # B, 1, seq_length, 64, 2, 2
+
+        for i, block in enumerate(self.double_blocks):
+            if self.blocks_to_swap:
+                self.offloader_double.wait_for_block(i)
+
+            # the guidance replaced by FFN output
+            img_mod = mod_vectors_dict.pop(f"double_blocks.{i}.img_mod.lin")
+            txt_mod = mod_vectors_dict.pop(f"double_blocks.{i}.txt_mod.lin")
+            double_mod = [img_mod, txt_mod]
+            del img_mod, txt_mod
+
+            img, txt = block(img=img, txt=txt, pe=pe, distill_vec=double_mod, txt_seq_len=txt_seq_len)
+            del double_mod
+
+            if self.blocks_to_swap:
+                self.offloader_double.submit_move_blocks(self.double_blocks, i)
+
+        img = torch.cat((img, txt), 1)
+        del txt
+
+        for i, block in enumerate(self.single_blocks):
+            if self.blocks_to_swap:
+                self.offloader_single.wait_for_block(i)
+
+            single_mod = mod_vectors_dict.pop(f"single_blocks.{i}.modulation.lin")
+            img = block(img, pe=pe, distill_vec=single_mod, txt_seq_len=txt_seq_len)
+            del single_mod
+
+            if self.blocks_to_swap:
+                self.offloader_single.submit_move_blocks(self.single_blocks, i)
+
+        img = img[:, :-max_txt_len, ...]
+        final_mod = mod_vectors_dict["final_layer.adaLN_modulation.1"]
+        img = self.final_layer(img, distill_vec=final_mod)  # (N, T, patch_size ** 2 * out_channels)
+        return img

library/custom_offloading_utils.py

@@ -0,0 +1,338 @@
+from concurrent.futures import ThreadPoolExecutor
+import gc
+import time
+from typing import Any, Optional, Union, Callable, Tuple
+import torch
+import torch.nn as nn
+
+
+# Keep these functions here for portability, and private to avoid confusion with the ones in device_utils.py
+def _clean_memory_on_device(device: torch.device):
+    r"""
+    Clean memory on the specified device, will be called from training scripts.
+    """
+    gc.collect()
+
+    # device may "cuda" or "cuda:0", so we need to check the type of device
+    if device.type == "cuda":
+        torch.cuda.empty_cache()
+    if device.type == "xpu":
+        torch.xpu.empty_cache()
+    if device.type == "mps":
+        torch.mps.empty_cache()
+
+
+def _synchronize_device(device: torch.device):
+    if device.type == "cuda":
+        torch.cuda.synchronize()
+    elif device.type == "xpu":
+        torch.xpu.synchronize()
+    elif device.type == "mps":
+        torch.mps.synchronize()
+
+
+def swap_weight_devices_cuda(device: torch.device, layer_to_cpu: nn.Module, layer_to_cuda: nn.Module):
+    assert layer_to_cpu.__class__ == layer_to_cuda.__class__
+
+    weight_swap_jobs: list[Tuple[nn.Module, nn.Module, torch.Tensor, torch.Tensor]] = []
+
+    # This is not working for all cases (e.g. SD3), so we need to find the corresponding modules
+    # for module_to_cpu, module_to_cuda in zip(layer_to_cpu.modules(), layer_to_cuda.modules()):
+    #     print(module_to_cpu.__class__, module_to_cuda.__class__)
+    #     if hasattr(module_to_cpu, "weight") and module_to_cpu.weight is not None:
+    #         weight_swap_jobs.append((module_to_cpu, module_to_cuda, module_to_cpu.weight.data, module_to_cuda.weight.data))
+
+    modules_to_cpu = {k: v for k, v in layer_to_cpu.named_modules()}
+    for module_to_cuda_name, module_to_cuda in layer_to_cuda.named_modules():
+        if hasattr(module_to_cuda, "weight") and module_to_cuda.weight is not None:
+            module_to_cpu = modules_to_cpu.get(module_to_cuda_name, None)
+            if module_to_cpu is not None and module_to_cpu.weight.shape == module_to_cuda.weight.shape:
+                weight_swap_jobs.append((module_to_cpu, module_to_cuda, module_to_cpu.weight.data, module_to_cuda.weight.data))
+            else:
+                if module_to_cuda.weight.data.device.type != device.type:
+                    # print(
+                    #     f"Module {module_to_cuda_name} not found in CPU model or shape mismatch, so not swapping and moving to device"
+                    # )
+                    module_to_cuda.weight.data = module_to_cuda.weight.data.to(device)
+
+    torch.cuda.current_stream().synchronize()  # this prevents the illegal loss value
+
+    stream = torch.Stream(device="cuda")
+    with torch.cuda.stream(stream):
+        # cuda to cpu
+        for module_to_cpu, module_to_cuda, cuda_data_view, cpu_data_view in weight_swap_jobs:
+            cuda_data_view.record_stream(stream)
+            module_to_cpu.weight.data = cuda_data_view.data.to("cpu", non_blocking=True)
+
+        stream.synchronize()
+
+        # cpu to cuda
+        for module_to_cpu, module_to_cuda, cuda_data_view, cpu_data_view in weight_swap_jobs:
+            cuda_data_view.copy_(module_to_cuda.weight.data, non_blocking=True)
+            module_to_cuda.weight.data = cuda_data_view
+
+    stream.synchronize()
+    torch.cuda.current_stream().synchronize()  # this prevents the illegal loss value
+
+
+def swap_weight_devices_no_cuda(device: torch.device, layer_to_cpu: nn.Module, layer_to_cuda: nn.Module):
+    """
+    not tested
+    """
+    assert layer_to_cpu.__class__ == layer_to_cuda.__class__
+
+    weight_swap_jobs: list[Tuple[nn.Module, nn.Module, torch.Tensor, torch.Tensor]] = []
+    for module_to_cpu, module_to_cuda in zip(layer_to_cpu.modules(), layer_to_cuda.modules()):
+        if hasattr(module_to_cpu, "weight") and module_to_cpu.weight is not None:
+            weight_swap_jobs.append((module_to_cpu, module_to_cuda, module_to_cpu.weight.data, module_to_cuda.weight.data))
+
+    # device to cpu
+    for module_to_cpu, module_to_cuda, cuda_data_view, cpu_data_view in weight_swap_jobs:
+        module_to_cpu.weight.data = cuda_data_view.data.to("cpu", non_blocking=True)
+
+    _synchronize_device(device)
+
+    # cpu to device
+    for module_to_cpu, module_to_cuda, cuda_data_view, cpu_data_view in weight_swap_jobs:
+        cuda_data_view.copy_(module_to_cuda.weight.data, non_blocking=True)
+        module_to_cuda.weight.data = cuda_data_view
+
+    _synchronize_device(device)
+
+
+def weighs_to_device(layer: nn.Module, device: torch.device):
+    for module in layer.modules():
+        if hasattr(module, "weight") and module.weight is not None:
+            module.weight.data = module.weight.data.to(device, non_blocking=True)
+
+
+class Offloader:
+    """
+    common offloading class
+    """
+
+    def __init__(self, num_blocks: int, blocks_to_swap: int, device: torch.device, debug: bool = False):
+        self.num_blocks = num_blocks
+        self.blocks_to_swap = blocks_to_swap
+        self.device = device
+        self.debug = debug
+
+        self.thread_pool = ThreadPoolExecutor(max_workers=1)
+        self.futures = {}
+        self.cuda_available = device.type == "cuda"
+
+    def swap_weight_devices(self, block_to_cpu: nn.Module, block_to_cuda: nn.Module):
+        if self.cuda_available:
+            swap_weight_devices_cuda(self.device, block_to_cpu, block_to_cuda)
+        else:
+            swap_weight_devices_no_cuda(self.device, block_to_cpu, block_to_cuda)
+
+    def _submit_move_blocks(self, blocks, block_idx_to_cpu, block_idx_to_cuda):
+        def move_blocks(bidx_to_cpu, block_to_cpu, bidx_to_cuda, block_to_cuda):
+            if self.debug:
+                start_time = time.perf_counter()
+                print(f"Move block {bidx_to_cpu} to CPU and block {bidx_to_cuda} to {'CUDA' if self.cuda_available else 'device'}")
+
+            self.swap_weight_devices(block_to_cpu, block_to_cuda)
+
+            if self.debug:
+                print(f"Moved blocks {bidx_to_cpu} and {bidx_to_cuda} in {time.perf_counter() - start_time:.2f}s")
+            return bidx_to_cpu, bidx_to_cuda  # , event
+
+        block_to_cpu = blocks[block_idx_to_cpu]
+        block_to_cuda = blocks[block_idx_to_cuda]
+
+        self.futures[block_idx_to_cuda] = self.thread_pool.submit(
+            move_blocks, block_idx_to_cpu, block_to_cpu, block_idx_to_cuda, block_to_cuda
+        )
+
+    def _wait_blocks_move(self, block_idx):
+        if block_idx not in self.futures:
+            return
+
+        if self.debug:
+            print(f"Wait for block {block_idx}")
+            start_time = time.perf_counter()
+
+        future = self.futures.pop(block_idx)
+        _, bidx_to_cuda = future.result()
+
+        assert block_idx == bidx_to_cuda, f"Block index mismatch: {block_idx} != {bidx_to_cuda}"
+
+        if self.debug:
+            print(f"Waited for block {block_idx}: {time.perf_counter() - start_time:.2f}s")
+
+
+# Gradient tensors
+_grad_t = Union[tuple[torch.Tensor, ...], torch.Tensor]
+
+
+class ModelOffloader(Offloader):
+    """
+    supports forward offloading
+    """
+
+    def __init__(
+        self,
+        blocks: Union[list[nn.Module], nn.ModuleList],
+        blocks_to_swap: int,
+        device: torch.device,
+        supports_backward: bool = True,
+        debug: bool = False,
+    ):
+        super().__init__(len(blocks), blocks_to_swap, device, debug)
+
+        self.supports_backward = supports_backward
+        self.forward_only = not supports_backward  # forward only offloading: can be changed to True for inference
+
+        if self.supports_backward:
+            # register backward hooks
+            self.remove_handles = []
+            for i, block in enumerate(blocks):
+                hook = self.create_backward_hook(blocks, i)
+                if hook is not None:
+                    handle = block.register_full_backward_hook(hook)
+                    self.remove_handles.append(handle)
+
+    def set_forward_only(self, forward_only: bool):
+        self.forward_only = forward_only
+
+    def __del__(self):
+        if self.supports_backward:
+            for handle in self.remove_handles:
+                handle.remove()
+
+    def create_backward_hook(
+        self, blocks: Union[list[nn.Module], nn.ModuleList], block_index: int
+    ) -> Optional[Callable[[nn.Module, _grad_t, _grad_t], Union[None, _grad_t]]]:
+        # -1 for 0-based index
+        num_blocks_propagated = self.num_blocks - block_index - 1
+        swapping = num_blocks_propagated > 0 and num_blocks_propagated <= self.blocks_to_swap
+        waiting = block_index > 0 and block_index <= self.blocks_to_swap
+
+        if not swapping and not waiting:
+            return None
+
+        # create  hook
+        block_idx_to_cpu = self.num_blocks - num_blocks_propagated
+        block_idx_to_cuda = self.blocks_to_swap - num_blocks_propagated
+        block_idx_to_wait = block_index - 1
+
+        def backward_hook(module: nn.Module, grad_input: _grad_t, grad_output: _grad_t):
+            if self.debug:
+                print(f"Backward hook for block {block_index}")
+
+            if swapping:
+                self._submit_move_blocks(blocks, block_idx_to_cpu, block_idx_to_cuda)
+            if waiting:
+                self._wait_blocks_move(block_idx_to_wait)
+            return None
+
+        return backward_hook
+
+    def prepare_block_devices_before_forward(self, blocks: Union[list[nn.Module], nn.ModuleList]):
+        if self.blocks_to_swap is None or self.blocks_to_swap == 0:
+            return
+
+        if self.debug:
+            print(f"Prepare block devices before forward")
+
+        for b in blocks[0 : self.num_blocks - self.blocks_to_swap]:
+            b.to(self.device)
+            weighs_to_device(b, self.device)  # make sure weights are on device
+
+        for b in blocks[self.num_blocks - self.blocks_to_swap :]:
+            b.to(self.device)  # move block to device first. this makes sure that buffers (non weights) are on the device
+            weighs_to_device(b, torch.device("cpu"))  # make sure weights are on cpu
+
+        _synchronize_device(self.device)
+        _clean_memory_on_device(self.device)
+
+    def wait_for_block(self, block_idx: int):
+        if self.blocks_to_swap is None or self.blocks_to_swap == 0:
+            return
+        self._wait_blocks_move(block_idx)
+
+    def submit_move_blocks(self, blocks: Union[list[nn.Module], nn.ModuleList], block_idx: int):
+        # check if blocks_to_swap is enabled
+        if self.blocks_to_swap is None or self.blocks_to_swap == 0:
+            return
+
+        # if backward is enabled, we do not swap blocks in forward pass more than blocks_to_swap, because it should be on GPU
+        if not self.forward_only and block_idx >= self.blocks_to_swap:
+            return
+
+        block_idx_to_cpu = block_idx
+        block_idx_to_cuda = self.num_blocks - self.blocks_to_swap + block_idx
+        # this works for forward-only offloading. move upstream blocks to cuda
+        block_idx_to_cuda = block_idx_to_cuda % self.num_blocks
+        self._submit_move_blocks(blocks, block_idx_to_cpu, block_idx_to_cuda)
+
+
+# endregion
+
+# region cpu offload utils
+
+
+def to_device(x: Any, device: torch.device) -> Any:
+    if isinstance(x, torch.Tensor):
+        return x.to(device)
+    elif isinstance(x, list):
+        return [to_device(elem, device) for elem in x]
+    elif isinstance(x, tuple):
+        return tuple(to_device(elem, device) for elem in x)
+    elif isinstance(x, dict):
+        return {k: to_device(v, device) for k, v in x.items()}
+    else:
+        return x
+
+
+def to_cpu(x: Any) -> Any:
+    """
+    Recursively moves torch.Tensor objects (and containers thereof) to CPU.
+
+    Args:
+        x: A torch.Tensor, or a (possibly nested) list, tuple, or dict containing tensors.
+
+    Returns:
+        The same structure as x, with all torch.Tensor objects moved to CPU.
+        Non-tensor objects are returned unchanged.
+    """
+    if isinstance(x, torch.Tensor):
+        return x.cpu()
+    elif isinstance(x, list):
+        return [to_cpu(elem) for elem in x]
+    elif isinstance(x, tuple):
+        return tuple(to_cpu(elem) for elem in x)
+    elif isinstance(x, dict):
+        return {k: to_cpu(v) for k, v in x.items()}
+    else:
+        return x
+
+
+def create_cpu_offloading_wrapper(func: Callable, device: torch.device) -> Callable:
+    """
+    Create a wrapper function that offloads inputs to CPU before calling the original function
+    and moves outputs back to the specified device.
+
+    Args:
+        func: The original function to wrap.
+        device: The device to move outputs back to.
+
+    Returns:
+        A wrapped function that offloads inputs to CPU and moves outputs back to the specified device.
+    """
+
+    def wrapper(orig_func: Callable) -> Callable:
+        def custom_forward(*inputs):
+            nonlocal device, orig_func
+            cuda_inputs = to_device(inputs, device)
+            outputs = orig_func(*cuda_inputs)
+            return to_cpu(outputs)
+
+        return custom_forward
+
+    return wrapper(func)
+
+
+# endregion

library/custom_train_functions.py

@@ -1,8 +1,16 @@
+from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
 import torch
 import argparse
 import random
 import re
+from torch.types import Number
 from typing import List, Optional, Union
+from .utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
 
 
 def prepare_scheduler_for_custom_training(noise_scheduler, device):
@@ -21,7 +29,7 @@ def prepare_scheduler_for_custom_training(noise_scheduler, device):
 
 def fix_noise_scheduler_betas_for_zero_terminal_snr(noise_scheduler):
     # fix beta: zero terminal SNR
-    print(f"fix noise scheduler betas: https://arxiv.org/abs/2305.08891")
+    logger.info(f"fix noise scheduler betas: https://arxiv.org/abs/2305.08891")
 
     def enforce_zero_terminal_snr(betas):
         # Convert betas to alphas_bar_sqrt
@@ -49,53 +57,58 @@ def fix_noise_scheduler_betas_for_zero_terminal_snr(noise_scheduler):
     alphas = 1.0 - betas
     alphas_cumprod = torch.cumprod(alphas, dim=0)
 
-    # print("original:", noise_scheduler.betas)
-    # print("fixed:", betas)
+    # logger.info(f"original: {noise_scheduler.betas}")
+    # logger.info(f"fixed: {betas}")
 
     noise_scheduler.betas = betas
     noise_scheduler.alphas = alphas
     noise_scheduler.alphas_cumprod = alphas_cumprod
 
 
-def apply_snr_weight(loss, timesteps, noise_scheduler, gamma, v_prediction=False):
+def apply_snr_weight(loss: torch.Tensor, timesteps: torch.IntTensor, noise_scheduler: DDPMScheduler, gamma: Number, v_prediction=False):
     snr = torch.stack([noise_scheduler.all_snr[t] for t in timesteps])
     min_snr_gamma = torch.minimum(snr, torch.full_like(snr, gamma))
     if v_prediction:
-        snr_weight = torch.div(min_snr_gamma, snr+1).float().to(loss.device)
+        snr_weight = torch.div(min_snr_gamma, snr + 1).float().to(loss.device)
     else:
         snr_weight = torch.div(min_snr_gamma, snr).float().to(loss.device)
     loss = loss * snr_weight
     return loss
 
 
-def scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler):
+def scale_v_prediction_loss_like_noise_prediction(loss: torch.Tensor, timesteps: torch.IntTensor, noise_scheduler: DDPMScheduler):
     scale = get_snr_scale(timesteps, noise_scheduler)
     loss = loss * scale
     return loss
 
 
-def get_snr_scale(timesteps, noise_scheduler):
+def get_snr_scale(timesteps: torch.IntTensor, noise_scheduler: DDPMScheduler):
     snr_t = torch.stack([noise_scheduler.all_snr[t] for t in timesteps])  # batch_size
     snr_t = torch.minimum(snr_t, torch.ones_like(snr_t) * 1000)  # if timestep is 0, snr_t is inf, so limit it to 1000
     scale = snr_t / (snr_t + 1)
     # # show debug info
-    # print(f"timesteps: {timesteps}, snr_t: {snr_t}, scale: {scale}")
+    # logger.info(f"timesteps: {timesteps}, snr_t: {snr_t}, scale: {scale}")
     return scale
 
 
-def add_v_prediction_like_loss(loss, timesteps, noise_scheduler, v_pred_like_loss):
+def add_v_prediction_like_loss(loss: torch.Tensor, timesteps: torch.IntTensor, noise_scheduler: DDPMScheduler, v_pred_like_loss: torch.Tensor):
     scale = get_snr_scale(timesteps, noise_scheduler)
-    # print(f"add v-prediction like loss: {v_pred_like_loss}, scale: {scale}, loss: {loss}, time: {timesteps}")
+    # logger.info(f"add v-prediction like loss: {v_pred_like_loss}, scale: {scale}, loss: {loss}, time: {timesteps}")
     loss = loss + loss / scale * v_pred_like_loss
     return loss
 
-def apply_debiased_estimation(loss, timesteps, noise_scheduler):
+
+def apply_debiased_estimation(loss: torch.Tensor, timesteps: torch.IntTensor, noise_scheduler: DDPMScheduler, v_prediction=False):
     snr_t = torch.stack([noise_scheduler.all_snr[t] for t in timesteps])  # batch_size
     snr_t = torch.minimum(snr_t, torch.ones_like(snr_t) * 1000)  # if timestep is 0, snr_t is inf, so limit it to 1000
-    weight = 1/torch.sqrt(snr_t)
+    if v_prediction:
+        weight = 1 / (snr_t + 1)
+    else:
+        weight = 1 / torch.sqrt(snr_t)
     loss = weight * loss
     return loss
 
+
 # TODO train_utilと分散しているのでどちらかに寄せる
 
 
@@ -268,7 +281,7 @@ def get_prompts_with_weights(tokenizer, prompt: List[str], max_length: int):
         tokens.append(text_token)
         weights.append(text_weight)
     if truncated:
-        print("Prompt was truncated. Try to shorten the prompt or increase max_embeddings_multiples")
+        logger.warning("Prompt was truncated. Try to shorten the prompt or increase max_embeddings_multiples")
     return tokens, weights
 
 
@@ -442,7 +455,7 @@ def get_weighted_text_embeddings(
 
 
 # https://wandb.ai/johnowhitaker/multires_noise/reports/Multi-Resolution-Noise-for-Diffusion-Model-Training--VmlldzozNjYyOTU2
-def pyramid_noise_like(noise, device, iterations=6, discount=0.4):
+def pyramid_noise_like(noise, device, iterations=6, discount=0.4) -> torch.FloatTensor:
     b, c, w, h = noise.shape  # EDIT: w and h get over-written, rename for a different variant!
     u = torch.nn.Upsample(size=(w, h), mode="bilinear").to(device)
     for i in range(iterations):
@@ -455,7 +468,7 @@ def pyramid_noise_like(noise, device, iterations=6, discount=0.4):
 
 
 # https://www.crosslabs.org//blog/diffusion-with-offset-noise
-def apply_noise_offset(latents, noise, noise_offset, adaptive_noise_scale):
+def apply_noise_offset(latents, noise, noise_offset, adaptive_noise_scale) -> torch.FloatTensor:
     if noise_offset is None:
         return noise
     if adaptive_noise_scale is not None:
@@ -471,6 +484,25 @@ def apply_noise_offset(latents, noise, noise_offset, adaptive_noise_scale):
     return noise
 
 
+def apply_masked_loss(loss, batch) -> torch.FloatTensor:
+    if "conditioning_images" in batch:
+        # conditioning image is -1 to 1. we need to convert it to 0 to 1
+        mask_image = batch["conditioning_images"].to(dtype=loss.dtype)[:, 0].unsqueeze(1)  # use R channel
+        mask_image = mask_image / 2 + 0.5
+        # print(f"conditioning_image: {mask_image.shape}")
+    elif "alpha_masks" in batch and batch["alpha_masks"] is not None:
+        # alpha mask is 0 to 1
+        mask_image = batch["alpha_masks"].to(dtype=loss.dtype).unsqueeze(1) # add channel dimension
+        # print(f"mask_image: {mask_image.shape}, {mask_image.mean()}")
+    else:
+        return loss
+
+    # resize to the same size as the loss
+    mask_image = torch.nn.functional.interpolate(mask_image, size=loss.shape[2:], mode="area")
+    loss = loss * mask_image
+    return loss
+
+
 """
 ##########################################
 # Perlin Noise

library/deepspeed_utils.py

@@ -0,0 +1,180 @@
+import os
+import argparse
+import torch
+from accelerate import DeepSpeedPlugin, Accelerator
+
+from .utils import setup_logging
+
+from .device_utils import get_preferred_device
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+def add_deepspeed_arguments(parser: argparse.ArgumentParser):
+    # DeepSpeed Arguments. https://huggingface.co/docs/accelerate/usage_guides/deepspeed
+    parser.add_argument("--deepspeed", action="store_true", help="enable deepspeed training")
+    parser.add_argument("--zero_stage", type=int, default=2, choices=[0, 1, 2, 3], help="Possible options are 0,1,2,3.")
+    parser.add_argument(
+        "--offload_optimizer_device",
+        type=str,
+        default=None,
+        choices=[None, "cpu", "nvme"],
+        help="Possible options are none|cpu|nvme. Only applicable with ZeRO Stages 2 and 3.",
+    )
+    parser.add_argument(
+        "--offload_optimizer_nvme_path",
+        type=str,
+        default=None,
+        help="Possible options are /nvme|/local_nvme. Only applicable with ZeRO Stage 3.",
+    )
+    parser.add_argument(
+        "--offload_param_device",
+        type=str,
+        default=None,
+        choices=[None, "cpu", "nvme"],
+        help="Possible options are none|cpu|nvme. Only applicable with ZeRO Stage 3.",
+    )
+    parser.add_argument(
+        "--offload_param_nvme_path",
+        type=str,
+        default=None,
+        help="Possible options are /nvme|/local_nvme. Only applicable with ZeRO Stage 3.",
+    )
+    parser.add_argument(
+        "--zero3_init_flag",
+        action="store_true",
+        help="Flag to indicate whether to enable `deepspeed.zero.Init` for constructing massive models."
+        "Only applicable with ZeRO Stage-3.",
+    )
+    parser.add_argument(
+        "--zero3_save_16bit_model",
+        action="store_true",
+        help="Flag to indicate whether to save 16-bit model. Only applicable with ZeRO Stage-3.",
+    )
+    parser.add_argument(
+        "--fp16_master_weights_and_gradients",
+        action="store_true",
+        help="fp16_master_and_gradients requires optimizer to support keeping fp16 master and gradients while keeping the optimizer states in fp32.",
+    )
+
+
+def prepare_deepspeed_args(args: argparse.Namespace):
+    if not args.deepspeed:
+        return
+
+    # To avoid RuntimeError: DataLoader worker exited unexpectedly with exit code 1.
+    args.max_data_loader_n_workers = 1
+
+
+def prepare_deepspeed_plugin(args: argparse.Namespace):
+    if not args.deepspeed:
+        return None
+
+    try:
+        import deepspeed
+    except ImportError as e:
+        logger.error(
+            "deepspeed is not installed. please install deepspeed in your environment with following command. DS_BUILD_OPS=0 pip install deepspeed"
+        )
+        exit(1)
+
+    deepspeed_plugin = DeepSpeedPlugin(
+        zero_stage=args.zero_stage,
+        gradient_accumulation_steps=args.gradient_accumulation_steps,
+        gradient_clipping=args.max_grad_norm,
+        offload_optimizer_device=args.offload_optimizer_device,
+        offload_optimizer_nvme_path=args.offload_optimizer_nvme_path,
+        offload_param_device=args.offload_param_device,
+        offload_param_nvme_path=args.offload_param_nvme_path,
+        zero3_init_flag=args.zero3_init_flag,
+        zero3_save_16bit_model=args.zero3_save_16bit_model,
+    )
+    deepspeed_plugin.deepspeed_config["train_micro_batch_size_per_gpu"] = args.train_batch_size
+    deepspeed_plugin.deepspeed_config["train_batch_size"] = (
+        args.train_batch_size * args.gradient_accumulation_steps * int(os.environ["WORLD_SIZE"])
+    )
+    
+    deepspeed_plugin.set_mixed_precision(args.mixed_precision)
+    if args.mixed_precision.lower() == "fp16":
+        deepspeed_plugin.deepspeed_config["fp16"]["initial_scale_power"] = 0  # preventing overflow.
+    if args.full_fp16 or args.fp16_master_weights_and_gradients:
+        if args.offload_optimizer_device == "cpu" and args.zero_stage == 2:
+            deepspeed_plugin.deepspeed_config["fp16"]["fp16_master_weights_and_grads"] = True
+            logger.info("[DeepSpeed] full fp16 enable.")
+        else:
+            logger.info(
+                "[DeepSpeed]full fp16, fp16_master_weights_and_grads currently only supported using ZeRO-Offload with DeepSpeedCPUAdam on ZeRO-2 stage."
+            )
+
+    if args.offload_optimizer_device is not None:
+        logger.info("[DeepSpeed] start to manually build cpu_adam.")
+        deepspeed.ops.op_builder.CPUAdamBuilder().load()
+        logger.info("[DeepSpeed] building cpu_adam done.")
+
+    return deepspeed_plugin
+
+
+# Accelerate library does not support multiple models for deepspeed. So, we need to wrap multiple models into a single model.
+def prepare_deepspeed_model(args: argparse.Namespace, **models):
+    # remove None from models
+    models = {k: v for k, v in models.items() if v is not None}
+
+    class DeepSpeedWrapper(torch.nn.Module):
+        def __init__(self, **kw_models) -> None:
+            super().__init__()
+            
+            self.models = torch.nn.ModuleDict()
+            
+            wrap_model_forward_with_torch_autocast = args.mixed_precision is not "no"
+
+            for key, model in kw_models.items():
+                if isinstance(model, list):
+                    model = torch.nn.ModuleList(model)
+                                            
+                if wrap_model_forward_with_torch_autocast:
+                    model = self.__wrap_model_with_torch_autocast(model)  
+                
+                assert isinstance(
+                    model, torch.nn.Module
+                ), f"model must be an instance of torch.nn.Module, but got {key} is {type(model)}"
+
+                self.models.update(torch.nn.ModuleDict({key: model}))
+
+        def __wrap_model_with_torch_autocast(self, model):
+            if isinstance(model, torch.nn.ModuleList):
+                model = torch.nn.ModuleList([self.__wrap_model_forward_with_torch_autocast(m) for m in model])
+            else:
+                model = self.__wrap_model_forward_with_torch_autocast(model)
+            return model
+
+        def __wrap_model_forward_with_torch_autocast(self, model):
+            
+            assert hasattr(model, "forward"), f"model must have a forward method."
+
+            forward_fn = model.forward
+
+            def forward(*args, **kwargs):
+                try:
+                    device_type = model.device.type
+                except AttributeError:
+                    logger.warning(
+                            "[DeepSpeed] model.device is not available. Using get_preferred_device() "
+                            "to determine the device_type for torch.autocast()."
+                    )                    
+                    device_type = get_preferred_device().type
+
+                with torch.autocast(device_type = device_type):
+                    return forward_fn(*args, **kwargs)
+
+            model.forward = forward
+            return model
+        
+        def get_models(self):
+            return self.models
+        
+
+    ds_model = DeepSpeedWrapper(**models)
+    return ds_model

library/device_utils.py

@@ -0,0 +1,108 @@
+import functools
+import gc
+from typing import Optional, Union
+
+import torch
+
+
+try:
+    # intel gpu support for pytorch older than 2.5
+    # ipex is not needed after pytorch 2.5
+    import intel_extension_for_pytorch as ipex  # noqa
+except Exception:
+    pass
+
+
+try:
+    HAS_CUDA = torch.cuda.is_available()
+except Exception:
+    HAS_CUDA = False
+
+try:
+    HAS_MPS = torch.backends.mps.is_available()
+except Exception:
+    HAS_MPS = False
+
+try:
+    HAS_XPU = torch.xpu.is_available()
+except Exception:
+    HAS_XPU = False
+
+
+def clean_memory():
+    gc.collect()
+    if HAS_CUDA:
+        torch.cuda.empty_cache()
+    if HAS_XPU:
+        torch.xpu.empty_cache()
+    if HAS_MPS:
+        torch.mps.empty_cache()
+
+
+def clean_memory_on_device(device: Optional[Union[str, torch.device]]):
+    r"""
+    Clean memory on the specified device, will be called from training scripts.
+    """
+    gc.collect()
+    if device is None:
+        return
+    if isinstance(device, str):
+        device = torch.device(device)
+    # device may "cuda" or "cuda:0", so we need to check the type of device
+    if device.type == "cuda":
+        torch.cuda.empty_cache()
+    if device.type == "xpu":
+        torch.xpu.empty_cache()
+    if device.type == "mps":
+        torch.mps.empty_cache()
+
+
+def synchronize_device(device: Optional[Union[str, torch.device]]):
+    if device is None:
+        return
+    if isinstance(device, str):
+        device = torch.device(device)
+    if device.type == "cuda":
+        torch.cuda.synchronize()
+    elif device.type == "xpu":
+        torch.xpu.synchronize()
+    elif device.type == "mps":
+        torch.mps.synchronize()
+
+
+@functools.lru_cache(maxsize=None)
+def get_preferred_device() -> torch.device:
+    r"""
+    Do not call this function from training scripts. Use accelerator.device instead.
+    """
+    if HAS_CUDA:
+        device = torch.device("cuda")
+    elif HAS_XPU:
+        device = torch.device("xpu")
+    elif HAS_MPS:
+        device = torch.device("mps")
+    else:
+        device = torch.device("cpu")
+    print(f"get_preferred_device() -> {device}")
+    return device
+
+
+def init_ipex():
+    """
+    Apply IPEX to CUDA hijacks using `library.ipex.ipex_init`.
+
+    This function should run right after importing torch and before doing anything else.
+
+    If xpu is not available, this function does nothing.
+    """
+    try:
+        if HAS_XPU:
+            from library.ipex import ipex_init
+
+            is_initialized, error_message = ipex_init()
+            if not is_initialized:
+                print("failed to initialize ipex:", error_message)
+        else:
+            return
+    except Exception as e:
+        print("failed to initialize ipex:", e)

library/flux_train_utils.py

@@ -0,0 +1,690 @@
+import argparse
+import math
+import os
+import numpy as np
+import toml
+import json
+import time
+from typing import Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from accelerate import Accelerator, PartialState
+from transformers import CLIPTextModel
+from tqdm import tqdm
+from PIL import Image
+from safetensors.torch import save_file
+
+from library import flux_models, flux_utils, strategy_base, train_util
+from library.device_utils import init_ipex, clean_memory_on_device
+from library.safetensors_utils import mem_eff_save_file
+
+init_ipex()
+
+from .utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+# region sample images
+
+
+def sample_images(
+    accelerator: Accelerator,
+    args: argparse.Namespace,
+    epoch,
+    steps,
+    flux,
+    ae,
+    text_encoders,
+    sample_prompts_te_outputs,
+    prompt_replacement=None,
+    controlnet=None,
+):
+    if steps == 0:
+        if not args.sample_at_first:
+            return
+    else:
+        if args.sample_every_n_steps is None and args.sample_every_n_epochs is None:
+            return
+        if args.sample_every_n_epochs is not None:
+            # sample_every_n_steps は無視する
+            if epoch is None or epoch % args.sample_every_n_epochs != 0:
+                return
+        else:
+            if steps % args.sample_every_n_steps != 0 or epoch is not None:  # steps is not divisible or end of epoch
+                return
+
+    logger.info("")
+    logger.info(f"generating sample images at step / サンプル画像生成 ステップ: {steps}")
+    if not os.path.isfile(args.sample_prompts) and sample_prompts_te_outputs is None:
+        logger.error(f"No prompt file / プロンプトファイルがありません: {args.sample_prompts}")
+        return
+
+    distributed_state = PartialState()  # for multi gpu distributed inference. this is a singleton, so it's safe to use it here
+
+    # unwrap unet and text_encoder(s)
+    flux = accelerator.unwrap_model(flux)
+    if text_encoders is not None:
+        text_encoders = [(accelerator.unwrap_model(te) if te is not None else None) for te in text_encoders]
+    if controlnet is not None:
+        controlnet = accelerator.unwrap_model(controlnet)
+    # print([(te.parameters().__next__().device if te is not None else None) for te in text_encoders])
+
+    prompts = train_util.load_prompts(args.sample_prompts)
+
+    save_dir = args.output_dir + "/sample"
+    os.makedirs(save_dir, exist_ok=True)
+
+    # save random state to restore later
+    rng_state = torch.get_rng_state()
+    cuda_rng_state = None
+    try:
+        cuda_rng_state = torch.cuda.get_rng_state() if torch.cuda.is_available() else None
+    except Exception:
+        pass
+
+    if distributed_state.num_processes <= 1:
+        # If only one device is available, just use the original prompt list. We don't need to care about the distribution of prompts.
+        with torch.no_grad(), accelerator.autocast():
+            for prompt_dict in prompts:
+                sample_image_inference(
+                    accelerator,
+                    args,
+                    flux,
+                    text_encoders,
+                    ae,
+                    save_dir,
+                    prompt_dict,
+                    epoch,
+                    steps,
+                    sample_prompts_te_outputs,
+                    prompt_replacement,
+                    controlnet,
+                )
+    else:
+        # Creating list with N elements, where each element is a list of prompt_dicts, and N is the number of processes available (number of devices available)
+        # prompt_dicts are assigned to lists based on order of processes, to attempt to time the image creation time to match enum order. Probably only works when steps and sampler are identical.
+        per_process_prompts = []  # list of lists
+        for i in range(distributed_state.num_processes):
+            per_process_prompts.append(prompts[i :: distributed_state.num_processes])
+
+        with torch.no_grad():
+            with distributed_state.split_between_processes(per_process_prompts) as prompt_dict_lists:
+                for prompt_dict in prompt_dict_lists[0]:
+                    sample_image_inference(
+                        accelerator,
+                        args,
+                        flux,
+                        text_encoders,
+                        ae,
+                        save_dir,
+                        prompt_dict,
+                        epoch,
+                        steps,
+                        sample_prompts_te_outputs,
+                        prompt_replacement,
+                        controlnet,
+                    )
+
+    torch.set_rng_state(rng_state)
+    if cuda_rng_state is not None:
+        torch.cuda.set_rng_state(cuda_rng_state)
+
+    clean_memory_on_device(accelerator.device)
+
+
+def sample_image_inference(
+    accelerator: Accelerator,
+    args: argparse.Namespace,
+    flux: flux_models.Flux,
+    text_encoders: Optional[List[CLIPTextModel]],
+    ae: flux_models.AutoEncoder,
+    save_dir,
+    prompt_dict,
+    epoch,
+    steps,
+    sample_prompts_te_outputs,
+    prompt_replacement,
+    controlnet,
+):
+    assert isinstance(prompt_dict, dict)
+    negative_prompt = prompt_dict.get("negative_prompt")
+    sample_steps = prompt_dict.get("sample_steps", 20)
+    width = prompt_dict.get("width", 512)
+    height = prompt_dict.get("height", 512)
+    emb_guidance_scale = prompt_dict.get("guidance_scale", 3.5)
+    cfg_scale = prompt_dict.get("scale", 1.0)
+    seed = prompt_dict.get("seed")
+    controlnet_image = prompt_dict.get("controlnet_image")
+    prompt: str = prompt_dict.get("prompt", "")
+    # sampler_name: str = prompt_dict.get("sample_sampler", args.sample_sampler)
+
+    if prompt_replacement is not None:
+        prompt = prompt.replace(prompt_replacement[0], prompt_replacement[1])
+        if negative_prompt is not None:
+            negative_prompt = negative_prompt.replace(prompt_replacement[0], prompt_replacement[1])
+
+    if seed is not None:
+        torch.manual_seed(seed)
+        torch.cuda.manual_seed(seed)
+    else:
+        # True random sample image generation
+        torch.seed()
+        torch.cuda.seed()
+
+    if negative_prompt is None:
+        negative_prompt = ""
+    height = max(64, height - height % 16)  # round to divisible by 16
+    width = max(64, width - width % 16)  # round to divisible by 16
+    logger.info(f"prompt: {prompt}")
+    if cfg_scale != 1.0:
+        logger.info(f"negative_prompt: {negative_prompt}")
+    elif negative_prompt != "":
+        logger.info(f"negative prompt is ignored because scale is 1.0")
+    logger.info(f"height: {height}")
+    logger.info(f"width: {width}")
+    logger.info(f"sample_steps: {sample_steps}")
+    logger.info(f"embedded guidance scale: {emb_guidance_scale}")
+    if cfg_scale != 1.0:
+        logger.info(f"CFG scale: {cfg_scale}")
+    # logger.info(f"sample_sampler: {sampler_name}")
+    if seed is not None:
+        logger.info(f"seed: {seed}")
+
+    # encode prompts
+    tokenize_strategy = strategy_base.TokenizeStrategy.get_strategy()
+    encoding_strategy = strategy_base.TextEncodingStrategy.get_strategy()
+
+    def encode_prompt(prpt):
+        text_encoder_conds = []
+        if sample_prompts_te_outputs and prpt in sample_prompts_te_outputs:
+            text_encoder_conds = sample_prompts_te_outputs[prpt]
+            print(f"Using cached text encoder outputs for prompt: {prpt}")
+        if text_encoders is not None:
+            print(f"Encoding prompt: {prpt}")
+            tokens_and_masks = tokenize_strategy.tokenize(prpt)
+            # strategy has apply_t5_attn_mask option
+            encoded_text_encoder_conds = encoding_strategy.encode_tokens(tokenize_strategy, text_encoders, tokens_and_masks)
+
+            # if text_encoder_conds is not cached, use encoded_text_encoder_conds
+            if len(text_encoder_conds) == 0:
+                text_encoder_conds = encoded_text_encoder_conds
+            else:
+                # if encoded_text_encoder_conds is not None, update cached text_encoder_conds
+                for i in range(len(encoded_text_encoder_conds)):
+                    if encoded_text_encoder_conds[i] is not None:
+                        text_encoder_conds[i] = encoded_text_encoder_conds[i]
+        return text_encoder_conds
+
+    l_pooled, t5_out, txt_ids, t5_attn_mask = encode_prompt(prompt)
+    # encode negative prompts
+    if cfg_scale != 1.0:
+        neg_l_pooled, neg_t5_out, _, neg_t5_attn_mask = encode_prompt(negative_prompt)
+        neg_t5_attn_mask = (
+            neg_t5_attn_mask.to(accelerator.device) if args.apply_t5_attn_mask and neg_t5_attn_mask is not None else None
+        )
+        neg_cond = (cfg_scale, neg_l_pooled, neg_t5_out, neg_t5_attn_mask)
+    else:
+        neg_cond = None
+
+    # sample image
+    weight_dtype = ae.dtype  # TOFO give dtype as argument
+    packed_latent_height = height // 16
+    packed_latent_width = width // 16
+    noise = torch.randn(
+        1,
+        packed_latent_height * packed_latent_width,
+        16 * 2 * 2,
+        device=accelerator.device,
+        dtype=weight_dtype,
+        generator=torch.Generator(device=accelerator.device).manual_seed(seed) if seed is not None else None,
+    )
+    timesteps = get_schedule(sample_steps, noise.shape[1], shift=True)  # Chroma can use shift=True
+    img_ids = flux_utils.prepare_img_ids(1, packed_latent_height, packed_latent_width).to(accelerator.device, weight_dtype)
+    t5_attn_mask = t5_attn_mask.to(accelerator.device) if args.apply_t5_attn_mask else None
+
+    if controlnet_image is not None:
+        controlnet_image = Image.open(controlnet_image).convert("RGB")
+        controlnet_image = controlnet_image.resize((width, height), Image.LANCZOS)
+        controlnet_image = torch.from_numpy((np.array(controlnet_image) / 127.5) - 1)
+        controlnet_image = controlnet_image.permute(2, 0, 1).unsqueeze(0).to(weight_dtype).to(accelerator.device)
+
+    with accelerator.autocast(), torch.no_grad():
+        x = denoise(
+            flux,
+            noise,
+            img_ids,
+            t5_out,
+            txt_ids,
+            l_pooled,
+            timesteps=timesteps,
+            guidance=emb_guidance_scale,
+            t5_attn_mask=t5_attn_mask,
+            controlnet=controlnet,
+            controlnet_img=controlnet_image,
+            neg_cond=neg_cond,
+        )
+
+    x = flux_utils.unpack_latents(x, packed_latent_height, packed_latent_width)
+
+    # latent to image
+    clean_memory_on_device(accelerator.device)
+    org_vae_device = ae.device  # will be on cpu
+    ae.to(accelerator.device)  # distributed_state.device is same as accelerator.device
+    with accelerator.autocast(), torch.no_grad():
+        x = ae.decode(x)
+    ae.to(org_vae_device)
+    clean_memory_on_device(accelerator.device)
+
+    x = x.clamp(-1, 1)
+    x = x.permute(0, 2, 3, 1)
+    image = Image.fromarray((127.5 * (x + 1.0)).float().cpu().numpy().astype(np.uint8)[0])
+
+    # adding accelerator.wait_for_everyone() here should sync up and ensure that sample images are saved in the same order as the original prompt list
+    # but adding 'enum' to the filename should be enough
+
+    ts_str = time.strftime("%Y%m%d%H%M%S", time.localtime())
+    num_suffix = f"e{epoch:06d}" if epoch is not None else f"{steps:06d}"
+    seed_suffix = "" if seed is None else f"_{seed}"
+    i: int = prompt_dict["enum"]
+    img_filename = f"{'' if args.output_name is None else args.output_name + '_'}{num_suffix}_{i:02d}_{ts_str}{seed_suffix}.png"
+    image.save(os.path.join(save_dir, img_filename))
+
+    # send images to wandb if enabled
+    if "wandb" in [tracker.name for tracker in accelerator.trackers]:
+        wandb_tracker = accelerator.get_tracker("wandb")
+
+        import wandb
+
+        # not to commit images to avoid inconsistency between training and logging steps
+        wandb_tracker.log({f"sample_{i}": wandb.Image(image, caption=prompt)}, commit=False)  # positive prompt as a caption
+
+
+def time_shift(mu: float, sigma: float, t: torch.Tensor):
+    return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
+
+
+def get_lin_function(x1: float = 256, y1: float = 0.5, x2: float = 4096, y2: float = 1.15) -> Callable[[float], float]:
+    m = (y2 - y1) / (x2 - x1)
+    b = y1 - m * x1
+    return lambda x: m * x + b
+
+
+def get_schedule(
+    num_steps: int,
+    image_seq_len: int,
+    base_shift: float = 0.5,
+    max_shift: float = 1.15,
+    shift: bool = True,
+) -> list[float]:
+    # extra step for zero
+    timesteps = torch.linspace(1, 0, num_steps + 1)
+
+    # shifting the schedule to favor high timesteps for higher signal images
+    if shift:
+        # eastimate mu based on linear estimation between two points
+        mu = get_lin_function(y1=base_shift, y2=max_shift)(image_seq_len)
+        timesteps = time_shift(mu, 1.0, timesteps)
+
+    return timesteps.tolist()
+
+
+def denoise(
+    model: flux_models.Flux,
+    img: torch.Tensor,
+    img_ids: torch.Tensor,
+    txt: torch.Tensor,  # t5_out
+    txt_ids: torch.Tensor,
+    vec: torch.Tensor,  # l_pooled
+    timesteps: list[float],
+    guidance: float = 4.0,
+    t5_attn_mask: Optional[torch.Tensor] = None,
+    controlnet: Optional[flux_models.ControlNetFlux] = None,
+    controlnet_img: Optional[torch.Tensor] = None,
+    neg_cond: Optional[Tuple[float, torch.Tensor, torch.Tensor, torch.Tensor]] = None,
+):
+    # this is ignored for schnell
+    guidance_vec = torch.full((img.shape[0],), guidance, device=img.device, dtype=img.dtype)
+    do_cfg = neg_cond is not None
+
+    for t_curr, t_prev in zip(tqdm(timesteps[:-1]), timesteps[1:]):
+        t_vec = torch.full((img.shape[0],), t_curr, dtype=img.dtype, device=img.device)
+        model.prepare_block_swap_before_forward()
+
+        if controlnet is not None:
+            block_samples, block_single_samples = controlnet(
+                img=img,
+                img_ids=img_ids,
+                controlnet_cond=controlnet_img,
+                txt=txt,
+                txt_ids=txt_ids,
+                y=vec,
+                timesteps=t_vec,
+                guidance=guidance_vec,
+                txt_attention_mask=t5_attn_mask,
+            )
+        else:
+            block_samples = None
+            block_single_samples = None
+
+        if not do_cfg:
+            pred = model(
+                img=img,
+                img_ids=img_ids,
+                txt=txt,
+                txt_ids=txt_ids,
+                y=vec,
+                block_controlnet_hidden_states=block_samples,
+                block_controlnet_single_hidden_states=block_single_samples,
+                timesteps=t_vec,
+                guidance=guidance_vec,
+                txt_attention_mask=t5_attn_mask,
+            )
+
+            img = img + (t_prev - t_curr) * pred
+        else:
+            cfg_scale, neg_l_pooled, neg_t5_out, neg_t5_attn_mask = neg_cond
+            nc_c_t5_attn_mask = None if t5_attn_mask is None else torch.cat([neg_t5_attn_mask, t5_attn_mask], dim=0)
+
+            # TODO is it ok to use the same block samples for both cond and uncond?
+            block_samples = None if block_samples is None else torch.cat([block_samples, block_samples], dim=0)
+            block_single_samples = (
+                None if block_single_samples is None else torch.cat([block_single_samples, block_single_samples], dim=0)
+            )
+
+            nc_c_pred = model(
+                img=torch.cat([img, img], dim=0),
+                img_ids=torch.cat([img_ids, img_ids], dim=0),
+                txt=torch.cat([neg_t5_out, txt], dim=0),
+                txt_ids=torch.cat([txt_ids, txt_ids], dim=0),
+                y=torch.cat([neg_l_pooled, vec], dim=0),
+                block_controlnet_hidden_states=block_samples,
+                block_controlnet_single_hidden_states=block_single_samples,
+                timesteps=t_vec.repeat(2),
+                guidance=guidance_vec.repeat(2),
+                txt_attention_mask=nc_c_t5_attn_mask,
+            )
+            neg_pred, pred = torch.chunk(nc_c_pred, 2, dim=0)
+            pred = neg_pred + (pred - neg_pred) * cfg_scale
+
+            img = img + (t_prev - t_curr) * pred
+
+    model.prepare_block_swap_before_forward()
+    return img
+
+
+# endregion
+
+
+# region train
+def get_sigmas(noise_scheduler, timesteps, device, n_dim=4, dtype=torch.float32):
+    sigmas = noise_scheduler.sigmas.to(device=device, dtype=dtype)
+    schedule_timesteps = noise_scheduler.timesteps.to(device)
+    timesteps = timesteps.to(device)
+    step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
+
+    sigma = sigmas[step_indices].flatten()
+    return sigma
+
+
+def compute_density_for_timestep_sampling(
+    weighting_scheme: str, batch_size: int, logit_mean: float = None, logit_std: float = None, mode_scale: float = None
+):
+    """Compute the density for sampling the timesteps when doing SD3 training.
+
+    Courtesy: This was contributed by Rafie Walker in https://github.com/huggingface/diffusers/pull/8528.
+
+    SD3 paper reference: https://arxiv.org/abs/2403.03206v1.
+    """
+    if weighting_scheme == "logit_normal":
+        # See 3.1 in the SD3 paper ($rf/lognorm(0.00,1.00)$).
+        u = torch.normal(mean=logit_mean, std=logit_std, size=(batch_size,), device="cpu")
+        u = torch.nn.functional.sigmoid(u)
+    elif weighting_scheme == "mode":
+        u = torch.rand(size=(batch_size,), device="cpu")
+        u = 1 - u - mode_scale * (torch.cos(math.pi * u / 2) ** 2 - 1 + u)
+    else:
+        u = torch.rand(size=(batch_size,), device="cpu")
+    return u
+
+
+def compute_loss_weighting_for_sd3(weighting_scheme: str, sigmas=None):
+    """Computes loss weighting scheme for SD3 training.
+
+    Courtesy: This was contributed by Rafie Walker in https://github.com/huggingface/diffusers/pull/8528.
+
+    SD3 paper reference: https://arxiv.org/abs/2403.03206v1.
+    """
+    if weighting_scheme == "sigma_sqrt":
+        weighting = (sigmas**-2.0).float()
+    elif weighting_scheme == "cosmap":
+        bot = 1 - 2 * sigmas + 2 * sigmas**2
+        weighting = 2 / (math.pi * bot)
+    else:
+        weighting = torch.ones_like(sigmas)
+    return weighting
+
+
+def get_noisy_model_input_and_timesteps(
+    args, noise_scheduler, latents: torch.Tensor, noise: torch.Tensor, device, dtype
+) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    bsz, _, h, w = latents.shape
+    assert bsz > 0, "Batch size not large enough"
+    num_timesteps = noise_scheduler.config.num_train_timesteps
+    if args.timestep_sampling == "uniform" or args.timestep_sampling == "sigmoid":
+        # Simple random sigma-based noise sampling
+        if args.timestep_sampling == "sigmoid":
+            # https://github.com/XLabs-AI/x-flux/tree/main
+            sigmas = torch.sigmoid(args.sigmoid_scale * torch.randn((bsz,), device=device))
+        else:
+            sigmas = torch.rand((bsz,), device=device)
+
+        timesteps = sigmas * num_timesteps
+    elif args.timestep_sampling == "shift":
+        shift = args.discrete_flow_shift
+        sigmas = torch.randn(bsz, device=device)
+        sigmas = sigmas * args.sigmoid_scale  # larger scale for more uniform sampling
+        sigmas = sigmas.sigmoid()
+        sigmas = (sigmas * shift) / (1 + (shift - 1) * sigmas)
+        timesteps = sigmas * num_timesteps
+    elif args.timestep_sampling == "flux_shift":
+        sigmas = torch.randn(bsz, device=device)
+        sigmas = sigmas * args.sigmoid_scale  # larger scale for more uniform sampling
+        sigmas = sigmas.sigmoid()
+        mu = get_lin_function(y1=0.5, y2=1.15)((h // 2) * (w // 2))  # we are pre-packed so must adjust for packed size
+        sigmas = time_shift(mu, 1.0, sigmas)
+        timesteps = sigmas * num_timesteps
+    else:
+        # Sample a random timestep for each image
+        # for weighting schemes where we sample timesteps non-uniformly
+        u = compute_density_for_timestep_sampling(
+            weighting_scheme=args.weighting_scheme,
+            batch_size=bsz,
+            logit_mean=args.logit_mean,
+            logit_std=args.logit_std,
+            mode_scale=args.mode_scale,
+        )
+        indices = (u * num_timesteps).long()
+        timesteps = noise_scheduler.timesteps[indices].to(device=device)
+        sigmas = get_sigmas(noise_scheduler, timesteps, device, n_dim=latents.ndim, dtype=dtype)
+
+    # Broadcast sigmas to latent shape
+    sigmas = sigmas.view(-1, 1, 1, 1)
+
+    # Add noise to the latents according to the noise magnitude at each timestep
+    # (this is the forward diffusion process)
+    if args.ip_noise_gamma:
+        xi = torch.randn_like(latents, device=latents.device, dtype=dtype)
+        if args.ip_noise_gamma_random_strength:
+            ip_noise_gamma = torch.rand(1, device=latents.device, dtype=dtype) * args.ip_noise_gamma
+        else:
+            ip_noise_gamma = args.ip_noise_gamma
+        noisy_model_input = (1.0 - sigmas) * latents + sigmas * (noise + ip_noise_gamma * xi)
+    else:
+        noisy_model_input = (1.0 - sigmas) * latents + sigmas * noise
+
+    return noisy_model_input.to(dtype), timesteps.to(dtype), sigmas
+
+
+def apply_model_prediction_type(args, model_pred, noisy_model_input, sigmas):
+    weighting = None
+    if args.model_prediction_type == "raw":
+        pass
+    elif args.model_prediction_type == "additive":
+        # add the model_pred to the noisy_model_input
+        model_pred = model_pred + noisy_model_input
+    elif args.model_prediction_type == "sigma_scaled":
+        # apply sigma scaling
+        model_pred = model_pred * (-sigmas) + noisy_model_input
+
+        # these weighting schemes use a uniform timestep sampling
+        # and instead post-weight the loss
+        weighting = compute_loss_weighting_for_sd3(weighting_scheme=args.weighting_scheme, sigmas=sigmas)
+
+    return model_pred, weighting
+
+
+def save_models(
+    ckpt_path: str,
+    flux: flux_models.Flux,
+    sai_metadata: Optional[dict],
+    save_dtype: Optional[torch.dtype] = None,
+    use_mem_eff_save: bool = False,
+):
+    state_dict = {}
+
+    def update_sd(prefix, sd):
+        for k, v in sd.items():
+            key = prefix + k
+            if save_dtype is not None and v.dtype != save_dtype:
+                v = v.detach().clone().to("cpu").to(save_dtype)
+            state_dict[key] = v
+
+    update_sd("", flux.state_dict())
+
+    if not use_mem_eff_save:
+        save_file(state_dict, ckpt_path, metadata=sai_metadata)
+    else:
+        mem_eff_save_file(state_dict, ckpt_path, metadata=sai_metadata)
+
+
+def save_flux_model_on_train_end(
+    args: argparse.Namespace, save_dtype: torch.dtype, epoch: int, global_step: int, flux: flux_models.Flux
+):
+    def sd_saver(ckpt_file, epoch_no, global_step):
+        sai_metadata = train_util.get_sai_model_spec(None, args, False, False, False, is_stable_diffusion_ckpt=True, flux="dev")
+        save_models(ckpt_file, flux, sai_metadata, save_dtype, args.mem_eff_save)
+
+    train_util.save_sd_model_on_train_end_common(args, True, True, epoch, global_step, sd_saver, None)
+
+
+# epochとstepの保存、メタデータにepoch/stepが含まれ引数が同じになるため、統合している
+# on_epoch_end: Trueならepoch終了時、Falseならstep経過時
+def save_flux_model_on_epoch_end_or_stepwise(
+    args: argparse.Namespace,
+    on_epoch_end: bool,
+    accelerator,
+    save_dtype: torch.dtype,
+    epoch: int,
+    num_train_epochs: int,
+    global_step: int,
+    flux: flux_models.Flux,
+):
+    def sd_saver(ckpt_file, epoch_no, global_step):
+        sai_metadata = train_util.get_sai_model_spec(None, args, False, False, False, is_stable_diffusion_ckpt=True, flux="dev")
+        save_models(ckpt_file, flux, sai_metadata, save_dtype, args.mem_eff_save)
+
+    train_util.save_sd_model_on_epoch_end_or_stepwise_common(
+        args,
+        on_epoch_end,
+        accelerator,
+        True,
+        True,
+        epoch,
+        num_train_epochs,
+        global_step,
+        sd_saver,
+        None,
+    )
+
+
+# endregion
+
+
+def add_flux_train_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--clip_l",
+        type=str,
+        help="path to clip_l (*.sft or *.safetensors), should be float16 / clip_lのパス（*.sftまたは*.safetensors）、float16が前提",
+    )
+    parser.add_argument(
+        "--t5xxl",
+        type=str,
+        help="path to t5xxl (*.sft or *.safetensors), should be float16 / t5xxlのパス（*.sftまたは*.safetensors）、float16が前提",
+    )
+    parser.add_argument("--ae", type=str, help="path to ae (*.sft or *.safetensors) / aeのパス（*.sftまたは*.safetensors）")
+    parser.add_argument(
+        "--controlnet_model_name_or_path",
+        type=str,
+        default=None,
+        help="path to controlnet (*.sft or *.safetensors) / controlnetのパス（*.sftまたは*.safetensors）",
+    )
+    parser.add_argument(
+        "--t5xxl_max_token_length",
+        type=int,
+        default=None,
+        help="maximum token length for T5-XXL. if omitted, 256 for schnell and 512 for dev"
+        " / T5-XXLの最大トークン長。省略された場合、schnellの場合は256、devの場合は512",
+    )
+    parser.add_argument(
+        "--apply_t5_attn_mask",
+        action="store_true",
+        help="apply attention mask to T5-XXL encode and FLUX double blocks / T5-XXLエンコードとFLUXダブルブロックにアテンションマスクを適用する",
+    )
+
+    parser.add_argument(
+        "--guidance_scale",
+        type=float,
+        default=3.5,
+        help="the FLUX.1 dev variant is a guidance distilled model",
+    )
+
+    parser.add_argument(
+        "--timestep_sampling",
+        choices=["sigma", "uniform", "sigmoid", "shift", "flux_shift"],
+        default="sigma",
+        help="Method to sample timesteps: sigma-based, uniform random, sigmoid of random normal, shift of sigmoid and FLUX.1 shifting."
+        " / タイムステップをサンプリングする方法：sigma、random uniform、random normalのsigmoid、sigmoidのシフト、FLUX.1のシフト。",
+    )
+    parser.add_argument(
+        "--sigmoid_scale",
+        type=float,
+        default=1.0,
+        help='Scale factor for sigmoid timestep sampling (only used when timestep-sampling is "sigmoid"). / sigmoidタイムステップサンプリングの倍率（timestep-samplingが"sigmoid"の場合のみ有効）。',
+    )
+    parser.add_argument(
+        "--model_prediction_type",
+        choices=["raw", "additive", "sigma_scaled"],
+        default="sigma_scaled",
+        help="How to interpret and process the model prediction: "
+        "raw (use as is), additive (add to noisy input), sigma_scaled (apply sigma scaling)."
+        " / モデル予測の解釈と処理方法："
+        "raw（そのまま使用）、additive（ノイズ入力に加算）、sigma_scaled（シグマスケーリングを適用）。",
+    )
+    parser.add_argument(
+        "--discrete_flow_shift",
+        type=float,
+        default=3.0,
+        help="Discrete flow shift for the Euler Discrete Scheduler, default is 3.0. / Euler Discrete Schedulerの離散フローシフト、デフォルトは3.0。",
+    )
+
+    parser.add_argument(
+        "--model_type",
+        type=str,
+        choices=["flux", "chroma"],
+        default="flux",
+        help="Model type to use for training / トレーニングに使用するモデルタイプ：flux or chroma (default: flux)",
+    )

library/flux_utils.py

@@ -0,0 +1,563 @@
+import json
+import os
+from dataclasses import replace
+from typing import List, Optional, Tuple, Union
+
+import einops
+import torch
+from accelerate import init_empty_weights
+from safetensors import safe_open
+from safetensors.torch import load_file
+from transformers import CLIPConfig, CLIPTextModel, T5Config, T5EncoderModel
+
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+from library import flux_models
+from library.safetensors_utils import load_safetensors
+
+MODEL_VERSION_FLUX_V1 = "flux1"
+MODEL_NAME_DEV = "dev"
+MODEL_NAME_SCHNELL = "schnell"
+MODEL_VERSION_CHROMA = "chroma"
+
+
+def analyze_checkpoint_state(ckpt_path: str) -> Tuple[bool, bool, Tuple[int, int], List[str]]:
+    """
+    チェックポイントの状態を分析し、DiffusersかBFLか、devかschnellか、ブロック数を計算して返す。
+
+    Args:
+        ckpt_path (str): チェックポイントファイルまたはディレクトリのパス。
+
+    Returns:
+        Tuple[bool, bool, Tuple[int, int], List[str]]:
+            - bool: Diffusersかどうかを示すフラグ。
+            - bool: Schnellかどうかを示すフラグ。
+            - Tuple[int, int]: ダブルブロックとシングルブロックの数。
+            - List[str]: チェックポイントに含まれるキーのリスト。
+    """
+    # check the state dict: Diffusers or BFL, dev or schnell, number of blocks
+    logger.info(f"Checking the state dict: Diffusers or BFL, dev or schnell")
+
+    if os.path.isdir(ckpt_path):  # if ckpt_path is a directory, it is Diffusers
+        ckpt_path = os.path.join(ckpt_path, "transformer", "diffusion_pytorch_model-00001-of-00003.safetensors")
+    if "00001-of-00003" in ckpt_path:
+        ckpt_paths = [ckpt_path.replace("00001-of-00003", f"0000{i}-of-00003") for i in range(1, 4)]
+    else:
+        ckpt_paths = [ckpt_path]
+
+    keys = []
+    for ckpt_path in ckpt_paths:
+        with safe_open(ckpt_path, framework="pt") as f:
+            keys.extend(f.keys())
+
+    # if the key has annoying prefix, remove it
+    if keys[0].startswith("model.diffusion_model."):
+        keys = [key.replace("model.diffusion_model.", "") for key in keys]
+
+    is_diffusers = "transformer_blocks.0.attn.add_k_proj.bias" in keys
+    is_schnell = not ("guidance_in.in_layer.bias" in keys or "time_text_embed.guidance_embedder.linear_1.bias" in keys)
+
+    # check number of double and single blocks
+    if not is_diffusers:
+        max_double_block_index = max(
+            [int(key.split(".")[1]) for key in keys if key.startswith("double_blocks.") and key.endswith(".img_attn.proj.bias")]
+        )
+        max_single_block_index = max(
+            [int(key.split(".")[1]) for key in keys if key.startswith("single_blocks.") and key.endswith(".modulation.lin.bias")]
+        )
+    else:
+        max_double_block_index = max(
+            [
+                int(key.split(".")[1])
+                for key in keys
+                if key.startswith("transformer_blocks.") and key.endswith(".attn.add_k_proj.bias")
+            ]
+        )
+        max_single_block_index = max(
+            [
+                int(key.split(".")[1])
+                for key in keys
+                if key.startswith("single_transformer_blocks.") and key.endswith(".attn.to_k.bias")
+            ]
+        )
+
+    num_double_blocks = max_double_block_index + 1
+    num_single_blocks = max_single_block_index + 1
+
+    return is_diffusers, is_schnell, (num_double_blocks, num_single_blocks), ckpt_paths
+
+
+def load_flow_model(
+    ckpt_path: str,
+    dtype: Optional[torch.dtype],
+    device: Union[str, torch.device],
+    disable_mmap: bool = False,
+    model_type: str = "flux",
+) -> Tuple[bool, flux_models.Flux]:
+    if model_type == "flux":
+        is_diffusers, is_schnell, (num_double_blocks, num_single_blocks), ckpt_paths = analyze_checkpoint_state(ckpt_path)
+        name = MODEL_NAME_DEV if not is_schnell else MODEL_NAME_SCHNELL
+
+        # build model
+        logger.info(f"Building Flux model {name} from {'Diffusers' if is_diffusers else 'BFL'} checkpoint")
+        with torch.device("meta"):
+            params = flux_models.configs[name].params
+
+            # set the number of blocks
+            if params.depth != num_double_blocks:
+                logger.info(f"Setting the number of double blocks from {params.depth} to {num_double_blocks}")
+                params = replace(params, depth=num_double_blocks)
+            if params.depth_single_blocks != num_single_blocks:
+                logger.info(f"Setting the number of single blocks from {params.depth_single_blocks} to {num_single_blocks}")
+                params = replace(params, depth_single_blocks=num_single_blocks)
+
+            model = flux_models.Flux(params)
+            if dtype is not None:
+                model = model.to(dtype)
+
+        # load_sft doesn't support torch.device
+        logger.info(f"Loading state dict from {ckpt_path}")
+        sd = {}
+        for ckpt_path in ckpt_paths:
+            sd.update(load_safetensors(ckpt_path, device=device, disable_mmap=disable_mmap, dtype=dtype))
+
+        # convert Diffusers to BFL
+        if is_diffusers:
+            logger.info("Converting Diffusers to BFL")
+            sd = convert_diffusers_sd_to_bfl(sd, num_double_blocks, num_single_blocks)
+            logger.info("Converted Diffusers to BFL")
+
+        # if the key has annoying prefix, remove it
+        for key in list(sd.keys()):
+            new_key = key.replace("model.diffusion_model.", "")
+            if new_key == key:
+                break  # the model doesn't have annoying prefix
+            sd[new_key] = sd.pop(key)
+
+        info = model.load_state_dict(sd, strict=False, assign=True)
+        logger.info(f"Loaded Flux: {info}")
+        return is_schnell, model
+
+    elif model_type == "chroma":
+        from . import chroma_models
+
+        # build model
+        logger.info("Building Chroma model")
+        with torch.device("meta"):
+            model = chroma_models.Chroma(chroma_models.chroma_params)
+            if dtype is not None:
+                model = model.to(dtype)
+
+        # load_sft doesn't support torch.device
+        logger.info(f"Loading state dict from {ckpt_path}")
+        sd = load_safetensors(ckpt_path, device=str(device), disable_mmap=disable_mmap, dtype=dtype)
+
+        # if the key has annoying prefix, remove it
+        for key in list(sd.keys()):
+            new_key = key.replace("model.diffusion_model.", "")
+            if new_key == key:
+                break  # the model doesn't have annoying prefix
+            sd[new_key] = sd.pop(key)
+
+        info = model.load_state_dict(sd, strict=False, assign=True)
+        logger.info(f"Loaded Chroma: {info}")
+        is_schnell = False  # Chroma is not schnell
+        return is_schnell, model
+
+    else:
+        raise ValueError(f"Unsupported model_type: {model_type}. Supported types are 'flux' and 'chroma'.")
+
+
+def load_ae(
+    ckpt_path: str, dtype: torch.dtype, device: Union[str, torch.device], disable_mmap: bool = False
+) -> flux_models.AutoEncoder:
+    logger.info("Building AutoEncoder")
+    with torch.device("meta"):
+        # dev and schnell have the same AE params
+        ae = flux_models.AutoEncoder(flux_models.configs[MODEL_NAME_DEV].ae_params).to(dtype)
+
+    logger.info(f"Loading state dict from {ckpt_path}")
+    sd = load_safetensors(ckpt_path, device=str(device), disable_mmap=disable_mmap, dtype=dtype)
+    info = ae.load_state_dict(sd, strict=False, assign=True)
+    logger.info(f"Loaded AE: {info}")
+    return ae
+
+
+def load_controlnet(
+    ckpt_path: Optional[str], is_schnell: bool, dtype: torch.dtype, device: Union[str, torch.device], disable_mmap: bool = False
+):
+    logger.info("Building ControlNet")
+    name = MODEL_NAME_DEV if not is_schnell else MODEL_NAME_SCHNELL
+    with torch.device(device):
+        controlnet = flux_models.ControlNetFlux(flux_models.configs[name].params).to(dtype)
+
+    if ckpt_path is not None:
+        logger.info(f"Loading state dict from {ckpt_path}")
+        sd = load_safetensors(ckpt_path, device=str(device), disable_mmap=disable_mmap, dtype=dtype)
+        info = controlnet.load_state_dict(sd, strict=False, assign=True)
+        logger.info(f"Loaded ControlNet: {info}")
+    return controlnet
+
+
+def dummy_clip_l() -> torch.nn.Module:
+    """
+    Returns a dummy CLIP-L model with the output shape of (N, 77, 768).
+    """
+    return DummyCLIPL()
+
+
+class DummyTextModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.embeddings = torch.nn.Parameter(torch.zeros(1))
+
+
+class DummyCLIPL(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.output_shape = (77, 1)  # Note: The original code had (77, 768), but we use (77, 1) for the dummy output
+
+        # dtype and device from these parameters. train_network.py accesses them
+        self.dummy_param = torch.nn.Parameter(torch.zeros(1))
+        self.dummy_param_2 = torch.nn.Parameter(torch.zeros(1))
+        self.dummy_param_3 = torch.nn.Parameter(torch.zeros(1))
+        self.text_model = DummyTextModel()
+
+    @property
+    def device(self):
+        return self.dummy_param.device
+
+    @property
+    def dtype(self):
+        return self.dummy_param.dtype
+
+    def forward(self, *args, **kwargs):
+        """
+        Returns a dummy output with the shape of (N, 77, 768).
+        """
+        batch_size = args[0].shape[0] if args else 1
+        return {"pooler_output": torch.zeros(batch_size, *self.output_shape, device=self.device, dtype=self.dtype)}
+
+
+def load_clip_l(
+    ckpt_path: Optional[str],
+    dtype: torch.dtype,
+    device: Union[str, torch.device],
+    disable_mmap: bool = False,
+    state_dict: Optional[dict] = None,
+) -> CLIPTextModel:
+    logger.info("Building CLIP-L")
+    CLIPL_CONFIG = {
+        "_name_or_path": "clip-vit-large-patch14/",
+        "architectures": ["CLIPModel"],
+        "initializer_factor": 1.0,
+        "logit_scale_init_value": 2.6592,
+        "model_type": "clip",
+        "projection_dim": 768,
+        # "text_config": {
+        "_name_or_path": "",
+        "add_cross_attention": False,
+        "architectures": None,
+        "attention_dropout": 0.0,
+        "bad_words_ids": None,
+        "bos_token_id": 0,
+        "chunk_size_feed_forward": 0,
+        "cross_attention_hidden_size": None,
+        "decoder_start_token_id": None,
+        "diversity_penalty": 0.0,
+        "do_sample": False,
+        "dropout": 0.0,
+        "early_stopping": False,
+        "encoder_no_repeat_ngram_size": 0,
+        "eos_token_id": 2,
+        "finetuning_task": None,
+        "forced_bos_token_id": None,
+        "forced_eos_token_id": None,
+        "hidden_act": "quick_gelu",
+        "hidden_size": 768,
+        "id2label": {"0": "LABEL_0", "1": "LABEL_1"},
+        "initializer_factor": 1.0,
+        "initializer_range": 0.02,
+        "intermediate_size": 3072,
+        "is_decoder": False,
+        "is_encoder_decoder": False,
+        "label2id": {"LABEL_0": 0, "LABEL_1": 1},
+        "layer_norm_eps": 1e-05,
+        "length_penalty": 1.0,
+        "max_length": 20,
+        "max_position_embeddings": 77,
+        "min_length": 0,
+        "model_type": "clip_text_model",
+        "no_repeat_ngram_size": 0,
+        "num_attention_heads": 12,
+        "num_beam_groups": 1,
+        "num_beams": 1,
+        "num_hidden_layers": 12,
+        "num_return_sequences": 1,
+        "output_attentions": False,
+        "output_hidden_states": False,
+        "output_scores": False,
+        "pad_token_id": 1,
+        "prefix": None,
+        "problem_type": None,
+        "projection_dim": 768,
+        "pruned_heads": {},
+        "remove_invalid_values": False,
+        "repetition_penalty": 1.0,
+        "return_dict": True,
+        "return_dict_in_generate": False,
+        "sep_token_id": None,
+        "task_specific_params": None,
+        "temperature": 1.0,
+        "tie_encoder_decoder": False,
+        "tie_word_embeddings": True,
+        "tokenizer_class": None,
+        "top_k": 50,
+        "top_p": 1.0,
+        "torch_dtype": None,
+        "torchscript": False,
+        "transformers_version": "4.16.0.dev0",
+        "use_bfloat16": False,
+        "vocab_size": 49408,
+        "hidden_act": "gelu",
+        "hidden_size": 1280,
+        "intermediate_size": 5120,
+        "num_attention_heads": 20,
+        "num_hidden_layers": 32,
+        # },
+        # "text_config_dict": {
+        "hidden_size": 768,
+        "intermediate_size": 3072,
+        "num_attention_heads": 12,
+        "num_hidden_layers": 12,
+        "projection_dim": 768,
+        # },
+        # "torch_dtype": "float32",
+        # "transformers_version": None,
+    }
+    config = CLIPConfig(**CLIPL_CONFIG)
+    with init_empty_weights():
+        clip = CLIPTextModel._from_config(config)
+
+    if state_dict is not None:
+        sd = state_dict
+    else:
+        logger.info(f"Loading state dict from {ckpt_path}")
+        sd = load_safetensors(ckpt_path, device=str(device), disable_mmap=disable_mmap, dtype=dtype)
+    info = clip.load_state_dict(sd, strict=False, assign=True)
+    logger.info(f"Loaded CLIP-L: {info}")
+    return clip
+
+
+def load_t5xxl(
+    ckpt_path: str,
+    dtype: Optional[torch.dtype],
+    device: Union[str, torch.device],
+    disable_mmap: bool = False,
+    state_dict: Optional[dict] = None,
+) -> T5EncoderModel:
+    T5_CONFIG_JSON = """
+{
+  "architectures": [
+    "T5EncoderModel"
+  ],
+  "classifier_dropout": 0.0,
+  "d_ff": 10240,
+  "d_kv": 64,
+  "d_model": 4096,
+  "decoder_start_token_id": 0,
+  "dense_act_fn": "gelu_new",
+  "dropout_rate": 0.1,
+  "eos_token_id": 1,
+  "feed_forward_proj": "gated-gelu",
+  "initializer_factor": 1.0,
+  "is_encoder_decoder": true,
+  "is_gated_act": true,
+  "layer_norm_epsilon": 1e-06,
+  "model_type": "t5",
+  "num_decoder_layers": 24,
+  "num_heads": 64,
+  "num_layers": 24,
+  "output_past": true,
+  "pad_token_id": 0,
+  "relative_attention_max_distance": 128,
+  "relative_attention_num_buckets": 32,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.41.2",
+  "use_cache": true,
+  "vocab_size": 32128
+}
+"""
+    config = json.loads(T5_CONFIG_JSON)
+    config = T5Config(**config)
+    with init_empty_weights():
+        t5xxl = T5EncoderModel._from_config(config)
+
+    if state_dict is not None:
+        sd = state_dict
+    else:
+        logger.info(f"Loading state dict from {ckpt_path}")
+        sd = load_safetensors(ckpt_path, device=str(device), disable_mmap=disable_mmap, dtype=dtype)
+    info = t5xxl.load_state_dict(sd, strict=False, assign=True)
+    logger.info(f"Loaded T5xxl: {info}")
+    return t5xxl
+
+
+def get_t5xxl_actual_dtype(t5xxl: T5EncoderModel) -> torch.dtype:
+    # nn.Embedding is the first layer, but it could be casted to bfloat16 or float32
+    return t5xxl.encoder.block[0].layer[0].SelfAttention.q.weight.dtype
+
+
+def prepare_img_ids(batch_size: int, packed_latent_height: int, packed_latent_width: int):
+    img_ids = torch.zeros(packed_latent_height, packed_latent_width, 3)
+    img_ids[..., 1] = img_ids[..., 1] + torch.arange(packed_latent_height)[:, None]
+    img_ids[..., 2] = img_ids[..., 2] + torch.arange(packed_latent_width)[None, :]
+    img_ids = einops.repeat(img_ids, "h w c -> b (h w) c", b=batch_size)
+    return img_ids
+
+
+def unpack_latents(x: torch.Tensor, packed_latent_height: int, packed_latent_width: int) -> torch.Tensor:
+    """
+    x: [b (h w) (c ph pw)] -> [b c (h ph) (w pw)], ph=2, pw=2
+    """
+    x = einops.rearrange(x, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=packed_latent_height, w=packed_latent_width, ph=2, pw=2)
+    return x
+
+
+def pack_latents(x: torch.Tensor) -> torch.Tensor:
+    """
+    x: [b c (h ph) (w pw)] -> [b (h w) (c ph pw)], ph=2, pw=2
+    """
+    x = einops.rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
+    return x
+
+
+# region Diffusers
+
+NUM_DOUBLE_BLOCKS = 19
+NUM_SINGLE_BLOCKS = 38
+
+BFL_TO_DIFFUSERS_MAP = {
+    "time_in.in_layer.weight": ["time_text_embed.timestep_embedder.linear_1.weight"],
+    "time_in.in_layer.bias": ["time_text_embed.timestep_embedder.linear_1.bias"],
+    "time_in.out_layer.weight": ["time_text_embed.timestep_embedder.linear_2.weight"],
+    "time_in.out_layer.bias": ["time_text_embed.timestep_embedder.linear_2.bias"],
+    "vector_in.in_layer.weight": ["time_text_embed.text_embedder.linear_1.weight"],
+    "vector_in.in_layer.bias": ["time_text_embed.text_embedder.linear_1.bias"],
+    "vector_in.out_layer.weight": ["time_text_embed.text_embedder.linear_2.weight"],
+    "vector_in.out_layer.bias": ["time_text_embed.text_embedder.linear_2.bias"],
+    "guidance_in.in_layer.weight": ["time_text_embed.guidance_embedder.linear_1.weight"],
+    "guidance_in.in_layer.bias": ["time_text_embed.guidance_embedder.linear_1.bias"],
+    "guidance_in.out_layer.weight": ["time_text_embed.guidance_embedder.linear_2.weight"],
+    "guidance_in.out_layer.bias": ["time_text_embed.guidance_embedder.linear_2.bias"],
+    "txt_in.weight": ["context_embedder.weight"],
+    "txt_in.bias": ["context_embedder.bias"],
+    "img_in.weight": ["x_embedder.weight"],
+    "img_in.bias": ["x_embedder.bias"],
+    "double_blocks.().img_mod.lin.weight": ["norm1.linear.weight"],
+    "double_blocks.().img_mod.lin.bias": ["norm1.linear.bias"],
+    "double_blocks.().txt_mod.lin.weight": ["norm1_context.linear.weight"],
+    "double_blocks.().txt_mod.lin.bias": ["norm1_context.linear.bias"],
+    "double_blocks.().img_attn.qkv.weight": ["attn.to_q.weight", "attn.to_k.weight", "attn.to_v.weight"],
+    "double_blocks.().img_attn.qkv.bias": ["attn.to_q.bias", "attn.to_k.bias", "attn.to_v.bias"],
+    "double_blocks.().txt_attn.qkv.weight": ["attn.add_q_proj.weight", "attn.add_k_proj.weight", "attn.add_v_proj.weight"],
+    "double_blocks.().txt_attn.qkv.bias": ["attn.add_q_proj.bias", "attn.add_k_proj.bias", "attn.add_v_proj.bias"],
+    "double_blocks.().img_attn.norm.query_norm.scale": ["attn.norm_q.weight"],
+    "double_blocks.().img_attn.norm.key_norm.scale": ["attn.norm_k.weight"],
+    "double_blocks.().txt_attn.norm.query_norm.scale": ["attn.norm_added_q.weight"],
+    "double_blocks.().txt_attn.norm.key_norm.scale": ["attn.norm_added_k.weight"],
+    "double_blocks.().img_mlp.0.weight": ["ff.net.0.proj.weight"],
+    "double_blocks.().img_mlp.0.bias": ["ff.net.0.proj.bias"],
+    "double_blocks.().img_mlp.2.weight": ["ff.net.2.weight"],
+    "double_blocks.().img_mlp.2.bias": ["ff.net.2.bias"],
+    "double_blocks.().txt_mlp.0.weight": ["ff_context.net.0.proj.weight"],
+    "double_blocks.().txt_mlp.0.bias": ["ff_context.net.0.proj.bias"],
+    "double_blocks.().txt_mlp.2.weight": ["ff_context.net.2.weight"],
+    "double_blocks.().txt_mlp.2.bias": ["ff_context.net.2.bias"],
+    "double_blocks.().img_attn.proj.weight": ["attn.to_out.0.weight"],
+    "double_blocks.().img_attn.proj.bias": ["attn.to_out.0.bias"],
+    "double_blocks.().txt_attn.proj.weight": ["attn.to_add_out.weight"],
+    "double_blocks.().txt_attn.proj.bias": ["attn.to_add_out.bias"],
+    "single_blocks.().modulation.lin.weight": ["norm.linear.weight"],
+    "single_blocks.().modulation.lin.bias": ["norm.linear.bias"],
+    "single_blocks.().linear1.weight": ["attn.to_q.weight", "attn.to_k.weight", "attn.to_v.weight", "proj_mlp.weight"],
+    "single_blocks.().linear1.bias": ["attn.to_q.bias", "attn.to_k.bias", "attn.to_v.bias", "proj_mlp.bias"],
+    "single_blocks.().linear2.weight": ["proj_out.weight"],
+    "single_blocks.().norm.query_norm.scale": ["attn.norm_q.weight"],
+    "single_blocks.().norm.key_norm.scale": ["attn.norm_k.weight"],
+    "single_blocks.().linear2.weight": ["proj_out.weight"],
+    "single_blocks.().linear2.bias": ["proj_out.bias"],
+    "final_layer.linear.weight": ["proj_out.weight"],
+    "final_layer.linear.bias": ["proj_out.bias"],
+    "final_layer.adaLN_modulation.1.weight": ["norm_out.linear.weight"],
+    "final_layer.adaLN_modulation.1.bias": ["norm_out.linear.bias"],
+}
+
+
+def make_diffusers_to_bfl_map(num_double_blocks: int, num_single_blocks: int) -> dict[str, tuple[int, str]]:
+    # make reverse map from diffusers map
+    diffusers_to_bfl_map = {}  # key: diffusers_key, value: (index, bfl_key)
+    for b in range(num_double_blocks):
+        for key, weights in BFL_TO_DIFFUSERS_MAP.items():
+            if key.startswith("double_blocks."):
+                block_prefix = f"transformer_blocks.{b}."
+                for i, weight in enumerate(weights):
+                    diffusers_to_bfl_map[f"{block_prefix}{weight}"] = (i, key.replace("()", f"{b}"))
+    for b in range(num_single_blocks):
+        for key, weights in BFL_TO_DIFFUSERS_MAP.items():
+            if key.startswith("single_blocks."):
+                block_prefix = f"single_transformer_blocks.{b}."
+                for i, weight in enumerate(weights):
+                    diffusers_to_bfl_map[f"{block_prefix}{weight}"] = (i, key.replace("()", f"{b}"))
+    for key, weights in BFL_TO_DIFFUSERS_MAP.items():
+        if not (key.startswith("double_blocks.") or key.startswith("single_blocks.")):
+            for i, weight in enumerate(weights):
+                diffusers_to_bfl_map[weight] = (i, key)
+    return diffusers_to_bfl_map
+
+
+def convert_diffusers_sd_to_bfl(
+    diffusers_sd: dict[str, torch.Tensor], num_double_blocks: int = NUM_DOUBLE_BLOCKS, num_single_blocks: int = NUM_SINGLE_BLOCKS
+) -> dict[str, torch.Tensor]:
+    diffusers_to_bfl_map = make_diffusers_to_bfl_map(num_double_blocks, num_single_blocks)
+
+    # iterate over three safetensors files to reduce memory usage
+    flux_sd = {}
+    for diffusers_key, tensor in diffusers_sd.items():
+        if diffusers_key in diffusers_to_bfl_map:
+            index, bfl_key = diffusers_to_bfl_map[diffusers_key]
+            if bfl_key not in flux_sd:
+                flux_sd[bfl_key] = []
+            flux_sd[bfl_key].append((index, tensor))
+        else:
+            logger.error(f"Error: Key not found in diffusers_to_bfl_map: {diffusers_key}")
+            raise KeyError(f"Key not found in diffusers_to_bfl_map: {diffusers_key}")
+
+    # concat tensors if multiple tensors are mapped to a single key, sort by index
+    for key, values in flux_sd.items():
+        if len(values) == 1:
+            flux_sd[key] = values[0][1]
+        else:
+            flux_sd[key] = torch.cat([value[1] for value in sorted(values, key=lambda x: x[0])])
+
+    # special case for final_layer.adaLN_modulation.1.weight and final_layer.adaLN_modulation.1.bias
+    def swap_scale_shift(weight):
+        shift, scale = weight.chunk(2, dim=0)
+        new_weight = torch.cat([scale, shift], dim=0)
+        return new_weight
+
+    if "final_layer.adaLN_modulation.1.weight" in flux_sd:
+        flux_sd["final_layer.adaLN_modulation.1.weight"] = swap_scale_shift(flux_sd["final_layer.adaLN_modulation.1.weight"])
+    if "final_layer.adaLN_modulation.1.bias" in flux_sd:
+        flux_sd["final_layer.adaLN_modulation.1.bias"] = swap_scale_shift(flux_sd["final_layer.adaLN_modulation.1.bias"])
+
+    return flux_sd
+
+
+# endregion

library/fp8_optimization_utils.py

@@ -0,0 +1,469 @@
+import os
+from typing import List, Optional, Union
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+import logging
+
+from tqdm import tqdm
+
+from library.device_utils import clean_memory_on_device
+from library.safetensors_utils import MemoryEfficientSafeOpen
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+def calculate_fp8_maxval(exp_bits=4, mantissa_bits=3, sign_bits=1):
+    """
+    Calculate the maximum representable value in FP8 format.
+    Default is E4M3 format (4-bit exponent, 3-bit mantissa, 1-bit sign). Only supports E4M3 and E5M2 with sign bit.
+
+    Args:
+        exp_bits (int): Number of exponent bits
+        mantissa_bits (int): Number of mantissa bits
+        sign_bits (int): Number of sign bits (0 or 1)
+
+    Returns:
+        float: Maximum value representable in FP8 format
+    """
+    assert exp_bits + mantissa_bits + sign_bits == 8, "Total bits must be 8"
+    if exp_bits == 4 and mantissa_bits == 3 and sign_bits == 1:
+        return torch.finfo(torch.float8_e4m3fn).max
+    elif exp_bits == 5 and mantissa_bits == 2 and sign_bits == 1:
+        return torch.finfo(torch.float8_e5m2).max
+    else:
+        raise ValueError(f"Unsupported FP8 format: E{exp_bits}M{mantissa_bits} with sign_bits={sign_bits}")
+
+
+# The following is a manual calculation method (wrong implementation for E5M2), kept for reference.
+"""
+# Calculate exponent bias
+bias = 2 ** (exp_bits - 1) - 1
+
+# Calculate maximum mantissa value
+mantissa_max = 1.0
+for i in range(mantissa_bits - 1):
+    mantissa_max += 2 ** -(i + 1)
+
+# Calculate maximum value
+max_value = mantissa_max * (2 ** (2**exp_bits - 1 - bias))
+
+return max_value
+"""
+
+
+def quantize_fp8(tensor, scale, fp8_dtype, max_value, min_value):
+    """
+    Quantize a tensor to FP8 format using PyTorch's native FP8 dtype support.
+
+    Args:
+        tensor (torch.Tensor): Tensor to quantize
+        scale (float or torch.Tensor): Scale factor
+        fp8_dtype (torch.dtype): Target FP8 dtype (torch.float8_e4m3fn or torch.float8_e5m2)
+        max_value (float): Maximum representable value in FP8
+        min_value (float): Minimum representable value in FP8
+
+    Returns:
+        torch.Tensor: Quantized tensor in FP8 format
+    """
+    tensor = tensor.to(torch.float32)  # ensure tensor is in float32 for division
+
+    # Create scaled tensor
+    tensor = torch.div(tensor, scale).nan_to_num_(0.0)  # handle NaN values, equivalent to nonzero_mask in previous function
+
+    # Clamp tensor to range
+    tensor = tensor.clamp_(min=min_value, max=max_value)
+
+    # Convert to FP8 dtype
+    tensor = tensor.to(fp8_dtype)
+
+    return tensor
+
+
+def optimize_state_dict_with_fp8(
+    state_dict: dict,
+    calc_device: Union[str, torch.device],
+    target_layer_keys: Optional[list[str]] = None,
+    exclude_layer_keys: Optional[list[str]] = None,
+    exp_bits: int = 4,
+    mantissa_bits: int = 3,
+    move_to_device: bool = False,
+    quantization_mode: str = "block",
+    block_size: Optional[int] = 64,
+):
+    """
+    Optimize Linear layer weights in a model's state dict to FP8 format. The state dict is modified in-place.
+    This function is a static version of load_safetensors_with_fp8_optimization without loading from files.
+
+    Args:
+        state_dict (dict): State dict to optimize, replaced in-place
+        calc_device (str): Device to quantize tensors on
+        target_layer_keys (list, optional): Layer key patterns to target (None for all Linear layers)
+        exclude_layer_keys (list, optional): Layer key patterns to exclude
+        exp_bits (int): Number of exponent bits
+        mantissa_bits (int): Number of mantissa bits
+        move_to_device (bool): Move optimized tensors to the calculating device
+
+    Returns:
+        dict: FP8 optimized state dict
+    """
+    if exp_bits == 4 and mantissa_bits == 3:
+        fp8_dtype = torch.float8_e4m3fn
+    elif exp_bits == 5 and mantissa_bits == 2:
+        fp8_dtype = torch.float8_e5m2
+    else:
+        raise ValueError(f"Unsupported FP8 format: E{exp_bits}M{mantissa_bits}")
+
+    # Calculate FP8 max value
+    max_value = calculate_fp8_maxval(exp_bits, mantissa_bits)
+    min_value = -max_value  # this function supports only signed FP8
+
+    # Create optimized state dict
+    optimized_count = 0
+
+    # Enumerate tarket keys
+    target_state_dict_keys = []
+    for key in state_dict.keys():
+        # Check if it's a weight key and matches target patterns
+        is_target = (target_layer_keys is None or any(pattern in key for pattern in target_layer_keys)) and key.endswith(".weight")
+        is_excluded = exclude_layer_keys is not None and any(pattern in key for pattern in exclude_layer_keys)
+        is_target = is_target and not is_excluded
+
+        if is_target and isinstance(state_dict[key], torch.Tensor):
+            target_state_dict_keys.append(key)
+
+    # Process each key
+    for key in tqdm(target_state_dict_keys):
+        value = state_dict[key]
+
+        # Save original device and dtype
+        original_device = value.device
+        original_dtype = value.dtype
+
+        # Move to calculation device
+        if calc_device is not None:
+            value = value.to(calc_device)
+
+        quantized_weight, scale_tensor = quantize_weight(key, value, fp8_dtype, max_value, min_value, quantization_mode, block_size)
+
+        # Add to state dict using original key for weight and new key for scale
+        fp8_key = key  # Maintain original key
+        scale_key = key.replace(".weight", ".scale_weight")
+
+        if not move_to_device:
+            quantized_weight = quantized_weight.to(original_device)
+
+        # keep scale shape: [1] or [out,1] or [out, num_blocks, 1]. We can determine the quantization mode from the shape of scale_weight in the patched model.
+        scale_tensor = scale_tensor.to(dtype=original_dtype, device=quantized_weight.device)
+
+        state_dict[fp8_key] = quantized_weight
+        state_dict[scale_key] = scale_tensor
+
+        optimized_count += 1
+
+        if calc_device is not None:  # optimized_count % 10 == 0 and
+            # free memory on calculation device
+            clean_memory_on_device(calc_device)
+
+    logger.info(f"Number of optimized Linear layers: {optimized_count}")
+    return state_dict
+
+
+def quantize_weight(
+    key: str,
+    tensor: torch.Tensor,
+    fp8_dtype: torch.dtype,
+    max_value: float,
+    min_value: float,
+    quantization_mode: str = "block",
+    block_size: int = 64,
+):
+    original_shape = tensor.shape
+
+    # Determine quantization mode
+    if quantization_mode == "block":
+        if tensor.ndim != 2:
+            quantization_mode = "tensor"  # fallback to per-tensor
+        else:
+            out_features, in_features = tensor.shape
+            if in_features % block_size != 0:
+                quantization_mode = "channel"  # fallback to per-channel
+                logger.warning(
+                    f"Layer {key} with shape {tensor.shape} is not divisible by block_size {block_size}, fallback to per-channel quantization."
+                )
+            else:
+                num_blocks = in_features // block_size
+                tensor = tensor.contiguous().view(out_features, num_blocks, block_size)  # [out, num_blocks, block_size]
+    elif quantization_mode == "channel":
+        if tensor.ndim != 2:
+            quantization_mode = "tensor"  # fallback to per-tensor
+
+    # Calculate scale factor (per-tensor or per-output-channel with percentile or max)
+    # value shape is expected to be [out_features, in_features] for Linear weights
+    if quantization_mode == "channel" or quantization_mode == "block":
+        # row-wise percentile to avoid being dominated by outliers
+        # result shape: [out_features, 1] or [out_features, num_blocks, 1]
+        scale_dim = 1 if quantization_mode == "channel" else 2
+        abs_w = torch.abs(tensor)
+
+        # shape: [out_features, 1] or [out_features, num_blocks, 1]
+        row_max = torch.max(abs_w, dim=scale_dim, keepdim=True).values
+        scale = row_max / max_value
+
+    else:
+        # per-tensor
+        tensor_max = torch.max(torch.abs(tensor).view(-1))
+        scale = tensor_max / max_value
+
+    # numerical safety
+    scale = torch.clamp(scale, min=1e-8)
+    scale = scale.to(torch.float32)  # ensure scale is in float32 for division
+
+    # Quantize weight to FP8 (scale can be scalar or [out,1], broadcasting works)
+    quantized_weight = quantize_fp8(tensor, scale, fp8_dtype, max_value, min_value)
+
+    # If block-wise, restore original shape
+    if quantization_mode == "block":
+        quantized_weight = quantized_weight.view(original_shape)  # restore to original shape [out, in]
+
+    return quantized_weight, scale
+
+
+def load_safetensors_with_fp8_optimization(
+    model_files: List[str],
+    calc_device: Union[str, torch.device],
+    target_layer_keys=None,
+    exclude_layer_keys=None,
+    exp_bits=4,
+    mantissa_bits=3,
+    move_to_device=False,
+    weight_hook=None,
+    quantization_mode: str = "block",
+    block_size: Optional[int] = 64,
+) -> dict:
+    """
+    Load weight tensors from safetensors files and merge LoRA weights into the state dict with explicit FP8 optimization.
+
+    Args:
+        model_files (list[str]): List of model files to load
+        calc_device (str or torch.device): Device to quantize tensors on
+        target_layer_keys (list, optional): Layer key patterns to target for optimization (None for all Linear layers)
+        exclude_layer_keys (list, optional): Layer key patterns to exclude from optimization
+        exp_bits (int): Number of exponent bits
+        mantissa_bits (int): Number of mantissa bits
+        move_to_device (bool): Move optimized tensors to the calculating device
+        weight_hook (callable, optional): Function to apply to each weight tensor before optimization
+        quantization_mode (str): Quantization mode, "tensor", "channel", or "block"
+        block_size (int, optional): Block size for block-wise quantization (used if quantization_mode is "block")
+
+    Returns:
+        dict: FP8 optimized state dict
+    """
+    if exp_bits == 4 and mantissa_bits == 3:
+        fp8_dtype = torch.float8_e4m3fn
+    elif exp_bits == 5 and mantissa_bits == 2:
+        fp8_dtype = torch.float8_e5m2
+    else:
+        raise ValueError(f"Unsupported FP8 format: E{exp_bits}M{mantissa_bits}")
+
+    # Calculate FP8 max value
+    max_value = calculate_fp8_maxval(exp_bits, mantissa_bits)
+    min_value = -max_value  # this function supports only signed FP8
+
+    # Define function to determine if a key is a target key. target means fp8 optimization, not for weight hook.
+    def is_target_key(key):
+        # Check if weight key matches target patterns and does not match exclude patterns
+        is_target = (target_layer_keys is None or any(pattern in key for pattern in target_layer_keys)) and key.endswith(".weight")
+        is_excluded = exclude_layer_keys is not None and any(pattern in key for pattern in exclude_layer_keys)
+        return is_target and not is_excluded
+
+    # Create optimized state dict
+    optimized_count = 0
+
+    # Process each file
+    state_dict = {}
+    for model_file in model_files:
+        with MemoryEfficientSafeOpen(model_file) as f:
+            keys = f.keys()
+            for key in tqdm(keys, desc=f"Loading {os.path.basename(model_file)}", unit="key"):
+                value = f.get_tensor(key)
+
+                # Save original device
+                original_device = value.device  # usually cpu
+
+                if weight_hook is not None:
+                    # Apply weight hook if provided
+                    value = weight_hook(key, value, keep_on_calc_device=(calc_device is not None))
+
+                if not is_target_key(key):
+                    target_device = calc_device if (calc_device is not None and move_to_device) else original_device
+                    value = value.to(target_device)
+                    state_dict[key] = value
+                    continue
+
+                # Move to calculation device
+                if calc_device is not None:
+                    value = value.to(calc_device)
+
+                original_dtype = value.dtype
+                quantized_weight, scale_tensor = quantize_weight(
+                    key, value, fp8_dtype, max_value, min_value, quantization_mode, block_size
+                )
+
+                # Add to state dict using original key for weight and new key for scale
+                fp8_key = key  # Maintain original key
+                scale_key = key.replace(".weight", ".scale_weight")
+                assert fp8_key != scale_key, "FP8 key and scale key must be different"
+
+                if not move_to_device:
+                    quantized_weight = quantized_weight.to(original_device)
+
+                # keep scale shape: [1] or [out,1] or [out, num_blocks, 1]. We can determine the quantization mode from the shape of scale_weight in the patched model.
+                scale_tensor = scale_tensor.to(dtype=original_dtype, device=quantized_weight.device)
+
+                state_dict[fp8_key] = quantized_weight
+                state_dict[scale_key] = scale_tensor
+
+                optimized_count += 1
+
+                if calc_device is not None and optimized_count % 10 == 0:
+                    # free memory on calculation device
+                    clean_memory_on_device(calc_device)
+
+    logger.info(f"Number of optimized Linear layers: {optimized_count}")
+    return state_dict
+
+
+def fp8_linear_forward_patch(self: nn.Linear, x, use_scaled_mm=False, max_value=None):
+    """
+    Patched forward method for Linear layers with FP8 weights.
+
+    Args:
+        self: Linear layer instance
+        x (torch.Tensor): Input tensor
+        use_scaled_mm (bool): Use scaled_mm for FP8 Linear layers, requires SM 8.9+ (RTX 40 series)
+        max_value (float): Maximum value for FP8 quantization. If None, no quantization is applied for input tensor.
+
+    Returns:
+        torch.Tensor: Result of linear transformation
+    """
+    if use_scaled_mm:
+        # **not tested**
+        # _scaled_mm only works for per-tensor scale for now (per-channel scale does not work in certain cases)
+        if self.scale_weight.ndim != 1:
+            raise ValueError("scaled_mm only supports per-tensor scale_weight for now.")
+
+        input_dtype = x.dtype
+        original_weight_dtype = self.scale_weight.dtype
+        target_dtype = self.weight.dtype
+        # assert x.ndim == 3, "Input tensor must be 3D (batch_size, seq_len, hidden_dim)"
+
+        if max_value is None:
+            # no input quantization
+            scale_x = torch.tensor(1.0, dtype=torch.float32, device=x.device)
+        else:
+            # calculate scale factor for input tensor
+            scale_x = (torch.max(torch.abs(x.flatten())) / max_value).to(torch.float32)
+
+            # quantize input tensor to FP8: this seems to consume a lot of memory
+            fp8_max_value = torch.finfo(target_dtype).max
+            fp8_min_value = torch.finfo(target_dtype).min
+            x = quantize_fp8(x, scale_x, target_dtype, fp8_max_value, fp8_min_value)
+
+        original_shape = x.shape
+        x = x.reshape(-1, x.shape[-1]).to(target_dtype)
+
+        weight = self.weight.t()
+        scale_weight = self.scale_weight.to(torch.float32)
+
+        if self.bias is not None:
+            # float32 is not supported with bias in scaled_mm
+            o = torch._scaled_mm(x, weight, out_dtype=original_weight_dtype, bias=self.bias, scale_a=scale_x, scale_b=scale_weight)
+        else:
+            o = torch._scaled_mm(x, weight, out_dtype=input_dtype, scale_a=scale_x, scale_b=scale_weight)
+
+        o = o.reshape(original_shape[0], original_shape[1], -1) if x.ndim == 3 else o.reshape(original_shape[0], -1)
+        return o.to(input_dtype)
+
+    else:
+        # Dequantize the weight
+        original_dtype = self.scale_weight.dtype
+        if self.scale_weight.ndim < 3:
+            # per-tensor or per-channel quantization, we can broadcast
+            dequantized_weight = self.weight.to(original_dtype) * self.scale_weight
+        else:
+            # block-wise quantization, need to reshape weight to match scale shape for broadcasting
+            out_features, num_blocks, _ = self.scale_weight.shape
+            dequantized_weight = self.weight.to(original_dtype).contiguous().view(out_features, num_blocks, -1)
+            dequantized_weight = dequantized_weight * self.scale_weight
+            dequantized_weight = dequantized_weight.view(self.weight.shape)
+
+        # Perform linear transformation
+        if self.bias is not None:
+            output = F.linear(x, dequantized_weight, self.bias)
+        else:
+            output = F.linear(x, dequantized_weight)
+
+        return output
+
+
+def apply_fp8_monkey_patch(model, optimized_state_dict, use_scaled_mm=False):
+    """
+    Apply monkey patching to a model using FP8 optimized state dict.
+
+    Args:
+        model (nn.Module): Model instance to patch
+        optimized_state_dict (dict): FP8 optimized state dict
+        use_scaled_mm (bool): Use scaled_mm for FP8 Linear layers, requires SM 8.9+ (RTX 40 series)
+
+    Returns:
+        nn.Module: The patched model (same instance, modified in-place)
+    """
+    # # Calculate FP8 float8_e5m2 max value
+    # max_value = calculate_fp8_maxval(5, 2)
+    max_value = None  # do not quantize input tensor
+
+    # Find all scale keys to identify FP8-optimized layers
+    scale_keys = [k for k in optimized_state_dict.keys() if k.endswith(".scale_weight")]
+
+    # Enumerate patched layers
+    patched_module_paths = set()
+    scale_shape_info = {}
+    for scale_key in scale_keys:
+        # Extract module path from scale key (remove .scale_weight)
+        module_path = scale_key.rsplit(".scale_weight", 1)[0]
+        patched_module_paths.add(module_path)
+
+        # Store scale shape information
+        scale_shape_info[module_path] = optimized_state_dict[scale_key].shape
+
+    patched_count = 0
+
+    # Apply monkey patch to each layer with FP8 weights
+    for name, module in model.named_modules():
+        # Check if this module has a corresponding scale_weight
+        has_scale = name in patched_module_paths
+
+        # Apply patch if it's a Linear layer with FP8 scale
+        if isinstance(module, nn.Linear) and has_scale:
+            # register the scale_weight as a buffer to load the state_dict
+            # module.register_buffer("scale_weight", torch.tensor(1.0, dtype=module.weight.dtype))
+            scale_shape = scale_shape_info[name]
+            module.register_buffer("scale_weight", torch.ones(scale_shape, dtype=module.weight.dtype))
+
+            # Create a new forward method with the patched version.
+            def new_forward(self, x):
+                return fp8_linear_forward_patch(self, x, use_scaled_mm, max_value)
+
+            # Bind method to module
+            module.forward = new_forward.__get__(module, type(module))
+
+            patched_count += 1
+
+    logger.info(f"Number of monkey-patched Linear layers: {patched_count}")
+    return model

library/huggingface_util.py

@@ -4,7 +4,10 @@ from pathlib import Path
 import argparse
 import os
 from library.utils import fire_in_thread
-
+from library.utils import setup_logging
+setup_logging()
+import logging
+logger = logging.getLogger(__name__)
 
 def exists_repo(repo_id: str, repo_type: str, revision: str = "main", token: str = None):
     api = HfApi(
@@ -33,9 +36,9 @@ def upload(
         try:
             api.create_repo(repo_id=repo_id, repo_type=repo_type, private=private)
         except Exception as e:  # とりあえずRepositoryNotFoundErrorは確認したが他にあると困るので
-            print("===========================================")
-            print(f"failed to create HuggingFace repo / HuggingFaceのリポジトリの作成に失敗しました : {e}")
-            print("===========================================")
+            logger.error("===========================================")
+            logger.error(f"failed to create HuggingFace repo / HuggingFaceのリポジトリの作成に失敗しました : {e}")
+            logger.error("===========================================")
 
     is_folder = (type(src) == str and os.path.isdir(src)) or (isinstance(src, Path) and src.is_dir())
 
@@ -56,9 +59,9 @@ def upload(
                     path_in_repo=path_in_repo,
                 )
         except Exception as e:  # RuntimeErrorを確認済みだが他にあると困るので
-            print("===========================================")
-            print(f"failed to upload to HuggingFace / HuggingFaceへのアップロードに失敗しました : {e}")
-            print("===========================================")
+            logger.error("===========================================")
+            logger.error(f"failed to upload to HuggingFace / HuggingFaceへのアップロードに失敗しました : {e}")
+            logger.error("===========================================")
 
     if args.async_upload and not force_sync_upload:
         fire_in_thread(uploader)

library/hunyuan_image_models.py

@@ -0,0 +1,489 @@
+# Original work: https://github.com/Tencent-Hunyuan/HunyuanImage-2.1
+# Re-implemented for license compliance for sd-scripts.
+
+from typing import Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+from accelerate import init_empty_weights
+
+from library import custom_offloading_utils
+from library.attention import AttentionParams
+from library.fp8_optimization_utils import apply_fp8_monkey_patch
+from library.lora_utils import load_safetensors_with_lora_and_fp8
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+from library.hunyuan_image_modules import (
+    SingleTokenRefiner,
+    ByT5Mapper,
+    PatchEmbed2D,
+    TimestepEmbedder,
+    MMDoubleStreamBlock,
+    MMSingleStreamBlock,
+    FinalLayer,
+)
+from library.hunyuan_image_utils import get_nd_rotary_pos_embed
+
+FP8_OPTIMIZATION_TARGET_KEYS = ["double_blocks", "single_blocks"]
+# FP8_OPTIMIZATION_EXCLUDE_KEYS = ["norm", "_mod", "_emb"]  # , "modulation"
+FP8_OPTIMIZATION_EXCLUDE_KEYS = ["norm", "_emb"]  # , "modulation", "_mod"
+
+# full exclude 24.2GB
+# norm and _emb 19.7GB
+# fp8 cast 19.7GB
+
+
+# region DiT Model
+class HYImageDiffusionTransformer(nn.Module):
+    """
+    HunyuanImage-2.1 Diffusion Transformer.
+
+    A multimodal transformer for image generation with text conditioning,
+    featuring separate double-stream and single-stream processing blocks.
+
+    Args:
+        attn_mode: Attention implementation mode ("torch" or "sageattn").
+    """
+
+    def __init__(self, attn_mode: str = "torch", split_attn: bool = False):
+        super().__init__()
+
+        # Fixed architecture parameters for HunyuanImage-2.1
+        self.patch_size = [1, 1]  # 1x1 patch size (no spatial downsampling)
+        self.in_channels = 64  # Input latent channels
+        self.out_channels = 64  # Output latent channels
+        self.unpatchify_channels = self.out_channels
+        self.guidance_embed = False  # Guidance embedding disabled
+        self.rope_dim_list = [64, 64]  # RoPE dimensions for 2D positional encoding
+        self.rope_theta = 256  # RoPE frequency scaling
+        self.use_attention_mask = True
+        self.text_projection = "single_refiner"
+        self.hidden_size = 3584  # Model dimension
+        self.heads_num = 28  # Number of attention heads
+
+        # Architecture configuration
+        mm_double_blocks_depth = 20  # Double-stream transformer blocks
+        mm_single_blocks_depth = 40  # Single-stream transformer blocks
+        mlp_width_ratio = 4  # MLP expansion ratio
+        text_states_dim = 3584  # Text encoder output dimension
+        guidance_embed = False  # No guidance embedding
+
+        # Layer configuration
+        mlp_act_type: str = "gelu_tanh"  # MLP activation function
+        qkv_bias: bool = True  # Use bias in QKV projections
+        qk_norm: bool = True  # Apply QK normalization
+        qk_norm_type: str = "rms"  # RMS normalization type
+
+        self.attn_mode = attn_mode
+        self.split_attn = split_attn
+
+        # ByT5 character-level text encoder mapping
+        self.byt5_in = ByT5Mapper(in_dim=1472, out_dim=2048, hidden_dim=2048, out_dim1=self.hidden_size, use_residual=False)
+
+        # Image latent patch embedding
+        self.img_in = PatchEmbed2D(self.patch_size, self.in_channels, self.hidden_size)
+
+        # Text token refinement with cross-attention
+        self.txt_in = SingleTokenRefiner(text_states_dim, self.hidden_size, self.heads_num, depth=2)
+
+        # Timestep embedding for diffusion process
+        self.time_in = TimestepEmbedder(self.hidden_size, nn.SiLU)
+
+        # MeanFlow not supported in this implementation
+        self.time_r_in = None
+
+        # Guidance embedding (disabled for non-distilled model)
+        self.guidance_in = TimestepEmbedder(self.hidden_size, nn.SiLU) if guidance_embed else None
+
+        # Double-stream blocks: separate image and text processing
+        self.double_blocks = nn.ModuleList(
+            [
+                MMDoubleStreamBlock(
+                    self.hidden_size,
+                    self.heads_num,
+                    mlp_width_ratio=mlp_width_ratio,
+                    mlp_act_type=mlp_act_type,
+                    qk_norm=qk_norm,
+                    qk_norm_type=qk_norm_type,
+                    qkv_bias=qkv_bias,
+                )
+                for _ in range(mm_double_blocks_depth)
+            ]
+        )
+
+        # Single-stream blocks: joint processing of concatenated features
+        self.single_blocks = nn.ModuleList(
+            [
+                MMSingleStreamBlock(
+                    self.hidden_size,
+                    self.heads_num,
+                    mlp_width_ratio=mlp_width_ratio,
+                    mlp_act_type=mlp_act_type,
+                    qk_norm=qk_norm,
+                    qk_norm_type=qk_norm_type,
+                )
+                for _ in range(mm_single_blocks_depth)
+            ]
+        )
+
+        self.final_layer = FinalLayer(self.hidden_size, self.patch_size, self.out_channels, nn.SiLU)
+
+        self.gradient_checkpointing = False
+        self.cpu_offload_checkpointing = False
+        self.blocks_to_swap = None
+
+        self.offloader_double = None
+        self.offloader_single = None
+        self.num_double_blocks = len(self.double_blocks)
+        self.num_single_blocks = len(self.single_blocks)
+
+    @property
+    def device(self):
+        return next(self.parameters()).device
+
+    @property
+    def dtype(self):
+        return next(self.parameters()).dtype
+
+    def enable_gradient_checkpointing(self, cpu_offload: bool = False):
+        self.gradient_checkpointing = True
+        self.cpu_offload_checkpointing = cpu_offload
+
+        for block in self.double_blocks + self.single_blocks:
+            block.enable_gradient_checkpointing(cpu_offload=cpu_offload)
+
+        print(f"HunyuanImage-2.1: Gradient checkpointing enabled. CPU offload: {cpu_offload}")
+
+    def disable_gradient_checkpointing(self):
+        self.gradient_checkpointing = False
+        self.cpu_offload_checkpointing = False
+
+        for block in self.double_blocks + self.single_blocks:
+            block.disable_gradient_checkpointing()
+
+        print("HunyuanImage-2.1: Gradient checkpointing disabled.")
+
+    def enable_block_swap(self, num_blocks: int, device: torch.device, supports_backward: bool = False):
+        self.blocks_to_swap = num_blocks
+        double_blocks_to_swap = num_blocks // 2
+        single_blocks_to_swap = (num_blocks - double_blocks_to_swap) * 2
+
+        assert double_blocks_to_swap <= self.num_double_blocks - 2 and single_blocks_to_swap <= self.num_single_blocks - 2, (
+            f"Cannot swap more than {self.num_double_blocks - 2} double blocks and {self.num_single_blocks - 2} single blocks. "
+            f"Requested {double_blocks_to_swap} double blocks and {single_blocks_to_swap} single blocks."
+        )
+
+        self.offloader_double = custom_offloading_utils.ModelOffloader(
+            self.double_blocks, double_blocks_to_swap, device, supports_backward=supports_backward
+        )
+        self.offloader_single = custom_offloading_utils.ModelOffloader(
+            self.single_blocks, single_blocks_to_swap, device, supports_backward=supports_backward
+        )
+        # , debug=True
+        print(
+            f"HunyuanImage-2.1: Block swap enabled. Swapping {num_blocks} blocks, double blocks: {double_blocks_to_swap}, single blocks: {single_blocks_to_swap}."
+        )
+
+    def switch_block_swap_for_inference(self):
+        if self.blocks_to_swap:
+            self.offloader_double.set_forward_only(True)
+            self.offloader_single.set_forward_only(True)
+            self.prepare_block_swap_before_forward()
+            print(f"HunyuanImage-2.1: Block swap set to forward only.")
+
+    def switch_block_swap_for_training(self):
+        if self.blocks_to_swap:
+            self.offloader_double.set_forward_only(False)
+            self.offloader_single.set_forward_only(False)
+            self.prepare_block_swap_before_forward()
+            print(f"HunyuanImage-2.1: Block swap set to forward and backward.")
+
+    def move_to_device_except_swap_blocks(self, device: torch.device):
+        # assume model is on cpu. do not move blocks to device to reduce temporary memory usage
+        if self.blocks_to_swap:
+            save_double_blocks = self.double_blocks
+            save_single_blocks = self.single_blocks
+            self.double_blocks = nn.ModuleList()
+            self.single_blocks = nn.ModuleList()
+
+        self.to(device)
+
+        if self.blocks_to_swap:
+            self.double_blocks = save_double_blocks
+            self.single_blocks = save_single_blocks
+
+    def prepare_block_swap_before_forward(self):
+        if self.blocks_to_swap is None or self.blocks_to_swap == 0:
+            return
+        self.offloader_double.prepare_block_devices_before_forward(self.double_blocks)
+        self.offloader_single.prepare_block_devices_before_forward(self.single_blocks)
+
+    def get_rotary_pos_embed(self, rope_sizes):
+        """
+        Generate 2D rotary position embeddings for image tokens.
+
+        Args:
+            rope_sizes: Tuple of (height, width) for spatial dimensions.
+
+        Returns:
+            Tuple of (freqs_cos, freqs_sin) tensors for rotary position encoding.
+        """
+        freqs_cos, freqs_sin = get_nd_rotary_pos_embed(self.rope_dim_list, rope_sizes, theta=self.rope_theta)
+        return freqs_cos, freqs_sin
+
+    def reorder_txt_token(
+        self, byt5_txt: torch.Tensor, txt: torch.Tensor, byt5_text_mask: torch.Tensor, text_mask: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, list[int]]:
+        """
+        Combine and reorder ByT5 character-level and word-level text embeddings.
+
+        Concatenates valid tokens from both encoders and creates appropriate masks.
+
+        Args:
+            byt5_txt: ByT5 character-level embeddings [B, L1, D].
+            txt: Word-level text embeddings [B, L2, D].
+            byt5_text_mask: Valid token mask for ByT5 [B, L1].
+            text_mask: Valid token mask for word tokens [B, L2].
+
+        Returns:
+            Tuple of (reordered_embeddings, combined_mask, sequence_lengths).
+        """
+        # Process each batch element separately to handle variable sequence lengths
+
+        reorder_txt = []
+        reorder_mask = []
+
+        txt_lens = []
+        for i in range(text_mask.shape[0]):
+            byt5_text_mask_i = byt5_text_mask[i].bool()
+            text_mask_i = text_mask[i].bool()
+            byt5_text_length = byt5_text_mask_i.sum()
+            text_length = text_mask_i.sum()
+            assert byt5_text_length == byt5_text_mask_i[:byt5_text_length].sum()
+            assert text_length == text_mask_i[:text_length].sum()
+
+            byt5_txt_i = byt5_txt[i]
+            txt_i = txt[i]
+            reorder_txt_i = torch.cat(
+                [byt5_txt_i[:byt5_text_length], txt_i[:text_length], byt5_txt_i[byt5_text_length:], txt_i[text_length:]], dim=0
+            )
+
+            reorder_mask_i = torch.zeros(
+                byt5_text_mask_i.shape[0] + text_mask_i.shape[0], dtype=torch.bool, device=byt5_text_mask_i.device
+            )
+            reorder_mask_i[: byt5_text_length + text_length] = True
+
+            reorder_txt.append(reorder_txt_i)
+            reorder_mask.append(reorder_mask_i)
+            txt_lens.append(byt5_text_length + text_length)
+
+        reorder_txt = torch.stack(reorder_txt)
+        reorder_mask = torch.stack(reorder_mask).to(dtype=torch.int64)
+
+        return reorder_txt, reorder_mask, txt_lens
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        text_states: torch.Tensor,
+        encoder_attention_mask: torch.Tensor,
+        byt5_text_states: Optional[torch.Tensor] = None,
+        byt5_text_mask: Optional[torch.Tensor] = None,
+        rotary_pos_emb_cache: Optional[Dict[Tuple[int, int], Tuple[torch.Tensor, torch.Tensor]]] = None,
+    ) -> torch.Tensor:
+        """
+        Forward pass through the HunyuanImage diffusion transformer.
+
+        Args:
+            hidden_states: Input image latents [B, C, H, W].
+            timestep: Diffusion timestep [B].
+            text_states: Word-level text embeddings [B, L, D].
+            encoder_attention_mask: Text attention mask [B, L].
+            byt5_text_states: ByT5 character-level embeddings [B, L_byt5, D_byt5].
+            byt5_text_mask: ByT5 attention mask [B, L_byt5].
+
+        Returns:
+            Tuple of (denoised_image, spatial_shape).
+        """
+        img = x = hidden_states
+        text_mask = encoder_attention_mask
+        t = timestep
+        txt = text_states
+
+        # Calculate spatial dimensions for rotary position embeddings
+        _, _, oh, ow = x.shape
+        th, tw = oh, ow  # Height and width (patch_size=[1,1] means no spatial downsampling)
+        if rotary_pos_emb_cache is not None:
+            if (th, tw) in rotary_pos_emb_cache:
+                freqs_cis = rotary_pos_emb_cache[(th, tw)]
+                freqs_cis = (freqs_cis[0].to(img.device), freqs_cis[1].to(img.device))
+            else:
+                freqs_cis = self.get_rotary_pos_embed((th, tw))
+                rotary_pos_emb_cache[(th, tw)] = (freqs_cis[0].cpu(), freqs_cis[1].cpu())
+        else:
+            freqs_cis = self.get_rotary_pos_embed((th, tw))
+
+        # Reshape image latents to sequence format: [B, C, H, W] -> [B, H*W, C]
+        img = self.img_in(img)
+
+        # Generate timestep conditioning vector
+        vec = self.time_in(t)
+
+        # MeanFlow and guidance embedding not used in this configuration
+
+        # Process text tokens through refinement layers
+        txt_attn_params = AttentionParams.create_attention_params_from_mask(self.attn_mode, self.split_attn, 0, text_mask)
+        txt = self.txt_in(txt, t, txt_attn_params)
+
+        # Integrate character-level ByT5 features with word-level tokens
+        # Use variable length sequences with sequence lengths
+        byt5_txt = self.byt5_in(byt5_text_states)
+        txt, text_mask, txt_lens = self.reorder_txt_token(byt5_txt, txt, byt5_text_mask, text_mask)
+
+        # Trim sequences to maximum length in the batch
+        img_seq_len = img.shape[1]
+        max_txt_len = max(txt_lens)
+        txt = txt[:, :max_txt_len, :]
+        text_mask = text_mask[:, :max_txt_len]
+
+        attn_params = AttentionParams.create_attention_params_from_mask(self.attn_mode, self.split_attn, img_seq_len, text_mask)
+
+        input_device = img.device
+
+        # Process through double-stream blocks (separate image/text attention)
+        for index, block in enumerate(self.double_blocks):
+            if self.blocks_to_swap:
+                self.offloader_double.wait_for_block(index)
+            img, txt = block(img, txt, vec, freqs_cis, attn_params)
+            if self.blocks_to_swap:
+                self.offloader_double.submit_move_blocks(self.double_blocks, index)
+
+        # Concatenate image and text tokens for joint processing
+        x = torch.cat((img, txt), 1)
+
+        # Process through single-stream blocks (joint attention)
+        for index, block in enumerate(self.single_blocks):
+            if self.blocks_to_swap:
+                self.offloader_single.wait_for_block(index)
+            x = block(x, vec, freqs_cis, attn_params)
+            if self.blocks_to_swap:
+                self.offloader_single.submit_move_blocks(self.single_blocks, index)
+
+        x = x.to(input_device)
+        vec = vec.to(input_device)
+
+        img = x[:, :img_seq_len, ...]
+        del x
+
+        # Apply final projection to output space
+        img = self.final_layer(img, vec)
+        del vec
+
+        # Reshape from sequence to spatial format: [B, L, C] -> [B, C, H, W]
+        img = self.unpatchify_2d(img, th, tw)
+        return img
+
+    def unpatchify_2d(self, x, h, w):
+        """
+        Convert sequence format back to spatial image format.
+
+        Args:
+            x: Input tensor [B, H*W, C].
+            h: Height dimension.
+            w: Width dimension.
+
+        Returns:
+            Spatial tensor [B, C, H, W].
+        """
+        c = self.unpatchify_channels
+
+        x = x.reshape(shape=(x.shape[0], h, w, c))
+        imgs = x.permute(0, 3, 1, 2)
+        return imgs
+
+
+# endregion
+
+# region Model Utils
+
+
+def create_model(attn_mode: str, split_attn: bool, dtype: Optional[torch.dtype]) -> HYImageDiffusionTransformer:
+    with init_empty_weights():
+        model = HYImageDiffusionTransformer(attn_mode=attn_mode, split_attn=split_attn)
+        if dtype is not None:
+            model.to(dtype)
+    return model
+
+
+def load_hunyuan_image_model(
+    device: Union[str, torch.device],
+    dit_path: str,
+    attn_mode: str,
+    split_attn: bool,
+    loading_device: Union[str, torch.device],
+    dit_weight_dtype: Optional[torch.dtype],
+    fp8_scaled: bool = False,
+    lora_weights_list: Optional[Dict[str, torch.Tensor]] = None,
+    lora_multipliers: Optional[list[float]] = None,
+) -> HYImageDiffusionTransformer:
+    """
+    Load a HunyuanImage model from the specified checkpoint.
+
+    Args:
+        device (Union[str, torch.device]): Device for optimization or merging
+        dit_path (str): Path to the DiT model checkpoint.
+        attn_mode (str): Attention mode to use, e.g., "torch", "flash", etc.
+        split_attn (bool): Whether to use split attention.
+        loading_device (Union[str, torch.device]): Device to load the model weights on.
+        dit_weight_dtype (Optional[torch.dtype]): Data type of the DiT weights.
+            If None, it will be loaded as is (same as the state_dict) or scaled for fp8. if not None, model weights will be casted to this dtype.
+        fp8_scaled (bool): Whether to use fp8 scaling for the model weights.
+        lora_weights_list (Optional[Dict[str, torch.Tensor]]): LoRA weights to apply, if any.
+        lora_multipliers (Optional[List[float]]): LoRA multipliers for the weights, if any.
+    """
+    # dit_weight_dtype is None for fp8_scaled
+    assert (not fp8_scaled and dit_weight_dtype is not None) or (fp8_scaled and dit_weight_dtype is None)
+
+    device = torch.device(device)
+    loading_device = torch.device(loading_device)
+
+    model = create_model(attn_mode, split_attn, dit_weight_dtype)
+
+    # load model weights with dynamic fp8 optimization and LoRA merging if needed
+    logger.info(f"Loading DiT model from {dit_path}, device={loading_device}")
+
+    sd = load_safetensors_with_lora_and_fp8(
+        model_files=dit_path,
+        lora_weights_list=lora_weights_list,
+        lora_multipliers=lora_multipliers,
+        fp8_optimization=fp8_scaled,
+        calc_device=device,
+        move_to_device=(loading_device == device),
+        dit_weight_dtype=dit_weight_dtype,
+        target_keys=FP8_OPTIMIZATION_TARGET_KEYS,
+        exclude_keys=FP8_OPTIMIZATION_EXCLUDE_KEYS,
+    )
+
+    if fp8_scaled:
+        apply_fp8_monkey_patch(model, sd, use_scaled_mm=False)
+
+        if loading_device.type != "cpu":
+            # make sure all the model weights are on the loading_device
+            logger.info(f"Moving weights to {loading_device}")
+            for key in sd.keys():
+                sd[key] = sd[key].to(loading_device)
+
+    info = model.load_state_dict(sd, strict=True, assign=True)
+    logger.info(f"Loaded DiT model from {dit_path}, info={info}")
+
+    return model
+
+
+# endregion

library/hunyuan_image_modules.py

@@ -0,0 +1,863 @@
+# Original work: https://github.com/Tencent-Hunyuan/HunyuanImage-2.1
+# Re-implemented for license compliance for sd-scripts.
+
+from typing import Tuple, Callable
+import torch
+import torch.nn as nn
+from einops import rearrange
+
+from library import custom_offloading_utils
+from library.attention import AttentionParams, attention
+from library.hunyuan_image_utils import timestep_embedding, apply_rotary_emb, _to_tuple, apply_gate, modulate
+from library.attention import attention
+
+# region Modules
+
+
+class ByT5Mapper(nn.Module):
+    """
+    Maps ByT5 character-level encoder outputs to transformer hidden space.
+
+    Applies layer normalization, two MLP layers with GELU activation,
+    and optional residual connection.
+
+    Args:
+        in_dim: Input dimension from ByT5 encoder (1472 for ByT5-large).
+        out_dim: Intermediate dimension after first projection.
+        hidden_dim: Hidden dimension for MLP layer.
+        out_dim1: Final output dimension matching transformer hidden size.
+        use_residual: Whether to add residual connection (requires in_dim == out_dim).
+    """
+
+    def __init__(self, in_dim, out_dim, hidden_dim, out_dim1, use_residual=True):
+        super().__init__()
+        if use_residual:
+            assert in_dim == out_dim
+        self.layernorm = nn.LayerNorm(in_dim)
+        self.fc1 = nn.Linear(in_dim, hidden_dim)
+        self.fc2 = nn.Linear(hidden_dim, out_dim)
+        self.fc3 = nn.Linear(out_dim, out_dim1)
+        self.use_residual = use_residual
+        self.act_fn = nn.GELU()
+
+    def forward(self, x):
+        """
+        Transform ByT5 embeddings to transformer space.
+
+        Args:
+            x: Input ByT5 embeddings [..., in_dim].
+
+        Returns:
+            Transformed embeddings [..., out_dim1].
+        """
+        residual = x if self.use_residual else None
+        x = self.layernorm(x)
+        x = self.fc1(x)
+        x = self.act_fn(x)
+        x = self.fc2(x)
+        x = self.act_fn(x)
+        x = self.fc3(x)
+        if self.use_residual:
+            x = x + residual
+        return x
+
+
+class PatchEmbed2D(nn.Module):
+    """
+    2D patch embedding layer for converting image latents to transformer tokens.
+
+    Uses 2D convolution to project image patches to embedding space.
+    For HunyuanImage-2.1, patch_size=[1,1] means no spatial downsampling.
+
+    Args:
+        patch_size: Spatial size of patches (int or tuple).
+        in_chans: Number of input channels.
+        embed_dim: Output embedding dimension.
+    """
+
+    def __init__(self, patch_size=16, in_chans=3, embed_dim=768):
+        super().__init__()
+        self.patch_size = tuple(patch_size)
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=self.patch_size, stride=self.patch_size, bias=True)
+        self.norm = nn.Identity()  # No normalization layer used
+
+    def forward(self, x):
+        x = self.proj(x)
+        x = x.flatten(2).transpose(1, 2)
+        x = self.norm(x)
+        return x
+
+
+class TimestepEmbedder(nn.Module):
+    """
+    Embeds scalar diffusion timesteps into vector representations.
+
+    Uses sinusoidal encoding followed by a two-layer MLP.
+
+    Args:
+        hidden_size: Output embedding dimension.
+        act_layer: Activation function class (e.g., nn.SiLU).
+        frequency_embedding_size: Dimension of sinusoidal encoding.
+        max_period: Maximum period for sinusoidal frequencies.
+        out_size: Output dimension (defaults to hidden_size).
+    """
+
+    def __init__(self, hidden_size, act_layer, frequency_embedding_size=256, max_period=10000, out_size=None):
+        super().__init__()
+        self.frequency_embedding_size = frequency_embedding_size
+        self.max_period = max_period
+        if out_size is None:
+            out_size = hidden_size
+
+        self.mlp = nn.Sequential(
+            nn.Linear(frequency_embedding_size, hidden_size, bias=True), act_layer(), nn.Linear(hidden_size, out_size, bias=True)
+        )
+
+    def forward(self, t):
+        t_freq = timestep_embedding(t, self.frequency_embedding_size, self.max_period).type(self.mlp[0].weight.dtype)
+        return self.mlp(t_freq)
+
+
+class TextProjection(nn.Module):
+    """
+    Projects text embeddings through a two-layer MLP.
+
+    Used for context-aware representation computation in token refinement.
+
+    Args:
+        in_channels: Input feature dimension.
+        hidden_size: Hidden and output dimension.
+        act_layer: Activation function class.
+    """
+
+    def __init__(self, in_channels, hidden_size, act_layer):
+        super().__init__()
+        self.linear_1 = nn.Linear(in_features=in_channels, out_features=hidden_size, bias=True)
+        self.act_1 = act_layer()
+        self.linear_2 = nn.Linear(in_features=hidden_size, out_features=hidden_size, bias=True)
+
+    def forward(self, caption):
+        hidden_states = self.linear_1(caption)
+        hidden_states = self.act_1(hidden_states)
+        hidden_states = self.linear_2(hidden_states)
+        return hidden_states
+
+
+class MLP(nn.Module):
+    """
+    Multi-layer perceptron with configurable activation and normalization.
+
+    Standard two-layer MLP with optional dropout and intermediate normalization.
+
+    Args:
+        in_channels: Input feature dimension.
+        hidden_channels: Hidden layer dimension (defaults to in_channels).
+        out_features: Output dimension (defaults to in_channels).
+        act_layer: Activation function class.
+        norm_layer: Optional normalization layer class.
+        bias: Whether to use bias (can be bool or tuple for each layer).
+        drop: Dropout rate (can be float or tuple for each layer).
+        use_conv: Whether to use convolution instead of linear (not supported).
+    """
+
+    def __init__(
+        self,
+        in_channels,
+        hidden_channels=None,
+        out_features=None,
+        act_layer=nn.GELU,
+        norm_layer=None,
+        bias=True,
+        drop=0.0,
+        use_conv=False,
+    ):
+        super().__init__()
+        assert not use_conv, "Convolutional MLP not supported in this implementation."
+
+        out_features = out_features or in_channels
+        hidden_channels = hidden_channels or in_channels
+        bias = _to_tuple(bias, 2)
+        drop_probs = _to_tuple(drop, 2)
+
+        self.fc1 = nn.Linear(in_channels, hidden_channels, bias=bias[0])
+        self.act = act_layer()
+        self.drop1 = nn.Dropout(drop_probs[0])
+        self.norm = norm_layer(hidden_channels) if norm_layer is not None else nn.Identity()
+        self.fc2 = nn.Linear(hidden_channels, out_features, bias=bias[1])
+        self.drop2 = nn.Dropout(drop_probs[1])
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop1(x)
+        x = self.norm(x)
+        x = self.fc2(x)
+        x = self.drop2(x)
+        return x
+
+
+class IndividualTokenRefinerBlock(nn.Module):
+    """
+    Single transformer block for individual token refinement.
+
+    Applies self-attention and MLP with adaptive layer normalization (AdaLN)
+    conditioned on timestep and context information.
+
+    Args:
+        hidden_size: Model dimension.
+        heads_num: Number of attention heads.
+        mlp_width_ratio: MLP expansion ratio.
+        mlp_drop_rate: MLP dropout rate.
+        act_type: Activation function (only "silu" supported).
+        qk_norm: QK normalization flag (must be False).
+        qk_norm_type: QK normalization type (only "layer" supported).
+        qkv_bias: Use bias in QKV projections.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        heads_num: int,
+        mlp_width_ratio: float = 4.0,
+        mlp_drop_rate: float = 0.0,
+        act_type: str = "silu",
+        qk_norm: bool = False,
+        qk_norm_type: str = "layer",
+        qkv_bias: bool = True,
+    ):
+        super().__init__()
+        assert qk_norm_type == "layer", "Only layer normalization supported for QK norm."
+        assert act_type == "silu", "Only SiLU activation supported."
+        assert not qk_norm, "QK normalization must be disabled."
+
+        self.heads_num = heads_num
+        mlp_hidden_dim = int(hidden_size * mlp_width_ratio)
+
+        self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=True, eps=1e-6)
+        self.self_attn_qkv = nn.Linear(hidden_size, hidden_size * 3, bias=qkv_bias)
+
+        self.self_attn_q_norm = nn.Identity()
+        self.self_attn_k_norm = nn.Identity()
+        self.self_attn_proj = nn.Linear(hidden_size, hidden_size, bias=qkv_bias)
+
+        self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=True, eps=1e-6)
+        self.mlp = MLP(in_channels=hidden_size, hidden_channels=mlp_hidden_dim, act_layer=nn.SiLU, drop=mlp_drop_rate)
+
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(hidden_size, 2 * hidden_size, bias=True),
+        )
+
+    def forward(self, x: torch.Tensor, c: torch.Tensor, attn_params: AttentionParams) -> torch.Tensor:
+        """
+        Apply self-attention and MLP with adaptive conditioning.
+
+        Args:
+            x: Input token embeddings [B, L, C].
+            c: Combined conditioning vector [B, C].
+            attn_params: Attention parameters including sequence lengths.
+
+        Returns:
+            Refined token embeddings [B, L, C].
+        """
+        gate_msa, gate_mlp = self.adaLN_modulation(c).chunk(2, dim=1)
+        norm_x = self.norm1(x)
+        qkv = self.self_attn_qkv(norm_x)
+        del norm_x
+        q, k, v = rearrange(qkv, "B L (K H D) -> K B L H D", K=3, H=self.heads_num)
+        del qkv
+        q = self.self_attn_q_norm(q).to(v)
+        k = self.self_attn_k_norm(k).to(v)
+        qkv = [q, k, v]
+        del q, k, v
+        attn = attention(qkv, attn_params=attn_params)
+
+        x = x + apply_gate(self.self_attn_proj(attn), gate_msa)
+        x = x + apply_gate(self.mlp(self.norm2(x)), gate_mlp)
+        return x
+
+
+class IndividualTokenRefiner(nn.Module):
+    """
+    Stack of token refinement blocks with self-attention.
+
+    Processes tokens individually with adaptive layer normalization.
+
+    Args:
+        hidden_size: Model dimension.
+        heads_num: Number of attention heads.
+        depth: Number of refinement blocks.
+        mlp_width_ratio: MLP expansion ratio.
+        mlp_drop_rate: MLP dropout rate.
+        act_type: Activation function type.
+        qk_norm: QK normalization flag.
+        qk_norm_type: QK normalization type.
+        qkv_bias: Use bias in QKV projections.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        heads_num: int,
+        depth: int,
+        mlp_width_ratio: float = 4.0,
+        mlp_drop_rate: float = 0.0,
+        act_type: str = "silu",
+        qk_norm: bool = False,
+        qk_norm_type: str = "layer",
+        qkv_bias: bool = True,
+    ):
+        super().__init__()
+        self.blocks = nn.ModuleList(
+            [
+                IndividualTokenRefinerBlock(
+                    hidden_size=hidden_size,
+                    heads_num=heads_num,
+                    mlp_width_ratio=mlp_width_ratio,
+                    mlp_drop_rate=mlp_drop_rate,
+                    act_type=act_type,
+                    qk_norm=qk_norm,
+                    qk_norm_type=qk_norm_type,
+                    qkv_bias=qkv_bias,
+                )
+                for _ in range(depth)
+            ]
+        )
+
+    def forward(self, x: torch.Tensor, c: torch.LongTensor, attn_params: AttentionParams) -> torch.Tensor:
+        """
+        Apply sequential token refinement.
+
+        Args:
+            x: Input token embeddings [B, L, C].
+            c: Combined conditioning vector [B, C].
+            attn_params: Attention parameters including sequence lengths.
+
+        Returns:
+            Refined token embeddings [B, L, C].
+        """
+        for block in self.blocks:
+            x = block(x, c, attn_params)
+        return x
+
+
+class SingleTokenRefiner(nn.Module):
+    """
+    Text embedding refinement with timestep and context conditioning.
+
+    Projects input text embeddings and applies self-attention refinement
+    conditioned on diffusion timestep and aggregate text context.
+
+    Args:
+        in_channels: Input text embedding dimension.
+        hidden_size: Transformer hidden dimension.
+        heads_num: Number of attention heads.
+        depth: Number of refinement blocks.
+    """
+
+    def __init__(self, in_channels: int, hidden_size: int, heads_num: int, depth: int):
+        # Fixed architecture parameters for HunyuanImage-2.1
+        mlp_drop_rate: float = 0.0  # No MLP dropout
+        act_type: str = "silu"  # SiLU activation
+        mlp_width_ratio: float = 4.0  # 4x MLP expansion
+        qk_norm: bool = False  # No QK normalization
+        qk_norm_type: str = "layer"  # Layer norm type (unused)
+        qkv_bias: bool = True  # Use QKV bias
+
+        super().__init__()
+        self.input_embedder = nn.Linear(in_channels, hidden_size, bias=True)
+        act_layer = nn.SiLU
+        self.t_embedder = TimestepEmbedder(hidden_size, act_layer)
+        self.c_embedder = TextProjection(in_channels, hidden_size, act_layer)
+        self.individual_token_refiner = IndividualTokenRefiner(
+            hidden_size=hidden_size,
+            heads_num=heads_num,
+            depth=depth,
+            mlp_width_ratio=mlp_width_ratio,
+            mlp_drop_rate=mlp_drop_rate,
+            act_type=act_type,
+            qk_norm=qk_norm,
+            qk_norm_type=qk_norm_type,
+            qkv_bias=qkv_bias,
+        )
+
+    def forward(self, x: torch.Tensor, t: torch.LongTensor, attn_params: AttentionParams) -> torch.Tensor:
+        """
+        Refine text embeddings with timestep conditioning.
+
+        Args:
+            x: Input text embeddings [B, L, in_channels].
+            t: Diffusion timestep [B].
+            attn_params: Attention parameters including sequence lengths.
+
+        Returns:
+            Refined embeddings [B, L, hidden_size].
+        """
+        timestep_aware_representations = self.t_embedder(t)
+
+        # Compute context-aware representations by averaging valid tokens
+        txt_lens = attn_params.seqlens  # img_len is not used for SingleTokenRefiner
+        context_aware_representations = torch.stack([x[i, : txt_lens[i]].mean(dim=0) for i in range(x.shape[0])], dim=0)  # [B, C]
+
+        context_aware_representations = self.c_embedder(context_aware_representations)
+        c = timestep_aware_representations + context_aware_representations
+        del timestep_aware_representations, context_aware_representations
+        x = self.input_embedder(x)
+        x = self.individual_token_refiner(x, c, attn_params)
+        return x
+
+
+class FinalLayer(nn.Module):
+    """
+    Final output projection layer with adaptive layer normalization.
+
+    Projects transformer hidden states to output patch space with
+    timestep-conditioned modulation.
+
+    Args:
+        hidden_size: Input hidden dimension.
+        patch_size: Spatial patch size for output reshaping.
+        out_channels: Number of output channels.
+        act_layer: Activation function class.
+    """
+
+    def __init__(self, hidden_size, patch_size, out_channels, act_layer):
+        super().__init__()
+
+        # Layer normalization without learnable parameters
+        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        out_size = (patch_size[0] * patch_size[1]) * out_channels
+        self.linear = nn.Linear(hidden_size, out_size, bias=True)
+
+        # Adaptive layer normalization modulation
+        self.adaLN_modulation = nn.Sequential(
+            act_layer(),
+            nn.Linear(hidden_size, 2 * hidden_size, bias=True),
+        )
+
+    def forward(self, x, c):
+        shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
+        x = modulate(self.norm_final(x), shift=shift, scale=scale)
+        del shift, scale, c
+        x = self.linear(x)
+        return x
+
+
+class RMSNorm(nn.Module):
+    """
+    Root Mean Square Layer Normalization.
+
+    Normalizes input using RMS and applies learnable scaling.
+    More efficient than LayerNorm as it doesn't compute mean.
+
+    Args:
+        dim: Input feature dimension.
+        eps: Small value for numerical stability.
+    """
+
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.eps = eps
+        self.weight = nn.Parameter(torch.ones(dim))
+
+    def _norm(self, x):
+        """
+        Apply RMS normalization.
+
+        Args:
+            x: Input tensor.
+
+        Returns:
+            RMS normalized tensor.
+        """
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+    def reset_parameters(self):
+        self.weight.fill_(1)
+
+    def forward(self, x):
+        """
+        Apply RMSNorm with learnable scaling.
+
+        Args:
+            x: Input tensor.
+
+        Returns:
+            Normalized and scaled tensor.
+        """
+        output = self._norm(x.float()).type_as(x)
+        del x
+        # output = output * self.weight
+        # fp8 support
+        output = output * self.weight.to(output.dtype)
+        return output
+
+
+# kept for reference, not used in current implementation
+# class LinearWarpforSingle(nn.Module):
+#     """
+#     Linear layer wrapper for concatenating and projecting two inputs.
+
+#     Used in single-stream blocks to combine attention output with MLP features.
+
+#     Args:
+#         in_dim: Input dimension (sum of both input feature dimensions).
+#         out_dim: Output dimension.
+#         bias: Whether to use bias in linear projection.
+#     """
+
+#     def __init__(self, in_dim: int, out_dim: int, bias=False):
+#         super().__init__()
+#         self.fc = nn.Linear(in_dim, out_dim, bias=bias)
+
+#     def forward(self, x, y):
+#         """Concatenate inputs along feature dimension and project."""
+#         x = torch.cat([x.contiguous(), y.contiguous()], dim=2).contiguous()
+#         return self.fc(x)
+
+
+class ModulateDiT(nn.Module):
+    """
+    Timestep conditioning modulation layer.
+
+    Projects timestep embeddings to multiple modulation parameters
+    for adaptive layer normalization.
+
+    Args:
+        hidden_size: Input conditioning dimension.
+        factor: Number of modulation parameters to generate.
+        act_layer: Activation function class.
+    """
+
+    def __init__(self, hidden_size: int, factor: int, act_layer: Callable):
+        super().__init__()
+        self.act = act_layer()
+        self.linear = nn.Linear(hidden_size, factor * hidden_size, bias=True)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.linear(self.act(x))
+
+
+class MMDoubleStreamBlock(nn.Module):
+    """
+    Multimodal double-stream transformer block.
+
+    Processes image and text tokens separately with cross-modal attention.
+    Each stream has its own normalization and MLP layers but shares
+    attention computation for cross-modal interaction.
+
+    Args:
+        hidden_size: Model dimension.
+        heads_num: Number of attention heads.
+        mlp_width_ratio: MLP expansion ratio.
+        mlp_act_type: MLP activation function (only "gelu_tanh" supported).
+        qk_norm: QK normalization flag (must be True).
+        qk_norm_type: QK normalization type (only "rms" supported).
+        qkv_bias: Use bias in QKV projections.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        heads_num: int,
+        mlp_width_ratio: float,
+        mlp_act_type: str = "gelu_tanh",
+        qk_norm: bool = True,
+        qk_norm_type: str = "rms",
+        qkv_bias: bool = False,
+    ):
+        super().__init__()
+
+        assert mlp_act_type == "gelu_tanh", "Only GELU-tanh activation supported."
+        assert qk_norm_type == "rms", "Only RMS normalization supported."
+        assert qk_norm, "QK normalization must be enabled."
+
+        self.heads_num = heads_num
+        head_dim = hidden_size // heads_num
+        mlp_hidden_dim = int(hidden_size * mlp_width_ratio)
+
+        # Image stream processing components
+        self.img_mod = ModulateDiT(hidden_size, factor=6, act_layer=nn.SiLU)
+        self.img_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+
+        self.img_attn_qkv = nn.Linear(hidden_size, hidden_size * 3, bias=qkv_bias)
+
+        self.img_attn_q_norm = RMSNorm(head_dim, eps=1e-6)
+        self.img_attn_k_norm = RMSNorm(head_dim, eps=1e-6)
+        self.img_attn_proj = nn.Linear(hidden_size, hidden_size, bias=qkv_bias)
+
+        self.img_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.img_mlp = MLP(hidden_size, mlp_hidden_dim, act_layer=lambda: nn.GELU(approximate="tanh"), bias=True)
+
+        # Text stream processing components
+        self.txt_mod = ModulateDiT(hidden_size, factor=6, act_layer=nn.SiLU)
+        self.txt_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+
+        self.txt_attn_qkv = nn.Linear(hidden_size, hidden_size * 3, bias=qkv_bias)
+        self.txt_attn_q_norm = RMSNorm(head_dim, eps=1e-6)
+        self.txt_attn_k_norm = RMSNorm(head_dim, eps=1e-6)
+        self.txt_attn_proj = nn.Linear(hidden_size, hidden_size, bias=qkv_bias)
+
+        self.txt_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.txt_mlp = MLP(hidden_size, mlp_hidden_dim, act_layer=lambda: nn.GELU(approximate="tanh"), bias=True)
+
+        self.gradient_checkpointing = False
+        self.cpu_offload_checkpointing = False
+
+    def enable_gradient_checkpointing(self, cpu_offload: bool = False):
+        self.gradient_checkpointing = True
+        self.cpu_offload_checkpointing = cpu_offload
+
+    def disable_gradient_checkpointing(self):
+        self.gradient_checkpointing = False
+        self.cpu_offload_checkpointing = False
+
+    def _forward(
+        self, img: torch.Tensor, txt: torch.Tensor, vec: torch.Tensor, freqs_cis: tuple = None, attn_params: AttentionParams = None
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Extract modulation parameters for image and text streams
+        (img_mod1_shift, img_mod1_scale, img_mod1_gate, img_mod2_shift, img_mod2_scale, img_mod2_gate) = self.img_mod(vec).chunk(
+            6, dim=-1
+        )
+        (txt_mod1_shift, txt_mod1_scale, txt_mod1_gate, txt_mod2_shift, txt_mod2_scale, txt_mod2_gate) = self.txt_mod(vec).chunk(
+            6, dim=-1
+        )
+
+        # Process image stream for attention
+        img_modulated = self.img_norm1(img)
+        img_modulated = modulate(img_modulated, shift=img_mod1_shift, scale=img_mod1_scale)
+        del img_mod1_shift, img_mod1_scale
+
+        img_qkv = self.img_attn_qkv(img_modulated)
+        del img_modulated
+        img_q, img_k, img_v = img_qkv.chunk(3, dim=-1)
+        del img_qkv
+
+        img_q = rearrange(img_q, "B L (H D) -> B L H D", H=self.heads_num)
+        img_k = rearrange(img_k, "B L (H D) -> B L H D", H=self.heads_num)
+        img_v = rearrange(img_v, "B L (H D) -> B L H D", H=self.heads_num)
+
+        # Apply QK-Norm if enabled
+        img_q = self.img_attn_q_norm(img_q).to(img_v)
+        img_k = self.img_attn_k_norm(img_k).to(img_v)
+
+        # Apply rotary position embeddings to image tokens
+        if freqs_cis is not None:
+            img_q, img_k = apply_rotary_emb(img_q, img_k, freqs_cis, head_first=False)
+            del freqs_cis
+
+        # Process text stream for attention
+        txt_modulated = self.txt_norm1(txt)
+        txt_modulated = modulate(txt_modulated, shift=txt_mod1_shift, scale=txt_mod1_scale)
+
+        txt_qkv = self.txt_attn_qkv(txt_modulated)
+        del txt_modulated
+        txt_q, txt_k, txt_v = txt_qkv.chunk(3, dim=-1)
+        del txt_qkv
+
+        txt_q = rearrange(txt_q, "B L (H D) -> B L H D", H=self.heads_num)
+        txt_k = rearrange(txt_k, "B L (H D) -> B L H D", H=self.heads_num)
+        txt_v = rearrange(txt_v, "B L (H D) -> B L H D", H=self.heads_num)
+
+        # Apply QK-Norm if enabled
+        txt_q = self.txt_attn_q_norm(txt_q).to(txt_v)
+        txt_k = self.txt_attn_k_norm(txt_k).to(txt_v)
+
+        # Concatenate image and text tokens for joint attention
+        img_seq_len = img.shape[1]
+        q = torch.cat([img_q, txt_q], dim=1)
+        del img_q, txt_q
+        k = torch.cat([img_k, txt_k], dim=1)
+        del img_k, txt_k
+        v = torch.cat([img_v, txt_v], dim=1)
+        del img_v, txt_v
+
+        qkv = [q, k, v]
+        del q, k, v
+        attn = attention(qkv, attn_params=attn_params)
+        del qkv
+
+        # Split attention outputs back to separate streams
+        img_attn, txt_attn = (attn[:, :img_seq_len].contiguous(), attn[:, img_seq_len:].contiguous())
+        del attn
+
+        # Apply attention projection and residual connection for image stream
+        img = img + apply_gate(self.img_attn_proj(img_attn), gate=img_mod1_gate)
+        del img_attn, img_mod1_gate
+
+        # Apply MLP and residual connection for image stream
+        img = img + apply_gate(
+            self.img_mlp(modulate(self.img_norm2(img), shift=img_mod2_shift, scale=img_mod2_scale)),
+            gate=img_mod2_gate,
+        )
+        del img_mod2_shift, img_mod2_scale, img_mod2_gate
+
+        # Apply attention projection and residual connection for text stream
+        txt = txt + apply_gate(self.txt_attn_proj(txt_attn), gate=txt_mod1_gate)
+        del txt_attn, txt_mod1_gate
+
+        # Apply MLP and residual connection for text stream
+        txt = txt + apply_gate(
+            self.txt_mlp(modulate(self.txt_norm2(txt), shift=txt_mod2_shift, scale=txt_mod2_scale)),
+            gate=txt_mod2_gate,
+        )
+        del txt_mod2_shift, txt_mod2_scale, txt_mod2_gate
+
+        return img, txt
+
+    def forward(
+        self, img: torch.Tensor, txt: torch.Tensor, vec: torch.Tensor, freqs_cis: tuple = None, attn_params: AttentionParams = None
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        if self.gradient_checkpointing and self.training:
+            forward_fn = self._forward
+            if self.cpu_offload_checkpointing:
+                forward_fn = custom_offloading_utils.cpu_offload_wrapper(forward_fn, self.img_attn_qkv.weight.device)
+
+            return torch.utils.checkpoint.checkpoint(forward_fn, img, txt, vec, freqs_cis, attn_params, use_reentrant=False)
+        else:
+            return self._forward(img, txt, vec, freqs_cis, attn_params)
+
+
+class MMSingleStreamBlock(nn.Module):
+    """
+    Multimodal single-stream transformer block.
+
+    Processes concatenated image and text tokens jointly with shared attention.
+    Uses parallel linear layers for efficiency and applies RoPE only to image tokens.
+
+    Args:
+        hidden_size: Model dimension.
+        heads_num: Number of attention heads.
+        mlp_width_ratio: MLP expansion ratio.
+        mlp_act_type: MLP activation function (only "gelu_tanh" supported).
+        qk_norm: QK normalization flag (must be True).
+        qk_norm_type: QK normalization type (only "rms" supported).
+        qk_scale: Attention scaling factor (computed automatically if None).
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        heads_num: int,
+        mlp_width_ratio: float = 4.0,
+        mlp_act_type: str = "gelu_tanh",
+        qk_norm: bool = True,
+        qk_norm_type: str = "rms",
+        qk_scale: float = None,
+    ):
+        super().__init__()
+
+        assert mlp_act_type == "gelu_tanh", "Only GELU-tanh activation supported."
+        assert qk_norm_type == "rms", "Only RMS normalization supported."
+        assert qk_norm, "QK normalization must be enabled."
+
+        self.hidden_size = hidden_size
+        self.heads_num = heads_num
+        head_dim = hidden_size // heads_num
+        mlp_hidden_dim = int(hidden_size * mlp_width_ratio)
+        self.mlp_hidden_dim = mlp_hidden_dim
+        self.scale = qk_scale or head_dim**-0.5
+
+        # Parallel linear projections for efficiency
+        self.linear1 = nn.Linear(hidden_size, hidden_size * 3 + mlp_hidden_dim)
+
+        # Combined output projection
+        # self.linear2 = LinearWarpforSingle(hidden_size + mlp_hidden_dim, hidden_size, bias=True) # for reference
+        self.linear2 = nn.Linear(hidden_size + mlp_hidden_dim, hidden_size, bias=True)
+
+        # QK normalization layers
+        self.q_norm = RMSNorm(head_dim, eps=1e-6)
+        self.k_norm = RMSNorm(head_dim, eps=1e-6)
+
+        self.pre_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+
+        self.mlp_act = nn.GELU(approximate="tanh")
+        self.modulation = ModulateDiT(hidden_size, factor=3, act_layer=nn.SiLU)
+
+        self.gradient_checkpointing = False
+        self.cpu_offload_checkpointing = False
+
+    def enable_gradient_checkpointing(self, cpu_offload: bool = False):
+        self.gradient_checkpointing = True
+        self.cpu_offload_checkpointing = cpu_offload
+
+    def disable_gradient_checkpointing(self):
+        self.gradient_checkpointing = False
+        self.cpu_offload_checkpointing = False
+
+    def _forward(
+        self,
+        x: torch.Tensor,
+        vec: torch.Tensor,
+        freqs_cis: Tuple[torch.Tensor, torch.Tensor] = None,
+        attn_params: AttentionParams = None,
+    ) -> torch.Tensor:
+        # Extract modulation parameters
+        mod_shift, mod_scale, mod_gate = self.modulation(vec).chunk(3, dim=-1)
+        x_mod = modulate(self.pre_norm(x), shift=mod_shift, scale=mod_scale)
+
+        # Compute Q, K, V, and MLP input
+        qkv_mlp = self.linear1(x_mod)
+        del x_mod
+        q, k, v, mlp = qkv_mlp.split([self.hidden_size, self.hidden_size, self.hidden_size, self.mlp_hidden_dim], dim=-1)
+        del qkv_mlp
+
+        q = rearrange(q, "B L (H D) -> B L H D", H=self.heads_num)
+        k = rearrange(k, "B L (H D) -> B L H D", H=self.heads_num)
+        v = rearrange(v, "B L (H D) -> B L H D", H=self.heads_num)
+
+        # Apply QK-Norm if enabled
+        q = self.q_norm(q).to(v)
+        k = self.k_norm(k).to(v)
+
+        # Separate image and text tokens
+        img_q, txt_q = q[:, : attn_params.img_len, :, :], q[:, attn_params.img_len :, :, :]
+        del q
+        img_k, txt_k = k[:, : attn_params.img_len, :, :], k[:, attn_params.img_len :, :, :]
+        del k
+
+        # Apply rotary position embeddings only to image tokens
+        img_q, img_k = apply_rotary_emb(img_q, img_k, freqs_cis, head_first=False)
+        del freqs_cis
+
+        # Recombine and compute joint attention
+        q = torch.cat([img_q, txt_q], dim=1)
+        del img_q, txt_q
+        k = torch.cat([img_k, txt_k], dim=1)
+        del img_k, txt_k
+        # v = torch.cat([img_v, txt_v], dim=1)
+        # del img_v, txt_v
+        qkv = [q, k, v]
+        del q, k, v
+        attn = attention(qkv, attn_params=attn_params)
+        del qkv
+
+        # Combine attention and MLP outputs, apply gating
+        # output = self.linear2(attn, self.mlp_act(mlp))
+
+        mlp = self.mlp_act(mlp)
+        output = torch.cat([attn, mlp], dim=2).contiguous()
+        del attn, mlp
+        output = self.linear2(output)
+
+        return x + apply_gate(output, gate=mod_gate)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        vec: torch.Tensor,
+        freqs_cis: Tuple[torch.Tensor, torch.Tensor] = None,
+        attn_params: AttentionParams = None,
+    ) -> torch.Tensor:
+        if self.gradient_checkpointing and self.training:
+            forward_fn = self._forward
+            if self.cpu_offload_checkpointing:
+                forward_fn = custom_offloading_utils.create_cpu_offloading_wrapper(forward_fn, self.linear1.weight.device)
+
+            return torch.utils.checkpoint.checkpoint(forward_fn, x, vec, freqs_cis, attn_params, use_reentrant=False)
+        else:
+            return self._forward(x, vec, freqs_cis, attn_params)
+
+
+# endregion

library/hunyuan_image_text_encoder.py

@@ -0,0 +1,661 @@
+import json
+import re
+from typing import Tuple, Optional, Union
+import torch
+from transformers import (
+    AutoTokenizer,
+    Qwen2_5_VLConfig,
+    Qwen2_5_VLForConditionalGeneration,
+    Qwen2Tokenizer,
+    T5ForConditionalGeneration,
+    T5Config,
+    T5Tokenizer,
+)
+from transformers.models.t5.modeling_t5 import T5Stack
+from accelerate import init_empty_weights
+
+from library.safetensors_utils import load_safetensors
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+BYT5_TOKENIZER_PATH = "google/byt5-small"
+QWEN_2_5_VL_IMAGE_ID = "Qwen/Qwen2.5-VL-7B-Instruct"
+
+
+# Copy from Glyph-SDXL-V2
+
+COLOR_IDX_JSON = """{"white": 0, "black": 1, "darkslategray": 2, "dimgray": 3, "darkolivegreen": 4, "midnightblue": 5, "saddlebrown": 6, "sienna": 7, "whitesmoke": 8, "darkslateblue": 9, 
+"indianred": 10, "linen": 11, "maroon": 12, "khaki": 13, "sandybrown": 14, "gray": 15, "gainsboro": 16, "teal": 17, "peru": 18, "gold": 19, 
+"snow": 20, "firebrick": 21, "crimson": 22, "chocolate": 23, "tomato": 24, "brown": 25, "goldenrod": 26, "antiquewhite": 27, "rosybrown": 28, "steelblue": 29, 
+"floralwhite": 30, "seashell": 31, "darkgreen": 32, "oldlace": 33, "darkkhaki": 34, "burlywood": 35, "red": 36, "darkgray": 37, "orange": 38, "royalblue": 39, 
+"seagreen": 40, "lightgray": 41, "tan": 42, "coral": 43, "beige": 44, "palevioletred": 45, "wheat": 46, "lavender": 47, "darkcyan": 48, "slateblue": 49, 
+"slategray": 50, "orangered": 51, "silver": 52, "olivedrab": 53, "forestgreen": 54, "darkgoldenrod": 55, "ivory": 56, "darkorange": 57, "yellow": 58, "hotpink": 59, 
+"ghostwhite": 60, "lightcoral": 61, "indigo": 62, "bisque": 63, "darkred": 64, "darksalmon": 65, "lightslategray": 66, "dodgerblue": 67, "lightpink": 68, "mistyrose": 69, 
+"mediumvioletred": 70, "cadetblue": 71, "deeppink": 72, "salmon": 73, "palegoldenrod": 74, "blanchedalmond": 75, "lightseagreen": 76, "cornflowerblue": 77, "yellowgreen": 78, "greenyellow": 79, 
+"navajowhite": 80, "papayawhip": 81, "mediumslateblue": 82, "purple": 83, "blueviolet": 84, "pink": 85, "cornsilk": 86, "lightsalmon": 87, "mediumpurple": 88, "moccasin": 89, 
+"turquoise": 90, "mediumseagreen": 91, "lavenderblush": 92, "mediumblue": 93, "darkseagreen": 94, "mediumturquoise": 95, "paleturquoise": 96, "skyblue": 97, "lemonchiffon": 98, "olive": 99, 
+"peachpuff": 100, "lightyellow": 101, "lightsteelblue": 102, "mediumorchid": 103, "plum": 104, "darkturquoise": 105, "aliceblue": 106, "mediumaquamarine": 107, "orchid": 108, "powderblue": 109, 
+"blue": 110, "darkorchid": 111, "violet": 112, "lightskyblue": 113, "lightcyan": 114, "lightgoldenrodyellow": 115, "navy": 116, "thistle": 117, "honeydew": 118, "mintcream": 119, 
+"lightblue": 120, "darkblue": 121, "darkmagenta": 122, "deepskyblue": 123, "magenta": 124, "limegreen": 125, "darkviolet": 126, "cyan": 127, "palegreen": 128, "aquamarine": 129, 
+"lawngreen": 130, "lightgreen": 131, "azure": 132, "chartreuse": 133, "green": 134, "mediumspringgreen": 135, "lime": 136, "springgreen": 137}"""
+
+MULTILINGUAL_10_LANG_IDX_JSON = """{"en-Montserrat-Regular": 0, "en-Poppins-Italic": 1, "en-GlacialIndifference-Regular": 2, "en-OpenSans-ExtraBoldItalic": 3, "en-Montserrat-Bold": 4, "en-Now-Regular": 5, "en-Garet-Regular": 6, "en-LeagueSpartan-Bold": 7, "en-DMSans-Regular": 8, "en-OpenSauceOne-Regular": 9, 
+"en-OpenSans-ExtraBold": 10, "en-KGPrimaryPenmanship": 11, "en-Anton-Regular": 12, "en-Aileron-BlackItalic": 13, "en-Quicksand-Light": 14, "en-Roboto-BoldItalic": 15, "en-TheSeasons-It": 16, "en-Kollektif": 17, "en-Inter-BoldItalic": 18, "en-Poppins-Medium": 19, 
+"en-Poppins-Light": 20, "en-RoxboroughCF-RegularItalic": 21, "en-PlayfairDisplay-SemiBold": 22, "en-Agrandir-Italic": 23, "en-Lato-Regular": 24, "en-MoreSugarRegular": 25, "en-CanvaSans-RegularItalic": 26, "en-PublicSans-Italic": 27, "en-CodePro-NormalLC": 28, "en-Belleza-Regular": 29, 
+"en-JosefinSans-Bold": 30, "en-HKGrotesk-Bold": 31, "en-Telegraf-Medium": 32, "en-BrittanySignatureRegular": 33, "en-Raleway-ExtraBoldItalic": 34, "en-Mont-RegularItalic": 35, "en-Arimo-BoldItalic": 36, "en-Lora-Italic": 37, "en-ArchivoBlack-Regular": 38, "en-Poppins": 39, 
+"en-Barlow-Black": 40, "en-CormorantGaramond-Bold": 41, "en-LibreBaskerville-Regular": 42, "en-CanvaSchoolFontRegular": 43, "en-BebasNeueBold": 44, "en-LazydogRegular": 45, "en-FredokaOne-Regular": 46, "en-Horizon-Bold": 47, "en-Nourd-Regular": 48, "en-Hatton-Regular": 49, 
+"en-Nunito-ExtraBoldItalic": 50, "en-CerebriSans-Regular": 51, "en-Montserrat-Light": 52, "en-TenorSans": 53, "en-Norwester-Regular": 54, "en-ClearSans-Bold": 55, "en-Cardo-Regular": 56, "en-Alice-Regular": 57, "en-Oswald-Regular": 58, "en-Gaegu-Bold": 59, 
+"en-Muli-Black": 60, "en-TAN-PEARL-Regular": 61, "en-CooperHewitt-Book": 62, "en-Agrandir-Grand": 63, "en-BlackMango-Thin": 64, "en-DMSerifDisplay-Regular": 65, "en-Antonio-Bold": 66, "en-Sniglet-Regular": 67, "en-BeVietnam-Regular": 68, "en-NunitoSans10pt-BlackItalic": 69, 
+"en-AbhayaLibre-ExtraBold": 70, "en-Rubik-Regular": 71, "en-PPNeueMachina-Regular": 72, "en-TAN - MON CHERI-Regular": 73, "en-Jua-Regular": 74, "en-Playlist-Script": 75, "en-SourceSansPro-BoldItalic": 76, "en-MoonTime-Regular": 77, "en-Eczar-ExtraBold": 78, "en-Gatwick-Regular": 79, 
+"en-MonumentExtended-Regular": 80, "en-BarlowSemiCondensed-Regular": 81, "en-BarlowCondensed-Regular": 82, "en-Alegreya-Regular": 83, "en-DreamAvenue": 84, "en-RobotoCondensed-Italic": 85, "en-BobbyJones-Regular": 86, "en-Garet-ExtraBold": 87, "en-YesevaOne-Regular": 88, "en-Dosis-ExtraBold": 89, 
+"en-LeagueGothic-Regular": 90, "en-OpenSans-Italic": 91, "en-TANAEGEAN-Regular": 92, "en-Maharlika-Regular": 93, "en-MarykateRegular": 94, "en-Cinzel-Regular": 95, "en-Agrandir-Wide": 96, "en-Chewy-Regular": 97, "en-BodoniFLF-BoldItalic": 98, "en-Nunito-BlackItalic": 99, 
+"en-LilitaOne": 100, "en-HandyCasualCondensed-Regular": 101, "en-Ovo": 102, "en-Livvic-Regular": 103, "en-Agrandir-Narrow": 104, "en-CrimsonPro-Italic": 105, "en-AnonymousPro-Bold": 106, "en-NF-OneLittleFont-Bold": 107, "en-RedHatDisplay-BoldItalic": 108, "en-CodecPro-Regular": 109, 
+"en-HalimunRegular": 110, "en-LibreFranklin-Black": 111, "en-TeXGyreTermes-BoldItalic": 112, "en-Shrikhand-Regular": 113, "en-TTNormsPro-Italic": 114, "en-Gagalin-Regular": 115, "en-OpenSans-Bold": 116, "en-GreatVibes-Regular": 117, "en-Breathing": 118, "en-HeroLight-Regular": 119, 
+"en-KGPrimaryDots": 120, "en-Quicksand-Bold": 121, "en-Brice-ExtraLightSemiExpanded": 122, "en-Lato-BoldItalic": 123, "en-Fraunces9pt-Italic": 124, "en-AbrilFatface-Regular": 125, "en-BerkshireSwash-Regular": 126, "en-Atma-Bold": 127, "en-HolidayRegular": 128, "en-BebasNeueCyrillic": 129, 
+"en-IntroRust-Base": 130, "en-Gistesy": 131, "en-BDScript-Regular": 132, "en-ApricotsRegular": 133, "en-Prompt-Black": 134, "en-TAN MERINGUE": 135, "en-Sukar Regular": 136, "en-GentySans-Regular": 137, "en-NeueEinstellung-Normal": 138, "en-Garet-Bold": 139, 
+"en-FiraSans-Black": 140, "en-BantayogLight": 141, "en-NotoSerifDisplay-Black": 142, "en-TTChocolates-Regular": 143, "en-Ubuntu-Regular": 144, "en-Assistant-Bold": 145, "en-ABeeZee-Regular": 146, "en-LexendDeca-Regular": 147, "en-KingredSerif": 148, "en-Radley-Regular": 149, 
+"en-BrownSugar": 150, "en-MigraItalic-ExtraboldItalic": 151, "en-ChildosArabic-Regular": 152, "en-PeaceSans": 153, "en-LondrinaSolid-Black": 154, "en-SpaceMono-BoldItalic": 155, "en-RobotoMono-Light": 156, "en-CourierPrime-Regular": 157, "en-Alata-Regular": 158, "en-Amsterdam-One": 159, 
+"en-IreneFlorentina-Regular": 160, "en-CatchyMager": 161, "en-Alta_regular": 162, "en-ArticulatCF-Regular": 163, "en-Raleway-Regular": 164, "en-BrasikaDisplay": 165, "en-TANAngleton-Italic": 166, "en-NotoSerifDisplay-ExtraCondensedItalic": 167, "en-Bryndan Write": 168, "en-TTCommonsPro-It": 169, 
+"en-AlexBrush-Regular": 170, "en-Antic-Regular": 171, "en-TTHoves-Bold": 172, "en-DroidSerif": 173, "en-AblationRegular": 174, "en-Marcellus-Regular": 175, "en-Sanchez-Italic": 176, "en-JosefinSans": 177, "en-Afrah-Regular": 178, "en-PinyonScript": 179, 
+"en-TTInterphases-BoldItalic": 180, "en-Yellowtail-Regular": 181, "en-Gliker-Regular": 182, "en-BobbyJonesSoft-Regular": 183, "en-IBMPlexSans": 184, "en-Amsterdam-Three": 185, "en-Amsterdam-FourSlant": 186, "en-TTFors-Regular": 187, "en-Quattrocento": 188, "en-Sifonn-Basic": 189, 
+"en-AlegreyaSans-Black": 190, "en-Daydream": 191, "en-AristotelicaProTx-Rg": 192, "en-NotoSerif": 193, "en-EBGaramond-Italic": 194, "en-HammersmithOne-Regular": 195, "en-RobotoSlab-Regular": 196, "en-DO-Sans-Regular": 197, "en-KGPrimaryDotsLined": 198, "en-Blinker-Regular": 199, 
+"en-TAN NIMBUS": 200, "en-Blueberry-Regular": 201, "en-Rosario-Regular": 202, "en-Forum": 203, "en-MistrullyRegular": 204, "en-SourceSerifPro-Regular": 205, "en-Bugaki-Regular": 206, "en-CMUSerif-Roman": 207, "en-GulfsDisplay-NormalItalic": 208, "en-PTSans-Bold": 209, 
+"en-Sensei-Medium": 210, "en-SquadaOne-Regular": 211, "en-Arapey-Italic": 212, "en-Parisienne-Regular": 213, "en-Aleo-Italic": 214, "en-QuicheDisplay-Italic": 215, "en-RocaOne-It": 216, "en-Funtastic-Regular": 217, "en-PTSerif-BoldItalic": 218, "en-Muller-RegularItalic": 219, 
+"en-ArgentCF-Regular": 220, "en-Brightwall-Italic": 221, "en-Knewave-Regular": 222, "en-TYSerif-D": 223, "en-Agrandir-Tight": 224, "en-AlfaSlabOne-Regular": 225, "en-TANTangkiwood-Display": 226, "en-Kief-Montaser-Regular": 227, "en-Gotham-Book": 228, "en-JuliusSansOne-Regular": 229, 
+"en-CocoGothic-Italic": 230, "en-SairaCondensed-Regular": 231, "en-DellaRespira-Regular": 232, "en-Questrial-Regular": 233, "en-BukhariScript-Regular": 234, "en-HelveticaWorld-Bold": 235, "en-TANKINDRED-Display": 236, "en-CinzelDecorative-Regular": 237, "en-Vidaloka-Regular": 238, "en-AlegreyaSansSC-Black": 239, 
+"en-FeelingPassionate-Regular": 240, "en-QuincyCF-Regular": 241, "en-FiraCode-Regular": 242, "en-Genty-Regular": 243, "en-Nickainley-Normal": 244, "en-RubikOne-Regular": 245, "en-Gidole-Regular": 246, "en-Borsok": 247, "en-Gordita-RegularItalic": 248, "en-Scripter-Regular": 249, 
+"en-Buffalo-Regular": 250, "en-KleinText-Regular": 251, "en-Creepster-Regular": 252, "en-Arvo-Bold": 253, "en-GabrielSans-NormalItalic": 254, "en-Heebo-Black": 255, "en-LexendExa-Regular": 256, "en-BrixtonSansTC-Regular": 257, "en-GildaDisplay-Regular": 258, "en-ChunkFive-Roman": 259, 
+"en-Amaranth-BoldItalic": 260, "en-BubbleboddyNeue-Regular": 261, "en-MavenPro-Bold": 262, "en-TTDrugs-Italic": 263, "en-CyGrotesk-KeyRegular": 264, "en-VarelaRound-Regular": 265, "en-Ruda-Black": 266, "en-SafiraMarch": 267, "en-BloggerSans": 268, "en-TANHEADLINE-Regular": 269, 
+"en-SloopScriptPro-Regular": 270, "en-NeueMontreal-Regular": 271, "en-Schoolbell-Regular": 272, "en-SigherRegular": 273, "en-InriaSerif-Regular": 274, "en-JetBrainsMono-Regular": 275, "en-MADEEvolveSans": 276, "en-Dekko": 277, "en-Handyman-Regular": 278, "en-Aileron-BoldItalic": 279, 
+"en-Bright-Italic": 280, "en-Solway-Regular": 281, "en-Higuen-Regular": 282, "en-WedgesItalic": 283, "en-TANASHFORD-BOLD": 284, "en-IBMPlexMono": 285, "en-RacingSansOne-Regular": 286, "en-RegularBrush": 287, "en-OpenSans-LightItalic": 288, "en-SpecialElite-Regular": 289, 
+"en-FuturaLTPro-Medium": 290, "en-MaragsaDisplay": 291, "en-BigShouldersDisplay-Regular": 292, "en-BDSans-Regular": 293, "en-RasputinRegular": 294, "en-Yvesyvesdrawing-BoldItalic": 295, "en-Bitter-Regular": 296, "en-LuckiestGuy-Regular": 297, "en-CanvaSchoolFontDotted": 298, "en-TTFirsNeue-Italic": 299, 
+"en-Sunday-Regular": 300, "en-HKGothic-MediumItalic": 301, "en-CaveatBrush-Regular": 302, "en-HeliosExt": 303, "en-ArchitectsDaughter-Regular": 304, "en-Angelina": 305, "en-Calistoga-Regular": 306, "en-ArchivoNarrow-Regular": 307, "en-ObjectSans-MediumSlanted": 308, "en-AyrLucidityCondensed-Regular": 309, 
+"en-Nexa-RegularItalic": 310, "en-Lustria-Regular": 311, "en-Amsterdam-TwoSlant": 312, "en-Virtual-Regular": 313, "en-Brusher-Regular": 314, "en-NF-Lepetitcochon-Regular": 315, "en-TANTWINKLE": 316, "en-LeJour-Serif": 317, "en-Prata-Regular": 318, "en-PPWoodland-Regular": 319, 
+"en-PlayfairDisplay-BoldItalic": 320, "en-AmaticSC-Regular": 321, "en-Cabin-Regular": 322, "en-Manjari-Bold": 323, "en-MrDafoe-Regular": 324, "en-TTRamillas-Italic": 325, "en-Luckybones-Bold": 326, "en-DarkerGrotesque-Light": 327, "en-BellabooRegular": 328, "en-CormorantSC-Bold": 329, 
+"en-GochiHand-Regular": 330, "en-Atteron": 331, "en-RocaTwo-Lt": 332, "en-ZCOOLXiaoWei-Regular": 333, "en-TANSONGBIRD": 334, "en-HeadingNow-74Regular": 335, "en-Luthier-BoldItalic": 336, "en-Oregano-Regular": 337, "en-AyrTropikaIsland-Int": 338, "en-Mali-Regular": 339, 
+"en-DidactGothic-Regular": 340, "en-Lovelace-Regular": 341, "en-BakerieSmooth-Regular": 342, "en-CarterOne": 343, "en-HussarBd": 344, "en-OldStandard-Italic": 345, "en-TAN-ASTORIA-Display": 346, "en-rugratssans-Regular": 347, "en-BMHANNA": 348, "en-BetterSaturday": 349, 
+"en-AdigianaToybox": 350, "en-Sailors": 351, "en-PlayfairDisplaySC-Italic": 352, "en-Etna-Regular": 353, "en-Revive80Signature": 354, "en-CAGenerated": 355, "en-Poppins-Regular": 356, "en-Jonathan-Regular": 357, "en-Pacifico-Regular": 358, "en-Saira-Black": 359, 
+"en-Loubag-Regular": 360, "en-Decalotype-Black": 361, "en-Mansalva-Regular": 362, "en-Allura-Regular": 363, "en-ProximaNova-Bold": 364, "en-TANMIGNON-DISPLAY": 365, "en-ArsenicaAntiqua-Regular": 366, "en-BreulGroteskA-RegularItalic": 367, "en-HKModular-Bold": 368, "en-TANNightingale-Regular": 369, 
+"en-AristotelicaProCndTxt-Rg": 370, "en-Aprila-Regular": 371, "en-Tomorrow-Regular": 372, "en-AngellaWhite": 373, "en-KaushanScript-Regular": 374, "en-NotoSans": 375, "en-LeJour-Script": 376, "en-BrixtonTC-Regular": 377, "en-OleoScript-Regular": 378, "en-Cakerolli-Regular": 379, 
+"en-Lobster-Regular": 380, "en-FrunchySerif-Regular": 381, "en-PorcelainRegular": 382, "en-AlojaExtended": 383, "en-SergioTrendy-Italic": 384, "en-LovelaceText-Bold": 385, "en-Anaktoria": 386, "en-JimmyScript-Light": 387, "en-IBMPlexSerif": 388, "en-Marta": 389, 
+"en-Mango-Regular": 390, "en-Overpass-Italic": 391, "en-Hagrid-Regular": 392, "en-ElikaGorica": 393, "en-Amiko-Regular": 394, "en-EFCOBrookshire-Regular": 395, "en-Caladea-Regular": 396, "en-MoonlightBold": 397, "en-Staatliches-Regular": 398, "en-Helios-Bold": 399, 
+"en-Satisfy-Regular": 400, "en-NexaScript-Regular": 401, "en-Trocchi-Regular": 402, "en-March": 403, "en-IbarraRealNova-Regular": 404, "en-Nectarine-Regular": 405, "en-Overpass-Light": 406, "en-TruetypewriterPolyglOTT": 407, "en-Bangers-Regular": 408, "en-Lazord-BoldExpandedItalic": 409, 
+"en-Chloe-Regular": 410, "en-BaskervilleDisplayPT-Regular": 411, "en-Bright-Regular": 412, "en-Vollkorn-Regular": 413, "en-Harmattan": 414, "en-SortsMillGoudy-Regular": 415, "en-Biryani-Bold": 416, "en-SugoProDisplay-Italic": 417, "en-Lazord-BoldItalic": 418, "en-Alike-Regular": 419, 
+"en-PermanentMarker-Regular": 420, "en-Sacramento-Regular": 421, "en-HKGroteskPro-Italic": 422, "en-Aleo-BoldItalic": 423, "en-Noot": 424, "en-TANGARLAND-Regular": 425, "en-Twister": 426, "en-Arsenal-Italic": 427, "en-Bogart-Italic": 428, "en-BethEllen-Regular": 429, 
+"en-Caveat-Regular": 430, "en-BalsamiqSans-Bold": 431, "en-BreeSerif-Regular": 432, "en-CodecPro-ExtraBold": 433, "en-Pierson-Light": 434, "en-CyGrotesk-WideRegular": 435, "en-Lumios-Marker": 436, "en-Comfortaa-Bold": 437, "en-TraceFontRegular": 438, "en-RTL-AdamScript-Regular": 439, 
+"en-EastmanGrotesque-Italic": 440, "en-Kalam-Bold": 441, "en-ChauPhilomeneOne-Regular": 442, "en-Coiny-Regular": 443, "en-Lovera": 444, "en-Gellatio": 445, "en-TitilliumWeb-Bold": 446, "en-OilvareBase-Italic": 447, "en-Catamaran-Black": 448, "en-Anteb-Italic": 449, 
+"en-SueEllenFrancisco": 450, "en-SweetApricot": 451, "en-BrightSunshine": 452, "en-IM_FELL_Double_Pica_Italic": 453, "en-Granaina-limpia": 454, "en-TANPARFAIT": 455, "en-AcherusGrotesque-Regular": 456, "en-AwesomeLathusca-Italic": 457, "en-Signika-Bold": 458, "en-Andasia": 459, 
+"en-DO-AllCaps-Slanted": 460, "en-Zenaida-Regular": 461, "en-Fahkwang-Regular": 462, "en-Play-Regular": 463, "en-BERNIERRegular-Regular": 464, "en-PlumaThin-Regular": 465, "en-SportsWorld": 466, "en-Garet-Black": 467, "en-CarolloPlayscript-BlackItalic": 468, "en-Cheque-Regular": 469, 
+"en-SEGO": 470, "en-BobbyJones-Condensed": 471, "en-NexaSlab-RegularItalic": 472, "en-DancingScript-Regular": 473, "en-PaalalabasDisplayWideBETA": 474, "en-Magnolia-Script": 475, "en-OpunMai-400It": 476, "en-MadelynFill-Regular": 477, "en-ZingRust-Base": 478, "en-FingerPaint-Regular": 479, 
+"en-BostonAngel-Light": 480, "en-Gliker-RegularExpanded": 481, "en-Ahsing": 482, "en-Engagement-Regular": 483, "en-EyesomeScript": 484, "en-LibraSerifModern-Regular": 485, "en-London-Regular": 486, "en-AtkinsonHyperlegible-Regular": 487, "en-StadioNow-TextItalic": 488, "en-Aniyah": 489, 
+"en-ITCAvantGardePro-Bold": 490, "en-Comica-Regular": 491, "en-Coustard-Regular": 492, "en-Brice-BoldCondensed": 493, "en-TANNEWYORK-Bold": 494, "en-TANBUSTER-Bold": 495, "en-Alatsi-Regular": 496, "en-TYSerif-Book": 497, "en-Jingleberry": 498, "en-Rajdhani-Bold": 499, 
+"en-LobsterTwo-BoldItalic": 500, "en-BestLight-Medium": 501, "en-Hitchcut-Regular": 502, "en-GermaniaOne-Regular": 503, "en-Emitha-Script": 504, "en-LemonTuesday": 505, "en-Cubao_Free_Regular": 506, "en-MonterchiSerif-Regular": 507, "en-AllertaStencil-Regular": 508, "en-RTL-Sondos-Regular": 509, 
+"en-HomemadeApple-Regular": 510, "en-CosmicOcto-Medium": 511, "cn-HelloFont-FangHuaTi": 0, "cn-HelloFont-ID-DianFangSong-Bold": 1, "cn-HelloFont-ID-DianFangSong": 2, "cn-HelloFont-ID-DianHei-CEJ": 3, "cn-HelloFont-ID-DianHei-DEJ": 4, "cn-HelloFont-ID-DianHei-EEJ": 5, "cn-HelloFont-ID-DianHei-FEJ": 6, "cn-HelloFont-ID-DianHei-GEJ": 7, "cn-HelloFont-ID-DianKai-Bold": 8, "cn-HelloFont-ID-DianKai": 9, 
+"cn-HelloFont-WenYiHei": 10, "cn-Hellofont-ID-ChenYanXingKai": 11, "cn-Hellofont-ID-DaZiBao": 12, "cn-Hellofont-ID-DaoCaoRen": 13, "cn-Hellofont-ID-JianSong": 14, "cn-Hellofont-ID-JiangHuZhaoPaiHei": 15, "cn-Hellofont-ID-KeSong": 16, "cn-Hellofont-ID-LeYuanTi": 17, "cn-Hellofont-ID-Pinocchio": 18, "cn-Hellofont-ID-QiMiaoTi": 19, 
+"cn-Hellofont-ID-QingHuaKai": 20, "cn-Hellofont-ID-QingHuaXingKai": 21, "cn-Hellofont-ID-ShanShuiXingKai": 22, "cn-Hellofont-ID-ShouXieQiShu": 23, "cn-Hellofont-ID-ShouXieTongZhenTi": 24, "cn-Hellofont-ID-TengLingTi": 25, "cn-Hellofont-ID-XiaoLiShu": 26, "cn-Hellofont-ID-XuanZhenSong": 27, "cn-Hellofont-ID-ZhongLingXingKai": 28, "cn-HellofontIDJiaoTangTi": 29, 
+"cn-HellofontIDJiuZhuTi": 30, "cn-HuXiaoBao-SaoBao": 31, "cn-HuXiaoBo-NanShen": 32, "cn-HuXiaoBo-ZhenShuai": 33, "cn-SourceHanSansSC-Bold": 34, "cn-SourceHanSansSC-ExtraLight": 35, "cn-SourceHanSansSC-Heavy": 36, "cn-SourceHanSansSC-Light": 37, "cn-SourceHanSansSC-Medium": 38, "cn-SourceHanSansSC-Normal": 39, 
+"cn-SourceHanSansSC-Regular": 40, "cn-SourceHanSerifSC-Bold": 41, "cn-SourceHanSerifSC-ExtraLight": 42, "cn-SourceHanSerifSC-Heavy": 43, "cn-SourceHanSerifSC-Light": 44, "cn-SourceHanSerifSC-Medium": 45, "cn-SourceHanSerifSC-Regular": 46, "cn-SourceHanSerifSC-SemiBold": 47, "cn-xiaowei": 48, "cn-AaJianHaoTi": 49, 
+"cn-AlibabaPuHuiTi-Bold": 50, "cn-AlibabaPuHuiTi-Heavy": 51, "cn-AlibabaPuHuiTi-Light": 52, "cn-AlibabaPuHuiTi-Medium": 53, "cn-AlibabaPuHuiTi-Regular": 54, "cn-CanvaAcidBoldSC": 55, "cn-CanvaBreezeCN": 56, "cn-CanvaBumperCropSC": 57, "cn-CanvaCakeShopCN": 58, "cn-CanvaEndeavorBlackSC": 59, 
+"cn-CanvaJoyHeiCN": 60, "cn-CanvaLiCN": 61, "cn-CanvaOrientalBrushCN": 62, "cn-CanvaPoster": 63, "cn-CanvaQinfuCalligraphyCN": 64, "cn-CanvaSweetHeartCN": 65, "cn-CanvaSwordLikeDreamCN": 66, "cn-CanvaTangyuanHandwritingCN": 67, "cn-CanvaWanderWorldCN": 68, "cn-CanvaWenCN": 69, 
+"cn-DianZiChunYi": 70, "cn-GenSekiGothicTW-H": 71, "cn-GenWanMinTW-L": 72, "cn-GenYoMinTW-B": 73, "cn-GenYoMinTW-EL": 74, "cn-GenYoMinTW-H": 75, "cn-GenYoMinTW-M": 76, "cn-GenYoMinTW-R": 77, "cn-GenYoMinTW-SB": 78, "cn-HYQiHei-AZEJ": 79, 
+"cn-HYQiHei-EES": 80, "cn-HanaMinA": 81, "cn-HappyZcool-2016": 82, "cn-HelloFont ZJ KeKouKeAiTi": 83, "cn-HelloFont-ID-BoBoTi": 84, "cn-HelloFont-ID-FuGuHei-25": 85, "cn-HelloFont-ID-FuGuHei-35": 86, "cn-HelloFont-ID-FuGuHei-45": 87, "cn-HelloFont-ID-FuGuHei-55": 88, "cn-HelloFont-ID-FuGuHei-65": 89, 
+"cn-HelloFont-ID-FuGuHei-75": 90, "cn-HelloFont-ID-FuGuHei-85": 91, "cn-HelloFont-ID-HeiKa": 92, "cn-HelloFont-ID-HeiTang": 93, "cn-HelloFont-ID-JianSong-95": 94, "cn-HelloFont-ID-JueJiangHei-50": 95, "cn-HelloFont-ID-JueJiangHei-55": 96, "cn-HelloFont-ID-JueJiangHei-60": 97, "cn-HelloFont-ID-JueJiangHei-65": 98, "cn-HelloFont-ID-JueJiangHei-70": 99, 
+"cn-HelloFont-ID-JueJiangHei-75": 100, "cn-HelloFont-ID-JueJiangHei-80": 101, "cn-HelloFont-ID-KuHeiTi": 102, "cn-HelloFont-ID-LingDongTi": 103, "cn-HelloFont-ID-LingLiTi": 104, "cn-HelloFont-ID-MuFengTi": 105, "cn-HelloFont-ID-NaiNaiJiangTi": 106, "cn-HelloFont-ID-PangDu": 107, "cn-HelloFont-ID-ReLieTi": 108, "cn-HelloFont-ID-RouRun": 109, 
+"cn-HelloFont-ID-SaShuangShouXieTi": 110, "cn-HelloFont-ID-WangZheFengFan": 111, "cn-HelloFont-ID-YouQiTi": 112, "cn-Hellofont-ID-XiaLeTi": 113, "cn-Hellofont-ID-XianXiaTi": 114, "cn-HuXiaoBoKuHei": 115, "cn-IDDanMoXingKai": 116, "cn-IDJueJiangHei": 117, "cn-IDMeiLingTi": 118, "cn-IDQQSugar": 119, 
+"cn-LiuJianMaoCao-Regular": 120, "cn-LongCang-Regular": 121, "cn-MaShanZheng-Regular": 122, "cn-PangMenZhengDao-3": 123, "cn-PangMenZhengDao-Cu": 124, "cn-PangMenZhengDao": 125, "cn-SentyCaramel": 126, "cn-SourceHanSerifSC": 127, "cn-WenCang-Regular": 128, "cn-WenQuanYiMicroHei": 129, 
+"cn-XianErTi": 130, "cn-YRDZSTJF": 131, "cn-YS-HelloFont-BangBangTi": 132, "cn-ZCOOLKuaiLe-Regular": 133, "cn-ZCOOLQingKeHuangYou-Regular": 134, "cn-ZCOOLXiaoWei-Regular": 135, "cn-ZCOOL_KuHei": 136, "cn-ZhiMangXing-Regular": 137, "cn-baotuxiaobaiti": 138, "cn-jiangxizhuokai-Regular": 139, 
+"cn-zcool-gdh": 140, "cn-zcoolqingkehuangyouti-Regular": 141, "cn-zcoolwenyiti": 142, "jp-04KanjyukuGothic": 0, "jp-07LightNovelPOP": 1, "jp-07NikumaruFont": 2, "jp-07YasashisaAntique": 3, "jp-07YasashisaGothic": 4, "jp-BokutachinoGothic2Bold": 5, "jp-BokutachinoGothic2Regular": 6, "jp-CHI_SpeedyRight_full_211128-Regular": 7, "jp-CHI_SpeedyRight_italic_full_211127-Regular": 8, "jp-CP-Font": 9, 
+"jp-Canva_CezanneProN-B": 10, "jp-Canva_CezanneProN-M": 11, "jp-Canva_ChiaroStd-B": 12, "jp-Canva_CometStd-B": 13, "jp-Canva_DotMincho16Std-M": 14, "jp-Canva_GrecoStd-B": 15, "jp-Canva_GrecoStd-M": 16, "jp-Canva_LyraStd-DB": 17, "jp-Canva_MatisseHatsuhiPro-B": 18, "jp-Canva_MatisseHatsuhiPro-M": 19, 
+"jp-Canva_ModeMinAStd-B": 20, "jp-Canva_NewCezanneProN-B": 21, "jp-Canva_NewCezanneProN-M": 22, "jp-Canva_PearlStd-L": 23, "jp-Canva_RaglanStd-UB": 24, "jp-Canva_RailwayStd-B": 25, "jp-Canva_ReggaeStd-B": 26, "jp-Canva_RocknRollStd-DB": 27, "jp-Canva_RodinCattleyaPro-B": 28, "jp-Canva_RodinCattleyaPro-M": 29, 
+"jp-Canva_RodinCattleyaPro-UB": 30, "jp-Canva_RodinHimawariPro-B": 31, "jp-Canva_RodinHimawariPro-M": 32, "jp-Canva_RodinMariaPro-B": 33, "jp-Canva_RodinMariaPro-DB": 34, "jp-Canva_RodinProN-M": 35, "jp-Canva_ShadowTLStd-B": 36, "jp-Canva_StickStd-B": 37, "jp-Canva_TsukuAOldMinPr6N-B": 38, "jp-Canva_TsukuAOldMinPr6N-R": 39, 
+"jp-Canva_UtrilloPro-DB": 40, "jp-Canva_UtrilloPro-M": 41, "jp-Canva_YurukaStd-UB": 42, "jp-FGUIGEN": 43, "jp-GlowSansJ-Condensed-Heavy": 44, "jp-GlowSansJ-Condensed-Light": 45, "jp-GlowSansJ-Normal-Bold": 46, "jp-GlowSansJ-Normal-Light": 47, "jp-HannariMincho": 48, "jp-HarenosoraMincho": 49, 
+"jp-Jiyucho": 50, "jp-Kaiso-Makina-B": 51, "jp-Kaisotai-Next-UP-B": 52, "jp-KokoroMinchoutai": 53, "jp-Mamelon-3-Hi-Regular": 54, "jp-MotoyaAnemoneStd-W1": 55, "jp-MotoyaAnemoneStd-W5": 56, "jp-MotoyaAnticPro-W3": 57, "jp-MotoyaCedarStd-W3": 58, "jp-MotoyaCedarStd-W5": 59, 
+"jp-MotoyaGochikaStd-W4": 60, "jp-MotoyaGochikaStd-W8": 61, "jp-MotoyaGothicMiyabiStd-W6": 62, "jp-MotoyaGothicStd-W3": 63, "jp-MotoyaGothicStd-W5": 64, "jp-MotoyaKoinStd-W3": 65, "jp-MotoyaKyotaiStd-W2": 66, "jp-MotoyaKyotaiStd-W4": 67, "jp-MotoyaMaruStd-W3": 68, "jp-MotoyaMaruStd-W5": 69, 
+"jp-MotoyaMinchoMiyabiStd-W4": 70, "jp-MotoyaMinchoMiyabiStd-W6": 71, "jp-MotoyaMinchoModernStd-W4": 72, "jp-MotoyaMinchoModernStd-W6": 73, "jp-MotoyaMinchoStd-W3": 74, "jp-MotoyaMinchoStd-W5": 75, "jp-MotoyaReisyoStd-W2": 76, "jp-MotoyaReisyoStd-W6": 77, "jp-MotoyaTohitsuStd-W4": 78, "jp-MotoyaTohitsuStd-W6": 79, 
+"jp-MtySousyokuEmBcJis-W6": 80, "jp-MtySousyokuLiBcJis-W6": 81, "jp-Mushin": 82, "jp-NotoSansJP-Bold": 83, "jp-NotoSansJP-Regular": 84, "jp-NudMotoyaAporoStd-W3": 85, "jp-NudMotoyaAporoStd-W5": 86, "jp-NudMotoyaCedarStd-W3": 87, "jp-NudMotoyaCedarStd-W5": 88, "jp-NudMotoyaMaruStd-W3": 89, 
+"jp-NudMotoyaMaruStd-W5": 90, "jp-NudMotoyaMinchoStd-W5": 91, "jp-Ounen-mouhitsu": 92, "jp-Ronde-B-Square": 93, "jp-SMotoyaGyosyoStd-W5": 94, "jp-SMotoyaSinkaiStd-W3": 95, "jp-SMotoyaSinkaiStd-W5": 96, "jp-SourceHanSansJP-Bold": 97, "jp-SourceHanSansJP-Regular": 98, "jp-SourceHanSerifJP-Bold": 99, 
+"jp-SourceHanSerifJP-Regular": 100, "jp-TazuganeGothicStdN-Bold": 101, "jp-TazuganeGothicStdN-Regular": 102, "jp-TelopMinProN-B": 103, "jp-Togalite-Bold": 104, "jp-Togalite-Regular": 105, "jp-TsukuMinPr6N-E": 106, "jp-TsukuMinPr6N-M": 107, "jp-mikachan_o": 108, "jp-nagayama_kai": 109, 
+"jp-07LogoTypeGothic7": 110, "jp-07TetsubinGothic": 111, "jp-851CHIKARA-DZUYOKU-KANA-A": 112, "jp-ARMinchoJIS-Light": 113, "jp-ARMinchoJIS-Ultra": 114, "jp-ARPCrystalMinchoJIS-Medium": 115, "jp-ARPCrystalRGothicJIS-Medium": 116, "jp-ARShounanShinpitsuGyosyoJIS-Medium": 117, "jp-AozoraMincho-bold": 118, "jp-AozoraMinchoRegular": 119, 
+"jp-ArialUnicodeMS-Bold": 120, "jp-ArialUnicodeMS": 121, "jp-CanvaBreezeJP": 122, "jp-CanvaLiCN": 123, "jp-CanvaLiJP": 124, "jp-CanvaOrientalBrushCN": 125, "jp-CanvaQinfuCalligraphyJP": 126, "jp-CanvaSweetHeartJP": 127, "jp-CanvaWenJP": 128, "jp-Corporate-Logo-Bold": 129, 
+"jp-DelaGothicOne-Regular": 130, "jp-GN-Kin-iro_SansSerif": 131, "jp-GN-Koharuiro_Sunray": 132, "jp-GenEiGothicM-B": 133, "jp-GenEiGothicM-R": 134, "jp-GenJyuuGothic-Bold": 135, "jp-GenRyuMinTW-B": 136, "jp-GenRyuMinTW-R": 137, "jp-GenSekiGothicTW-B": 138, "jp-GenSekiGothicTW-R": 139, 
+"jp-GenSenRoundedTW-B": 140, "jp-GenSenRoundedTW-R": 141, "jp-GenShinGothic-Bold": 142, "jp-GenShinGothic-Normal": 143, "jp-GenWanMinTW-L": 144, "jp-GenYoGothicTW-B": 145, "jp-GenYoGothicTW-R": 146, "jp-GenYoMinTW-B": 147, "jp-GenYoMinTW-R": 148, "jp-HGBouquet": 149, 
+"jp-HanaMinA": 150, "jp-HanazomeFont": 151, "jp-HinaMincho-Regular": 152, "jp-Honoka-Antique-Maru": 153, "jp-Honoka-Mincho": 154, "jp-HuiFontP": 155, "jp-IPAexMincho": 156, "jp-JK-Gothic-L": 157, "jp-JK-Gothic-M": 158, "jp-JackeyFont": 159, 
+"jp-KaiseiTokumin-Bold": 160, "jp-KaiseiTokumin-Regular": 161, "jp-Keifont": 162, "jp-KiwiMaru-Regular": 163, "jp-Koku-Mincho-Regular": 164, "jp-MotoyaLMaru-W3-90ms-RKSJ-H": 165, "jp-NewTegomin-Regular": 166, "jp-NicoKaku": 167, "jp-NicoMoji+": 168, "jp-Otsutome_font-Bold": 169, 
+"jp-PottaOne-Regular": 170, "jp-RampartOne-Regular": 171, "jp-Senobi-Gothic-Bold": 172, "jp-Senobi-Gothic-Regular": 173, "jp-SmartFontUI-Proportional": 174, "jp-SoukouMincho": 175, "jp-TEST_Klee-DB": 176, "jp-TEST_Klee-M": 177, "jp-TEST_UDMincho-B": 178, "jp-TEST_UDMincho-L": 179, 
+"jp-TT_Akakane-EB": 180, "jp-Tanuki-Permanent-Marker": 181, "jp-TrainOne-Regular": 182, "jp-TsunagiGothic-Black": 183, "jp-Ume-Hy-Gothic": 184, "jp-Ume-P-Mincho": 185, "jp-WenQuanYiMicroHei": 186, "jp-XANO-mincho-U32": 187, "jp-YOzFontM90-Regular": 188, "jp-Yomogi-Regular": 189, 
+"jp-YujiBoku-Regular": 190, "jp-YujiSyuku-Regular": 191, "jp-ZenKakuGothicNew-Bold": 192, "jp-ZenKakuGothicNew-Regular": 193, "jp-ZenKurenaido-Regular": 194, "jp-ZenMaruGothic-Bold": 195, "jp-ZenMaruGothic-Regular": 196, "jp-darts-font": 197, "jp-irohakakuC-Bold": 198, "jp-irohakakuC-Medium": 199, 
+"jp-irohakakuC-Regular": 200, "jp-katyou": 201, "jp-mplus-1m-bold": 202, "jp-mplus-1m-regular": 203, "jp-mplus-1p-bold": 204, "jp-mplus-1p-regular": 205, "jp-rounded-mplus-1p-bold": 206, "jp-rounded-mplus-1p-regular": 207, "jp-timemachine-wa": 208, "jp-ttf-GenEiLateMin-Medium": 209, 
+"jp-uzura_font": 210, "kr-Arita-buri-Bold_OTF": 0, "kr-Arita-buri-HairLine_OTF": 1, "kr-Arita-buri-Light_OTF": 2, "kr-Arita-buri-Medium_OTF": 3, "kr-Arita-buri-SemiBold_OTF": 4, "kr-Canva_YDSunshineL": 5, "kr-Canva_YDSunshineM": 6, "kr-Canva_YoonGulimPro710": 7, "kr-Canva_YoonGulimPro730": 8, "kr-Canva_YoonGulimPro740": 9, 
+"kr-Canva_YoonGulimPro760": 10, "kr-Canva_YoonGulimPro770": 11, "kr-Canva_YoonGulimPro790": 12, "kr-CreHappB": 13, "kr-CreHappL": 14, "kr-CreHappM": 15, "kr-CreHappS": 16, "kr-OTAuroraB": 17, "kr-OTAuroraL": 18, "kr-OTAuroraR": 19, 
+"kr-OTDoldamgilB": 20, "kr-OTDoldamgilL": 21, "kr-OTDoldamgilR": 22, "kr-OTHamsterB": 23, "kr-OTHamsterL": 24, "kr-OTHamsterR": 25, "kr-OTHapchangdanB": 26, "kr-OTHapchangdanL": 27, "kr-OTHapchangdanR": 28, "kr-OTSupersizeBkBOX": 29, 
+"kr-SourceHanSansKR-Bold": 30, "kr-SourceHanSansKR-ExtraLight": 31, "kr-SourceHanSansKR-Heavy": 32, "kr-SourceHanSansKR-Light": 33, "kr-SourceHanSansKR-Medium": 34, "kr-SourceHanSansKR-Normal": 35, "kr-SourceHanSansKR-Regular": 36, "kr-SourceHanSansSC-Bold": 37, "kr-SourceHanSansSC-ExtraLight": 38, "kr-SourceHanSansSC-Heavy": 39, 
+"kr-SourceHanSansSC-Light": 40, "kr-SourceHanSansSC-Medium": 41, "kr-SourceHanSansSC-Normal": 42, "kr-SourceHanSansSC-Regular": 43, "kr-SourceHanSerifSC-Bold": 44, "kr-SourceHanSerifSC-SemiBold": 45, "kr-TDTDBubbleBubbleOTF": 46, "kr-TDTDConfusionOTF": 47, "kr-TDTDCuteAndCuteOTF": 48, "kr-TDTDEggTakOTF": 49, 
+"kr-TDTDEmotionalLetterOTF": 50, "kr-TDTDGalapagosOTF": 51, "kr-TDTDHappyHourOTF": 52, "kr-TDTDLatteOTF": 53, "kr-TDTDMoonLightOTF": 54, "kr-TDTDParkForestOTF": 55, "kr-TDTDPencilOTF": 56, "kr-TDTDSmileOTF": 57, "kr-TDTDSproutOTF": 58, "kr-TDTDSunshineOTF": 59, 
+"kr-TDTDWaferOTF": 60, "kr-777Chyaochyureu": 61, "kr-ArialUnicodeMS-Bold": 62, "kr-ArialUnicodeMS": 63, "kr-BMHANNA": 64, "kr-Baekmuk-Dotum": 65, "kr-BagelFatOne-Regular": 66, "kr-CoreBandi": 67, "kr-CoreBandiFace": 68, "kr-CoreBori": 69, 
+"kr-DoHyeon-Regular": 70, "kr-Dokdo-Regular": 71, "kr-Gaegu-Bold": 72, "kr-Gaegu-Light": 73, "kr-Gaegu-Regular": 74, "kr-GamjaFlower-Regular": 75, "kr-GasoekOne-Regular": 76, "kr-GothicA1-Black": 77, "kr-GothicA1-Bold": 78, "kr-GothicA1-ExtraBold": 79, 
+"kr-GothicA1-ExtraLight": 80, "kr-GothicA1-Light": 81, "kr-GothicA1-Medium": 82, "kr-GothicA1-Regular": 83, "kr-GothicA1-SemiBold": 84, "kr-GothicA1-Thin": 85, "kr-Gugi-Regular": 86, "kr-HiMelody-Regular": 87, "kr-Jua-Regular": 88, "kr-KirangHaerang-Regular": 89, 
+"kr-NanumBrush": 90, "kr-NanumPen": 91, "kr-NanumSquareRoundB": 92, "kr-NanumSquareRoundEB": 93, "kr-NanumSquareRoundL": 94, "kr-NanumSquareRoundR": 95, "kr-SeH-CB": 96, "kr-SeH-CBL": 97, "kr-SeH-CEB": 98, "kr-SeH-CL": 99, 
+"kr-SeH-CM": 100, "kr-SeN-CB": 101, "kr-SeN-CBL": 102, "kr-SeN-CEB": 103, "kr-SeN-CL": 104, "kr-SeN-CM": 105, "kr-Sunflower-Bold": 106, "kr-Sunflower-Light": 107, "kr-Sunflower-Medium": 108, "kr-TTClaytoyR": 109, 
+"kr-TTDalpangiR": 110, "kr-TTMamablockR": 111, "kr-TTNauidongmuR": 112, "kr-TTOktapbangR": 113, "kr-UhBeeMiMi": 114, "kr-UhBeeMiMiBold": 115, "kr-UhBeeSe_hyun": 116, "kr-UhBeeSe_hyunBold": 117, "kr-UhBeenamsoyoung": 118, "kr-UhBeenamsoyoungBold": 119, 
+"kr-WenQuanYiMicroHei": 120, "kr-YeonSung-Regular": 121}"""
+
+
+def add_special_token(tokenizer: T5Tokenizer, text_encoder: T5Stack):
+    """
+    Add special tokens for color and font to tokenizer and text encoder.
+
+    Args:
+        tokenizer: Huggingface tokenizer.
+        text_encoder: Huggingface T5 encoder.
+    """
+    idx_font_dict = json.loads(MULTILINGUAL_10_LANG_IDX_JSON)
+    idx_color_dict = json.loads(COLOR_IDX_JSON)
+
+    font_token = [f"<{font_code[:2]}-font-{idx_font_dict[font_code]}>" for font_code in idx_font_dict]
+    color_token = [f"<color-{i}>" for i in range(len(idx_color_dict))]
+    additional_special_tokens = []
+    additional_special_tokens += color_token
+    additional_special_tokens += font_token
+
+    tokenizer.add_tokens(additional_special_tokens, special_tokens=True)
+    # Set mean_resizing=False to avoid PyTorch LAPACK dependency
+    text_encoder.resize_token_embeddings(len(tokenizer), mean_resizing=False)
+
+
+def load_byt5(
+    ckpt_path: str,
+    dtype: Optional[torch.dtype],
+    device: Union[str, torch.device],
+    disable_mmap: bool = False,
+    state_dict: Optional[dict] = None,
+) -> Tuple[T5Stack, T5Tokenizer]:
+    BYT5_CONFIG_JSON = """
+{
+    "_name_or_path": "/home/patrick/t5/byt5-small",
+    "architectures": [
+        "T5ForConditionalGeneration"
+    ],
+    "d_ff": 3584,
+    "d_kv": 64,
+    "d_model": 1472,
+    "decoder_start_token_id": 0,
+    "dropout_rate": 0.1,
+    "eos_token_id": 1,
+    "feed_forward_proj": "gated-gelu",
+    "gradient_checkpointing": false,
+    "initializer_factor": 1.0,
+    "is_encoder_decoder": true,
+    "layer_norm_epsilon": 1e-06,
+    "model_type": "t5",
+    "num_decoder_layers": 4,
+    "num_heads": 6,
+    "num_layers": 12,
+    "pad_token_id": 0,
+    "relative_attention_num_buckets": 32,
+    "tie_word_embeddings": false,
+    "tokenizer_class": "ByT5Tokenizer",
+    "transformers_version": "4.7.0.dev0",
+    "use_cache": true,
+    "vocab_size": 384
+    }
+"""
+
+    logger.info(f"Loading BYT5 tokenizer from {BYT5_TOKENIZER_PATH}")
+    byt5_tokenizer = AutoTokenizer.from_pretrained(BYT5_TOKENIZER_PATH)
+
+    logger.info("Initializing BYT5 text encoder")
+    config = json.loads(BYT5_CONFIG_JSON)
+    config = T5Config(**config)
+    with init_empty_weights():
+        byt5_text_encoder = T5ForConditionalGeneration._from_config(config).get_encoder()
+
+    add_special_token(byt5_tokenizer, byt5_text_encoder)
+
+    if state_dict is not None:
+        sd = state_dict
+    else:
+        logger.info(f"Loading state dict from {ckpt_path}")
+        sd = load_safetensors(ckpt_path, device, disable_mmap=disable_mmap, dtype=dtype)
+
+    # remove "encoder." prefix
+    sd = {k[len("encoder.") :] if k.startswith("encoder.") else k: v for k, v in sd.items()}
+    sd["embed_tokens.weight"] = sd.pop("shared.weight")
+
+    info = byt5_text_encoder.load_state_dict(sd, strict=True, assign=True)
+    byt5_text_encoder.to(device)
+    byt5_text_encoder.eval()
+    logger.info(f"BYT5 text encoder loaded with info: {info}")
+
+    return byt5_tokenizer, byt5_text_encoder
+
+
+def load_qwen2_5_vl(
+    ckpt_path: str,
+    dtype: Optional[torch.dtype],
+    device: Union[str, torch.device],
+    disable_mmap: bool = False,
+    state_dict: Optional[dict] = None,
+) -> tuple[Qwen2Tokenizer, Qwen2_5_VLForConditionalGeneration]:
+    QWEN2_5_VL_CONFIG_JSON = """
+{
+  "architectures": [
+    "Qwen2_5_VLForConditionalGeneration"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "hidden_act": "silu",
+  "hidden_size": 3584,
+  "image_token_id": 151655,
+  "initializer_range": 0.02,
+  "intermediate_size": 18944,
+  "max_position_embeddings": 128000,
+  "max_window_layers": 28,
+  "model_type": "qwen2_5_vl",
+  "num_attention_heads": 28,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 4,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": {
+    "mrope_section": [
+      16,
+      24,
+      24
+    ],
+    "rope_type": "default",
+    "type": "default"
+  },
+  "rope_theta": 1000000.0,
+  "sliding_window": 32768,
+  "text_config": {
+    "architectures": [
+      "Qwen2_5_VLForConditionalGeneration"
+    ],
+    "attention_dropout": 0.0,
+    "bos_token_id": 151643,
+    "eos_token_id": 151645,
+    "hidden_act": "silu",
+    "hidden_size": 3584,
+    "image_token_id": null,
+    "initializer_range": 0.02,
+    "intermediate_size": 18944,
+    "layer_types": [
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention"
+    ],
+    "max_position_embeddings": 128000,
+    "max_window_layers": 28,
+    "model_type": "qwen2_5_vl_text",
+    "num_attention_heads": 28,
+    "num_hidden_layers": 28,
+    "num_key_value_heads": 4,
+    "rms_norm_eps": 1e-06,
+    "rope_scaling": {
+      "mrope_section": [
+        16,
+        24,
+        24
+      ],
+      "rope_type": "default",
+      "type": "default"
+    },
+    "rope_theta": 1000000.0,
+    "sliding_window": null,
+    "torch_dtype": "float32",
+    "use_cache": true,
+    "use_sliding_window": false,
+    "video_token_id": null,
+    "vision_end_token_id": 151653,
+    "vision_start_token_id": 151652,
+    "vision_token_id": 151654,
+    "vocab_size": 152064
+  },
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.53.1",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "video_token_id": 151656,
+  "vision_config": {
+    "depth": 32,
+    "fullatt_block_indexes": [
+      7,
+      15,
+      23,
+      31
+    ],
+    "hidden_act": "silu",
+    "hidden_size": 1280,
+    "in_channels": 3,
+    "in_chans": 3,
+    "initializer_range": 0.02,
+    "intermediate_size": 3420,
+    "model_type": "qwen2_5_vl",
+    "num_heads": 16,
+    "out_hidden_size": 3584,
+    "patch_size": 14,
+    "spatial_merge_size": 2,
+    "spatial_patch_size": 14,
+    "temporal_patch_size": 2,
+    "tokens_per_second": 2,
+    "torch_dtype": "float32",
+    "window_size": 112
+  },
+  "vision_end_token_id": 151653,
+  "vision_start_token_id": 151652,
+  "vision_token_id": 151654,
+  "vocab_size": 152064
+}
+"""
+    config = json.loads(QWEN2_5_VL_CONFIG_JSON)
+    config = Qwen2_5_VLConfig(**config)
+    with init_empty_weights():
+        qwen2_5_vl = Qwen2_5_VLForConditionalGeneration._from_config(config)
+
+    if state_dict is not None:
+        sd = state_dict
+    else:
+        logger.info(f"Loading state dict from {ckpt_path}")
+        sd = load_safetensors(ckpt_path, device, disable_mmap=disable_mmap, dtype=dtype)
+
+    # convert prefixes
+    for key in list(sd.keys()):
+        if key.startswith("model."):
+            new_key = key.replace("model.", "model.language_model.", 1)
+        elif key.startswith("visual."):
+            new_key = key.replace("visual.", "model.visual.", 1)
+        else:
+            continue
+        if key not in sd:
+            logger.warning(f"Key {key} not found in state dict, skipping.")
+            continue
+        sd[new_key] = sd.pop(key)
+
+    info = qwen2_5_vl.load_state_dict(sd, strict=True, assign=True)
+    logger.info(f"Loaded Qwen2.5-VL: {info}")
+    qwen2_5_vl.to(device)
+    qwen2_5_vl.eval()
+
+    if dtype is not None:
+        if dtype.itemsize == 1:  # fp8
+            org_dtype = torch.bfloat16  # model weight is fp8 in loading, but original dtype is bfloat16
+            logger.info(f"prepare Qwen2.5-VL for fp8: set to {dtype} from {org_dtype}")
+            qwen2_5_vl.to(dtype)
+
+            # prepare LLM for fp8
+            def prepare_fp8(vl_model: Qwen2_5_VLForConditionalGeneration, target_dtype):
+                def forward_hook(module):
+                    def forward(hidden_states):
+                        input_dtype = hidden_states.dtype
+                        hidden_states = hidden_states.to(torch.float32)
+                        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+                        hidden_states = hidden_states * torch.rsqrt(variance + module.variance_epsilon)
+                        # return module.weight.to(input_dtype) * hidden_states.to(input_dtype)
+                        return (module.weight.to(torch.float32) * hidden_states.to(torch.float32)).to(input_dtype)
+
+                    return forward
+
+                def decoder_forward_hook(module):
+                    def forward(
+                        hidden_states: torch.Tensor,
+                        attention_mask: Optional[torch.Tensor] = None,
+                        position_ids: Optional[torch.LongTensor] = None,
+                        past_key_value: Optional[tuple[torch.Tensor]] = None,
+                        output_attentions: Optional[bool] = False,
+                        use_cache: Optional[bool] = False,
+                        cache_position: Optional[torch.LongTensor] = None,
+                        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
+                        **kwargs,
+                    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
+
+                        residual = hidden_states
+
+                        hidden_states = module.input_layernorm(hidden_states)
+
+                        # Self Attention
+                        hidden_states, self_attn_weights = module.self_attn(
+                            hidden_states=hidden_states,
+                            attention_mask=attention_mask,
+                            position_ids=position_ids,
+                            past_key_value=past_key_value,
+                            output_attentions=output_attentions,
+                            use_cache=use_cache,
+                            cache_position=cache_position,
+                            position_embeddings=position_embeddings,
+                            **kwargs,
+                        )
+                        input_dtype = hidden_states.dtype
+                        hidden_states = residual.to(torch.float32) + hidden_states.to(torch.float32)
+                        hidden_states = hidden_states.to(input_dtype)
+
+                        # Fully Connected
+                        residual = hidden_states
+                        hidden_states = module.post_attention_layernorm(hidden_states)
+                        hidden_states = module.mlp(hidden_states)
+                        hidden_states = residual + hidden_states
+
+                        outputs = (hidden_states,)
+
+                        if output_attentions:
+                            outputs += (self_attn_weights,)
+
+                        return outputs
+
+                    return forward
+
+                for module in vl_model.modules():
+                    if module.__class__.__name__ in ["Embedding"]:
+                        # print("set", module.__class__.__name__, "to", target_dtype)
+                        module.to(target_dtype)
+                    if module.__class__.__name__ in ["Qwen2RMSNorm"]:
+                        # print("set", module.__class__.__name__, "hooks")
+                        module.forward = forward_hook(module)
+                    if module.__class__.__name__ in ["Qwen2_5_VLDecoderLayer"]:
+                        # print("set", module.__class__.__name__, "hooks")
+                        module.forward = decoder_forward_hook(module)
+                    if module.__class__.__name__ in ["Qwen2_5_VisionRotaryEmbedding"]:
+                        # print("set", module.__class__.__name__, "hooks")
+                        module.to(target_dtype)
+
+            prepare_fp8(qwen2_5_vl, org_dtype)
+
+        else:
+            logger.info(f"Setting Qwen2.5-VL to dtype: {dtype}")
+            qwen2_5_vl.to(dtype)
+
+    # Load tokenizer
+    logger.info(f"Loading tokenizer from {QWEN_2_5_VL_IMAGE_ID}")
+    tokenizer = Qwen2Tokenizer.from_pretrained(QWEN_2_5_VL_IMAGE_ID)
+    return tokenizer, qwen2_5_vl
+
+
+TOKENIZER_MAX_LENGTH = 1024
+PROMPT_TEMPLATE_ENCODE_START_IDX = 34
+
+
+def get_qwen_prompt_embeds(
+    tokenizer: Qwen2Tokenizer, vlm: Qwen2_5_VLForConditionalGeneration, prompt: Union[str, list[str]] = None
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    input_ids, mask = get_qwen_tokens(tokenizer, prompt)
+    return get_qwen_prompt_embeds_from_tokens(vlm, input_ids, mask)
+
+
+def get_qwen_tokens(tokenizer: Qwen2Tokenizer, prompt: Union[str, list[str]] = None) -> Tuple[torch.Tensor, torch.Tensor]:
+    tokenizer_max_length = TOKENIZER_MAX_LENGTH
+
+    # HunyuanImage-2.1 does not use "<|im_start|>assistant\n" in the prompt template
+    prompt_template_encode = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>"
+    # \n<|im_start|>assistant\n"
+    prompt_template_encode_start_idx = PROMPT_TEMPLATE_ENCODE_START_IDX
+    # default_sample_size = 128
+
+    prompt = [prompt] if isinstance(prompt, str) else prompt
+
+    template = prompt_template_encode
+    drop_idx = prompt_template_encode_start_idx
+    txt = [template.format(e) for e in prompt]
+    txt_tokens = tokenizer(txt, max_length=tokenizer_max_length + drop_idx, padding=True, truncation=True, return_tensors="pt")
+    return txt_tokens.input_ids, txt_tokens.attention_mask
+
+
+def get_qwen_prompt_embeds_from_tokens(
+    vlm: Qwen2_5_VLForConditionalGeneration, input_ids: torch.Tensor, attention_mask: torch.Tensor
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    tokenizer_max_length = TOKENIZER_MAX_LENGTH
+    drop_idx = PROMPT_TEMPLATE_ENCODE_START_IDX
+
+    device = vlm.device
+    dtype = vlm.dtype
+
+    input_ids = input_ids.to(device=device)
+    attention_mask = attention_mask.to(device=device)
+
+    if dtype.itemsize == 1:  # fp8
+        with torch.no_grad(), torch.autocast(device_type=device.type, dtype=torch.bfloat16, enabled=True):
+            encoder_hidden_states = vlm(input_ids=input_ids, attention_mask=attention_mask, output_hidden_states=True)
+    else:
+        with torch.no_grad(), torch.autocast(device_type=device.type, dtype=dtype, enabled=True):
+            encoder_hidden_states = vlm(input_ids=input_ids, attention_mask=attention_mask, output_hidden_states=True)
+
+    hidden_states = encoder_hidden_states.hidden_states[-3]  # use the 3rd last layer's hidden states for HunyuanImage-2.1
+    if hidden_states.shape[1] > tokenizer_max_length + drop_idx:
+        logger.warning(f"Hidden states shape {hidden_states.shape} exceeds max length {tokenizer_max_length + drop_idx}")
+
+    # --- Unnecessary complicated processing, keep for reference ---
+    # split_hidden_states = extract_masked_hidden(hidden_states, txt_tokens.attention_mask)
+    # split_hidden_states = [e[drop_idx:] for e in split_hidden_states]
+    # attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states]
+    # max_seq_len = max([e.size(0) for e in split_hidden_states])
+    # prompt_embeds = torch.stack([torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states])
+    # encoder_attention_mask = torch.stack([torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list])
+    # ----------------------------------------------------------
+
+    prompt_embeds = hidden_states[:, drop_idx:, :]
+    encoder_attention_mask = attention_mask[:, drop_idx:]
+    prompt_embeds = prompt_embeds.to(device=device)
+
+    return prompt_embeds, encoder_attention_mask
+
+
+def format_prompt(texts, styles):
+    """
+    Text "{text}" in {color}, {type}.
+    """
+
+    prompt = ""
+    for text, style in zip(texts, styles):
+        # color and style are always None in official implementation, so we only use text
+        text_prompt = f'Text "{text}"'
+        text_prompt += ". "
+        prompt = prompt + text_prompt
+    return prompt
+
+
+BYT5_MAX_LENGTH = 128
+
+
+def get_glyph_prompt_embeds(
+    tokenizer: T5Tokenizer, text_encoder: T5Stack, prompt: Optional[str] = None
+) -> Tuple[list[bool], torch.Tensor, torch.Tensor]:
+    byt5_tokens, byt5_text_mask = get_byt5_text_tokens(tokenizer, prompt)
+    return get_byt5_prompt_embeds_from_tokens(text_encoder, byt5_tokens, byt5_text_mask)
+
+
+def get_byt5_prompt_embeds_from_tokens(
+    text_encoder: T5Stack, byt5_text_ids: Optional[torch.Tensor], byt5_text_mask: Optional[torch.Tensor]
+) -> Tuple[list[bool], torch.Tensor, torch.Tensor]:
+    byt5_max_length = BYT5_MAX_LENGTH
+
+    if byt5_text_ids is None or byt5_text_mask is None or byt5_text_mask.sum() == 0:
+        return (
+            [False],
+            torch.zeros((1, byt5_max_length, 1472), device=text_encoder.device),
+            torch.zeros((1, byt5_max_length), device=text_encoder.device, dtype=torch.int64),
+        )
+
+    byt5_text_ids = byt5_text_ids.to(device=text_encoder.device)
+    byt5_text_mask = byt5_text_mask.to(device=text_encoder.device)
+
+    with torch.no_grad(), torch.autocast(device_type=text_encoder.device.type, dtype=text_encoder.dtype, enabled=True):
+        byt5_prompt_embeds = text_encoder(byt5_text_ids, attention_mask=byt5_text_mask.float())
+    byt5_emb = byt5_prompt_embeds[0]
+
+    return [True], byt5_emb, byt5_text_mask
+
+
+def get_byt5_text_tokens(tokenizer, prompt):
+    if not prompt:
+        return None, None
+
+    try:
+        text_prompt_texts = []
+        # pattern_quote_single = r"\'(.*?)\'"
+        pattern_quote_double = r"\"(.*?)\""
+        pattern_quote_chinese_single = r"‘(.*?)’"
+        pattern_quote_chinese_double = r"“(.*?)”"
+
+        # matches_quote_single = re.findall(pattern_quote_single, prompt)
+        matches_quote_double = re.findall(pattern_quote_double, prompt)
+        matches_quote_chinese_single = re.findall(pattern_quote_chinese_single, prompt)
+        matches_quote_chinese_double = re.findall(pattern_quote_chinese_double, prompt)
+
+        # text_prompt_texts.extend(matches_quote_single)
+        text_prompt_texts.extend(matches_quote_double)
+        text_prompt_texts.extend(matches_quote_chinese_single)
+        text_prompt_texts.extend(matches_quote_chinese_double)
+
+        if not text_prompt_texts:
+            return None, None
+
+        text_prompt_style_list = [{"color": None, "font-family": None} for _ in range(len(text_prompt_texts))]
+        glyph_text_formatted = format_prompt(text_prompt_texts, text_prompt_style_list)
+        logger.info(f"Glyph text formatted: {glyph_text_formatted}")
+
+        byt5_text_inputs = tokenizer(
+            glyph_text_formatted,
+            padding="max_length",
+            max_length=BYT5_MAX_LENGTH,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+
+        byt5_text_ids = byt5_text_inputs.input_ids
+        byt5_text_mask = byt5_text_inputs.attention_mask
+
+        return byt5_text_ids, byt5_text_mask
+
+    except Exception as e:
+        logger.warning(f"Warning: Error in glyph encoding, using fallback: {e}")
+        return None, None

library/hunyuan_image_utils.py

@@ -0,0 +1,525 @@
+# Original work: https://github.com/Tencent-Hunyuan/HunyuanImage-2.1
+# Re-implemented for license compliance for sd-scripts.
+
+import math
+from typing import Tuple, Union, Optional
+import torch
+
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+MODEL_VERSION_2_1 = "hunyuan-image-2.1"
+
+# region model
+
+
+def _to_tuple(x, dim=2):
+    """
+    Convert int or sequence to tuple of specified dimension.
+
+    Args:
+        x: Int or sequence to convert.
+        dim: Target dimension for tuple.
+
+    Returns:
+        Tuple of length dim.
+    """
+    if isinstance(x, int) or isinstance(x, float):
+        return (x,) * dim
+    elif len(x) == dim:
+        return x
+    else:
+        raise ValueError(f"Expected length {dim} or int, but got {x}")
+
+
+def get_meshgrid_nd(start, dim=2):
+    """
+    Generate n-dimensional coordinate meshgrid from 0 to grid_size.
+
+    Creates coordinate grids for each spatial dimension, useful for
+    generating position embeddings.
+
+    Args:
+        start: Grid size for each dimension (int or tuple).
+        dim: Number of spatial dimensions.
+
+    Returns:
+        Coordinate grid tensor [dim, *grid_size].
+    """
+    # Convert start to grid sizes
+    num = _to_tuple(start, dim=dim)
+    start = (0,) * dim
+    stop = num
+
+    # Generate coordinate arrays for each dimension
+    axis_grid = []
+    for i in range(dim):
+        a, b, n = start[i], stop[i], num[i]
+        g = torch.linspace(a, b, n + 1, dtype=torch.float32)[:n]
+        axis_grid.append(g)
+    grid = torch.meshgrid(*axis_grid, indexing="ij")  # dim x [W, H, D]
+    grid = torch.stack(grid, dim=0)  # [dim, W, H, D]
+
+    return grid
+
+
+def get_nd_rotary_pos_embed(rope_dim_list, start, theta=10000.0):
+    """
+    Generate n-dimensional rotary position embeddings for spatial tokens.
+
+    Creates RoPE embeddings for multi-dimensional positional encoding,
+    distributing head dimensions across spatial dimensions.
+
+    Args:
+        rope_dim_list: Dimensions allocated to each spatial axis (should sum to head_dim).
+        start: Spatial grid size for each dimension.
+        theta: Base frequency for RoPE computation.
+
+    Returns:
+        Tuple of (cos_freqs, sin_freqs) for rotary embedding [H*W, D/2].
+    """
+
+    grid = get_meshgrid_nd(start, dim=len(rope_dim_list))  # [3, W, H, D] / [2, W, H]
+
+    # Generate RoPE embeddings for each spatial dimension
+    embs = []
+    for i in range(len(rope_dim_list)):
+        emb = get_1d_rotary_pos_embed(rope_dim_list[i], grid[i].reshape(-1), theta)  # 2 x [WHD, rope_dim_list[i]]
+        embs.append(emb)
+
+    cos = torch.cat([emb[0] for emb in embs], dim=1)  # (WHD, D/2)
+    sin = torch.cat([emb[1] for emb in embs], dim=1)  # (WHD, D/2)
+    return cos, sin
+
+
+def get_1d_rotary_pos_embed(
+    dim: int, pos: Union[torch.FloatTensor, int], theta: float = 10000.0
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Generate 1D rotary position embeddings.
+
+    Args:
+        dim: Embedding dimension (must be even).
+        pos: Position indices [S] or scalar for sequence length.
+        theta: Base frequency for sinusoidal encoding.
+
+    Returns:
+        Tuple of (cos_freqs, sin_freqs) tensors [S, D].
+    """
+    if isinstance(pos, int):
+        pos = torch.arange(pos).float()
+
+    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))  # [D/2]
+    freqs = torch.outer(pos, freqs)  # [S, D/2]
+    freqs_cos = freqs.cos().repeat_interleave(2, dim=1)  # [S, D]
+    freqs_sin = freqs.sin().repeat_interleave(2, dim=1)  # [S, D]
+    return freqs_cos, freqs_sin
+
+
+def timestep_embedding(t, dim, max_period=10000):
+    """
+    Create sinusoidal timestep embeddings for diffusion models.
+
+    Converts scalar timesteps to high-dimensional embeddings using
+    sinusoidal encoding at different frequencies.
+
+    Args:
+        t: Timestep tensor [N].
+        dim: Output embedding dimension.
+        max_period: Maximum period for frequency computation.
+
+    Returns:
+        Timestep embeddings [N, dim].
+    """
+    half = dim // 2
+    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(device=t.device)
+    args = t[:, None].float() * freqs[None]
+    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+    if dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+
+
+def modulate(x, shift=None, scale=None):
+    """
+    Apply adaptive layer normalization modulation.
+
+    Applies scale and shift transformations for conditioning
+    in adaptive layer normalization.
+
+    Args:
+        x: Input tensor to modulate.
+        shift: Additive shift parameter (optional).
+        scale: Multiplicative scale parameter (optional).
+
+    Returns:
+        Modulated tensor x * (1 + scale) + shift.
+    """
+    if scale is None and shift is None:
+        return x
+    elif shift is None:
+        return x * (1 + scale.unsqueeze(1))
+    elif scale is None:
+        return x + shift.unsqueeze(1)
+    else:
+        return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+
+
+def apply_gate(x, gate=None, tanh=False):
+    """
+    Apply gating mechanism to tensor.
+
+    Multiplies input by gate values, optionally applying tanh activation.
+    Used in residual connections for adaptive control.
+
+    Args:
+        x: Input tensor to gate.
+        gate: Gating values (optional).
+        tanh: Whether to apply tanh to gate values.
+
+    Returns:
+        Gated tensor x * gate (with optional tanh).
+    """
+    if gate is None:
+        return x
+    if tanh:
+        return x * gate.unsqueeze(1).tanh()
+    else:
+        return x * gate.unsqueeze(1)
+
+
+def reshape_for_broadcast(
+    freqs_cis: Tuple[torch.Tensor, torch.Tensor],
+    x: torch.Tensor,
+    head_first=False,
+):
+    """
+    Reshape RoPE frequency tensors for broadcasting with attention tensors.
+
+    Args:
+        freqs_cis: Tuple of (cos_freqs, sin_freqs) tensors.
+        x: Target tensor for broadcasting compatibility.
+        head_first: Must be False (only supported layout).
+
+    Returns:
+        Reshaped (cos_freqs, sin_freqs) tensors ready for broadcasting.
+    """
+    assert not head_first, "Only head_first=False layout supported."
+    assert isinstance(freqs_cis, tuple), "Expected tuple of (cos, sin) frequency tensors."
+    assert x.ndim > 1, f"x should have at least 2 dimensions, but got {x.ndim}"
+
+    # Validate frequency tensor dimensions match target tensor
+    assert freqs_cis[0].shape == (
+        x.shape[1],
+        x.shape[-1],
+    ), f"Frequency tensor shape {freqs_cis[0].shape} incompatible with target shape {x.shape}"
+
+    shape = [d if i == 1 or i == x.ndim - 1 else 1 for i, d in enumerate(x.shape)]
+    return freqs_cis[0].view(*shape), freqs_cis[1].view(*shape)
+
+
+def rotate_half(x):
+    """
+    Rotate half the dimensions for RoPE computation.
+
+    Splits the last dimension in half and applies a 90-degree rotation
+    by swapping and negating components.
+
+    Args:
+        x: Input tensor [..., D] where D is even.
+
+    Returns:
+        Rotated tensor with same shape as input.
+    """
+    x_real, x_imag = x.float().reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
+    return torch.stack([-x_imag, x_real], dim=-1).flatten(3)
+
+
+def apply_rotary_emb(
+    xq: torch.Tensor, xk: torch.Tensor, freqs_cis: Tuple[torch.Tensor, torch.Tensor], head_first: bool = False
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Apply rotary position embeddings to query and key tensors.
+
+    Args:
+        xq: Query tensor [B, S, H, D].
+        xk: Key tensor [B, S, H, D].
+        freqs_cis: Tuple of (cos_freqs, sin_freqs) for rotation.
+        head_first: Whether head dimension precedes sequence dimension.
+
+    Returns:
+        Tuple of rotated (query, key) tensors.
+    """
+    device = xq.device
+    dtype = xq.dtype
+
+    cos, sin = reshape_for_broadcast(freqs_cis, xq, head_first)
+    cos, sin = cos.to(device), sin.to(device)
+
+    # Apply rotation: x' = x * cos + rotate_half(x) * sin
+    xq_out = (xq.float() * cos + rotate_half(xq.float()) * sin).to(dtype)
+    xk_out = (xk.float() * cos + rotate_half(xk.float()) * sin).to(dtype)
+
+    return xq_out, xk_out
+
+
+# endregion
+
+# region inference
+
+
+def get_timesteps_sigmas(sampling_steps: int, shift: float, device: torch.device) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Generate timesteps and sigmas for diffusion sampling.
+
+    Args:
+        sampling_steps: Number of sampling steps.
+        shift: Sigma shift parameter for schedule modification.
+        device: Target device for tensors.
+
+    Returns:
+        Tuple of (timesteps, sigmas) tensors.
+    """
+    sigmas = torch.linspace(1, 0, sampling_steps + 1)
+    sigmas = (shift * sigmas) / (1 + (shift - 1) * sigmas)
+    sigmas = sigmas.to(torch.float32)
+    timesteps = (sigmas[:-1] * 1000).to(dtype=torch.float32, device=device)
+    return timesteps, sigmas
+
+
+def step(latents, noise_pred, sigmas, step_i):
+    """
+    Perform a single diffusion sampling step.
+
+    Args:
+        latents: Current latent state.
+        noise_pred: Predicted noise.
+        sigmas: Noise schedule sigmas.
+        step_i: Current step index.
+
+    Returns:
+        Updated latents after the step.
+    """
+    return latents.float() - (sigmas[step_i] - sigmas[step_i + 1]) * noise_pred.float()
+
+
+# endregion
+
+
+# region AdaptiveProjectedGuidance
+
+
+class MomentumBuffer:
+    """
+    Exponential moving average buffer for APG momentum.
+    """
+
+    def __init__(self, momentum: float):
+        self.momentum = momentum
+        self.running_average = 0
+
+    def update(self, update_value: torch.Tensor):
+        new_average = self.momentum * self.running_average
+        self.running_average = update_value + new_average
+
+
+def normalized_guidance_apg(
+    pred_cond: torch.Tensor,
+    pred_uncond: torch.Tensor,
+    guidance_scale: float,
+    momentum_buffer: Optional[MomentumBuffer] = None,
+    eta: float = 1.0,
+    norm_threshold: float = 0.0,
+    use_original_formulation: bool = False,
+):
+    """
+    Apply normalized adaptive projected guidance.
+
+    Projects the guidance vector to reduce over-saturation while maintaining
+    directional control by decomposing into parallel and orthogonal components.
+
+    Args:
+        pred_cond: Conditional prediction.
+        pred_uncond: Unconditional prediction.
+        guidance_scale: Guidance scale factor.
+        momentum_buffer: Optional momentum buffer for temporal smoothing.
+        eta: Scaling factor for parallel component.
+        norm_threshold: Maximum norm for guidance vector clipping.
+        use_original_formulation: Whether to use original APG formulation.
+
+    Returns:
+        Guided prediction tensor.
+    """
+    diff = pred_cond - pred_uncond
+    dim = [-i for i in range(1, len(diff.shape))]  # All dimensions except batch
+
+    # Apply momentum smoothing if available
+    if momentum_buffer is not None:
+        momentum_buffer.update(diff)
+        diff = momentum_buffer.running_average
+
+    # Apply norm clipping if threshold is set
+    if norm_threshold > 0:
+        diff_norm = diff.norm(p=2, dim=dim, keepdim=True)
+        scale_factor = torch.minimum(torch.ones_like(diff_norm), norm_threshold / diff_norm)
+        diff = diff * scale_factor
+
+    # Project guidance vector into parallel and orthogonal components
+    v0, v1 = diff.double(), pred_cond.double()
+    v1 = torch.nn.functional.normalize(v1, dim=dim)
+    v0_parallel = (v0 * v1).sum(dim=dim, keepdim=True) * v1
+    v0_orthogonal = v0 - v0_parallel
+    diff_parallel, diff_orthogonal = v0_parallel.type_as(diff), v0_orthogonal.type_as(diff)
+
+    # Combine components with different scaling
+    normalized_update = diff_orthogonal + eta * diff_parallel
+    pred = pred_cond if use_original_formulation else pred_uncond
+    pred = pred + guidance_scale * normalized_update
+
+    return pred
+
+
+class AdaptiveProjectedGuidance:
+    """
+    Adaptive Projected Guidance for classifier-free guidance.
+
+    Implements APG which projects the guidance vector to reduce over-saturation
+    while maintaining directional control.
+    """
+
+    def __init__(
+        self,
+        guidance_scale: float = 7.5,
+        adaptive_projected_guidance_momentum: Optional[float] = None,
+        adaptive_projected_guidance_rescale: float = 15.0,
+        eta: float = 0.0,
+        guidance_rescale: float = 0.0,
+        use_original_formulation: bool = False,
+    ):
+        self.guidance_scale = guidance_scale
+        self.adaptive_projected_guidance_momentum = adaptive_projected_guidance_momentum
+        self.adaptive_projected_guidance_rescale = adaptive_projected_guidance_rescale
+        self.eta = eta
+        self.guidance_rescale = guidance_rescale
+        self.use_original_formulation = use_original_formulation
+        self.momentum_buffer = None
+
+    def __call__(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None, step=None) -> torch.Tensor:
+        if step == 0 and self.adaptive_projected_guidance_momentum is not None:
+            self.momentum_buffer = MomentumBuffer(self.adaptive_projected_guidance_momentum)
+
+        pred = normalized_guidance_apg(
+            pred_cond,
+            pred_uncond,
+            self.guidance_scale,
+            self.momentum_buffer,
+            self.eta,
+            self.adaptive_projected_guidance_rescale,
+            self.use_original_formulation,
+        )
+
+        if self.guidance_rescale > 0.0:
+            pred = rescale_noise_cfg(pred, pred_cond, self.guidance_rescale)
+
+        return pred
+
+
+def rescale_noise_cfg(guided_noise, conditional_noise, rescale_factor=0.0):
+    """
+    Rescale guided noise prediction to prevent overexposure and improve image quality.
+
+    This implementation addresses the overexposure issue described in "Common Diffusion Noise
+    Schedules and Sample Steps are Flawed" (https://arxiv.org/pdf/2305.08891.pdf) (Section 3.4).
+    The rescaling preserves the statistical properties of the conditional prediction while reducing artifacts.
+
+    Args:
+        guided_noise (torch.Tensor): Noise prediction from classifier-free guidance.
+        conditional_noise (torch.Tensor): Noise prediction from conditional model.
+        rescale_factor (float): Interpolation factor between original and rescaled predictions.
+                               0.0 = no rescaling, 1.0 = full rescaling.
+
+    Returns:
+        torch.Tensor: Rescaled noise prediction with reduced overexposure.
+    """
+    if rescale_factor == 0.0:
+        return guided_noise
+
+    # Calculate standard deviation across spatial dimensions for both predictions
+    spatial_dims = list(range(1, conditional_noise.ndim))
+    conditional_std = conditional_noise.std(dim=spatial_dims, keepdim=True)
+    guided_std = guided_noise.std(dim=spatial_dims, keepdim=True)
+
+    # Rescale guided noise to match conditional noise statistics
+    std_ratio = conditional_std / guided_std
+    rescaled_prediction = guided_noise * std_ratio
+
+    # Interpolate between original and rescaled predictions
+    final_prediction = rescale_factor * rescaled_prediction + (1.0 - rescale_factor) * guided_noise
+
+    return final_prediction
+
+
+def apply_classifier_free_guidance(
+    noise_pred_text: torch.Tensor,
+    noise_pred_uncond: torch.Tensor,
+    is_ocr: bool,
+    guidance_scale: float,
+    step: int,
+    apg_start_step_ocr: int = 38,
+    apg_start_step_general: int = 5,
+    cfg_guider_ocr: AdaptiveProjectedGuidance = None,
+    cfg_guider_general: AdaptiveProjectedGuidance = None,
+    guidance_rescale: float = 0.0,
+):
+    """
+    Apply classifier-free guidance with OCR-aware APG for batch_size=1.
+
+    Args:
+        noise_pred_text: Conditional noise prediction tensor [1, ...].
+        noise_pred_uncond: Unconditional noise prediction tensor [1, ...].
+        is_ocr: Whether this sample requires OCR-specific guidance.
+        guidance_scale: Guidance scale for CFG.
+        step: Current diffusion step index.
+        apg_start_step_ocr: Step to start APG for OCR regions.
+        apg_start_step_general: Step to start APG for general regions.
+        cfg_guider_ocr: APG guider for OCR regions.
+        cfg_guider_general: APG guider for general regions.
+
+    Returns:
+        Guided noise prediction tensor [1, ...].
+    """
+    if guidance_scale == 1.0:
+        return noise_pred_text
+
+    # Select appropriate guider and start step based on OCR requirement
+    if is_ocr:
+        cfg_guider = cfg_guider_ocr
+        apg_start_step = apg_start_step_ocr
+    else:
+        cfg_guider = cfg_guider_general
+        apg_start_step = apg_start_step_general
+
+    # Apply standard CFG or APG based on current step
+    if step <= apg_start_step:
+        # Standard classifier-free guidance
+        noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+        if guidance_rescale > 0.0:
+            # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+            noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale)
+
+        # Initialize APG guider state
+        _ = cfg_guider(noise_pred_text, noise_pred_uncond, step=step)
+    else:
+        # Use APG for guidance
+        noise_pred = cfg_guider(noise_pred_text, noise_pred_uncond, step=step)
+
+    return noise_pred
+
+
+# endregion

library/hunyuan_image_vae.py

@@ -0,0 +1,755 @@
+from typing import Optional, Tuple
+
+from einops import rearrange
+import numpy as np
+import torch
+from torch import Tensor, nn
+from torch.nn import Conv2d
+from diffusers.models.autoencoders.vae import DiagonalGaussianDistribution
+
+from library.safetensors_utils import load_safetensors
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+VAE_SCALE_FACTOR = 32  # 32x spatial compression
+
+LATENT_SCALING_FACTOR = 0.75289  # Latent scaling factor for Hunyuan Image-2.1
+
+
+def swish(x: Tensor) -> Tensor:
+    """Swish activation function: x * sigmoid(x)."""
+    return x * torch.sigmoid(x)
+
+
+class AttnBlock(nn.Module):
+    """Self-attention block using scaled dot-product attention."""
+
+    def __init__(self, in_channels: int, chunk_size: Optional[int] = None):
+        super().__init__()
+        self.in_channels = in_channels
+        self.norm = nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+        if chunk_size is None or chunk_size <= 0:
+            self.q = Conv2d(in_channels, in_channels, kernel_size=1)
+            self.k = Conv2d(in_channels, in_channels, kernel_size=1)
+            self.v = Conv2d(in_channels, in_channels, kernel_size=1)
+            self.proj_out = Conv2d(in_channels, in_channels, kernel_size=1)
+        else:
+            self.q = ChunkedConv2d(in_channels, in_channels, kernel_size=1, chunk_size=chunk_size)
+            self.k = ChunkedConv2d(in_channels, in_channels, kernel_size=1, chunk_size=chunk_size)
+            self.v = ChunkedConv2d(in_channels, in_channels, kernel_size=1, chunk_size=chunk_size)
+            self.proj_out = ChunkedConv2d(in_channels, in_channels, kernel_size=1, chunk_size=chunk_size)
+
+    def attention(self, x: Tensor) -> Tensor:
+        x = self.norm(x)
+        q = self.q(x)
+        k = self.k(x)
+        v = self.v(x)
+
+        b, c, h, w = q.shape
+        q = rearrange(q, "b c h w -> b (h w) c").contiguous()
+        k = rearrange(k, "b c h w -> b (h w) c").contiguous()
+        v = rearrange(v, "b c h w -> b (h w) c").contiguous()
+
+        x = nn.functional.scaled_dot_product_attention(q, k, v)
+        return rearrange(x, "b (h w) c -> b c h w", h=h, w=w, c=c, b=b)
+
+    def forward(self, x: Tensor) -> Tensor:
+        return x + self.proj_out(self.attention(x))
+
+
+class ChunkedConv2d(nn.Conv2d):
+    """
+    Convolutional layer that processes input in chunks to reduce memory usage.
+
+    Parameters
+    ----------
+    chunk_size : int, optional
+        Size of chunks to process at a time. Default is 64.
+    """
+
+    def __init__(self, *args, **kwargs):
+        if "chunk_size" in kwargs:
+            self.chunk_size = kwargs.pop("chunk_size", 64)
+        super().__init__(*args, **kwargs)
+        assert self.padding_mode == "zeros", "Only 'zeros' padding mode is supported."
+        assert self.dilation == (1, 1) and self.stride == (1, 1), "Only dilation=1 and stride=1 are supported."
+        assert self.groups == 1, "Only groups=1 is supported."
+        assert self.kernel_size[0] == self.kernel_size[1], "Only square kernels are supported."
+        assert (
+            self.padding[0] == self.padding[1] and self.padding[0] == self.kernel_size[0] // 2
+        ), "Only kernel_size//2 padding is supported."
+        self.original_padding = self.padding
+        self.padding = (0, 0)  # We handle padding manually in forward
+
+    def forward(self, x: Tensor) -> Tensor:
+        # If chunking is not needed, process normally. We chunk only along height dimension.
+        if self.chunk_size is None or x.shape[1] <= self.chunk_size:
+            self.padding = self.original_padding
+            x = super().forward(x)
+            self.padding = (0, 0)
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+            return x
+
+        # Process input in chunks to reduce memory usage
+        org_shape = x.shape
+
+        # If kernel size is not 1, we need to use overlapping chunks
+        overlap = self.kernel_size[0] // 2  # 1 for kernel size 3
+        step = self.chunk_size - overlap
+        y = torch.zeros((org_shape[0], self.out_channels, org_shape[2], org_shape[3]), dtype=x.dtype, device=x.device)
+        yi = 0
+        i = 0
+        while i < org_shape[2]:
+            si = i if i == 0 else i - overlap
+            ei = i + self.chunk_size
+
+            # Check last chunk. If remaining part is small, include it in last chunk
+            if ei > org_shape[2] or ei + step // 4 > org_shape[2]:
+                ei = org_shape[2]
+
+            chunk = x[:, :, : ei - si, :]
+            x = x[:, :, ei - si - overlap * 2 :, :]
+
+            # Pad chunk if needed: This is as the original Conv2d with padding
+            if i == 0:  # First chunk
+                # Pad except bottom
+                chunk = torch.nn.functional.pad(chunk, (overlap, overlap, overlap, 0), mode="constant", value=0)
+            elif ei == org_shape[2]:  # Last chunk
+                # Pad except top
+                chunk = torch.nn.functional.pad(chunk, (overlap, overlap, 0, overlap), mode="constant", value=0)
+            else:
+                # Pad left and right only
+                chunk = torch.nn.functional.pad(chunk, (overlap, overlap), mode="constant", value=0)
+
+            chunk = super().forward(chunk)
+            y[:, :, yi : yi + chunk.shape[2], :] = chunk
+            yi += chunk.shape[2]
+            del chunk
+
+            if ei == org_shape[2]:
+                break
+            i += step
+
+        assert yi == org_shape[2], f"yi={yi}, org_shape[2]={org_shape[2]}"
+
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()  # This helps reduce peak memory usage, but slows down a bit
+        return y
+
+
+class ResnetBlock(nn.Module):
+    """
+    Residual block with two convolutions, group normalization, and swish activation.
+    Includes skip connection with optional channel dimension matching.
+
+    Parameters
+    ----------
+    in_channels : int
+        Number of input channels.
+    out_channels : int
+        Number of output channels.
+    """
+
+    def __init__(self, in_channels: int, out_channels: int, chunk_size: Optional[int] = None):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        self.norm1 = nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+        self.norm2 = nn.GroupNorm(num_groups=32, num_channels=out_channels, eps=1e-6, affine=True)
+        if chunk_size is None or chunk_size <= 0:
+            self.conv1 = Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+            self.conv2 = Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
+
+            # Skip connection projection for channel dimension mismatch
+            if self.in_channels != self.out_channels:
+                self.nin_shortcut = Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+        else:
+            self.conv1 = ChunkedConv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1, chunk_size=chunk_size)
+            self.conv2 = ChunkedConv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, chunk_size=chunk_size)
+
+            # Skip connection projection for channel dimension mismatch
+            if self.in_channels != self.out_channels:
+                self.nin_shortcut = ChunkedConv2d(
+                    in_channels, out_channels, kernel_size=1, stride=1, padding=0, chunk_size=chunk_size
+                )
+
+    def forward(self, x: Tensor) -> Tensor:
+        h = x
+        # First convolution block
+        h = self.norm1(h)
+        h = swish(h)
+        h = self.conv1(h)
+        # Second convolution block
+        h = self.norm2(h)
+        h = swish(h)
+        h = self.conv2(h)
+
+        # Apply skip connection with optional projection
+        if self.in_channels != self.out_channels:
+            x = self.nin_shortcut(x)
+        return x + h
+
+
+class Downsample(nn.Module):
+    """
+    Spatial downsampling block that reduces resolution by 2x using convolution followed by
+    pixel rearrangement. Includes skip connection with grouped averaging.
+
+    Parameters
+    ----------
+    in_channels : int
+        Number of input channels.
+    out_channels : int
+        Number of output channels (must be divisible by 4).
+    """
+
+    def __init__(self, in_channels: int, out_channels: int, chunk_size: Optional[int] = None):
+        super().__init__()
+        factor = 4  # 2x2 spatial reduction factor
+        assert out_channels % factor == 0
+
+        if chunk_size is None or chunk_size <= 0:
+            self.conv = Conv2d(in_channels, out_channels // factor, kernel_size=3, stride=1, padding=1)
+        else:
+            self.conv = ChunkedConv2d(
+                in_channels, out_channels // factor, kernel_size=3, stride=1, padding=1, chunk_size=chunk_size
+            )
+        self.group_size = factor * in_channels // out_channels
+
+    def forward(self, x: Tensor) -> Tensor:
+        # Apply convolution and rearrange pixels for 2x downsampling
+        h = self.conv(x)
+        h = rearrange(h, "b c (h r1) (w r2) -> b (r1 r2 c) h w", r1=2, r2=2)
+
+        # Create skip connection with pixel rearrangement
+        shortcut = rearrange(x, "b c (h r1) (w r2) -> b (r1 r2 c) h w", r1=2, r2=2)
+        B, C, H, W = shortcut.shape
+        shortcut = shortcut.view(B, h.shape[1], self.group_size, H, W).mean(dim=2)
+
+        return h + shortcut
+
+
+class Upsample(nn.Module):
+    """
+    Spatial upsampling block that increases resolution by 2x using convolution followed by
+    pixel rearrangement. Includes skip connection with channel repetition.
+
+    Parameters
+    ----------
+    in_channels : int
+        Number of input channels.
+    out_channels : int
+        Number of output channels.
+    """
+
+    def __init__(self, in_channels: int, out_channels: int, chunk_size: Optional[int] = None):
+        super().__init__()
+        factor = 4  # 2x2 spatial expansion factor
+
+        if chunk_size is None or chunk_size <= 0:
+            self.conv = Conv2d(in_channels, out_channels * factor, kernel_size=3, stride=1, padding=1)
+        else:
+            self.conv = ChunkedConv2d(in_channels, out_channels * factor, kernel_size=3, stride=1, padding=1, chunk_size=chunk_size)
+
+        self.repeats = factor * out_channels // in_channels
+
+    def forward(self, x: Tensor) -> Tensor:
+        # Apply convolution and rearrange pixels for 2x upsampling
+        h = self.conv(x)
+        h = rearrange(h, "b (r1 r2 c) h w -> b c (h r1) (w r2)", r1=2, r2=2)
+
+        # Create skip connection with channel repetition
+        shortcut = x.repeat_interleave(repeats=self.repeats, dim=1)
+        shortcut = rearrange(shortcut, "b (r1 r2 c) h w -> b c (h r1) (w r2)", r1=2, r2=2)
+
+        return h + shortcut
+
+
+class Encoder(nn.Module):
+    """
+    VAE encoder that progressively downsamples input images to a latent representation.
+    Uses residual blocks, attention, and spatial downsampling.
+
+    Parameters
+    ----------
+    in_channels : int
+        Number of input image channels (e.g., 3 for RGB).
+    z_channels : int
+        Number of latent channels in the output.
+    block_out_channels : Tuple[int, ...]
+        Output channels for each downsampling block.
+    num_res_blocks : int
+        Number of residual blocks per downsampling stage.
+    ffactor_spatial : int
+        Total spatial downsampling factor (e.g., 32 for 32x compression).
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        z_channels: int,
+        block_out_channels: Tuple[int, ...],
+        num_res_blocks: int,
+        ffactor_spatial: int,
+        chunk_size: Optional[int] = None,
+    ):
+        super().__init__()
+        assert block_out_channels[-1] % (2 * z_channels) == 0
+
+        self.z_channels = z_channels
+        self.block_out_channels = block_out_channels
+        self.num_res_blocks = num_res_blocks
+
+        if chunk_size is None or chunk_size <= 0:
+            self.conv_in = Conv2d(in_channels, block_out_channels[0], kernel_size=3, stride=1, padding=1)
+        else:
+            self.conv_in = ChunkedConv2d(
+                in_channels, block_out_channels[0], kernel_size=3, stride=1, padding=1, chunk_size=chunk_size
+            )
+
+        self.down = nn.ModuleList()
+        block_in = block_out_channels[0]
+
+        # Build downsampling blocks
+        for i_level, ch in enumerate(block_out_channels):
+            block = nn.ModuleList()
+            block_out = ch
+
+            # Add residual blocks for this level
+            for _ in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in, out_channels=block_out, chunk_size=chunk_size))
+                block_in = block_out
+
+            down = nn.Module()
+            down.block = block
+
+            # Add spatial downsampling if needed
+            add_spatial_downsample = bool(i_level < np.log2(ffactor_spatial))
+            if add_spatial_downsample:
+                assert i_level < len(block_out_channels) - 1
+                block_out = block_out_channels[i_level + 1]
+                down.downsample = Downsample(block_in, block_out, chunk_size=chunk_size)
+                block_in = block_out
+
+            self.down.append(down)
+
+        # Middle blocks with attention
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in, out_channels=block_in, chunk_size=chunk_size)
+        self.mid.attn_1 = AttnBlock(block_in, chunk_size=chunk_size)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in, out_channels=block_in, chunk_size=chunk_size)
+
+        # Output layers
+        self.norm_out = nn.GroupNorm(num_groups=32, num_channels=block_in, eps=1e-6, affine=True)
+        if chunk_size is None or chunk_size <= 0:
+            self.conv_out = Conv2d(block_in, 2 * z_channels, kernel_size=3, stride=1, padding=1)
+        else:
+            self.conv_out = ChunkedConv2d(block_in, 2 * z_channels, kernel_size=3, stride=1, padding=1, chunk_size=chunk_size)
+
+    def forward(self, x: Tensor) -> Tensor:
+        # Initial convolution
+        h = self.conv_in(x)
+
+        # Progressive downsampling through blocks
+        for i_level in range(len(self.block_out_channels)):
+            # Apply residual blocks at this level
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](h)
+            # Apply spatial downsampling if available
+            if hasattr(self.down[i_level], "downsample"):
+                h = self.down[i_level].downsample(h)
+
+        # Middle processing with attention
+        h = self.mid.block_1(h)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h)
+
+        # Final output layers with skip connection
+        group_size = self.block_out_channels[-1] // (2 * self.z_channels)
+        shortcut = rearrange(h, "b (c r) h w -> b c r h w", r=group_size).mean(dim=2)
+        h = self.norm_out(h)
+        h = swish(h)
+        h = self.conv_out(h)
+        h += shortcut
+        return h
+
+
+class Decoder(nn.Module):
+    """
+    VAE decoder that progressively upsamples latent representations back to images.
+    Uses residual blocks, attention, and spatial upsampling.
+
+    Parameters
+    ----------
+    z_channels : int
+        Number of latent channels in the input.
+    out_channels : int
+        Number of output image channels (e.g., 3 for RGB).
+    block_out_channels : Tuple[int, ...]
+        Output channels for each upsampling block.
+    num_res_blocks : int
+        Number of residual blocks per upsampling stage.
+    ffactor_spatial : int
+        Total spatial upsampling factor (e.g., 32 for 32x expansion).
+    """
+
+    def __init__(
+        self,
+        z_channels: int,
+        out_channels: int,
+        block_out_channels: Tuple[int, ...],
+        num_res_blocks: int,
+        ffactor_spatial: int,
+        chunk_size: Optional[int] = None,
+    ):
+        super().__init__()
+        assert block_out_channels[0] % z_channels == 0
+
+        self.z_channels = z_channels
+        self.block_out_channels = block_out_channels
+        self.num_res_blocks = num_res_blocks
+
+        block_in = block_out_channels[0]
+        if chunk_size is None or chunk_size <= 0:
+            self.conv_in = Conv2d(z_channels, block_in, kernel_size=3, stride=1, padding=1)
+        else:
+            self.conv_in = ChunkedConv2d(z_channels, block_in, kernel_size=3, stride=1, padding=1, chunk_size=chunk_size)
+
+        # Middle blocks with attention
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in, out_channels=block_in, chunk_size=chunk_size)
+        self.mid.attn_1 = AttnBlock(block_in, chunk_size=chunk_size)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in, out_channels=block_in, chunk_size=chunk_size)
+
+        # Build upsampling blocks
+        self.up = nn.ModuleList()
+        for i_level, ch in enumerate(block_out_channels):
+            block = nn.ModuleList()
+            block_out = ch
+
+            # Add residual blocks for this level (extra block for decoder)
+            for _ in range(self.num_res_blocks + 1):
+                block.append(ResnetBlock(in_channels=block_in, out_channels=block_out, chunk_size=chunk_size))
+                block_in = block_out
+
+            up = nn.Module()
+            up.block = block
+
+            # Add spatial upsampling if needed
+            add_spatial_upsample = bool(i_level < np.log2(ffactor_spatial))
+            if add_spatial_upsample:
+                assert i_level < len(block_out_channels) - 1
+                block_out = block_out_channels[i_level + 1]
+                up.upsample = Upsample(block_in, block_out, chunk_size=chunk_size)
+                block_in = block_out
+
+            self.up.append(up)
+
+        # Output layers
+        self.norm_out = nn.GroupNorm(num_groups=32, num_channels=block_in, eps=1e-6, affine=True)
+        if chunk_size is None or chunk_size <= 0:
+            self.conv_out = Conv2d(block_in, out_channels, kernel_size=3, stride=1, padding=1)
+        else:
+            self.conv_out = ChunkedConv2d(block_in, out_channels, kernel_size=3, stride=1, padding=1, chunk_size=chunk_size)
+
+    def forward(self, z: Tensor) -> Tensor:
+        # Initial processing with skip connection
+        repeats = self.block_out_channels[0] // self.z_channels
+        h = self.conv_in(z) + z.repeat_interleave(repeats=repeats, dim=1)
+
+        # Middle processing with attention
+        h = self.mid.block_1(h)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h)
+
+        # Progressive upsampling through blocks
+        for i_level in range(len(self.block_out_channels)):
+            # Apply residual blocks at this level
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](h)
+            # Apply spatial upsampling if available
+            if hasattr(self.up[i_level], "upsample"):
+                h = self.up[i_level].upsample(h)
+
+        # Final output layers
+        h = self.norm_out(h)
+        h = swish(h)
+        h = self.conv_out(h)
+        return h
+
+
+class HunyuanVAE2D(nn.Module):
+    """
+    VAE model for Hunyuan Image-2.1 with spatial tiling support.
+
+    This VAE uses a fixed architecture optimized for the Hunyuan Image-2.1 model,
+    with 32x spatial compression and optional memory-efficient tiling for large images.
+    """
+
+    def __init__(self, chunk_size: Optional[int] = None):
+        super().__init__()
+
+        # Fixed configuration for Hunyuan Image-2.1
+        block_out_channels = (128, 256, 512, 512, 1024, 1024)
+        in_channels = 3  # RGB input
+        out_channels = 3  # RGB output
+        latent_channels = 64
+        layers_per_block = 2
+        ffactor_spatial = 32  # 32x spatial compression
+        sample_size = 384  # Minimum sample size for tiling
+        scaling_factor = LATENT_SCALING_FACTOR  # 0.75289  # Latent scaling factor
+
+        self.ffactor_spatial = ffactor_spatial
+        self.scaling_factor = scaling_factor
+
+        self.encoder = Encoder(
+            in_channels=in_channels,
+            z_channels=latent_channels,
+            block_out_channels=block_out_channels,
+            num_res_blocks=layers_per_block,
+            ffactor_spatial=ffactor_spatial,
+            chunk_size=chunk_size,
+        )
+
+        self.decoder = Decoder(
+            z_channels=latent_channels,
+            out_channels=out_channels,
+            block_out_channels=list(reversed(block_out_channels)),
+            num_res_blocks=layers_per_block,
+            ffactor_spatial=ffactor_spatial,
+            chunk_size=chunk_size,
+        )
+
+        # Spatial tiling configuration for memory efficiency
+        self.use_spatial_tiling = False
+        self.tile_sample_min_size = sample_size
+        self.tile_latent_min_size = sample_size // ffactor_spatial
+        self.tile_overlap_factor = 0.25  # 25% overlap between tiles
+
+    @property
+    def dtype(self):
+        """Get the data type of the model parameters."""
+        return next(self.encoder.parameters()).dtype
+
+    @property
+    def device(self):
+        """Get the device of the model parameters."""
+        return next(self.encoder.parameters()).device
+
+    def enable_spatial_tiling(self, use_tiling: bool = True):
+        """Enable or disable spatial tiling."""
+        self.use_spatial_tiling = use_tiling
+
+    def disable_spatial_tiling(self):
+        """Disable spatial tiling."""
+        self.use_spatial_tiling = False
+
+    def enable_tiling(self, use_tiling: bool = True):
+        """Enable or disable spatial tiling (alias for enable_spatial_tiling)."""
+        self.enable_spatial_tiling(use_tiling)
+
+    def disable_tiling(self):
+        """Disable spatial tiling (alias for disable_spatial_tiling)."""
+        self.disable_spatial_tiling()
+
+    def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        """
+        Blend two tensors horizontally with smooth transition.
+
+        Parameters
+        ----------
+        a : torch.Tensor
+            Left tensor.
+        b : torch.Tensor
+            Right tensor.
+        blend_extent : int
+            Number of columns to blend.
+        """
+        blend_extent = min(a.shape[-1], b.shape[-1], blend_extent)
+        for x in range(blend_extent):
+            b[:, :, :, x] = a[:, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, x] * (x / blend_extent)
+        return b
+
+    def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        """
+        Blend two tensors vertically with smooth transition.
+
+        Parameters
+        ----------
+        a : torch.Tensor
+            Top tensor.
+        b : torch.Tensor
+            Bottom tensor.
+        blend_extent : int
+            Number of rows to blend.
+        """
+        blend_extent = min(a.shape[-2], b.shape[-2], blend_extent)
+        for y in range(blend_extent):
+            b[:, :, y, :] = a[:, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, y, :] * (y / blend_extent)
+        return b
+
+    def spatial_tiled_encode(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Encode large images using spatial tiling to reduce memory usage.
+        Tiles are processed independently and blended at boundaries.
+
+        Parameters
+        ----------
+        x : torch.Tensor
+            Input tensor of shape (B, C, T, H, W) or (B, C, H, W).
+        """
+        # Handle 5D input (B, C, T, H, W) by removing time dimension
+        original_ndim = x.ndim
+        if original_ndim == 5:
+            x = x.squeeze(2)
+
+        B, C, H, W = x.shape
+        overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor))
+        blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor)
+        row_limit = self.tile_latent_min_size - blend_extent
+
+        rows = []
+        for i in range(0, H, overlap_size):
+            row = []
+            for j in range(0, W, overlap_size):
+                tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size]
+                tile = self.encoder(tile)
+                row.append(tile)
+            rows.append(row)
+
+        result_rows = []
+        for i, row in enumerate(rows):
+            result_row = []
+            for j, tile in enumerate(row):
+                if i > 0:
+                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
+                if j > 0:
+                    tile = self.blend_h(row[j - 1], tile, blend_extent)
+                result_row.append(tile[:, :, :row_limit, :row_limit])
+            result_rows.append(torch.cat(result_row, dim=-1))
+
+        moments = torch.cat(result_rows, dim=-2)
+        return moments
+
+    def spatial_tiled_decode(self, z: torch.Tensor) -> torch.Tensor:
+        """
+        Decode large latents using spatial tiling to reduce memory usage.
+        Tiles are processed independently and blended at boundaries.
+
+        Parameters
+        ----------
+        z : torch.Tensor
+            Latent tensor of shape (B, C, H, W).
+        """
+        B, C, H, W = z.shape
+        overlap_size = int(self.tile_latent_min_size * (1 - self.tile_overlap_factor))
+        blend_extent = int(self.tile_sample_min_size * self.tile_overlap_factor)
+        row_limit = self.tile_sample_min_size - blend_extent
+
+        rows = []
+        for i in range(0, H, overlap_size):
+            row = []
+            for j in range(0, W, overlap_size):
+                tile = z[:, :, :, i : i + self.tile_latent_min_size, j : j + self.tile_latent_min_size]
+                decoded = self.decoder(tile)
+                row.append(decoded)
+            rows.append(row)
+
+        result_rows = []
+        for i, row in enumerate(rows):
+            result_row = []
+            for j, tile in enumerate(row):
+                if i > 0:
+                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
+                if j > 0:
+                    tile = self.blend_h(row[j - 1], tile, blend_extent)
+                result_row.append(tile[:, :, :, :row_limit, :row_limit])
+            result_rows.append(torch.cat(result_row, dim=-1))
+
+        dec = torch.cat(result_rows, dim=-2)
+        return dec
+
+    def encode(self, x: Tensor) -> DiagonalGaussianDistribution:
+        """
+        Encode input images to latent representation.
+        Uses spatial tiling for large images if enabled.
+
+        Parameters
+        ----------
+        x : Tensor
+            Input image tensor of shape (B, C, H, W) or (B, C, T, H, W).
+
+        Returns
+        -------
+        DiagonalGaussianDistribution
+            Latent distribution with mean and logvar.
+        """
+        # Handle 5D input (B, C, T, H, W) by removing time dimension
+        original_ndim = x.ndim
+        if original_ndim == 5:
+            x = x.squeeze(2)
+
+        # Use tiling for large images to reduce memory usage
+        if self.use_spatial_tiling and (x.shape[-1] > self.tile_sample_min_size or x.shape[-2] > self.tile_sample_min_size):
+            h = self.spatial_tiled_encode(x)
+        else:
+            h = self.encoder(x)
+
+        # Restore time dimension if input was 5D
+        if original_ndim == 5:
+            h = h.unsqueeze(2)
+
+        posterior = DiagonalGaussianDistribution(h)
+        return posterior
+
+    def decode(self, z: Tensor):
+        """
+        Decode latent representation back to images.
+        Uses spatial tiling for large latents if enabled.
+
+        Parameters
+        ----------
+        z : Tensor
+            Latent tensor of shape (B, C, H, W) or (B, C, T, H, W).
+
+        Returns
+        -------
+        Tensor
+            Decoded image tensor.
+        """
+        # Handle 5D input (B, C, T, H, W) by removing time dimension
+        original_ndim = z.ndim
+        if original_ndim == 5:
+            z = z.squeeze(2)
+
+        # Use tiling for large latents to reduce memory usage
+        if self.use_spatial_tiling and (z.shape[-1] > self.tile_latent_min_size or z.shape[-2] > self.tile_latent_min_size):
+            decoded = self.spatial_tiled_decode(z)
+        else:
+            decoded = self.decoder(z)
+
+        # Restore time dimension if input was 5D
+        if original_ndim == 5:
+            decoded = decoded.unsqueeze(2)
+
+        return decoded
+
+
+def load_vae(vae_path: str, device: torch.device, disable_mmap: bool = False, chunk_size: Optional[int] = None) -> HunyuanVAE2D:
+    logger.info(f"Initializing VAE with chunk_size={chunk_size}")
+    vae = HunyuanVAE2D(chunk_size=chunk_size)
+
+    logger.info(f"Loading VAE from {vae_path}")
+    state_dict = load_safetensors(vae_path, device=device, disable_mmap=disable_mmap)
+    info = vae.load_state_dict(state_dict, strict=True, assign=True)
+    logger.info(f"Loaded VAE: {info}")
+
+    vae.to(device)
+    return vae

library/ipex/__init__.py

@@ -1,170 +1,204 @@
 import os
 import sys
-import contextlib
 import torch
-import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
+try:
+    import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
+    has_ipex = True
+except Exception:
+    has_ipex = False
 from .hijacks import ipex_hijacks
 
+torch_version = float(torch.__version__[:3])
+
 # pylint: disable=protected-access, missing-function-docstring, line-too-long
 
 def ipex_init(): # pylint: disable=too-many-statements
     try:
-        # Replace cuda with xpu:
-        torch.cuda.current_device = torch.xpu.current_device
-        torch.cuda.current_stream = torch.xpu.current_stream
-        torch.cuda.device = torch.xpu.device
-        torch.cuda.device_count = torch.xpu.device_count
-        torch.cuda.device_of = torch.xpu.device_of
-        torch.cuda.get_device_name = torch.xpu.get_device_name
-        torch.cuda.get_device_properties = torch.xpu.get_device_properties
-        torch.cuda.init = torch.xpu.init
-        torch.cuda.is_available = torch.xpu.is_available
-        torch.cuda.is_initialized = torch.xpu.is_initialized
-        torch.cuda.is_current_stream_capturing = lambda: False
-        torch.cuda.set_device = torch.xpu.set_device
-        torch.cuda.stream = torch.xpu.stream
-        torch.cuda.synchronize = torch.xpu.synchronize
-        torch.cuda.Event = torch.xpu.Event
-        torch.cuda.Stream = torch.xpu.Stream
-        torch.cuda.FloatTensor = torch.xpu.FloatTensor
-        torch.Tensor.cuda = torch.Tensor.xpu
-        torch.Tensor.is_cuda = torch.Tensor.is_xpu
-        torch.UntypedStorage.cuda = torch.UntypedStorage.xpu
-        torch.cuda._initialization_lock = torch.xpu.lazy_init._initialization_lock
-        torch.cuda._initialized = torch.xpu.lazy_init._initialized
-        torch.cuda._lazy_seed_tracker = torch.xpu.lazy_init._lazy_seed_tracker
-        torch.cuda._queued_calls = torch.xpu.lazy_init._queued_calls
-        torch.cuda._tls = torch.xpu.lazy_init._tls
-        torch.cuda.threading = torch.xpu.lazy_init.threading
-        torch.cuda.traceback = torch.xpu.lazy_init.traceback
-        torch.cuda.Optional = torch.xpu.Optional
-        torch.cuda.__cached__ = torch.xpu.__cached__
-        torch.cuda.__loader__ = torch.xpu.__loader__
-        torch.cuda.ComplexFloatStorage = torch.xpu.ComplexFloatStorage
-        torch.cuda.Tuple = torch.xpu.Tuple
-        torch.cuda.streams = torch.xpu.streams
-        torch.cuda._lazy_new = torch.xpu._lazy_new
-        torch.cuda.FloatStorage = torch.xpu.FloatStorage
-        torch.cuda.Any = torch.xpu.Any
-        torch.cuda.__doc__ = torch.xpu.__doc__
-        torch.cuda.default_generators = torch.xpu.default_generators
-        torch.cuda.HalfTensor = torch.xpu.HalfTensor
-        torch.cuda._get_device_index = torch.xpu._get_device_index
-        torch.cuda.__path__ = torch.xpu.__path__
-        torch.cuda.Device = torch.xpu.Device
-        torch.cuda.IntTensor = torch.xpu.IntTensor
-        torch.cuda.ByteStorage = torch.xpu.ByteStorage
-        torch.cuda.set_stream = torch.xpu.set_stream
-        torch.cuda.BoolStorage = torch.xpu.BoolStorage
-        torch.cuda.os = torch.xpu.os
-        torch.cuda.torch = torch.xpu.torch
-        torch.cuda.BFloat16Storage = torch.xpu.BFloat16Storage
-        torch.cuda.Union = torch.xpu.Union
-        torch.cuda.DoubleTensor = torch.xpu.DoubleTensor
-        torch.cuda.ShortTensor = torch.xpu.ShortTensor
-        torch.cuda.LongTensor = torch.xpu.LongTensor
-        torch.cuda.IntStorage = torch.xpu.IntStorage
-        torch.cuda.LongStorage = torch.xpu.LongStorage
-        torch.cuda.__annotations__ = torch.xpu.__annotations__
-        torch.cuda.__package__ = torch.xpu.__package__
-        torch.cuda.__builtins__ = torch.xpu.__builtins__
-        torch.cuda.CharTensor = torch.xpu.CharTensor
-        torch.cuda.List = torch.xpu.List
-        torch.cuda._lazy_init = torch.xpu._lazy_init
-        torch.cuda.BFloat16Tensor = torch.xpu.BFloat16Tensor
-        torch.cuda.DoubleStorage = torch.xpu.DoubleStorage
-        torch.cuda.ByteTensor = torch.xpu.ByteTensor
-        torch.cuda.StreamContext = torch.xpu.StreamContext
-        torch.cuda.ComplexDoubleStorage = torch.xpu.ComplexDoubleStorage
-        torch.cuda.ShortStorage = torch.xpu.ShortStorage
-        torch.cuda._lazy_call = torch.xpu._lazy_call
-        torch.cuda.HalfStorage = torch.xpu.HalfStorage
-        torch.cuda.random = torch.xpu.random
-        torch.cuda._device = torch.xpu._device
-        torch.cuda.classproperty = torch.xpu.classproperty
-        torch.cuda.__name__ = torch.xpu.__name__
-        torch.cuda._device_t = torch.xpu._device_t
-        torch.cuda.warnings = torch.xpu.warnings
-        torch.cuda.__spec__ = torch.xpu.__spec__
-        torch.cuda.BoolTensor = torch.xpu.BoolTensor
-        torch.cuda.CharStorage = torch.xpu.CharStorage
-        torch.cuda.__file__ = torch.xpu.__file__
-        torch.cuda._is_in_bad_fork = torch.xpu.lazy_init._is_in_bad_fork
-        # torch.cuda.is_current_stream_capturing = torch.xpu.is_current_stream_capturing
-
-        # Memory:
-        torch.cuda.memory = torch.xpu.memory
-        if 'linux' in sys.platform and "WSL2" in os.popen("uname -a").read():
-            torch.xpu.empty_cache = lambda: None
-        torch.cuda.empty_cache = torch.xpu.empty_cache
-        torch.cuda.memory_stats = torch.xpu.memory_stats
-        torch.cuda.memory_summary = torch.xpu.memory_summary
-        torch.cuda.memory_snapshot = torch.xpu.memory_snapshot
-        torch.cuda.memory_allocated = torch.xpu.memory_allocated
-        torch.cuda.max_memory_allocated = torch.xpu.max_memory_allocated
-        torch.cuda.memory_reserved = torch.xpu.memory_reserved
-        torch.cuda.memory_cached = torch.xpu.memory_reserved
-        torch.cuda.max_memory_reserved = torch.xpu.max_memory_reserved
-        torch.cuda.max_memory_cached = torch.xpu.max_memory_reserved
-        torch.cuda.reset_peak_memory_stats = torch.xpu.reset_peak_memory_stats
-        torch.cuda.reset_max_memory_cached = torch.xpu.reset_peak_memory_stats
-        torch.cuda.reset_max_memory_allocated = torch.xpu.reset_peak_memory_stats
-        torch.cuda.memory_stats_as_nested_dict = torch.xpu.memory_stats_as_nested_dict
-        torch.cuda.reset_accumulated_memory_stats = torch.xpu.reset_accumulated_memory_stats
-
-        # RNG:
-        torch.cuda.get_rng_state = torch.xpu.get_rng_state
-        torch.cuda.get_rng_state_all = torch.xpu.get_rng_state_all
-        torch.cuda.set_rng_state = torch.xpu.set_rng_state
-        torch.cuda.set_rng_state_all = torch.xpu.set_rng_state_all
-        torch.cuda.manual_seed = torch.xpu.manual_seed
-        torch.cuda.manual_seed_all = torch.xpu.manual_seed_all
-        torch.cuda.seed = torch.xpu.seed
-        torch.cuda.seed_all = torch.xpu.seed_all
-        torch.cuda.initial_seed = torch.xpu.initial_seed
-
-        # AMP:
-        torch.cuda.amp = torch.xpu.amp
-        if not hasattr(torch.cuda.amp, "common"):
-            torch.cuda.amp.common = contextlib.nullcontext()
-        torch.cuda.amp.common.amp_definitely_not_available = lambda: False
-        try:
-            torch.cuda.amp.GradScaler = torch.xpu.amp.GradScaler
-        except Exception: # pylint: disable=broad-exception-caught
+        if hasattr(torch, "cuda") and hasattr(torch.cuda, "is_xpu_hijacked") and torch.cuda.is_xpu_hijacked:
+            return True, "Skipping IPEX hijack"
+        else:
             try:
-                from .gradscaler import gradscaler_init # pylint: disable=import-outside-toplevel, import-error
-                gradscaler_init()
-                torch.cuda.amp.GradScaler = torch.xpu.amp.GradScaler
-            except Exception: # pylint: disable=broad-exception-caught
-                torch.cuda.amp.GradScaler = ipex.cpu.autocast._grad_scaler.GradScaler
+                # force xpu device on torch compile and triton
+                # import inductor utils to get around lazy import
+                from torch._inductor import utils as torch_inductor_utils # pylint: disable=import-error, unused-import # noqa: F401
+                torch._inductor.utils.GPU_TYPES = ["xpu"]
+                torch._inductor.utils.get_gpu_type = lambda *args, **kwargs: "xpu"
+                from triton import backends as triton_backends # pylint: disable=import-error
+                triton_backends.backends["nvidia"].driver.is_active = lambda *args, **kwargs: False
+            except Exception:
+                pass
+            # Replace cuda with xpu:
+            torch.cuda.current_device = torch.xpu.current_device
+            torch.cuda.current_stream = torch.xpu.current_stream
+            torch.cuda.device = torch.xpu.device
+            torch.cuda.device_count = torch.xpu.device_count
+            torch.cuda.device_of = torch.xpu.device_of
+            torch.cuda.get_device_name = torch.xpu.get_device_name
+            torch.cuda.get_device_properties = torch.xpu.get_device_properties
+            torch.cuda.init = torch.xpu.init
+            torch.cuda.is_available = torch.xpu.is_available
+            torch.cuda.is_initialized = torch.xpu.is_initialized
+            torch.cuda.is_current_stream_capturing = lambda: False
+            torch.cuda.stream = torch.xpu.stream
+            torch.cuda.Event = torch.xpu.Event
+            torch.cuda.Stream = torch.xpu.Stream
+            torch.Tensor.cuda = torch.Tensor.xpu
+            torch.Tensor.is_cuda = torch.Tensor.is_xpu
+            torch.nn.Module.cuda = torch.nn.Module.xpu
+            torch.cuda.Optional = torch.xpu.Optional
+            torch.cuda.__cached__ = torch.xpu.__cached__
+            torch.cuda.__loader__ = torch.xpu.__loader__
+            torch.cuda.streams = torch.xpu.streams
+            torch.cuda.Any = torch.xpu.Any
+            torch.cuda.__doc__ = torch.xpu.__doc__
+            torch.cuda.default_generators = torch.xpu.default_generators
+            torch.cuda._get_device_index = torch.xpu._get_device_index
+            torch.cuda.__path__ = torch.xpu.__path__
+            torch.cuda.set_stream = torch.xpu.set_stream
+            torch.cuda.torch = torch.xpu.torch
+            torch.cuda.Union = torch.xpu.Union
+            torch.cuda.__annotations__ = torch.xpu.__annotations__
+            torch.cuda.__package__ = torch.xpu.__package__
+            torch.cuda.__builtins__ = torch.xpu.__builtins__
+            torch.cuda._lazy_init = torch.xpu._lazy_init
+            torch.cuda.StreamContext = torch.xpu.StreamContext
+            torch.cuda._lazy_call = torch.xpu._lazy_call
+            torch.cuda.random = torch.xpu.random
+            torch.cuda._device = torch.xpu._device
+            torch.cuda.__name__ = torch.xpu.__name__
+            torch.cuda._device_t = torch.xpu._device_t
+            torch.cuda.__spec__ = torch.xpu.__spec__
+            torch.cuda.__file__ = torch.xpu.__file__
+            # torch.cuda.is_current_stream_capturing = torch.xpu.is_current_stream_capturing
 
-        # C
-        torch._C._cuda_getCurrentRawStream = ipex._C._getCurrentStream
-        ipex._C._DeviceProperties.multi_processor_count = ipex._C._DeviceProperties.gpu_eu_count
-        ipex._C._DeviceProperties.major = 2023
-        ipex._C._DeviceProperties.minor = 2
+            if torch_version < 2.3:
+                torch.cuda._initialization_lock = torch.xpu.lazy_init._initialization_lock
+                torch.cuda._initialized = torch.xpu.lazy_init._initialized
+                torch.cuda._is_in_bad_fork = torch.xpu.lazy_init._is_in_bad_fork
+                torch.cuda._lazy_seed_tracker = torch.xpu.lazy_init._lazy_seed_tracker
+                torch.cuda._queued_calls = torch.xpu.lazy_init._queued_calls
+                torch.cuda._tls = torch.xpu.lazy_init._tls
+                torch.cuda.threading = torch.xpu.lazy_init.threading
+                torch.cuda.traceback = torch.xpu.lazy_init.traceback
+                torch.cuda._lazy_new = torch.xpu._lazy_new
 
-        # Fix functions with ipex:
-        torch.cuda.mem_get_info = lambda device=None: [(torch.xpu.get_device_properties(device).total_memory - torch.xpu.memory_reserved(device)), torch.xpu.get_device_properties(device).total_memory]
-        torch._utils._get_available_device_type = lambda: "xpu"
-        torch.has_cuda = True
-        torch.cuda.has_half = True
-        torch.cuda.is_bf16_supported = lambda *args, **kwargs: True
-        torch.cuda.is_fp16_supported = lambda *args, **kwargs: True
-        torch.version.cuda = "11.7"
-        torch.cuda.get_device_capability = lambda *args, **kwargs: [11,7]
-        torch.cuda.get_device_properties.major = 11
-        torch.cuda.get_device_properties.minor = 7
-        torch.cuda.ipc_collect = lambda *args, **kwargs: None
-        torch.cuda.utilization = lambda *args, **kwargs: 0
+                torch.cuda.FloatTensor = torch.xpu.FloatTensor
+                torch.cuda.FloatStorage = torch.xpu.FloatStorage
+                torch.cuda.BFloat16Tensor = torch.xpu.BFloat16Tensor
+                torch.cuda.BFloat16Storage = torch.xpu.BFloat16Storage
+                torch.cuda.HalfTensor = torch.xpu.HalfTensor
+                torch.cuda.HalfStorage = torch.xpu.HalfStorage
+                torch.cuda.ByteTensor = torch.xpu.ByteTensor
+                torch.cuda.ByteStorage = torch.xpu.ByteStorage
+                torch.cuda.DoubleTensor = torch.xpu.DoubleTensor
+                torch.cuda.DoubleStorage = torch.xpu.DoubleStorage
+                torch.cuda.ShortTensor = torch.xpu.ShortTensor
+                torch.cuda.ShortStorage = torch.xpu.ShortStorage
+                torch.cuda.LongTensor = torch.xpu.LongTensor
+                torch.cuda.LongStorage = torch.xpu.LongStorage
+                torch.cuda.IntTensor = torch.xpu.IntTensor
+                torch.cuda.IntStorage = torch.xpu.IntStorage
+                torch.cuda.CharTensor = torch.xpu.CharTensor
+                torch.cuda.CharStorage = torch.xpu.CharStorage
+                torch.cuda.BoolTensor = torch.xpu.BoolTensor
+                torch.cuda.BoolStorage = torch.xpu.BoolStorage
+                torch.cuda.ComplexFloatStorage = torch.xpu.ComplexFloatStorage
+                torch.cuda.ComplexDoubleStorage = torch.xpu.ComplexDoubleStorage
+            else:
+                torch.cuda._initialization_lock = torch.xpu._initialization_lock
+                torch.cuda._initialized = torch.xpu._initialized
+                torch.cuda._is_in_bad_fork = torch.xpu._is_in_bad_fork
+                torch.cuda._lazy_seed_tracker = torch.xpu._lazy_seed_tracker
+                torch.cuda._queued_calls = torch.xpu._queued_calls
+                torch.cuda._tls = torch.xpu._tls
+                torch.cuda.threading = torch.xpu.threading
+                torch.cuda.traceback = torch.xpu.traceback
 
-        ipex_hijacks()
-        if not torch.xpu.has_fp64_dtype():
+            if torch_version < 2.5:
+                torch.cuda.os = torch.xpu.os
+                torch.cuda.Device = torch.xpu.Device
+                torch.cuda.warnings = torch.xpu.warnings
+                torch.cuda.classproperty = torch.xpu.classproperty
+                torch.UntypedStorage.cuda = torch.UntypedStorage.xpu
+
+            if torch_version < 2.7:
+                torch.cuda.Tuple = torch.xpu.Tuple
+                torch.cuda.List = torch.xpu.List
+
+
+            # Memory:
+            if 'linux' in sys.platform and "WSL2" in os.popen("uname -a").read():
+                torch.xpu.empty_cache = lambda: None
+            torch.cuda.empty_cache = torch.xpu.empty_cache
+
+            if has_ipex:
+                torch.cuda.memory_summary = torch.xpu.memory_summary
+                torch.cuda.memory_snapshot = torch.xpu.memory_snapshot
+            torch.cuda.memory = torch.xpu.memory
+            torch.cuda.memory_stats = torch.xpu.memory_stats
+            torch.cuda.memory_allocated = torch.xpu.memory_allocated
+            torch.cuda.max_memory_allocated = torch.xpu.max_memory_allocated
+            torch.cuda.memory_reserved = torch.xpu.memory_reserved
+            torch.cuda.memory_cached = torch.xpu.memory_reserved
+            torch.cuda.max_memory_reserved = torch.xpu.max_memory_reserved
+            torch.cuda.max_memory_cached = torch.xpu.max_memory_reserved
+            torch.cuda.reset_peak_memory_stats = torch.xpu.reset_peak_memory_stats
+            torch.cuda.reset_max_memory_cached = torch.xpu.reset_peak_memory_stats
+            torch.cuda.reset_max_memory_allocated = torch.xpu.reset_peak_memory_stats
+            torch.cuda.memory_stats_as_nested_dict = torch.xpu.memory_stats_as_nested_dict
+            torch.cuda.reset_accumulated_memory_stats = torch.xpu.reset_accumulated_memory_stats
+
+            # RNG:
+            torch.cuda.get_rng_state = torch.xpu.get_rng_state
+            torch.cuda.get_rng_state_all = torch.xpu.get_rng_state_all
+            torch.cuda.set_rng_state = torch.xpu.set_rng_state
+            torch.cuda.set_rng_state_all = torch.xpu.set_rng_state_all
+            torch.cuda.manual_seed = torch.xpu.manual_seed
+            torch.cuda.manual_seed_all = torch.xpu.manual_seed_all
+            torch.cuda.seed = torch.xpu.seed
+            torch.cuda.seed_all = torch.xpu.seed_all
+            torch.cuda.initial_seed = torch.xpu.initial_seed
+
+            # C
+            if torch_version < 2.3:
+                torch._C._cuda_getCurrentRawStream = ipex._C._getCurrentRawStream
+                ipex._C._DeviceProperties.multi_processor_count = ipex._C._DeviceProperties.gpu_subslice_count
+                ipex._C._DeviceProperties.major = 12
+                ipex._C._DeviceProperties.minor = 1
+                ipex._C._DeviceProperties.L2_cache_size = 16*1024*1024 # A770 and A750
+            else:
+                torch._C._cuda_getCurrentRawStream = torch._C._xpu_getCurrentRawStream
+                torch._C._XpuDeviceProperties.multi_processor_count = torch._C._XpuDeviceProperties.gpu_subslice_count
+                torch._C._XpuDeviceProperties.major = 12
+                torch._C._XpuDeviceProperties.minor = 1
+                torch._C._XpuDeviceProperties.L2_cache_size = 16*1024*1024 # A770 and A750
+
+            # Fix functions with ipex:
+            # torch.xpu.mem_get_info always returns the total memory as free memory
+            torch.xpu.mem_get_info = lambda device=None: [(torch.xpu.get_device_properties(device).total_memory - torch.xpu.memory_reserved(device)), torch.xpu.get_device_properties(device).total_memory]
+            torch.cuda.mem_get_info = torch.xpu.mem_get_info
+            torch._utils._get_available_device_type = lambda: "xpu"
+            torch.has_cuda = True
+            torch.cuda.has_half = True
+            torch.cuda.is_bf16_supported = getattr(torch.xpu, "is_bf16_supported", lambda *args, **kwargs: True)
+            torch.cuda.is_fp16_supported = lambda *args, **kwargs: True
+            torch.backends.cuda.is_built = lambda *args, **kwargs: True
+            torch.version.cuda = "12.1"
+            torch.cuda.get_arch_list = getattr(torch.xpu, "get_arch_list", lambda: ["pvc", "dg2", "ats-m150"])
+            torch.cuda.get_device_capability = lambda *args, **kwargs: (12,1)
+            torch.cuda.get_device_properties.major = 12
+            torch.cuda.get_device_properties.minor = 1
+            torch.cuda.get_device_properties.L2_cache_size = 16*1024*1024 # A770 and A750
+            torch.cuda.ipc_collect = lambda *args, **kwargs: None
+            torch.cuda.utilization = lambda *args, **kwargs: 0
+
+            device_supports_fp64 = ipex_hijacks()
             try:
                 from .diffusers import ipex_diffusers
-                ipex_diffusers()
+                ipex_diffusers(device_supports_fp64=device_supports_fp64)
             except Exception: # pylint: disable=broad-exception-caught
                 pass
+            torch.cuda.is_xpu_hijacked = True
     except Exception as e:
         return False, e
     return True, None

library/ipex/attention.py

@@ -1,175 +1,119 @@
 import os
 import torch
-import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
-from functools import cache
+from functools import cache, wraps
 
 # pylint: disable=protected-access, missing-function-docstring, line-too-long
 
-# ARC GPUs can't allocate more than 4GB to a single block so we slice the attetion layers
+# ARC GPUs can't allocate more than 4GB to a single block so we slice the attention layers
 
-sdpa_slice_trigger_rate = float(os.environ.get('IPEX_SDPA_SLICE_TRIGGER_RATE', 4))
-attention_slice_rate = float(os.environ.get('IPEX_ATTENTION_SLICE_RATE', 4))
+sdpa_slice_trigger_rate = float(os.environ.get('IPEX_SDPA_SLICE_TRIGGER_RATE', 1))
+attention_slice_rate = float(os.environ.get('IPEX_ATTENTION_SLICE_RATE', 0.5))
 
 # Find something divisible with the input_tokens
 @cache
-def find_slice_size(slice_size, slice_block_size):
-    while (slice_size * slice_block_size) > attention_slice_rate:
-        slice_size = slice_size // 2
-        if slice_size <= 1:
-            slice_size = 1
-            break
-    return slice_size
+def find_split_size(original_size, slice_block_size, slice_rate=2):
+    split_size = original_size
+    while True:
+        if (split_size * slice_block_size) <= slice_rate and original_size % split_size == 0:
+            return split_size
+        split_size = split_size - 1
+        if split_size <= 1:
+            return 1
+    return split_size
+
 
 # Find slice sizes for SDPA
 @cache
-def find_sdpa_slice_sizes(query_shape, query_element_size):
-    if len(query_shape) == 3:
-        batch_size_attention, query_tokens, shape_three = query_shape
-        shape_four = 1
-    else:
-        batch_size_attention, query_tokens, shape_three, shape_four = query_shape
+def find_sdpa_slice_sizes(query_shape, key_shape, query_element_size, slice_rate=2, trigger_rate=3):
+    batch_size, attn_heads, query_len, _ = query_shape
+    _, _, key_len, _ = key_shape
 
-    slice_block_size = query_tokens * shape_three * shape_four / 1024 / 1024 * query_element_size
-    block_size = batch_size_attention * slice_block_size
+    slice_batch_size = attn_heads * (query_len * key_len) * query_element_size / 1024 / 1024 / 1024
 
-    split_slice_size = batch_size_attention
-    split_2_slice_size = query_tokens
-    split_3_slice_size = shape_three
+    split_batch_size = batch_size
+    split_head_size = attn_heads
+    split_query_size = query_len
 
-    do_split = False
-    do_split_2 = False
-    do_split_3 = False
+    do_batch_split = False
+    do_head_split = False
+    do_query_split = False
 
-    if block_size > sdpa_slice_trigger_rate:
-        do_split = True
-        split_slice_size = find_slice_size(split_slice_size, slice_block_size)
-        if split_slice_size * slice_block_size > attention_slice_rate:
-            slice_2_block_size = split_slice_size * shape_three * shape_four / 1024 / 1024 * query_element_size
-            do_split_2 = True
-            split_2_slice_size = find_slice_size(split_2_slice_size, slice_2_block_size)
-            if split_2_slice_size * slice_2_block_size > attention_slice_rate:
-                slice_3_block_size = split_slice_size * split_2_slice_size * shape_four / 1024 / 1024 * query_element_size
-                do_split_3 = True
-                split_3_slice_size = find_slice_size(split_3_slice_size, slice_3_block_size)
+    if batch_size * slice_batch_size >= trigger_rate:
+        do_batch_split = True
+        split_batch_size = find_split_size(batch_size, slice_batch_size, slice_rate=slice_rate)
 
-    return do_split, do_split_2, do_split_3, split_slice_size, split_2_slice_size, split_3_slice_size
+        if split_batch_size * slice_batch_size > slice_rate:
+            slice_head_size = split_batch_size * (query_len * key_len) * query_element_size / 1024 / 1024 / 1024
+            do_head_split = True
+            split_head_size = find_split_size(attn_heads, slice_head_size, slice_rate=slice_rate)
 
-# Find slice sizes for BMM
-@cache
-def find_bmm_slice_sizes(input_shape, input_element_size, mat2_shape):
-    batch_size_attention, input_tokens, mat2_atten_shape = input_shape[0], input_shape[1], mat2_shape[2]
-    slice_block_size = input_tokens * mat2_atten_shape / 1024 / 1024 * input_element_size
-    block_size = batch_size_attention * slice_block_size
+            if split_head_size * slice_head_size > slice_rate:
+                slice_query_size = split_batch_size * split_head_size * (key_len) * query_element_size / 1024 / 1024 / 1024
+                do_query_split = True
+                split_query_size = find_split_size(query_len, slice_query_size, slice_rate=slice_rate)
 
-    split_slice_size = batch_size_attention
-    split_2_slice_size = input_tokens
-    split_3_slice_size = mat2_atten_shape
+    return do_batch_split, do_head_split, do_query_split, split_batch_size, split_head_size, split_query_size
 
-    do_split = False
-    do_split_2 = False
-    do_split_3 = False
-
-    if block_size > attention_slice_rate:
-        do_split = True
-        split_slice_size = find_slice_size(split_slice_size, slice_block_size)
-        if split_slice_size * slice_block_size > attention_slice_rate:
-            slice_2_block_size = split_slice_size * mat2_atten_shape / 1024 / 1024 * input_element_size
-            do_split_2 = True
-            split_2_slice_size = find_slice_size(split_2_slice_size, slice_2_block_size)
-            if split_2_slice_size * slice_2_block_size > attention_slice_rate:
-                slice_3_block_size = split_slice_size * split_2_slice_size / 1024 / 1024 * input_element_size
-                do_split_3 = True
-                split_3_slice_size = find_slice_size(split_3_slice_size, slice_3_block_size)
-
-    return do_split, do_split_2, do_split_3, split_slice_size, split_2_slice_size, split_3_slice_size
-
-
-original_torch_bmm = torch.bmm
-def torch_bmm_32_bit(input, mat2, *, out=None):
-    if input.device.type != "xpu":
-        return original_torch_bmm(input, mat2, out=out)
-    do_split, do_split_2, do_split_3, split_slice_size, split_2_slice_size, split_3_slice_size = find_bmm_slice_sizes(input.shape, input.element_size(), mat2.shape)
-
-    # Slice BMM
-    if do_split:
-        batch_size_attention, input_tokens, mat2_atten_shape = input.shape[0], input.shape[1], mat2.shape[2]
-        hidden_states = torch.zeros(input.shape[0], input.shape[1], mat2.shape[2], device=input.device, dtype=input.dtype)
-        for i in range(batch_size_attention // split_slice_size):
-            start_idx = i * split_slice_size
-            end_idx = (i + 1) * split_slice_size
-            if do_split_2:
-                for i2 in range(input_tokens // split_2_slice_size): # pylint: disable=invalid-name
-                    start_idx_2 = i2 * split_2_slice_size
-                    end_idx_2 = (i2 + 1) * split_2_slice_size
-                    if do_split_3:
-                        for i3 in range(mat2_atten_shape // split_3_slice_size): # pylint: disable=invalid-name
-                            start_idx_3 = i3 * split_3_slice_size
-                            end_idx_3 = (i3 + 1) * split_3_slice_size
-                            hidden_states[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3] = original_torch_bmm(
-                                input[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3],
-                                mat2[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3],
-                                out=out
-                            )
-                    else:
-                        hidden_states[start_idx:end_idx, start_idx_2:end_idx_2] = original_torch_bmm(
-                            input[start_idx:end_idx, start_idx_2:end_idx_2],
-                            mat2[start_idx:end_idx, start_idx_2:end_idx_2],
-                            out=out
-                        )
-            else:
-                hidden_states[start_idx:end_idx] = original_torch_bmm(
-                    input[start_idx:end_idx],
-                    mat2[start_idx:end_idx],
-                    out=out
-                )
-    else:
-        return original_torch_bmm(input, mat2, out=out)
-    return hidden_states
 
 original_scaled_dot_product_attention = torch.nn.functional.scaled_dot_product_attention
-def scaled_dot_product_attention_32_bit(query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False):
+@wraps(torch.nn.functional.scaled_dot_product_attention)
+def dynamic_scaled_dot_product_attention(query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False, **kwargs):
     if query.device.type != "xpu":
-        return original_scaled_dot_product_attention(query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal)
-    do_split, do_split_2, do_split_3, split_slice_size, split_2_slice_size, split_3_slice_size = find_sdpa_slice_sizes(query.shape, query.element_size())
+        return original_scaled_dot_product_attention(query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal, **kwargs)
+    is_unsqueezed = False
+    if query.dim() == 3:
+        query = query.unsqueeze(0)
+        is_unsqueezed = True
+        if key.dim() == 3:
+            key = key.unsqueeze(0)
+        if value.dim() == 3:
+            value = value.unsqueeze(0)
+    do_batch_split, do_head_split, do_query_split, split_batch_size, split_head_size, split_query_size = find_sdpa_slice_sizes(query.shape, key.shape, query.element_size(), slice_rate=attention_slice_rate, trigger_rate=sdpa_slice_trigger_rate)
 
     # Slice SDPA
-    if do_split:
-        batch_size_attention, query_tokens, shape_three = query.shape[0], query.shape[1], query.shape[2]
-        hidden_states = torch.zeros(query.shape, device=query.device, dtype=query.dtype)
-        for i in range(batch_size_attention // split_slice_size):
-            start_idx = i * split_slice_size
-            end_idx = (i + 1) * split_slice_size
-            if do_split_2:
-                for i2 in range(query_tokens // split_2_slice_size): # pylint: disable=invalid-name
-                    start_idx_2 = i2 * split_2_slice_size
-                    end_idx_2 = (i2 + 1) * split_2_slice_size
-                    if do_split_3:
-                        for i3 in range(shape_three // split_3_slice_size): # pylint: disable=invalid-name
-                            start_idx_3 = i3 * split_3_slice_size
-                            end_idx_3 = (i3 + 1) * split_3_slice_size
-                            hidden_states[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3] = original_scaled_dot_product_attention(
-                                query[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3],
-                                key[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3],
-                                value[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3],
-                                attn_mask=attn_mask[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3] if attn_mask is not None else attn_mask,
-                                dropout_p=dropout_p, is_causal=is_causal
+    if do_batch_split:
+        batch_size, attn_heads, query_len, _ = query.shape
+        _, _, _, head_dim = value.shape
+        hidden_states = torch.zeros((batch_size, attn_heads, query_len, head_dim), device=query.device, dtype=query.dtype)
+        if attn_mask is not None:
+            attn_mask = attn_mask.expand((query.shape[0], query.shape[1], query.shape[2], key.shape[-2]))
+        for ib in range(batch_size // split_batch_size):
+            start_idx = ib * split_batch_size
+            end_idx = (ib + 1) * split_batch_size
+            if do_head_split:
+                for ih in range(attn_heads // split_head_size): # pylint: disable=invalid-name
+                    start_idx_h = ih * split_head_size
+                    end_idx_h = (ih + 1) * split_head_size
+                    if do_query_split:
+                        for iq in range(query_len // split_query_size): # pylint: disable=invalid-name
+                            start_idx_q = iq * split_query_size
+                            end_idx_q = (iq + 1) * split_query_size
+                            hidden_states[start_idx:end_idx, start_idx_h:end_idx_h, start_idx_q:end_idx_q, :] = original_scaled_dot_product_attention(
+                                query[start_idx:end_idx, start_idx_h:end_idx_h, start_idx_q:end_idx_q, :],
+                                key[start_idx:end_idx, start_idx_h:end_idx_h, :, :],
+                                value[start_idx:end_idx, start_idx_h:end_idx_h, :, :],
+                                attn_mask=attn_mask[start_idx:end_idx, start_idx_h:end_idx_h, start_idx_q:end_idx_q, :] if attn_mask is not None else attn_mask,
+                                dropout_p=dropout_p, is_causal=is_causal, **kwargs
                             )
                     else:
-                        hidden_states[start_idx:end_idx, start_idx_2:end_idx_2] = original_scaled_dot_product_attention(
-                            query[start_idx:end_idx, start_idx_2:end_idx_2],
-                            key[start_idx:end_idx, start_idx_2:end_idx_2],
-                            value[start_idx:end_idx, start_idx_2:end_idx_2],
-                            attn_mask=attn_mask[start_idx:end_idx, start_idx_2:end_idx_2] if attn_mask is not None else attn_mask,
-                            dropout_p=dropout_p, is_causal=is_causal
+                        hidden_states[start_idx:end_idx, start_idx_h:end_idx_h, :, :] = original_scaled_dot_product_attention(
+                            query[start_idx:end_idx, start_idx_h:end_idx_h, :, :],
+                            key[start_idx:end_idx, start_idx_h:end_idx_h, :, :],
+                            value[start_idx:end_idx, start_idx_h:end_idx_h, :, :],
+                            attn_mask=attn_mask[start_idx:end_idx, start_idx_h:end_idx_h, :, :] if attn_mask is not None else attn_mask,
+                            dropout_p=dropout_p, is_causal=is_causal, **kwargs
                         )
             else:
-                hidden_states[start_idx:end_idx] = original_scaled_dot_product_attention(
-                    query[start_idx:end_idx],
-                    key[start_idx:end_idx],
-                    value[start_idx:end_idx],
-                    attn_mask=attn_mask[start_idx:end_idx] if attn_mask is not None else attn_mask,
-                    dropout_p=dropout_p, is_causal=is_causal
+                hidden_states[start_idx:end_idx, :, :, :] = original_scaled_dot_product_attention(
+                    query[start_idx:end_idx, :, :, :],
+                    key[start_idx:end_idx, :, :, :],
+                    value[start_idx:end_idx, :, :, :],
+                    attn_mask=attn_mask[start_idx:end_idx, :, :, :] if attn_mask is not None else attn_mask,
+                    dropout_p=dropout_p, is_causal=is_causal, **kwargs
                 )
+        torch.xpu.synchronize(query.device)
     else:
-        return original_scaled_dot_product_attention(query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal)
+        hidden_states = original_scaled_dot_product_attention(query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal, **kwargs)
+    if is_unsqueezed:
+        hidden_states = hidden_states.squeeze(0)
     return hidden_states

library/ipex/diffusers.py

@@ -1,310 +1,126 @@
-import os
+from functools import wraps
 import torch
-import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
-import diffusers #0.24.0 # pylint: disable=import-error
-from diffusers.models.attention_processor import Attention
-from diffusers.utils import USE_PEFT_BACKEND
-from functools import cache
+import diffusers # pylint: disable=import-error
+from diffusers.utils import torch_utils # pylint: disable=import-error, unused-import # noqa: F401
 
 # pylint: disable=protected-access, missing-function-docstring, line-too-long
 
-attention_slice_rate = float(os.environ.get('IPEX_ATTENTION_SLICE_RATE', 4))
 
-@cache
-def find_slice_size(slice_size, slice_block_size):
-    while (slice_size * slice_block_size) > attention_slice_rate:
-        slice_size = slice_size // 2
-        if slice_size <= 1:
-            slice_size = 1
-            break
-    return slice_size
-
-@cache
-def find_attention_slice_sizes(query_shape, query_element_size, query_device_type, slice_size=None):
-    if len(query_shape) == 3:
-        batch_size_attention, query_tokens, shape_three = query_shape
-        shape_four = 1
-    else:
-        batch_size_attention, query_tokens, shape_three, shape_four = query_shape
-    if slice_size is not None:
-        batch_size_attention = slice_size
-
-    slice_block_size = query_tokens * shape_three * shape_four / 1024 / 1024 * query_element_size
-    block_size = batch_size_attention * slice_block_size
-
-    split_slice_size = batch_size_attention
-    split_2_slice_size = query_tokens
-    split_3_slice_size = shape_three
-
-    do_split = False
-    do_split_2 = False
-    do_split_3 = False
-
-    if query_device_type != "xpu":
-        return do_split, do_split_2, do_split_3, split_slice_size, split_2_slice_size, split_3_slice_size
-
-    if block_size > attention_slice_rate:
-        do_split = True
-        split_slice_size = find_slice_size(split_slice_size, slice_block_size)
-        if split_slice_size * slice_block_size > attention_slice_rate:
-            slice_2_block_size = split_slice_size * shape_three * shape_four / 1024 / 1024 * query_element_size
-            do_split_2 = True
-            split_2_slice_size = find_slice_size(split_2_slice_size, slice_2_block_size)
-            if split_2_slice_size * slice_2_block_size > attention_slice_rate:
-                slice_3_block_size = split_slice_size * split_2_slice_size * shape_four / 1024 / 1024 * query_element_size
-                do_split_3 = True
-                split_3_slice_size = find_slice_size(split_3_slice_size, slice_3_block_size)
-
-    return do_split, do_split_2, do_split_3, split_slice_size, split_2_slice_size, split_3_slice_size
-
-class SlicedAttnProcessor: # pylint: disable=too-few-public-methods
-    r"""
-    Processor for implementing sliced attention.
-
-    Args:
-        slice_size (`int`, *optional*):
-            The number of steps to compute attention. Uses as many slices as `attention_head_dim // slice_size`, and
-            `attention_head_dim` must be a multiple of the `slice_size`.
-    """
-
-    def __init__(self, slice_size):
-        self.slice_size = slice_size
-
-    def __call__(self, attn: Attention, hidden_states: torch.FloatTensor,
-    encoder_hidden_states=None, attention_mask=None) -> torch.FloatTensor: # pylint: disable=too-many-statements, too-many-locals, too-many-branches
-
-        residual = hidden_states
-
-        input_ndim = hidden_states.ndim
-
-        if input_ndim == 4:
-            batch_size, channel, height, width = hidden_states.shape
-            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
-
-        batch_size, sequence_length, _ = (
-            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
-        )
-        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
-
-        if attn.group_norm is not None:
-            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
-
-        query = attn.to_q(hidden_states)
-        dim = query.shape[-1]
-        query = attn.head_to_batch_dim(query)
-
-        if encoder_hidden_states is None:
-            encoder_hidden_states = hidden_states
-        elif attn.norm_cross:
-            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
-
-        key = attn.to_k(encoder_hidden_states)
-        value = attn.to_v(encoder_hidden_states)
-        key = attn.head_to_batch_dim(key)
-        value = attn.head_to_batch_dim(value)
-
-        batch_size_attention, query_tokens, shape_three = query.shape
-        hidden_states = torch.zeros(
-            (batch_size_attention, query_tokens, dim // attn.heads), device=query.device, dtype=query.dtype
-        )
-
-        ####################################################################
-        # ARC GPUs can't allocate more than 4GB to a single block, Slice it:
-        _, do_split_2, do_split_3, split_slice_size, split_2_slice_size, split_3_slice_size = find_attention_slice_sizes(query.shape, query.element_size(), query.device.type, slice_size=self.slice_size)
-
-        for i in range(batch_size_attention // split_slice_size):
-            start_idx = i * split_slice_size
-            end_idx = (i + 1) * split_slice_size
-            if do_split_2:
-                for i2 in range(query_tokens // split_2_slice_size): # pylint: disable=invalid-name
-                    start_idx_2 = i2 * split_2_slice_size
-                    end_idx_2 = (i2 + 1) * split_2_slice_size
-                    if do_split_3:
-                        for i3 in range(shape_three // split_3_slice_size): # pylint: disable=invalid-name
-                            start_idx_3 = i3 * split_3_slice_size
-                            end_idx_3 = (i3 + 1) * split_3_slice_size
-
-                            query_slice = query[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3]
-                            key_slice = key[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3]
-                            attn_mask_slice = attention_mask[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3] if attention_mask is not None else None
-
-                            attn_slice = attn.get_attention_scores(query_slice, key_slice, attn_mask_slice)
-                            del query_slice
-                            del key_slice
-                            del attn_mask_slice
-                            attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3])
-
-                            hidden_states[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3] = attn_slice
-                            del attn_slice
-                    else:
-                        query_slice = query[start_idx:end_idx, start_idx_2:end_idx_2]
-                        key_slice = key[start_idx:end_idx, start_idx_2:end_idx_2]
-                        attn_mask_slice = attention_mask[start_idx:end_idx, start_idx_2:end_idx_2] if attention_mask is not None else None
-
-                        attn_slice = attn.get_attention_scores(query_slice, key_slice, attn_mask_slice)
-                        del query_slice
-                        del key_slice
-                        del attn_mask_slice
-                        attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx, start_idx_2:end_idx_2])
-
-                        hidden_states[start_idx:end_idx, start_idx_2:end_idx_2] = attn_slice
-                        del attn_slice
-            else:
-                query_slice = query[start_idx:end_idx]
-                key_slice = key[start_idx:end_idx]
-                attn_mask_slice = attention_mask[start_idx:end_idx] if attention_mask is not None else None
-
-                attn_slice = attn.get_attention_scores(query_slice, key_slice, attn_mask_slice)
-                del query_slice
-                del key_slice
-                del attn_mask_slice
-                attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx])
-
-                hidden_states[start_idx:end_idx] = attn_slice
-                del attn_slice
-        ####################################################################
-
-        hidden_states = attn.batch_to_head_dim(hidden_states)
-
-        # linear proj
-        hidden_states = attn.to_out[0](hidden_states)
-        # dropout
-        hidden_states = attn.to_out[1](hidden_states)
-
-        if input_ndim == 4:
-            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
-
-        if attn.residual_connection:
-            hidden_states = hidden_states + residual
-
-        hidden_states = hidden_states / attn.rescale_output_factor
-
-        return hidden_states
+# Diffusers FreeU
+# Diffusers is imported before ipex hijacks so fourier_filter needs hijacking too
+original_fourier_filter = diffusers.utils.torch_utils.fourier_filter
+@wraps(diffusers.utils.torch_utils.fourier_filter)
+def fourier_filter(x_in, threshold, scale):
+    return_dtype = x_in.dtype
+    return original_fourier_filter(x_in.to(dtype=torch.float32), threshold, scale).to(dtype=return_dtype)
 
 
-class AttnProcessor:
-    r"""
-    Default processor for performing attention-related computations.
-    """
+# fp64 error
+class FluxPosEmbed(torch.nn.Module):
+    def __init__(self, theta: int, axes_dim):
+        super().__init__()
+        self.theta = theta
+        self.axes_dim = axes_dim
 
-    def __call__(self, attn: Attention, hidden_states: torch.FloatTensor,
-    encoder_hidden_states=None, attention_mask=None,
-    temb=None, scale: float = 1.0) -> torch.Tensor: # pylint: disable=too-many-statements, too-many-locals, too-many-branches
+    def forward(self, ids: torch.Tensor) -> torch.Tensor:
+        n_axes = ids.shape[-1]
+        cos_out = []
+        sin_out = []
+        pos = ids.float()
+        for i in range(n_axes):
+            cos, sin = diffusers.models.embeddings.get_1d_rotary_pos_embed(
+                self.axes_dim[i],
+                pos[:, i],
+                theta=self.theta,
+                repeat_interleave_real=True,
+                use_real=True,
+                freqs_dtype=torch.float32,
+            )
+            cos_out.append(cos)
+            sin_out.append(sin)
+        freqs_cos = torch.cat(cos_out, dim=-1).to(ids.device)
+        freqs_sin = torch.cat(sin_out, dim=-1).to(ids.device)
+        return freqs_cos, freqs_sin
 
-        residual = hidden_states
 
-        args = () if USE_PEFT_BACKEND else (scale,)
+def hidream_rope(pos: torch.Tensor, dim: int, theta: int) -> torch.Tensor:
+    assert dim % 2 == 0, "The dimension must be even."
+    return_device = pos.device
+    pos = pos.to("cpu")
 
-        if attn.spatial_norm is not None:
-            hidden_states = attn.spatial_norm(hidden_states, temb)
+    scale = torch.arange(0, dim, 2, dtype=torch.float64, device=pos.device) / dim
+    omega = 1.0 / (theta**scale)
 
-        input_ndim = hidden_states.ndim
+    batch_size, seq_length = pos.shape
+    out = torch.einsum("...n,d->...nd", pos, omega)
+    cos_out = torch.cos(out)
+    sin_out = torch.sin(out)
 
-        if input_ndim == 4:
-            batch_size, channel, height, width = hidden_states.shape
-            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+    stacked_out = torch.stack([cos_out, -sin_out, sin_out, cos_out], dim=-1)
+    out = stacked_out.view(batch_size, -1, dim // 2, 2, 2)
+    return out.to(return_device, dtype=torch.float32)
 
-        batch_size, sequence_length, _ = (
-            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
-        )
-        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
 
-        if attn.group_norm is not None:
-            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos, output_type="np"):
+    if output_type == "np":
+        return diffusers.models.embeddings.get_1d_sincos_pos_embed_from_grid_np(embed_dim=embed_dim, pos=pos)
+    if embed_dim % 2 != 0:
+        raise ValueError("embed_dim must be divisible by 2")
 
-        query = attn.to_q(hidden_states, *args)
+    omega = torch.arange(embed_dim // 2, device=pos.device, dtype=torch.float32)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / 10000**omega  # (D/2,)
 
-        if encoder_hidden_states is None:
-            encoder_hidden_states = hidden_states
-        elif attn.norm_cross:
-            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+    pos = pos.reshape(-1)  # (M,)
+    out = torch.outer(pos, omega)  # (M, D/2), outer product
 
-        key = attn.to_k(encoder_hidden_states, *args)
-        value = attn.to_v(encoder_hidden_states, *args)
+    emb_sin = torch.sin(out)  # (M, D/2)
+    emb_cos = torch.cos(out)  # (M, D/2)
 
-        query = attn.head_to_batch_dim(query)
-        key = attn.head_to_batch_dim(key)
-        value = attn.head_to_batch_dim(value)
+    emb = torch.concat([emb_sin, emb_cos], dim=1)  # (M, D)
+    return emb
 
-        ####################################################################
-        # ARC GPUs can't allocate more than 4GB to a single block, Slice it:
-        batch_size_attention, query_tokens, shape_three = query.shape[0], query.shape[1], query.shape[2]
-        hidden_states = torch.zeros(query.shape, device=query.device, dtype=query.dtype)
-        do_split, do_split_2, do_split_3, split_slice_size, split_2_slice_size, split_3_slice_size = find_attention_slice_sizes(query.shape, query.element_size(), query.device.type)
 
-        if do_split:
-            for i in range(batch_size_attention // split_slice_size):
-                start_idx = i * split_slice_size
-                end_idx = (i + 1) * split_slice_size
-                if do_split_2:
-                    for i2 in range(query_tokens // split_2_slice_size): # pylint: disable=invalid-name
-                        start_idx_2 = i2 * split_2_slice_size
-                        end_idx_2 = (i2 + 1) * split_2_slice_size
-                        if do_split_3:
-                            for i3 in range(shape_three // split_3_slice_size): # pylint: disable=invalid-name
-                                start_idx_3 = i3 * split_3_slice_size
-                                end_idx_3 = (i3 + 1) * split_3_slice_size
+def apply_rotary_emb(x, freqs_cis, use_real: bool = True, use_real_unbind_dim: int = -1):
+    if use_real:
+        cos, sin = freqs_cis  # [S, D]
+        cos = cos[None, None]
+        sin = sin[None, None]
+        cos, sin = cos.to(x.device), sin.to(x.device)
 
-                                query_slice = query[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3]
-                                key_slice = key[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3]
-                                attn_mask_slice = attention_mask[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3] if attention_mask is not None else None
-
-                                attn_slice = attn.get_attention_scores(query_slice, key_slice, attn_mask_slice)
-                                del query_slice
-                                del key_slice
-                                del attn_mask_slice
-                                attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3])
-
-                                hidden_states[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3] = attn_slice
-                                del attn_slice
-                        else:
-                            query_slice = query[start_idx:end_idx, start_idx_2:end_idx_2]
-                            key_slice = key[start_idx:end_idx, start_idx_2:end_idx_2]
-                            attn_mask_slice = attention_mask[start_idx:end_idx, start_idx_2:end_idx_2] if attention_mask is not None else None
-
-                            attn_slice = attn.get_attention_scores(query_slice, key_slice, attn_mask_slice)
-                            del query_slice
-                            del key_slice
-                            del attn_mask_slice
-                            attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx, start_idx_2:end_idx_2])
-
-                            hidden_states[start_idx:end_idx, start_idx_2:end_idx_2] = attn_slice
-                            del attn_slice
-                else:
-                    query_slice = query[start_idx:end_idx]
-                    key_slice = key[start_idx:end_idx]
-                    attn_mask_slice = attention_mask[start_idx:end_idx] if attention_mask is not None else None
-
-                    attn_slice = attn.get_attention_scores(query_slice, key_slice, attn_mask_slice)
-                    del query_slice
-                    del key_slice
-                    del attn_mask_slice
-                    attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx])
-
-                    hidden_states[start_idx:end_idx] = attn_slice
-                    del attn_slice
+        if use_real_unbind_dim == -1:
+            # Used for flux, cogvideox, hunyuan-dit
+            x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
+            x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
+        elif use_real_unbind_dim == -2:
+            # Used for Stable Audio, OmniGen, CogView4 and Cosmos
+            x_real, x_imag = x.reshape(*x.shape[:-1], 2, -1).unbind(-2)  # [B, S, H, D//2]
+            x_rotated = torch.cat([-x_imag, x_real], dim=-1)
         else:
-            attention_probs = attn.get_attention_scores(query, key, attention_mask)
-            hidden_states = torch.bmm(attention_probs, value)
-        ####################################################################
-        hidden_states = attn.batch_to_head_dim(hidden_states)
+            raise ValueError(f"`use_real_unbind_dim={use_real_unbind_dim}` but should be -1 or -2.")
 
-        # linear proj
-        hidden_states = attn.to_out[0](hidden_states, *args)
-        # dropout
-        hidden_states = attn.to_out[1](hidden_states)
+        out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
+        return out
+    else:
+        # used for lumina
+        # force cpu with Alchemist
+        x_rotated = torch.view_as_complex(x.to("cpu").float().reshape(*x.shape[:-1], -1, 2))
+        freqs_cis = freqs_cis.to("cpu").unsqueeze(2)
+        x_out = torch.view_as_real(x_rotated * freqs_cis).flatten(3)
+        return x_out.type_as(x).to(x.device)
 
-        if input_ndim == 4:
-            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
 
-        if attn.residual_connection:
-            hidden_states = hidden_states + residual
-
-        hidden_states = hidden_states / attn.rescale_output_factor
-
-        return hidden_states
-
-def ipex_diffusers():
-    #ARC GPUs can't allocate more than 4GB to a single block:
-    diffusers.models.attention_processor.SlicedAttnProcessor = SlicedAttnProcessor
-    diffusers.models.attention_processor.AttnProcessor = AttnProcessor
+def ipex_diffusers(device_supports_fp64=False):
+    diffusers.utils.torch_utils.fourier_filter = fourier_filter
+    if not device_supports_fp64:
+        # get around lazy imports
+        from diffusers.models import embeddings as diffusers_embeddings # pylint: disable=import-error, unused-import # noqa: F401
+        from diffusers.models import transformers as diffusers_transformers # pylint: disable=import-error, unused-import # noqa: F401
+        from diffusers.models import controlnets as diffusers_controlnets # pylint: disable=import-error, unused-import # noqa: F401
+        diffusers.models.embeddings.get_1d_sincos_pos_embed_from_grid = get_1d_sincos_pos_embed_from_grid
+        diffusers.models.embeddings.FluxPosEmbed = FluxPosEmbed
+        diffusers.models.embeddings.apply_rotary_emb = apply_rotary_emb
+        diffusers.models.transformers.transformer_flux.FluxPosEmbed = FluxPosEmbed
+        diffusers.models.transformers.transformer_lumina2.apply_rotary_emb = apply_rotary_emb
+        diffusers.models.controlnets.controlnet_flux.FluxPosEmbed = FluxPosEmbed
+        diffusers.models.transformers.transformer_hidream_image.rope = hidream_rope

library/ipex/gradscaler.py

@@ -1,183 +0,0 @@
-from collections import defaultdict
-import torch
-import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
-import intel_extension_for_pytorch._C as core # pylint: disable=import-error, unused-import
-
-# pylint: disable=protected-access, missing-function-docstring, line-too-long
-
-device_supports_fp64 = torch.xpu.has_fp64_dtype()
-OptState = ipex.cpu.autocast._grad_scaler.OptState
-_MultiDeviceReplicator = ipex.cpu.autocast._grad_scaler._MultiDeviceReplicator
-_refresh_per_optimizer_state = ipex.cpu.autocast._grad_scaler._refresh_per_optimizer_state
-
-def _unscale_grads_(self, optimizer, inv_scale, found_inf, allow_fp16): # pylint: disable=unused-argument
-    per_device_inv_scale = _MultiDeviceReplicator(inv_scale)
-    per_device_found_inf = _MultiDeviceReplicator(found_inf)
-
-    # To set up _amp_foreach_non_finite_check_and_unscale_, split grads by device and dtype.
-    # There could be hundreds of grads, so we'd like to iterate through them just once.
-    # However, we don't know their devices or dtypes in advance.
-
-    # https://stackoverflow.com/questions/5029934/defaultdict-of-defaultdict
-    # Google says mypy struggles with defaultdicts type annotations.
-    per_device_and_dtype_grads = defaultdict(lambda: defaultdict(list))  # type: ignore[var-annotated]
-    # sync grad to master weight
-    if hasattr(optimizer, "sync_grad"):
-        optimizer.sync_grad()
-    with torch.no_grad():
-        for group in optimizer.param_groups:
-            for param in group["params"]:
-                if param.grad is None:
-                    continue
-                if (not allow_fp16) and param.grad.dtype == torch.float16:
-                    raise ValueError("Attempting to unscale FP16 gradients.")
-                if param.grad.is_sparse:
-                    # is_coalesced() == False means the sparse grad has values with duplicate indices.
-                    # coalesce() deduplicates indices and adds all values that have the same index.
-                    # For scaled fp16 values, there's a good chance coalescing will cause overflow,
-                    # so we should check the coalesced _values().
-                    if param.grad.dtype is torch.float16:
-                        param.grad = param.grad.coalesce()
-                    to_unscale = param.grad._values()
-                else:
-                    to_unscale = param.grad
-
-                # -: is there a way to split by device and dtype without appending in the inner loop?
-                to_unscale = to_unscale.to("cpu")
-                per_device_and_dtype_grads[to_unscale.device][
-                    to_unscale.dtype
-                ].append(to_unscale)
-
-        for _, per_dtype_grads in per_device_and_dtype_grads.items():
-            for grads in per_dtype_grads.values():
-                core._amp_foreach_non_finite_check_and_unscale_(
-                    grads,
-                    per_device_found_inf.get("cpu"),
-                    per_device_inv_scale.get("cpu"),
-                )
-
-    return per_device_found_inf._per_device_tensors
-
-def unscale_(self, optimizer):
-    """
-    Divides ("unscales") the optimizer's gradient tensors by the scale factor.
-    :meth:`unscale_` is optional, serving cases where you need to
-    :ref:`modify or inspect gradients<working-with-unscaled-gradients>`
-    between the backward pass(es) and :meth:`step`.
-    If :meth:`unscale_` is not called explicitly,  gradients will be unscaled  automatically during :meth:`step`.
-    Simple example, using :meth:`unscale_` to enable clipping of unscaled gradients::
-        ...
-        scaler.scale(loss).backward()
-        scaler.unscale_(optimizer)
-        torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm)
-        scaler.step(optimizer)
-        scaler.update()
-    Args:
-        optimizer (torch.optim.Optimizer):  Optimizer that owns the gradients to be unscaled.
-    .. warning::
-        :meth:`unscale_` should only be called once per optimizer per :meth:`step` call,
-        and only after all gradients for that optimizer's assigned parameters have been accumulated.
-        Calling :meth:`unscale_` twice for a given optimizer between each :meth:`step` triggers a RuntimeError.
-    .. warning::
-        :meth:`unscale_` may unscale sparse gradients out of place, replacing the ``.grad`` attribute.
-    """
-    if not self._enabled:
-        return
-
-    self._check_scale_growth_tracker("unscale_")
-
-    optimizer_state = self._per_optimizer_states[id(optimizer)]
-
-    if optimizer_state["stage"] is OptState.UNSCALED: # pylint: disable=no-else-raise
-        raise RuntimeError(
-            "unscale_() has already been called on this optimizer since the last update()."
-        )
-    elif optimizer_state["stage"] is OptState.STEPPED:
-        raise RuntimeError("unscale_() is being called after step().")
-
-    # FP32 division can be imprecise for certain compile options, so we carry out the reciprocal in FP64.
-    assert self._scale is not None
-    if device_supports_fp64:
-        inv_scale = self._scale.double().reciprocal().float()
-    else:
-        inv_scale = self._scale.to("cpu").double().reciprocal().float().to(self._scale.device)
-    found_inf = torch.full(
-        (1,), 0.0, dtype=torch.float32, device=self._scale.device
-    )
-
-    optimizer_state["found_inf_per_device"] = self._unscale_grads_(
-        optimizer, inv_scale, found_inf, False
-    )
-    optimizer_state["stage"] = OptState.UNSCALED
-
-def update(self, new_scale=None):
-    """
-    Updates the scale factor.
-    If any optimizer steps were skipped the scale is multiplied by ``backoff_factor``
-    to reduce it. If ``growth_interval`` unskipped iterations occurred consecutively,
-    the scale is multiplied by ``growth_factor`` to increase it.
-    Passing ``new_scale`` sets the new scale value manually. (``new_scale`` is not
-    used directly, it's used to fill GradScaler's internal scale tensor. So if
-    ``new_scale`` was a tensor, later in-place changes to that tensor will not further
-    affect the scale GradScaler uses internally.)
-    Args:
-        new_scale (float or :class:`torch.FloatTensor`, optional, default=None):  New scale factor.
-    .. warning::
-        :meth:`update` should only be called at the end of the iteration, after ``scaler.step(optimizer)`` has
-        been invoked for all optimizers used this iteration.
-    """
-    if not self._enabled:
-        return
-
-    _scale, _growth_tracker = self._check_scale_growth_tracker("update")
-
-    if new_scale is not None:
-        # Accept a new user-defined scale.
-        if isinstance(new_scale, float):
-            self._scale.fill_(new_scale)  # type: ignore[union-attr]
-        else:
-            reason = "new_scale should be a float or a 1-element torch.FloatTensor with requires_grad=False."
-            assert isinstance(new_scale, torch.FloatTensor), reason  # type: ignore[attr-defined]
-            assert new_scale.numel() == 1, reason
-            assert new_scale.requires_grad is False, reason
-            self._scale.copy_(new_scale)  # type: ignore[union-attr]
-    else:
-        # Consume shared inf/nan data collected from optimizers to update the scale.
-        # If all found_inf tensors are on the same device as self._scale, this operation is asynchronous.
-        found_infs = [
-            found_inf.to(device="cpu", non_blocking=True)
-            for state in self._per_optimizer_states.values()
-            for found_inf in state["found_inf_per_device"].values()
-        ]
-
-        assert len(found_infs) > 0, "No inf checks were recorded prior to update."
-
-        found_inf_combined = found_infs[0]
-        if len(found_infs) > 1:
-            for i in range(1, len(found_infs)):
-                found_inf_combined += found_infs[i]
-
-        to_device = _scale.device
-        _scale = _scale.to("cpu")
-        _growth_tracker = _growth_tracker.to("cpu")
-
-        core._amp_update_scale_(
-            _scale,
-            _growth_tracker,
-            found_inf_combined,
-            self._growth_factor,
-            self._backoff_factor,
-            self._growth_interval,
-        )
-
-        _scale = _scale.to(to_device)
-        _growth_tracker = _growth_tracker.to(to_device)
-    # To prepare for next iteration, clear the data collected from optimizers this iteration.
-    self._per_optimizer_states = defaultdict(_refresh_per_optimizer_state)
-
-def gradscaler_init():
-    torch.xpu.amp.GradScaler = ipex.cpu.autocast._grad_scaler.GradScaler
-    torch.xpu.amp.GradScaler._unscale_grads_ = _unscale_grads_
-    torch.xpu.amp.GradScaler.unscale_ = unscale_
-    torch.xpu.amp.GradScaler.update = update
-    return torch.xpu.amp.GradScaler

library/ipex/hijacks.py

@@ -1,6 +1,26 @@
-import contextlib
+import os
+from functools import wraps
+from contextlib import nullcontext
 import torch
-import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
+import numpy as np
+
+torch_version = float(torch.__version__[:3])
+current_xpu_device = f"xpu:{torch.xpu.current_device()}"
+device_supports_fp64 = torch.xpu.has_fp64_dtype() if hasattr(torch.xpu, "has_fp64_dtype") else torch.xpu.get_device_properties(current_xpu_device).has_fp64
+
+if os.environ.get('IPEX_FORCE_ATTENTION_SLICE', '0') == '0':
+    if (torch.xpu.get_device_properties(current_xpu_device).total_memory / 1024 / 1024 / 1024) > 4.1:
+        try:
+            x = torch.ones((33000,33000), dtype=torch.float32, device=current_xpu_device)
+            del x
+            torch.xpu.empty_cache()
+            use_dynamic_attention = False
+        except Exception:
+            use_dynamic_attention = True
+    else:
+        use_dynamic_attention = True
+else:
+    use_dynamic_attention = bool(os.environ.get('IPEX_FORCE_ATTENTION_SLICE', '0') == '1')
 
 # pylint: disable=protected-access, missing-function-docstring, line-too-long, unnecessary-lambda, no-else-return
 
@@ -8,34 +28,71 @@ class DummyDataParallel(torch.nn.Module): # pylint: disable=missing-class-docstr
     def __new__(cls, module, device_ids=None, output_device=None, dim=0): # pylint: disable=unused-argument
         if isinstance(device_ids, list) and len(device_ids) > 1:
             print("IPEX backend doesn't support DataParallel on multiple XPU devices")
-        return module.to("xpu")
+        return module.to(f"xpu:{torch.xpu.current_device()}")
 
 def return_null_context(*args, **kwargs): # pylint: disable=unused-argument
-    return contextlib.nullcontext()
+    return nullcontext()
 
 @property
 def is_cuda(self):
-    return self.device.type == 'xpu' or self.device.type == 'cuda'
+    return self.device.type == "xpu" or self.device.type == "cuda"
 
-def check_device(device):
-    return bool((isinstance(device, torch.device) and device.type == "cuda") or (isinstance(device, str) and "cuda" in device) or isinstance(device, int))
+def check_device_type(device, device_type: str) -> bool:
+    if device is None or type(device) not in {str, int, torch.device}:
+        return False
+    else:
+        return bool(torch.device(device).type == device_type)
 
-def return_xpu(device):
-    return f"xpu:{device.split(':')[-1]}" if isinstance(device, str) and ":" in device else f"xpu:{device}" if isinstance(device, int) else torch.device("xpu") if isinstance(device, torch.device) else "xpu"
+def check_cuda(device) -> bool:
+    return bool(isinstance(device, int) or check_device_type(device, "cuda"))
+
+def return_xpu(device): # keep the device instance type, aka return string if the input is string
+    return f"xpu:{torch.xpu.current_device()}" if device is None else f"xpu:{device.split(':')[-1]}" if isinstance(device, str) and ":" in device else f"xpu:{device}" if isinstance(device, int) else torch.device(f"xpu:{device.index}" if device.index is not None else "xpu") if isinstance(device, torch.device) else "xpu"
 
 
 # Autocast
-original_autocast = torch.autocast
-def ipex_autocast(*args, **kwargs):
-    if len(args) > 0 and args[0] == "cuda":
-        return original_autocast("xpu", *args[1:], **kwargs)
+original_autocast_init = torch.amp.autocast_mode.autocast.__init__
+@wraps(torch.amp.autocast_mode.autocast.__init__)
+def autocast_init(self, device_type=None, dtype=None, enabled=True, cache_enabled=None):
+    if device_type is None or check_cuda(device_type):
+        return original_autocast_init(self, device_type="xpu", dtype=dtype, enabled=enabled, cache_enabled=cache_enabled)
     else:
-        return original_autocast(*args, **kwargs)
+        return original_autocast_init(self, device_type=device_type, dtype=dtype, enabled=enabled, cache_enabled=cache_enabled)
+
+
+original_grad_scaler_init = torch.amp.grad_scaler.GradScaler.__init__
+@wraps(torch.amp.grad_scaler.GradScaler.__init__)
+def GradScaler_init(self, device: str = None, init_scale: float = 2.0**16, growth_factor: float = 2.0, backoff_factor: float = 0.5, growth_interval: int = 2000, enabled: bool = True):
+    if device is None or check_cuda(device):
+        return original_grad_scaler_init(self, device=return_xpu(device), init_scale=init_scale, growth_factor=growth_factor, backoff_factor=backoff_factor, growth_interval=growth_interval, enabled=enabled)
+    else:
+        return original_grad_scaler_init(self, device=device, init_scale=init_scale, growth_factor=growth_factor, backoff_factor=backoff_factor, growth_interval=growth_interval, enabled=enabled)
+
+
+original_is_autocast_enabled = torch.is_autocast_enabled
+@wraps(torch.is_autocast_enabled)
+def torch_is_autocast_enabled(device_type=None):
+    if device_type is None or check_cuda(device_type):
+        return original_is_autocast_enabled(return_xpu(device_type))
+    else:
+        return original_is_autocast_enabled(device_type)
+
+
+original_get_autocast_dtype = torch.get_autocast_dtype
+@wraps(torch.get_autocast_dtype)
+def torch_get_autocast_dtype(device_type=None):
+    if device_type is None or check_cuda(device_type) or check_device_type(device_type, "xpu"):
+        return torch.bfloat16
+    else:
+        return original_get_autocast_dtype(device_type)
+
 
 # Latent Antialias CPU Offload:
+# IPEX 2.5 and above has partial support but doesn't really work most of the time.
 original_interpolate = torch.nn.functional.interpolate
+@wraps(torch.nn.functional.interpolate)
 def interpolate(tensor, size=None, scale_factor=None, mode='nearest', align_corners=None, recompute_scale_factor=None, antialias=False): # pylint: disable=too-many-arguments
-    if antialias or align_corners is not None:
+    if mode in {'bicubic', 'bilinear'}:
         return_device = tensor.device
         return_dtype = tensor.dtype
         return original_interpolate(tensor.to("cpu", dtype=torch.float32), size=size, scale_factor=scale_factor, mode=mode,
@@ -44,42 +101,71 @@ def interpolate(tensor, size=None, scale_factor=None, mode='nearest', align_corn
         return original_interpolate(tensor, size=size, scale_factor=scale_factor, mode=mode,
         align_corners=align_corners, recompute_scale_factor=recompute_scale_factor, antialias=antialias)
 
+
 # Diffusers Float64 (Alchemist GPUs doesn't support 64 bit):
 original_from_numpy = torch.from_numpy
+@wraps(torch.from_numpy)
 def from_numpy(ndarray):
     if ndarray.dtype == float:
-        return original_from_numpy(ndarray.astype('float32'))
+        return original_from_numpy(ndarray.astype("float32"))
     else:
         return original_from_numpy(ndarray)
 
-if torch.xpu.has_fp64_dtype():
-    original_torch_bmm = torch.bmm
+original_as_tensor = torch.as_tensor
+@wraps(torch.as_tensor)
+def as_tensor(data, dtype=None, device=None):
+    if check_cuda(device):
+        device = return_xpu(device)
+    if isinstance(data, np.ndarray) and data.dtype == float and not check_device_type(device, "cpu"):
+        return original_as_tensor(data, dtype=torch.float32, device=device)
+    else:
+        return original_as_tensor(data, dtype=dtype, device=device)
+
+
+if not use_dynamic_attention:
     original_scaled_dot_product_attention = torch.nn.functional.scaled_dot_product_attention
 else:
     # 32 bit attention workarounds for Alchemist:
     try:
-        from .attention import torch_bmm_32_bit as original_torch_bmm
-        from .attention import scaled_dot_product_attention_32_bit as original_scaled_dot_product_attention
+        from .attention import dynamic_scaled_dot_product_attention as original_scaled_dot_product_attention
     except Exception: # pylint: disable=broad-exception-caught
-        original_torch_bmm = torch.bmm
         original_scaled_dot_product_attention = torch.nn.functional.scaled_dot_product_attention
 
-
-# Data Type Errors:
-def torch_bmm(input, mat2, *, out=None):
-    if input.dtype != mat2.dtype:
-        mat2 = mat2.to(input.dtype)
-    return original_torch_bmm(input, mat2, out=out)
-
-def scaled_dot_product_attention(query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False):
+@wraps(torch.nn.functional.scaled_dot_product_attention)
+def scaled_dot_product_attention(query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False, **kwargs):
     if query.dtype != key.dtype:
         key = key.to(dtype=query.dtype)
     if query.dtype != value.dtype:
         value = value.to(dtype=query.dtype)
-    return original_scaled_dot_product_attention(query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal)
+    if attn_mask is not None and query.dtype != attn_mask.dtype:
+        attn_mask = attn_mask.to(dtype=query.dtype)
+    return original_scaled_dot_product_attention(query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal, **kwargs)
+
+# Data Type Errors:
+original_torch_bmm = torch.bmm
+@wraps(torch.bmm)
+def torch_bmm(input, mat2, *, out=None):
+    if input.dtype != mat2.dtype:
+        mat2 = mat2.to(dtype=input.dtype)
+    return original_torch_bmm(input, mat2, out=out)
+
+# Diffusers FreeU
+original_fft_fftn = torch.fft.fftn
+@wraps(torch.fft.fftn)
+def fft_fftn(input, s=None, dim=None, norm=None, *, out=None):
+    return_dtype = input.dtype
+    return original_fft_fftn(input.to(dtype=torch.float32), s=s, dim=dim, norm=norm, out=out).to(dtype=return_dtype)
+
+# Diffusers FreeU
+original_fft_ifftn = torch.fft.ifftn
+@wraps(torch.fft.ifftn)
+def fft_ifftn(input, s=None, dim=None, norm=None, *, out=None):
+    return_dtype = input.dtype
+    return original_fft_ifftn(input.to(dtype=torch.float32), s=s, dim=dim, norm=norm, out=out).to(dtype=return_dtype)
 
 # A1111 FP16
 original_functional_group_norm = torch.nn.functional.group_norm
+@wraps(torch.nn.functional.group_norm)
 def functional_group_norm(input, num_groups, weight=None, bias=None, eps=1e-05):
     if weight is not None and input.dtype != weight.data.dtype:
         input = input.to(dtype=weight.data.dtype)
@@ -89,6 +175,7 @@ def functional_group_norm(input, num_groups, weight=None, bias=None, eps=1e-05):
 
 # A1111 BF16
 original_functional_layer_norm = torch.nn.functional.layer_norm
+@wraps(torch.nn.functional.layer_norm)
 def functional_layer_norm(input, normalized_shape, weight=None, bias=None, eps=1e-05):
     if weight is not None and input.dtype != weight.data.dtype:
         input = input.to(dtype=weight.data.dtype)
@@ -98,6 +185,7 @@ def functional_layer_norm(input, normalized_shape, weight=None, bias=None, eps=1
 
 # Training
 original_functional_linear = torch.nn.functional.linear
+@wraps(torch.nn.functional.linear)
 def functional_linear(input, weight, bias=None):
     if input.dtype != weight.data.dtype:
         input = input.to(dtype=weight.data.dtype)
@@ -105,7 +193,17 @@ def functional_linear(input, weight, bias=None):
         bias.data = bias.data.to(dtype=weight.data.dtype)
     return original_functional_linear(input, weight, bias=bias)
 
+original_functional_conv1d = torch.nn.functional.conv1d
+@wraps(torch.nn.functional.conv1d)
+def functional_conv1d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1):
+    if input.dtype != weight.data.dtype:
+        input = input.to(dtype=weight.data.dtype)
+    if bias is not None and bias.data.dtype != weight.data.dtype:
+        bias.data = bias.data.to(dtype=weight.data.dtype)
+    return original_functional_conv1d(input, weight, bias=bias, stride=stride, padding=padding, dilation=dilation, groups=groups)
+
 original_functional_conv2d = torch.nn.functional.conv2d
+@wraps(torch.nn.functional.conv2d)
 def functional_conv2d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1):
     if input.dtype != weight.data.dtype:
         input = input.to(dtype=weight.data.dtype)
@@ -113,16 +211,19 @@ def functional_conv2d(input, weight, bias=None, stride=1, padding=0, dilation=1,
         bias.data = bias.data.to(dtype=weight.data.dtype)
     return original_functional_conv2d(input, weight, bias=bias, stride=stride, padding=padding, dilation=dilation, groups=groups)
 
-# A1111 Embedding BF16
-original_torch_cat = torch.cat
-def torch_cat(tensor, *args, **kwargs):
-    if len(tensor) == 3 and (tensor[0].dtype != tensor[1].dtype or tensor[2].dtype != tensor[1].dtype):
-        return original_torch_cat([tensor[0].to(tensor[1].dtype), tensor[1], tensor[2].to(tensor[1].dtype)], *args, **kwargs)
-    else:
-        return original_torch_cat(tensor, *args, **kwargs)
+# LTX Video
+original_functional_conv3d = torch.nn.functional.conv3d
+@wraps(torch.nn.functional.conv3d)
+def functional_conv3d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1):
+    if input.dtype != weight.data.dtype:
+        input = input.to(dtype=weight.data.dtype)
+    if bias is not None and bias.data.dtype != weight.data.dtype:
+        bias.data = bias.data.to(dtype=weight.data.dtype)
+    return original_functional_conv3d(input, weight, bias=bias, stride=stride, padding=padding, dilation=dilation, groups=groups)
 
 # SwinIR BF16:
 original_functional_pad = torch.nn.functional.pad
+@wraps(torch.nn.functional.pad)
 def functional_pad(input, pad, mode='constant', value=None):
     if mode == 'reflect' and input.dtype == torch.bfloat16:
         return original_functional_pad(input.to(torch.float32), pad, mode=mode, value=value).to(dtype=torch.bfloat16)
@@ -131,118 +232,235 @@ def functional_pad(input, pad, mode='constant', value=None):
 
 
 original_torch_tensor = torch.tensor
-def torch_tensor(*args, device=None, **kwargs):
-    if check_device(device):
-        return original_torch_tensor(*args, device=return_xpu(device), **kwargs)
-    else:
-        return original_torch_tensor(*args, device=device, **kwargs)
+@wraps(torch.tensor)
+def torch_tensor(data, *args, dtype=None, device=None, **kwargs):
+    global device_supports_fp64
+    if check_cuda(device):
+        device = return_xpu(device)
+    if not device_supports_fp64:
+        if check_device_type(device, "xpu"):
+            if dtype == torch.float64:
+                dtype = torch.float32
+            elif dtype is None and (hasattr(data, "dtype") and (data.dtype == torch.float64 or data.dtype == float)):
+                dtype = torch.float32
+    return original_torch_tensor(data, *args, dtype=dtype, device=device, **kwargs)
 
-original_Tensor_to = torch.Tensor.to
+torch.Tensor.original_Tensor_to = torch.Tensor.to
+@wraps(torch.Tensor.to)
 def Tensor_to(self, device=None, *args, **kwargs):
-    if check_device(device):
-        return original_Tensor_to(self, return_xpu(device), *args, **kwargs)
+    if check_cuda(device):
+        return self.original_Tensor_to(return_xpu(device), *args, **kwargs)
     else:
-        return original_Tensor_to(self, device, *args, **kwargs)
+        return self.original_Tensor_to(device, *args, **kwargs)
 
 original_Tensor_cuda = torch.Tensor.cuda
+@wraps(torch.Tensor.cuda)
 def Tensor_cuda(self, device=None, *args, **kwargs):
-    if check_device(device):
-        return original_Tensor_cuda(self, return_xpu(device), *args, **kwargs)
+    if device is None or check_cuda(device):
+        return self.to(return_xpu(device), *args, **kwargs)
     else:
         return original_Tensor_cuda(self, device, *args, **kwargs)
 
+original_Tensor_pin_memory = torch.Tensor.pin_memory
+@wraps(torch.Tensor.pin_memory)
+def Tensor_pin_memory(self, device=None, *args, **kwargs):
+    if device is None or check_cuda(device):
+        return original_Tensor_pin_memory(self, return_xpu(device), *args, **kwargs)
+    else:
+        return original_Tensor_pin_memory(self, device, *args, **kwargs)
+
 original_UntypedStorage_init = torch.UntypedStorage.__init__
+@wraps(torch.UntypedStorage.__init__)
 def UntypedStorage_init(*args, device=None, **kwargs):
-    if check_device(device):
+    if check_cuda(device):
         return original_UntypedStorage_init(*args, device=return_xpu(device), **kwargs)
     else:
         return original_UntypedStorage_init(*args, device=device, **kwargs)
 
-original_UntypedStorage_cuda = torch.UntypedStorage.cuda
-def UntypedStorage_cuda(self, device=None, *args, **kwargs):
-    if check_device(device):
-        return original_UntypedStorage_cuda(self, return_xpu(device), *args, **kwargs)
-    else:
-        return original_UntypedStorage_cuda(self, device, *args, **kwargs)
+if torch_version >= 2.4:
+    original_UntypedStorage_to = torch.UntypedStorage.to
+    @wraps(torch.UntypedStorage.to)
+    def UntypedStorage_to(self, *args, device=None, **kwargs):
+        if check_cuda(device):
+            return original_UntypedStorage_to(self, *args, device=return_xpu(device), **kwargs)
+        else:
+            return original_UntypedStorage_to(self, *args, device=device, **kwargs)
+
+    original_UntypedStorage_cuda = torch.UntypedStorage.cuda
+    @wraps(torch.UntypedStorage.cuda)
+    def UntypedStorage_cuda(self, device=None, non_blocking=False, **kwargs):
+        if device is None or check_cuda(device):
+            return self.to(device=return_xpu(device), non_blocking=non_blocking, **kwargs)
+        else:
+            return original_UntypedStorage_cuda(self, device=device, non_blocking=non_blocking, **kwargs)
 
 original_torch_empty = torch.empty
+@wraps(torch.empty)
 def torch_empty(*args, device=None, **kwargs):
-    if check_device(device):
+    if check_cuda(device):
         return original_torch_empty(*args, device=return_xpu(device), **kwargs)
     else:
         return original_torch_empty(*args, device=device, **kwargs)
 
 original_torch_randn = torch.randn
-def torch_randn(*args, device=None, **kwargs):
-    if check_device(device):
+@wraps(torch.randn)
+def torch_randn(*args, device=None, dtype=None, **kwargs):
+    if dtype is bytes:
+        dtype = None
+    if check_cuda(device):
         return original_torch_randn(*args, device=return_xpu(device), **kwargs)
     else:
         return original_torch_randn(*args, device=device, **kwargs)
 
 original_torch_ones = torch.ones
+@wraps(torch.ones)
 def torch_ones(*args, device=None, **kwargs):
-    if check_device(device):
+    if check_cuda(device):
         return original_torch_ones(*args, device=return_xpu(device), **kwargs)
     else:
         return original_torch_ones(*args, device=device, **kwargs)
 
 original_torch_zeros = torch.zeros
+@wraps(torch.zeros)
 def torch_zeros(*args, device=None, **kwargs):
-    if check_device(device):
+    if check_cuda(device):
         return original_torch_zeros(*args, device=return_xpu(device), **kwargs)
     else:
         return original_torch_zeros(*args, device=device, **kwargs)
 
+original_torch_full = torch.full
+@wraps(torch.full)
+def torch_full(*args, device=None, **kwargs):
+    if check_cuda(device):
+        return original_torch_full(*args, device=return_xpu(device), **kwargs)
+    else:
+        return original_torch_full(*args, device=device, **kwargs)
+
 original_torch_linspace = torch.linspace
+@wraps(torch.linspace)
 def torch_linspace(*args, device=None, **kwargs):
-    if check_device(device):
+    if check_cuda(device):
         return original_torch_linspace(*args, device=return_xpu(device), **kwargs)
     else:
         return original_torch_linspace(*args, device=device, **kwargs)
 
-original_torch_Generator = torch.Generator
-def torch_Generator(device=None):
-    if check_device(device):
-        return original_torch_Generator(return_xpu(device))
+original_torch_eye = torch.eye
+@wraps(torch.eye)
+def torch_eye(*args, device=None, **kwargs):
+    if check_cuda(device):
+        return original_torch_eye(*args, device=return_xpu(device), **kwargs)
     else:
-        return original_torch_Generator(device)
+        return original_torch_eye(*args, device=device, **kwargs)
 
 original_torch_load = torch.load
-def torch_load(f, map_location=None, pickle_module=None, *, weights_only=False, mmap=None, **kwargs):
-    if check_device(map_location):
-        return original_torch_load(f, map_location=return_xpu(map_location), pickle_module=pickle_module, weights_only=weights_only, mmap=mmap, **kwargs)
+@wraps(torch.load)
+def torch_load(f, map_location=None, *args, **kwargs):
+    if map_location is None or check_cuda(map_location):
+        return original_torch_load(f, *args, map_location=return_xpu(map_location), **kwargs)
     else:
-        return original_torch_load(f, map_location=map_location, pickle_module=pickle_module, weights_only=weights_only, mmap=mmap, **kwargs)
+        return original_torch_load(f, *args, map_location=map_location, **kwargs)
+
+@wraps(torch.cuda.synchronize)
+def torch_cuda_synchronize(device=None):
+    if check_cuda(device):
+        return torch.xpu.synchronize(return_xpu(device))
+    else:
+        return torch.xpu.synchronize(device)
+
+@wraps(torch.cuda.device)
+def torch_cuda_device(device):
+    if check_cuda(device):
+        return torch.xpu.device(return_xpu(device))
+    else:
+        return torch.xpu.device(device)
+
+@wraps(torch.cuda.set_device)
+def torch_cuda_set_device(device):
+    if check_cuda(device):
+        torch.xpu.set_device(return_xpu(device))
+    else:
+        torch.xpu.set_device(device)
+
+# torch.Generator has to be a class for isinstance checks
+original_torch_Generator = torch.Generator
+class torch_Generator(original_torch_Generator):
+    def __new__(self, device=None):
+        # can't hijack __init__ because of C override so use return super().__new__
+        if check_cuda(device):
+            return super().__new__(self, return_xpu(device))
+        else:
+            return super().__new__(self, device)
+
 
 # Hijack Functions:
 def ipex_hijacks():
+    global device_supports_fp64
+    if torch_version >= 2.4:
+        torch.UntypedStorage.cuda = UntypedStorage_cuda
+        torch.UntypedStorage.to = UntypedStorage_to
     torch.tensor = torch_tensor
     torch.Tensor.to = Tensor_to
     torch.Tensor.cuda = Tensor_cuda
+    torch.Tensor.pin_memory = Tensor_pin_memory
     torch.UntypedStorage.__init__ = UntypedStorage_init
-    torch.UntypedStorage.cuda = UntypedStorage_cuda
     torch.empty = torch_empty
     torch.randn = torch_randn
     torch.ones = torch_ones
     torch.zeros = torch_zeros
+    torch.full = torch_full
     torch.linspace = torch_linspace
-    torch.Generator = torch_Generator
+    torch.eye = torch_eye
     torch.load = torch_load
+    torch.cuda.synchronize = torch_cuda_synchronize
+    torch.cuda.device = torch_cuda_device
+    torch.cuda.set_device = torch_cuda_set_device
+
+    torch.Generator = torch_Generator
+    torch._C.Generator = torch_Generator
 
     torch.backends.cuda.sdp_kernel = return_null_context
     torch.nn.DataParallel = DummyDataParallel
     torch.UntypedStorage.is_cuda = is_cuda
-    torch.autocast = ipex_autocast
+    torch.amp.autocast_mode.autocast.__init__ = autocast_init
 
+    torch.nn.functional.interpolate = interpolate
     torch.nn.functional.scaled_dot_product_attention = scaled_dot_product_attention
     torch.nn.functional.group_norm = functional_group_norm
     torch.nn.functional.layer_norm = functional_layer_norm
     torch.nn.functional.linear = functional_linear
+    torch.nn.functional.conv1d = functional_conv1d
     torch.nn.functional.conv2d = functional_conv2d
-    torch.nn.functional.interpolate = interpolate
+    torch.nn.functional.conv3d = functional_conv3d
     torch.nn.functional.pad = functional_pad
 
     torch.bmm = torch_bmm
-    torch.cat = torch_cat
-    if not torch.xpu.has_fp64_dtype():
+    torch.fft.fftn = fft_fftn
+    torch.fft.ifftn = fft_ifftn
+    if not device_supports_fp64:
         torch.from_numpy = from_numpy
+        torch.as_tensor = as_tensor
+
+    # AMP:
+    torch.amp.grad_scaler.GradScaler.__init__ = GradScaler_init
+    torch.is_autocast_enabled = torch_is_autocast_enabled
+    torch.get_autocast_gpu_dtype = torch_get_autocast_dtype
+    torch.get_autocast_dtype = torch_get_autocast_dtype
+
+    if hasattr(torch.xpu, "amp"):
+        if not hasattr(torch.xpu.amp, "custom_fwd"):
+            torch.xpu.amp.custom_fwd = torch.cuda.amp.custom_fwd
+            torch.xpu.amp.custom_bwd = torch.cuda.amp.custom_bwd
+        if not hasattr(torch.xpu.amp, "GradScaler"):
+            torch.xpu.amp.GradScaler = torch.amp.grad_scaler.GradScaler
+        torch.cuda.amp = torch.xpu.amp
+    else:
+        if not hasattr(torch.amp, "custom_fwd"):
+            torch.amp.custom_fwd = torch.cuda.amp.custom_fwd
+            torch.amp.custom_bwd = torch.cuda.amp.custom_bwd
+        torch.cuda.amp = torch.amp
+
+    if not hasattr(torch.cuda.amp, "common"):
+        torch.cuda.amp.common = nullcontext()
+    torch.cuda.amp.common.amp_definitely_not_available = lambda: False
+
+    return device_supports_fp64

library/jpeg_xl_util.py

@@ -0,0 +1,186 @@
+# Modified from https://github.com/Fraetor/jxl_decode Original license: MIT
+# Added partial read support for up to 200x speedup
+
+import os
+from typing import List, Tuple
+
+class JXLBitstream:
+    """
+    A stream of bits with methods for easy handling.
+    """
+
+    def __init__(self, file, offset: int = 0, offsets: List[List[int]] = None):
+        self.shift = 0
+        self.bitstream = bytearray()
+        self.file = file
+        self.offset = offset
+        self.offsets = offsets
+        if self.offsets:
+            self.offset = self.offsets[0][1]
+            self.previous_data_len = 0
+            self.index = 0
+        self.file.seek(self.offset)
+
+    def get_bits(self, length: int = 1) -> int:
+        if self.offsets and self.shift + length > self.previous_data_len + self.offsets[self.index][2]:
+            self.partial_to_read_length = length
+            if self.shift < self.previous_data_len + self.offsets[self.index][2]:
+                self.partial_read(0, length)
+            self.bitstream.extend(self.file.read(self.partial_to_read_length))
+        else:
+            self.bitstream.extend(self.file.read(length))
+        bitmask = 2**length - 1
+        bits = (int.from_bytes(self.bitstream, "little") >> self.shift) & bitmask
+        self.shift += length
+        return bits
+
+    def partial_read(self, current_length: int, length: int) -> None:
+        self.previous_data_len += self.offsets[self.index][2]
+        to_read_length = self.previous_data_len - (self.shift + current_length)
+        self.bitstream.extend(self.file.read(to_read_length))
+        current_length += to_read_length
+        self.partial_to_read_length -= to_read_length
+        self.index += 1
+        self.file.seek(self.offsets[self.index][1])
+        if self.shift + length > self.previous_data_len + self.offsets[self.index][2]:
+            self.partial_read(current_length, length)
+
+
+def decode_codestream(file, offset: int = 0, offsets: List[List[int]] = None) -> Tuple[int,int]:
+    """
+    Decodes the actual codestream.
+    JXL codestream specification: http://www-internal/2022/18181-1
+    """
+
+    # Convert codestream to int within an object to get some handy methods.
+    codestream = JXLBitstream(file, offset=offset, offsets=offsets)
+
+    # Skip signature
+    codestream.get_bits(16)
+
+    # SizeHeader
+    div8 = codestream.get_bits(1)
+    if div8:
+        height = 8 * (1 + codestream.get_bits(5))
+    else:
+        distribution = codestream.get_bits(2)
+        match distribution:
+            case 0:
+                height = 1 + codestream.get_bits(9)
+            case 1:
+                height = 1 + codestream.get_bits(13)
+            case 2:
+                height = 1 + codestream.get_bits(18)
+            case 3:
+                height = 1 + codestream.get_bits(30)
+    ratio = codestream.get_bits(3)
+    if div8 and not ratio:
+        width = 8 * (1 + codestream.get_bits(5))
+    elif not ratio:
+        distribution = codestream.get_bits(2)
+        match distribution:
+            case 0:
+                width = 1 + codestream.get_bits(9)
+            case 1:
+                width = 1 + codestream.get_bits(13)
+            case 2:
+                width = 1 + codestream.get_bits(18)
+            case 3:
+                width = 1 + codestream.get_bits(30)
+    else:
+        match ratio:
+            case 1:
+                width = height
+            case 2:
+                width = (height * 12) // 10
+            case 3:
+                width = (height * 4) // 3
+            case 4:
+                width = (height * 3) // 2
+            case 5:
+                width = (height * 16) // 9
+            case 6:
+                width = (height * 5) // 4
+            case 7:
+                width = (height * 2) // 1
+    return width, height
+
+
+def decode_container(file) -> Tuple[int,int]:
+    """
+    Parses the ISOBMFF container, extracts the codestream, and decodes it.
+    JXL container specification: http://www-internal/2022/18181-2
+    """
+
+    def parse_box(file, file_start: int) -> dict:
+        file.seek(file_start)
+        LBox = int.from_bytes(file.read(4), "big")
+        XLBox = None
+        if 1 < LBox <= 8:
+            raise ValueError(f"Invalid LBox at byte {file_start}.")
+        if LBox == 1:
+            file.seek(file_start + 8)
+            XLBox = int.from_bytes(file.read(8), "big")
+            if XLBox <= 16:
+                raise ValueError(f"Invalid XLBox at byte {file_start}.")
+        if XLBox:
+            header_length = 16
+            box_length = XLBox
+        else:
+            header_length = 8
+            if LBox == 0:
+                box_length = os.fstat(file.fileno()).st_size - file_start
+            else:
+                box_length = LBox
+        file.seek(file_start + 4)
+        box_type = file.read(4)
+        file.seek(file_start)
+        return {
+            "length": box_length,
+            "type": box_type,
+            "offset": header_length,
+        }
+
+    file.seek(0)
+    # Reject files missing required boxes. These two boxes are required to be at
+    # the start and contain no values, so we can manually check there presence.
+    # Signature box. (Redundant as has already been checked.)
+    if file.read(12) != bytes.fromhex("0000000C 4A584C20 0D0A870A"):
+        raise ValueError("Invalid signature box.")
+    # File Type box.
+    if file.read(20) != bytes.fromhex(
+        "00000014 66747970 6A786C20 00000000 6A786C20"
+    ):
+        raise ValueError("Invalid file type box.")
+
+    offset = 0
+    offsets = []
+    data_offset_not_found = True
+    container_pointer = 32
+    file_size = os.fstat(file.fileno()).st_size
+    while data_offset_not_found:
+        box = parse_box(file, container_pointer)
+        match box["type"]:
+            case b"jxlc":
+                offset = container_pointer + box["offset"]
+                data_offset_not_found = False
+            case b"jxlp":
+                file.seek(container_pointer + box["offset"])
+                index = int.from_bytes(file.read(4), "big")
+                offsets.append([index, container_pointer + box["offset"] + 4, box["length"] - box["offset"] - 4])
+        container_pointer += box["length"]
+        if container_pointer >= file_size:
+            data_offset_not_found = False
+
+    if offsets:
+        offsets.sort(key=lambda i: i[0])
+    file.seek(0)
+
+    return decode_codestream(file, offset=offset, offsets=offsets)
+
+
+def get_jxl_size(path: str) -> Tuple[int,int]:
+    with open(path, "rb") as file:
+        if file.read(2) == bytes.fromhex("FF0A"):
+            return decode_codestream(file)
+        return decode_container(file)

library/lora_utils.py

@@ -0,0 +1,246 @@
+import os
+import re
+from typing import Dict, List, Optional, Union
+import torch
+from tqdm import tqdm
+from library.device_utils import synchronize_device
+from library.fp8_optimization_utils import load_safetensors_with_fp8_optimization
+from library.safetensors_utils import MemoryEfficientSafeOpen
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+def filter_lora_state_dict(
+    weights_sd: Dict[str, torch.Tensor],
+    include_pattern: Optional[str] = None,
+    exclude_pattern: Optional[str] = None,
+) -> Dict[str, torch.Tensor]:
+    # apply include/exclude patterns
+    original_key_count = len(weights_sd.keys())
+    if include_pattern is not None:
+        regex_include = re.compile(include_pattern)
+        weights_sd = {k: v for k, v in weights_sd.items() if regex_include.search(k)}
+        logger.info(f"Filtered keys with include pattern {include_pattern}: {original_key_count} -> {len(weights_sd.keys())}")
+
+    if exclude_pattern is not None:
+        original_key_count_ex = len(weights_sd.keys())
+        regex_exclude = re.compile(exclude_pattern)
+        weights_sd = {k: v for k, v in weights_sd.items() if not regex_exclude.search(k)}
+        logger.info(f"Filtered keys with exclude pattern {exclude_pattern}: {original_key_count_ex} -> {len(weights_sd.keys())}")
+
+    if len(weights_sd) != original_key_count:
+        remaining_keys = list(set([k.split(".", 1)[0] for k in weights_sd.keys()]))
+        remaining_keys.sort()
+        logger.info(f"Remaining LoRA modules after filtering: {remaining_keys}")
+        if len(weights_sd) == 0:
+            logger.warning("No keys left after filtering.")
+
+    return weights_sd
+
+
+def load_safetensors_with_lora_and_fp8(
+    model_files: Union[str, List[str]],
+    lora_weights_list: Optional[Dict[str, torch.Tensor]],
+    lora_multipliers: Optional[List[float]],
+    fp8_optimization: bool,
+    calc_device: torch.device,
+    move_to_device: bool = False,
+    dit_weight_dtype: Optional[torch.dtype] = None,
+    target_keys: Optional[List[str]] = None,
+    exclude_keys: Optional[List[str]] = None,
+) -> dict[str, torch.Tensor]:
+    """
+    Merge LoRA weights into the state dict of a model with fp8 optimization if needed.
+
+    Args:
+        model_files (Union[str, List[str]]): Path to the model file or list of paths. If the path matches a pattern like `00001-of-00004`, it will load all files with the same prefix.
+        lora_weights_list (Optional[Dict[str, torch.Tensor]]): Dictionary of LoRA weight tensors to load.
+        lora_multipliers (Optional[List[float]]): List of multipliers for LoRA weights.
+        fp8_optimization (bool): Whether to apply FP8 optimization.
+        calc_device (torch.device): Device to calculate on.
+        move_to_device (bool): Whether to move tensors to the calculation device after loading.
+        target_keys (Optional[List[str]]): Keys to target for optimization.
+        exclude_keys (Optional[List[str]]): Keys to exclude from optimization.
+    """
+
+    # if the file name ends with 00001-of-00004 etc, we need to load the files with the same prefix
+    if isinstance(model_files, str):
+        model_files = [model_files]
+
+    extended_model_files = []
+    for model_file in model_files:
+        basename = os.path.basename(model_file)
+        match = re.match(r"^(.*?)(\d+)-of-(\d+)\.safetensors$", basename)
+        if match:
+            prefix = basename[: match.start(2)]
+            count = int(match.group(3))
+            state_dict = {}
+            for i in range(count):
+                filename = f"{prefix}{i + 1:05d}-of-{count:05d}.safetensors"
+                filepath = os.path.join(os.path.dirname(model_file), filename)
+                if os.path.exists(filepath):
+                    extended_model_files.append(filepath)
+                else:
+                    raise FileNotFoundError(f"File {filepath} not found")
+        else:
+            extended_model_files.append(model_file)
+    model_files = extended_model_files
+    logger.info(f"Loading model files: {model_files}")
+
+    # load LoRA weights
+    weight_hook = None
+    if lora_weights_list is None or len(lora_weights_list) == 0:
+        lora_weights_list = []
+        lora_multipliers = []
+        list_of_lora_weight_keys = []
+    else:
+        list_of_lora_weight_keys = []
+        for lora_sd in lora_weights_list:
+            lora_weight_keys = set(lora_sd.keys())
+            list_of_lora_weight_keys.append(lora_weight_keys)
+
+        if lora_multipliers is None:
+            lora_multipliers = [1.0] * len(lora_weights_list)
+        while len(lora_multipliers) < len(lora_weights_list):
+            lora_multipliers.append(1.0)
+        if len(lora_multipliers) > len(lora_weights_list):
+            lora_multipliers = lora_multipliers[: len(lora_weights_list)]
+
+        # Merge LoRA weights into the state dict
+        logger.info(f"Merging LoRA weights into state dict. multipliers: {lora_multipliers}")
+
+        # make hook for LoRA merging
+        def weight_hook_func(model_weight_key, model_weight, keep_on_calc_device=False):
+            nonlocal list_of_lora_weight_keys, lora_weights_list, lora_multipliers, calc_device
+
+            if not model_weight_key.endswith(".weight"):
+                return model_weight
+
+            original_device = model_weight.device
+            if original_device != calc_device:
+                model_weight = model_weight.to(calc_device)  # to make calculation faster
+
+            for lora_weight_keys, lora_sd, multiplier in zip(list_of_lora_weight_keys, lora_weights_list, lora_multipliers):
+                # check if this weight has LoRA weights
+                lora_name = model_weight_key.rsplit(".", 1)[0]  # remove trailing ".weight"
+                lora_name = "lora_unet_" + lora_name.replace(".", "_")
+                down_key = lora_name + ".lora_down.weight"
+                up_key = lora_name + ".lora_up.weight"
+                alpha_key = lora_name + ".alpha"
+                if down_key not in lora_weight_keys or up_key not in lora_weight_keys:
+                    continue
+
+                # get LoRA weights
+                down_weight = lora_sd[down_key]
+                up_weight = lora_sd[up_key]
+
+                dim = down_weight.size()[0]
+                alpha = lora_sd.get(alpha_key, dim)
+                scale = alpha / dim
+
+                down_weight = down_weight.to(calc_device)
+                up_weight = up_weight.to(calc_device)
+
+                # W <- W + U * D
+                if len(model_weight.size()) == 2:
+                    # linear
+                    if len(up_weight.size()) == 4:  # use linear projection mismatch
+                        up_weight = up_weight.squeeze(3).squeeze(2)
+                        down_weight = down_weight.squeeze(3).squeeze(2)
+                    model_weight = model_weight + multiplier * (up_weight @ down_weight) * scale
+                elif down_weight.size()[2:4] == (1, 1):
+                    # conv2d 1x1
+                    model_weight = (
+                        model_weight
+                        + multiplier
+                        * (up_weight.squeeze(3).squeeze(2) @ down_weight.squeeze(3).squeeze(2)).unsqueeze(2).unsqueeze(3)
+                        * scale
+                    )
+                else:
+                    # conv2d 3x3
+                    conved = torch.nn.functional.conv2d(down_weight.permute(1, 0, 2, 3), up_weight).permute(1, 0, 2, 3)
+                    # logger.info(conved.size(), weight.size(), module.stride, module.padding)
+                    model_weight = model_weight + multiplier * conved * scale
+
+                # remove LoRA keys from set
+                lora_weight_keys.remove(down_key)
+                lora_weight_keys.remove(up_key)
+                if alpha_key in lora_weight_keys:
+                    lora_weight_keys.remove(alpha_key)
+
+            if not keep_on_calc_device and original_device != calc_device:
+                model_weight = model_weight.to(original_device)  # move back to original device
+            return model_weight
+
+        weight_hook = weight_hook_func
+
+    state_dict = load_safetensors_with_fp8_optimization_and_hook(
+        model_files,
+        fp8_optimization,
+        calc_device,
+        move_to_device,
+        dit_weight_dtype,
+        target_keys,
+        exclude_keys,
+        weight_hook=weight_hook,
+    )
+
+    for lora_weight_keys in list_of_lora_weight_keys:
+        # check if all LoRA keys are used
+        if len(lora_weight_keys) > 0:
+            # if there are still LoRA keys left, it means they are not used in the model
+            # this is a warning, not an error
+            logger.warning(f"Warning: not all LoRA keys are used: {', '.join(lora_weight_keys)}")
+
+    return state_dict
+
+
+def load_safetensors_with_fp8_optimization_and_hook(
+    model_files: list[str],
+    fp8_optimization: bool,
+    calc_device: torch.device,
+    move_to_device: bool = False,
+    dit_weight_dtype: Optional[torch.dtype] = None,
+    target_keys: Optional[List[str]] = None,
+    exclude_keys: Optional[List[str]] = None,
+    weight_hook: callable = None,
+) -> dict[str, torch.Tensor]:
+    """
+    Load state dict from safetensors files and merge LoRA weights into the state dict with fp8 optimization if needed.
+    """
+    if fp8_optimization:
+        logger.info(
+            f"Loading state dict with FP8 optimization. Dtype of weight: {dit_weight_dtype}, hook enabled: {weight_hook is not None}"
+        )
+        # dit_weight_dtype is not used because we use fp8 optimization
+        state_dict = load_safetensors_with_fp8_optimization(
+            model_files, calc_device, target_keys, exclude_keys, move_to_device=move_to_device, weight_hook=weight_hook
+        )
+    else:
+        logger.info(
+            f"Loading state dict without FP8 optimization. Dtype of weight: {dit_weight_dtype}, hook enabled: {weight_hook is not None}"
+        )
+        state_dict = {}
+        for model_file in model_files:
+            with MemoryEfficientSafeOpen(model_file) as f:
+                for key in tqdm(f.keys(), desc=f"Loading {os.path.basename(model_file)}", leave=False):
+                    if weight_hook is None and move_to_device:
+                        value = f.get_tensor(key, device=calc_device, dtype=dit_weight_dtype)
+                    else:
+                        value = f.get_tensor(key)  # we cannot directly load to device because get_tensor does non-blocking transfer
+                        if weight_hook is not None:
+                            value = weight_hook(key, value, keep_on_calc_device=move_to_device)
+                        if move_to_device:
+                            value = value.to(calc_device, dtype=dit_weight_dtype, non_blocking=True)
+                        elif dit_weight_dtype is not None:
+                            value = value.to(dit_weight_dtype)
+
+                    state_dict[key] = value
+        if move_to_device:
+            synchronize_device(calc_device)
+
+    return state_dict

library/lpw_stable_diffusion.py

@@ -17,7 +17,6 @@ from diffusers.models import AutoencoderKL, UNet2DConditionModel
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker
 from diffusers.utils import logging
 
-
 try:
     from diffusers.utils import PIL_INTERPOLATION
 except ImportError:
@@ -626,7 +625,7 @@ class StableDiffusionLongPromptWeightingPipeline(StableDiffusionPipeline):
             raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
 
         if height % 8 != 0 or width % 8 != 0:
-            print(height, width)
+            logger.info(f'{height} {width}')
             raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
 
         if (callback_steps is None) or (

library/lumina_util.py

@@ -0,0 +1,259 @@
+import json
+import os
+from dataclasses import replace
+from typing import List, Optional, Tuple, Union
+
+import einops
+import torch
+from accelerate import init_empty_weights
+from safetensors import safe_open
+from safetensors.torch import load_file
+from transformers import Gemma2Config, Gemma2Model
+
+from library.utils import setup_logging
+from library import lumina_models, flux_models
+from library.safetensors_utils import load_safetensors
+import logging
+
+setup_logging()
+logger = logging.getLogger(__name__)
+
+MODEL_VERSION_LUMINA_V2 = "lumina2"
+
+
+def load_lumina_model(
+    ckpt_path: str,
+    dtype: Optional[torch.dtype],
+    device: torch.device,
+    disable_mmap: bool = False,
+    use_flash_attn: bool = False,
+    use_sage_attn: bool = False,
+):
+    """
+    Load the Lumina model from the checkpoint path.
+
+    Args:
+        ckpt_path (str): Path to the checkpoint.
+        dtype (torch.dtype): The data type for the model.
+        device (torch.device): The device to load the model on.
+        disable_mmap (bool, optional): Whether to disable mmap. Defaults to False.
+        use_flash_attn (bool, optional): Whether to use flash attention. Defaults to False.
+
+    Returns:
+        model (lumina_models.NextDiT): The loaded model.
+    """
+    logger.info("Building Lumina")
+    with torch.device("meta"):
+        model = lumina_models.NextDiT_2B_GQA_patch2_Adaln_Refiner(use_flash_attn=use_flash_attn, use_sage_attn=use_sage_attn).to(
+            dtype
+        )
+
+    logger.info(f"Loading state dict from {ckpt_path}")
+    state_dict = load_safetensors(ckpt_path, device=device, disable_mmap=disable_mmap, dtype=dtype)
+
+    # Neta-Lumina support
+    if "model.diffusion_model.cap_embedder.0.weight" in state_dict:
+        # remove "model.diffusion_model." prefix
+        filtered_state_dict = {
+            k.replace("model.diffusion_model.", ""): v for k, v in state_dict.items() if k.startswith("model.diffusion_model.")
+        }
+        state_dict = filtered_state_dict
+
+    info = model.load_state_dict(state_dict, strict=False, assign=True)
+    logger.info(f"Loaded Lumina: {info}")
+    return model
+
+
+def load_ae(
+    ckpt_path: str,
+    dtype: torch.dtype,
+    device: Union[str, torch.device],
+    disable_mmap: bool = False,
+) -> flux_models.AutoEncoder:
+    """
+    Load the AutoEncoder model from the checkpoint path.
+
+    Args:
+        ckpt_path (str): Path to the checkpoint.
+        dtype (torch.dtype): The data type for the model.
+        device (Union[str, torch.device]): The device to load the model on.
+        disable_mmap (bool, optional): Whether to disable mmap. Defaults to False.
+
+    Returns:
+        ae (flux_models.AutoEncoder): The loaded model.
+    """
+    logger.info("Building AutoEncoder")
+    with torch.device("meta"):
+        # dev and schnell have the same AE params
+        ae = flux_models.AutoEncoder(flux_models.configs["schnell"].ae_params).to(dtype)
+
+    logger.info(f"Loading state dict from {ckpt_path}")
+    sd = load_safetensors(ckpt_path, device=device, disable_mmap=disable_mmap, dtype=dtype)
+
+    # Neta-Lumina support
+    if "vae.decoder.conv_in.bias" in sd:
+        # remove "vae." prefix
+        filtered_sd = {k.replace("vae.", ""): v for k, v in sd.items() if k.startswith("vae.")}
+        sd = filtered_sd
+
+    info = ae.load_state_dict(sd, strict=False, assign=True)
+    logger.info(f"Loaded AE: {info}")
+    return ae
+
+
+def load_gemma2(
+    ckpt_path: Optional[str],
+    dtype: torch.dtype,
+    device: Union[str, torch.device],
+    disable_mmap: bool = False,
+    state_dict: Optional[dict] = None,
+) -> Gemma2Model:
+    """
+    Load the Gemma2 model from the checkpoint path.
+
+    Args:
+        ckpt_path (str): Path to the checkpoint.
+        dtype (torch.dtype): The data type for the model.
+        device (Union[str, torch.device]): The device to load the model on.
+        disable_mmap (bool, optional): Whether to disable mmap. Defaults to False.
+        state_dict (Optional[dict], optional): The state dict to load. Defaults to None.
+
+    Returns:
+        gemma2 (Gemma2Model): The loaded model
+    """
+    logger.info("Building Gemma2")
+    GEMMA2_CONFIG = {
+        "_name_or_path": "google/gemma-2-2b",
+        "architectures": ["Gemma2Model"],
+        "attention_bias": False,
+        "attention_dropout": 0.0,
+        "attn_logit_softcapping": 50.0,
+        "bos_token_id": 2,
+        "cache_implementation": "hybrid",
+        "eos_token_id": 1,
+        "final_logit_softcapping": 30.0,
+        "head_dim": 256,
+        "hidden_act": "gelu_pytorch_tanh",
+        "hidden_activation": "gelu_pytorch_tanh",
+        "hidden_size": 2304,
+        "initializer_range": 0.02,
+        "intermediate_size": 9216,
+        "max_position_embeddings": 8192,
+        "model_type": "gemma2",
+        "num_attention_heads": 8,
+        "num_hidden_layers": 26,
+        "num_key_value_heads": 4,
+        "pad_token_id": 0,
+        "query_pre_attn_scalar": 256,
+        "rms_norm_eps": 1e-06,
+        "rope_theta": 10000.0,
+        "sliding_window": 4096,
+        "torch_dtype": "float32",
+        "transformers_version": "4.44.2",
+        "use_cache": True,
+        "vocab_size": 256000,
+    }
+
+    config = Gemma2Config(**GEMMA2_CONFIG)
+    with init_empty_weights():
+        gemma2 = Gemma2Model._from_config(config)
+
+    if state_dict is not None:
+        sd = state_dict
+    else:
+        logger.info(f"Loading state dict from {ckpt_path}")
+        sd = load_safetensors(ckpt_path, device=str(device), disable_mmap=disable_mmap, dtype=dtype)
+
+    for key in list(sd.keys()):
+        new_key = key.replace("model.", "")
+        if new_key == key:
+            break  # the model doesn't have annoying prefix
+        sd[new_key] = sd.pop(key)
+
+    # Neta-Lumina support
+    if "text_encoders.gemma2_2b.logit_scale" in sd:
+        # remove "text_encoders.gemma2_2b.transformer.model." prefix
+        filtered_sd = {
+            k.replace("text_encoders.gemma2_2b.transformer.model.", ""): v
+            for k, v in sd.items()
+            if k.startswith("text_encoders.gemma2_2b.transformer.model.")
+        }
+        sd = filtered_sd
+
+    info = gemma2.load_state_dict(sd, strict=False, assign=True)
+    logger.info(f"Loaded Gemma2: {info}")
+    return gemma2
+
+
+def unpack_latents(x: torch.Tensor, packed_latent_height: int, packed_latent_width: int) -> torch.Tensor:
+    """
+    x: [b (h w) (c ph pw)] -> [b c (h ph) (w pw)], ph=2, pw=2
+    """
+    x = einops.rearrange(x, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=packed_latent_height, w=packed_latent_width, ph=2, pw=2)
+    return x
+
+
+def pack_latents(x: torch.Tensor) -> torch.Tensor:
+    """
+    x: [b c (h ph) (w pw)] -> [b (h w) (c ph pw)], ph=2, pw=2
+    """
+    x = einops.rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
+    return x
+
+
+DIFFUSERS_TO_ALPHA_VLLM_MAP: dict[str, str] = {
+    # Embedding layers
+    "time_caption_embed.caption_embedder.0.weight": "cap_embedder.0.weight",
+    "time_caption_embed.caption_embedder.1.weight": "cap_embedder.1.weight",
+    "text_embedder.1.bias": "cap_embedder.1.bias",
+    "patch_embedder.proj.weight": "x_embedder.weight",
+    "patch_embedder.proj.bias": "x_embedder.bias",
+    # Attention modulation
+    "transformer_blocks.().adaln_modulation.1.weight": "layers.().adaLN_modulation.1.weight",
+    "transformer_blocks.().adaln_modulation.1.bias": "layers.().adaLN_modulation.1.bias",
+    # Final layers
+    "final_adaln_modulation.1.weight": "final_layer.adaLN_modulation.1.weight",
+    "final_adaln_modulation.1.bias": "final_layer.adaLN_modulation.1.bias",
+    "final_linear.weight": "final_layer.linear.weight",
+    "final_linear.bias": "final_layer.linear.bias",
+    # Noise refiner
+    "single_transformer_blocks.().adaln_modulation.1.weight": "noise_refiner.().adaLN_modulation.1.weight",
+    "single_transformer_blocks.().adaln_modulation.1.bias": "noise_refiner.().adaLN_modulation.1.bias",
+    "single_transformer_blocks.().attn.to_qkv.weight": "noise_refiner.().attention.qkv.weight",
+    "single_transformer_blocks.().attn.to_out.0.weight": "noise_refiner.().attention.out.weight",
+    # Normalization
+    "transformer_blocks.().norm1.weight": "layers.().attention_norm1.weight",
+    "transformer_blocks.().norm2.weight": "layers.().attention_norm2.weight",
+    # FFN
+    "transformer_blocks.().ff.net.0.proj.weight": "layers.().feed_forward.w1.weight",
+    "transformer_blocks.().ff.net.2.weight": "layers.().feed_forward.w2.weight",
+    "transformer_blocks.().ff.net.4.weight": "layers.().feed_forward.w3.weight",
+}
+
+
+def convert_diffusers_sd_to_alpha_vllm(sd: dict, num_double_blocks: int) -> dict:
+    """Convert Diffusers checkpoint to Alpha-VLLM format"""
+    logger.info("Converting Diffusers checkpoint to Alpha-VLLM format")
+    new_sd = sd.copy()  # Preserve original keys
+
+    for diff_key, alpha_key in DIFFUSERS_TO_ALPHA_VLLM_MAP.items():
+        # Handle block-specific patterns
+        if "()." in diff_key:
+            for block_idx in range(num_double_blocks):
+                block_alpha_key = alpha_key.replace("().", f"{block_idx}.")
+                block_diff_key = diff_key.replace("().", f"{block_idx}.")
+
+                # Search for and convert block-specific keys
+                for input_key, value in list(sd.items()):
+                    if input_key == block_diff_key:
+                        new_sd[block_alpha_key] = value
+        else:
+            # Handle static keys
+            if diff_key in sd:
+                print(f"Replacing {diff_key} with {alpha_key}")
+                new_sd[alpha_key] = sd[diff_key]
+            else:
+                print(f"Not found: {diff_key}")
+
+    logger.info(f"Converted {len(new_sd)} keys to Alpha-VLLM format")
+    return new_sd

library/model_util.py

@@ -3,22 +3,23 @@
 
 import math
 import os
+
 import torch
+from library.device_utils import init_ipex
 
-try:
-    import intel_extension_for_pytorch as ipex
+init_ipex()
 
-    if torch.xpu.is_available():
-        from library.ipex import ipex_init
-
-        ipex_init()
-except Exception:
-    pass
 import diffusers
 from transformers import CLIPTextModel, CLIPTokenizer, CLIPTextConfig, logging
 from diffusers import AutoencoderKL, DDIMScheduler, StableDiffusionPipeline  # , UNet2DConditionModel
 from safetensors.torch import load_file, save_file
 from library.original_unet import UNet2DConditionModel
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
 
 # DiffUsers版StableDiffusionのモデルパラメータ
 NUM_TRAIN_TIMESTEPS = 1000
@@ -645,16 +646,15 @@ def convert_ldm_clip_checkpoint_v2(checkpoint, max_length):
             new_sd[key_pfx + "k_proj" + key_suffix] = values[1]
             new_sd[key_pfx + "v_proj" + key_suffix] = values[2]
 
-    # rename or add position_ids
+    # remove position_ids for newer transformer, which causes error :(
     ANOTHER_POSITION_IDS_KEY = "text_model.encoder.text_model.embeddings.position_ids"
     if ANOTHER_POSITION_IDS_KEY in new_sd:
         # waifu diffusion v1.4
-        position_ids = new_sd[ANOTHER_POSITION_IDS_KEY]
         del new_sd[ANOTHER_POSITION_IDS_KEY]
-    else:
-        position_ids = torch.Tensor([list(range(max_length))]).to(torch.int64)
 
-    new_sd["text_model.embeddings.position_ids"] = position_ids
+    if "text_model.embeddings.position_ids" in new_sd:
+        del new_sd["text_model.embeddings.position_ids"]
+
     return new_sd
 
 
@@ -950,7 +950,7 @@ def convert_vae_state_dict(vae_state_dict):
     for k, v in new_state_dict.items():
         for weight_name in weights_to_convert:
             if f"mid.attn_1.{weight_name}.weight" in k:
-                # print(f"Reshaping {k} for SD format: shape {v.shape} -> {v.shape} x 1 x 1")
+                # logger.info(f"Reshaping {k} for SD format: shape {v.shape} -> {v.shape} x 1 x 1")
                 new_state_dict[k] = reshape_weight_for_sd(v)
 
     return new_state_dict
@@ -977,7 +977,7 @@ def load_checkpoint_with_text_encoder_conversion(ckpt_path, device="cpu"):
         checkpoint = None
         state_dict = load_file(ckpt_path)  # , device) # may causes error
     else:
-        checkpoint = torch.load(ckpt_path, map_location=device)
+        checkpoint = torch.load(ckpt_path, map_location=device, weights_only=False)
         if "state_dict" in checkpoint:
             state_dict = checkpoint["state_dict"]
         else:
@@ -1008,7 +1008,7 @@ def load_models_from_stable_diffusion_checkpoint(v2, ckpt_path, device="cpu", dt
 
     unet = UNet2DConditionModel(**unet_config).to(device)
     info = unet.load_state_dict(converted_unet_checkpoint)
-    print("loading u-net:", info)
+    logger.info(f"loading u-net: {info}")
 
     # Convert the VAE model.
     vae_config = create_vae_diffusers_config()
@@ -1016,7 +1016,7 @@ def load_models_from_stable_diffusion_checkpoint(v2, ckpt_path, device="cpu", dt
 
     vae = AutoencoderKL(**vae_config).to(device)
     info = vae.load_state_dict(converted_vae_checkpoint)
-    print("loading vae:", info)
+    logger.info(f"loading vae: {info}")
 
     # convert text_model
     if v2:
@@ -1050,7 +1050,7 @@ def load_models_from_stable_diffusion_checkpoint(v2, ckpt_path, device="cpu", dt
         # logging.set_verbosity_error()  # don't show annoying warning
         # text_model = CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14").to(device)
         # logging.set_verbosity_warning()
-        # print(f"config: {text_model.config}")
+        # logger.info(f"config: {text_model.config}")
         cfg = CLIPTextConfig(
             vocab_size=49408,
             hidden_size=768,
@@ -1073,7 +1073,7 @@ def load_models_from_stable_diffusion_checkpoint(v2, ckpt_path, device="cpu", dt
         )
         text_model = CLIPTextModel._from_config(cfg)
         info = text_model.load_state_dict(converted_text_encoder_checkpoint)
-    print("loading text encoder:", info)
+    logger.info(f"loading text encoder: {info}")
 
     return text_model, vae, unet
 
@@ -1148,7 +1148,7 @@ def convert_text_encoder_state_dict_to_sd_v2(checkpoint, make_dummy_weights=Fals
 
     # 最後の層などを捏造するか
     if make_dummy_weights:
-        print("make dummy weights for resblock.23, text_projection and logit scale.")
+        logger.info("make dummy weights for resblock.23, text_projection and logit scale.")
         keys = list(new_sd.keys())
         for key in keys:
             if key.startswith("transformer.resblocks.22."):
@@ -1267,14 +1267,14 @@ VAE_PREFIX = "first_stage_model."
 
 
 def load_vae(vae_id, dtype):
-    print(f"load VAE: {vae_id}")
+    logger.info(f"load VAE: {vae_id}")
     if os.path.isdir(vae_id) or not os.path.isfile(vae_id):
         # Diffusers local/remote
         try:
             vae = AutoencoderKL.from_pretrained(vae_id, subfolder=None, torch_dtype=dtype)
         except EnvironmentError as e:
-            print(f"exception occurs in loading vae: {e}")
-            print("retry with subfolder='vae'")
+            logger.error(f"exception occurs in loading vae: {e}")
+            logger.error("retry with subfolder='vae'")
             vae = AutoencoderKL.from_pretrained(vae_id, subfolder="vae", torch_dtype=dtype)
         return vae
 
@@ -1346,13 +1346,13 @@ def make_bucket_resolutions(max_reso, min_size=256, max_size=1024, divisible=64)
 
 if __name__ == "__main__":
     resos = make_bucket_resolutions((512, 768))
-    print(len(resos))
-    print(resos)
+    logger.info(f"{len(resos)}")
+    logger.info(f"{resos}")
     aspect_ratios = [w / h for w, h in resos]
-    print(aspect_ratios)
+    logger.info(f"{aspect_ratios}")
 
     ars = set()
     for ar in aspect_ratios:
         if ar in ars:
-            print("error! duplicate ar:", ar)
+            logger.error(f"error! duplicate ar: {ar}")
         ars.add(ar)

library/original_unet.py

@@ -113,6 +113,12 @@ import torch
 from torch import nn
 from torch.nn import functional as F
 from einops import rearrange
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
 
 BLOCK_OUT_CHANNELS: Tuple[int] = (320, 640, 1280, 1280)
 TIMESTEP_INPUT_DIM = BLOCK_OUT_CHANNELS[0]
@@ -526,7 +532,9 @@ class DownBlock2D(nn.Module):
 
                     return custom_forward
 
-                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+                )
             else:
                 hidden_states = resnet(hidden_states, temb)
 
@@ -622,15 +630,9 @@ class CrossAttention(nn.Module):
                 hidden_states,
                 encoder_hidden_states,
                 attention_mask,
-            ) = translate_attention_names_from_diffusers(
-                hidden_states=hidden_states, context=context, mask=mask, **kwargs
-            )
+            ) = translate_attention_names_from_diffusers(hidden_states=hidden_states, context=context, mask=mask, **kwargs)
             return self.processor(
-                attn=self,
-                hidden_states=hidden_states,
-                encoder_hidden_states=context,
-                attention_mask=mask,
-                **kwargs
+                attn=self, hidden_states=hidden_states, encoder_hidden_states=context, attention_mask=mask, **kwargs
             )
         if self.use_memory_efficient_attention_xformers:
             return self.forward_memory_efficient_xformers(hidden_states, context, mask)
@@ -744,13 +746,14 @@ class CrossAttention(nn.Module):
         out = self.to_out[0](out)
         return out
 
+
 def translate_attention_names_from_diffusers(
     hidden_states: torch.FloatTensor,
     context: Optional[torch.FloatTensor] = None,
     mask: Optional[torch.FloatTensor] = None,
     # HF naming
     encoder_hidden_states: Optional[torch.FloatTensor] = None,
-    attention_mask: Optional[torch.FloatTensor] = None
+    attention_mask: Optional[torch.FloatTensor] = None,
 ):
     # translate from hugging face diffusers
     context = context if context is not None else encoder_hidden_states
@@ -760,6 +763,7 @@ def translate_attention_names_from_diffusers(
 
     return hidden_states, context, mask
 
+
 # feedforward
 class GEGLU(nn.Module):
     r"""
@@ -1011,9 +1015,11 @@ class CrossAttnDownBlock2D(nn.Module):
 
                     return custom_forward
 
-                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
                 hidden_states = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(attn, return_dict=False), hidden_states, encoder_hidden_states
+                    create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(attn, return_dict=False), hidden_states, encoder_hidden_states, use_reentrant=False
                 )[0]
             else:
                 hidden_states = resnet(hidden_states, temb)
@@ -1094,10 +1100,12 @@ class UNetMidBlock2DCrossAttn(nn.Module):
 
                 if attn is not None:
                     hidden_states = torch.utils.checkpoint.checkpoint(
-                        create_custom_forward(attn, return_dict=False), hidden_states, encoder_hidden_states
+                        create_custom_forward(attn, return_dict=False), hidden_states, encoder_hidden_states, use_reentrant=False
                     )[0]
 
-                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+                )
             else:
                 if attn is not None:
                     hidden_states = attn(hidden_states, encoder_hidden_states).sample
@@ -1197,7 +1205,9 @@ class UpBlock2D(nn.Module):
 
                     return custom_forward
 
-                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+                )
             else:
                 hidden_states = resnet(hidden_states, temb)
 
@@ -1262,9 +1272,9 @@ class CrossAttnUpBlock2D(nn.Module):
         for attn in self.attentions:
             attn.set_use_memory_efficient_attention(xformers, mem_eff)
 
-    def set_use_sdpa(self, spda):
+    def set_use_sdpa(self, sdpa):
         for attn in self.attentions:
-            attn.set_use_sdpa(spda)
+            attn.set_use_sdpa(sdpa)
 
     def forward(
         self,
@@ -1292,9 +1302,11 @@ class CrossAttnUpBlock2D(nn.Module):
 
                     return custom_forward
 
-                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
                 hidden_states = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(attn, return_dict=False), hidden_states, encoder_hidden_states
+                    create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(attn, return_dict=False), hidden_states, encoder_hidden_states, use_reentrant=False
                 )[0]
             else:
                 hidden_states = resnet(hidden_states, temb)
@@ -1380,7 +1392,7 @@ class UNet2DConditionModel(nn.Module):
     ):
         super().__init__()
         assert sample_size is not None, "sample_size must be specified"
-        print(
+        logger.info(
             f"UNet2DConditionModel: {sample_size}, {attention_head_dim}, {cross_attention_dim}, {use_linear_projection}, {upcast_attention}"
         )
 
@@ -1514,7 +1526,7 @@ class UNet2DConditionModel(nn.Module):
     def set_gradient_checkpointing(self, value=False):
         modules = self.down_blocks + [self.mid_block] + self.up_blocks
         for module in modules:
-            print(module.__class__.__name__, module.gradient_checkpointing, "->", value)
+            logger.info(f"{module.__class__.__name__} {module.gradient_checkpointing} -> {value}")
             module.gradient_checkpointing = value
 
     # endregion
@@ -1709,14 +1721,14 @@ class InferUNet2DConditionModel:
 
     def set_deep_shrink(self, ds_depth_1, ds_timesteps_1=650, ds_depth_2=None, ds_timesteps_2=None, ds_ratio=0.5):
         if ds_depth_1 is None:
-            print("Deep Shrink is disabled.")
+            logger.info("Deep Shrink is disabled.")
             self.ds_depth_1 = None
             self.ds_timesteps_1 = None
             self.ds_depth_2 = None
             self.ds_timesteps_2 = None
             self.ds_ratio = None
         else:
-            print(
+            logger.info(
                 f"Deep Shrink is enabled: [depth={ds_depth_1}/{ds_depth_2}, timesteps={ds_timesteps_1}/{ds_timesteps_2}, ratio={ds_ratio}]"
             )
             self.ds_depth_1 = ds_depth_1

library/safetensors_utils.py

@@ -0,0 +1,351 @@
+import os
+import re
+import numpy as np
+import torch
+import json
+import struct
+from typing import Dict, Any, Union, Optional
+
+from safetensors.torch import load_file
+
+from library.device_utils import synchronize_device
+
+
+def mem_eff_save_file(tensors: Dict[str, torch.Tensor], filename: str, metadata: Dict[str, Any] = None):
+    """
+    memory efficient save file
+    """
+
+    _TYPES = {
+        torch.float64: "F64",
+        torch.float32: "F32",
+        torch.float16: "F16",
+        torch.bfloat16: "BF16",
+        torch.int64: "I64",
+        torch.int32: "I32",
+        torch.int16: "I16",
+        torch.int8: "I8",
+        torch.uint8: "U8",
+        torch.bool: "BOOL",
+        getattr(torch, "float8_e5m2", None): "F8_E5M2",
+        getattr(torch, "float8_e4m3fn", None): "F8_E4M3",
+    }
+    _ALIGN = 256
+
+    def validate_metadata(metadata: Dict[str, Any]) -> Dict[str, str]:
+        validated = {}
+        for key, value in metadata.items():
+            if not isinstance(key, str):
+                raise ValueError(f"Metadata key must be a string, got {type(key)}")
+            if not isinstance(value, str):
+                print(f"Warning: Metadata value for key '{key}' is not a string. Converting to string.")
+                validated[key] = str(value)
+            else:
+                validated[key] = value
+        return validated
+
+
+    header = {}
+    offset = 0
+    if metadata:
+        header["__metadata__"] = validate_metadata(metadata)
+    for k, v in tensors.items():
+        if v.numel() == 0:  # empty tensor
+            header[k] = {"dtype": _TYPES[v.dtype], "shape": list(v.shape), "data_offsets": [offset, offset]}
+        else:
+            size = v.numel() * v.element_size()
+            header[k] = {"dtype": _TYPES[v.dtype], "shape": list(v.shape), "data_offsets": [offset, offset + size]}
+            offset += size
+
+    hjson = json.dumps(header).encode("utf-8")
+    hjson += b" " * (-(len(hjson) + 8) % _ALIGN)
+
+    with open(filename, "wb") as f:
+        f.write(struct.pack("<Q", len(hjson)))
+        f.write(hjson)
+
+        for k, v in tensors.items():
+            if v.numel() == 0:
+                continue
+            if v.is_cuda:
+                # Direct GPU to disk save
+                with torch.cuda.device(v.device):
+                    if v.dim() == 0:  # if scalar, need to add a dimension to work with view
+                        v = v.unsqueeze(0)
+                    tensor_bytes = v.contiguous().view(torch.uint8)
+                    tensor_bytes.cpu().numpy().tofile(f)
+            else:
+                # CPU tensor save
+                if v.dim() == 0:  # if scalar, need to add a dimension to work with view
+                    v = v.unsqueeze(0)
+                v.contiguous().view(torch.uint8).numpy().tofile(f)
+
+
+class MemoryEfficientSafeOpen:
+    """Memory-efficient reader for safetensors files.
+
+    This class provides a memory-efficient way to read tensors from safetensors files
+    by using memory mapping for large tensors and avoiding unnecessary copies.
+    """
+
+    def __init__(self, filename):
+        """Initialize the SafeTensor reader.
+
+        Args:
+            filename (str): Path to the safetensors file to read.
+        """
+        self.filename = filename
+        self.file = open(filename, "rb")
+        self.header, self.header_size = self._read_header()
+
+    def __enter__(self):
+        """Enter context manager."""
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        """Exit context manager and close file."""
+        self.file.close()
+
+    def keys(self):
+        """Get all tensor keys in the file.
+
+        Returns:
+            list: List of tensor names (excludes metadata).
+        """
+        return [k for k in self.header.keys() if k != "__metadata__"]
+
+    def metadata(self) -> Dict[str, str]:
+        """Get metadata from the file.
+
+        Returns:
+            Dict[str, str]: Metadata dictionary.
+        """
+        return self.header.get("__metadata__", {})
+
+    def _read_header(self):
+        """Read and parse the header from the safetensors file.
+
+        Returns:
+            tuple: (header_dict, header_size) containing parsed header and its size.
+        """
+        # Read header size (8 bytes, little-endian unsigned long long)
+        header_size = struct.unpack("<Q", self.file.read(8))[0]
+        # Read and decode header JSON
+        header_json = self.file.read(header_size).decode("utf-8")
+        return json.loads(header_json), header_size
+
+    def get_tensor(self, key: str, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None):
+        """Load a tensor from the file with memory-efficient strategies.
+
+        **Note:**
+        If device is 'cuda' , the transfer to GPU is done efficiently using pinned memory and non-blocking transfer.
+        So you must ensure that the transfer is completed before using the tensor (e.g., by `torch.cuda.synchronize()`).
+
+        If the tensor is large (>10MB) and the target device is CUDA, memory mapping with numpy.memmap is used to avoid intermediate copies.
+
+        Args:
+            key (str): Name of the tensor to load.
+            device (Optional[torch.device]): Target device for the tensor.
+            dtype (Optional[torch.dtype]): Target dtype for the tensor.
+
+        Returns:
+            torch.Tensor: The loaded tensor.
+
+        Raises:
+            KeyError: If the tensor key is not found in the file.
+        """
+        if key not in self.header:
+            raise KeyError(f"Tensor '{key}' not found in the file")
+
+        metadata = self.header[key]
+        offset_start, offset_end = metadata["data_offsets"]
+        num_bytes = offset_end - offset_start
+
+        original_dtype = self._get_torch_dtype(metadata["dtype"])
+        target_dtype = dtype if dtype is not None else original_dtype
+
+        # Handle empty tensors
+        if num_bytes == 0:
+            return torch.empty(metadata["shape"], dtype=target_dtype, device=device)
+
+        # Determine if we should use pinned memory for GPU transfer
+        non_blocking = device is not None and device.type == "cuda"
+
+        # Calculate absolute file offset
+        tensor_offset = self.header_size + 8 + offset_start  # adjust offset by header size
+
+        # Memory mapping strategy for large tensors to GPU
+        # Use memmap for large tensors to avoid intermediate copies.
+        # If device is cpu, tensor is not copied to gpu, so using memmap locks the file, which is not desired.
+        # So we only use memmap if device is not cpu.
+        if num_bytes > 10 * 1024 * 1024 and device is not None and device.type != "cpu":
+            # Create memory map for zero-copy reading
+            mm = np.memmap(self.filename, mode="c", dtype=np.uint8, offset=tensor_offset, shape=(num_bytes,))
+            byte_tensor = torch.from_numpy(mm)  # zero copy
+            del mm
+
+            # Deserialize tensor (view and reshape)
+            cpu_tensor = self._deserialize_tensor(byte_tensor, metadata)  # view and reshape
+            del byte_tensor
+
+            # Transfer to target device and dtype
+            gpu_tensor = cpu_tensor.to(device=device, dtype=target_dtype, non_blocking=non_blocking)
+            del cpu_tensor
+            return gpu_tensor
+
+        # Standard file reading strategy for smaller tensors or CPU target
+        # seek to the specified position
+        self.file.seek(tensor_offset)
+
+        # read directly into a numpy array by numpy.fromfile without intermediate copy
+        numpy_array = np.fromfile(self.file, dtype=np.uint8, count=num_bytes)
+        byte_tensor = torch.from_numpy(numpy_array)
+        del numpy_array
+
+        # deserialize (view and reshape)
+        deserialized_tensor = self._deserialize_tensor(byte_tensor, metadata)
+        del byte_tensor
+
+        # cast to target dtype and move to device
+        return deserialized_tensor.to(device=device, dtype=target_dtype, non_blocking=non_blocking)
+
+    def _deserialize_tensor(self, byte_tensor: torch.Tensor, metadata: Dict):
+        """Deserialize byte tensor to the correct shape and dtype.
+
+        Args:
+            byte_tensor (torch.Tensor): Raw byte tensor from file.
+            metadata (Dict): Tensor metadata containing dtype and shape info.
+
+        Returns:
+            torch.Tensor: Deserialized tensor with correct shape and dtype.
+        """
+        dtype = self._get_torch_dtype(metadata["dtype"])
+        shape = metadata["shape"]
+
+        # Handle special float8 types
+        if metadata["dtype"] in ["F8_E5M2", "F8_E4M3"]:
+            return self._convert_float8(byte_tensor, metadata["dtype"], shape)
+
+        # Standard conversion: view as target dtype and reshape
+        return byte_tensor.view(dtype).reshape(shape)
+
+    @staticmethod
+    def _get_torch_dtype(dtype_str):
+        """Convert string dtype to PyTorch dtype.
+
+        Args:
+            dtype_str (str): String representation of the dtype.
+
+        Returns:
+            torch.dtype: Corresponding PyTorch dtype.
+        """
+        # Standard dtype mappings
+        dtype_map = {
+            "F64": torch.float64,
+            "F32": torch.float32,
+            "F16": torch.float16,
+            "BF16": torch.bfloat16,
+            "I64": torch.int64,
+            "I32": torch.int32,
+            "I16": torch.int16,
+            "I8": torch.int8,
+            "U8": torch.uint8,
+            "BOOL": torch.bool,
+        }
+        # Add float8 types if available in PyTorch version
+        if hasattr(torch, "float8_e5m2"):
+            dtype_map["F8_E5M2"] = torch.float8_e5m2
+        if hasattr(torch, "float8_e4m3fn"):
+            dtype_map["F8_E4M3"] = torch.float8_e4m3fn
+        return dtype_map.get(dtype_str)
+
+    @staticmethod
+    def _convert_float8(byte_tensor, dtype_str, shape):
+        """Convert byte tensor to float8 format if supported.
+
+        Args:
+            byte_tensor (torch.Tensor): Raw byte tensor.
+            dtype_str (str): Float8 dtype string ("F8_E5M2" or "F8_E4M3").
+            shape (tuple): Target tensor shape.
+
+        Returns:
+            torch.Tensor: Tensor with float8 dtype.
+
+        Raises:
+            ValueError: If float8 type is not supported in current PyTorch version.
+        """
+        # Convert to specific float8 types if available
+        if dtype_str == "F8_E5M2" and hasattr(torch, "float8_e5m2"):
+            return byte_tensor.view(torch.float8_e5m2).reshape(shape)
+        elif dtype_str == "F8_E4M3" and hasattr(torch, "float8_e4m3fn"):
+            return byte_tensor.view(torch.float8_e4m3fn).reshape(shape)
+        else:
+            # Float8 not supported in this PyTorch version
+            raise ValueError(f"Unsupported float8 type: {dtype_str} (upgrade PyTorch to support float8 types)")
+
+
+def load_safetensors(
+    path: str, device: Union[str, torch.device], disable_mmap: bool = False, dtype: Optional[torch.dtype] = None
+) -> dict[str, torch.Tensor]:
+    if disable_mmap:
+        # return safetensors.torch.load(open(path, "rb").read())
+        # use experimental loader
+        # logger.info(f"Loading without mmap (experimental)")
+        state_dict = {}
+        device = torch.device(device) if device is not None else None
+        with MemoryEfficientSafeOpen(path) as f:
+            for key in f.keys():
+                state_dict[key] = f.get_tensor(key, device=device, dtype=dtype)
+        synchronize_device(device)
+        return state_dict
+    else:
+        try:
+            state_dict = load_file(path, device=device)
+        except:
+            state_dict = load_file(path)  # prevent device invalid Error
+        if dtype is not None:
+            for key in state_dict.keys():
+                state_dict[key] = state_dict[key].to(dtype=dtype)
+        return state_dict
+
+
+def load_split_weights(
+    file_path: str, device: Union[str, torch.device] = "cpu", disable_mmap: bool = False, dtype: Optional[torch.dtype] = None
+) -> Dict[str, torch.Tensor]:
+    """
+    Load split weights from a file. If the file name ends with 00001-of-00004 etc, it will load all files with the same prefix.
+    dtype is as is, no conversion is done.
+    """
+    device = torch.device(device)
+
+    # if the file name ends with 00001-of-00004 etc, we need to load the files with the same prefix
+    basename = os.path.basename(file_path)
+    match = re.match(r"^(.*?)(\d+)-of-(\d+)\.safetensors$", basename)
+    if match:
+        prefix = basename[: match.start(2)]
+        count = int(match.group(3))
+        state_dict = {}
+        for i in range(count):
+            filename = f"{prefix}{i + 1:05d}-of-{count:05d}.safetensors"
+            filepath = os.path.join(os.path.dirname(file_path), filename)
+            if os.path.exists(filepath):
+                state_dict.update(load_safetensors(filepath, device=device, disable_mmap=disable_mmap, dtype=dtype))
+            else:
+                raise FileNotFoundError(f"File {filepath} not found")
+    else:
+        state_dict = load_safetensors(file_path, device=device, disable_mmap=disable_mmap, dtype=dtype)
+    return state_dict
+
+
+def find_key(safetensors_file: str, starts_with: Optional[str] = None, ends_with: Optional[str] = None) -> Optional[str]:
+    """
+    Find a key in a safetensors file that starts with `starts_with` and ends with `ends_with`.
+    If `starts_with` is None, it will match any key.
+    If `ends_with` is None, it will match any key.
+    Returns the first matching key or None if no key matches.
+    """
+    with MemoryEfficientSafeOpen(safetensors_file) as f:
+        for key in f.keys():
+            if (starts_with is None or key.startswith(starts_with)) and (ends_with is None or key.endswith(ends_with)):
+                return key
+    return None

library/sai_model_spec.py

@@ -1,10 +1,21 @@
 # based on https://github.com/Stability-AI/ModelSpec
 import datetime
 import hashlib
+import argparse
+import base64
+import logging
+import mimetypes
+import subprocess
+from dataclasses import dataclass, field
 from io import BytesIO
 import os
-from typing import List, Optional, Tuple, Union
+from typing import Union
 import safetensors
+from library.utils import setup_logging
+
+setup_logging()
+
+logger = logging.getLogger(__name__)
 
 r"""
 # Metadata Example
@@ -25,23 +36,32 @@ metadata = {
 """
 
 BASE_METADATA = {
-    # === Must ===
-    "modelspec.sai_model_spec": "1.0.0",  # Required version ID for the spec
+    # === MUST ===
+    "modelspec.sai_model_spec": "1.0.1",
     "modelspec.architecture": None,
     "modelspec.implementation": None,
     "modelspec.title": None,
     "modelspec.resolution": None,
-    # === Should ===
+    # === SHOULD ===
     "modelspec.description": None,
     "modelspec.author": None,
     "modelspec.date": None,
-    # === Can ===
+    "modelspec.hash_sha256": None,
+    # === CAN===
+    "modelspec.implementation_version": None,
     "modelspec.license": None,
+    "modelspec.usage_hint": None,
+    "modelspec.thumbnail": None,
     "modelspec.tags": None,
     "modelspec.merged_from": None,
+    "modelspec.trigger_phrase": None,
     "modelspec.prediction_type": None,
     "modelspec.timestep_range": None,
     "modelspec.encoder_layer": None,
+    "modelspec.preprocessor": None,
+    "modelspec.is_negative_embedding": None,
+    "modelspec.unet_dtype": None,
+    "modelspec.vae_dtype": None,
 }
 
 # 別に使うやつだけ定義
@@ -51,17 +71,270 @@ ARCH_SD_V1 = "stable-diffusion-v1"
 ARCH_SD_V2_512 = "stable-diffusion-v2-512"
 ARCH_SD_V2_768_V = "stable-diffusion-v2-768-v"
 ARCH_SD_XL_V1_BASE = "stable-diffusion-xl-v1-base"
+ARCH_SD3_M = "stable-diffusion-3"  # may be followed by "-m" or "-5-large" etc.
+# ARCH_SD3_UNKNOWN = "stable-diffusion-3"
+ARCH_FLUX_1_DEV = "flux-1-dev"
+ARCH_FLUX_1_SCHNELL = "flux-1-schnell"
+ARCH_FLUX_1_CHROMA = "chroma"  # for Flux Chroma
+ARCH_FLUX_1_UNKNOWN = "flux-1"
+ARCH_LUMINA_2 = "lumina-2"
+ARCH_LUMINA_UNKNOWN = "lumina"
+ARCH_HUNYUAN_IMAGE_2_1 = "hunyuan-image-2.1"
+ARCH_HUNYUAN_IMAGE_UNKNOWN = "hunyuan-image"
 
 ADAPTER_LORA = "lora"
 ADAPTER_TEXTUAL_INVERSION = "textual-inversion"
 
 IMPL_STABILITY_AI = "https://github.com/Stability-AI/generative-models"
+IMPL_COMFY_UI = "https://github.com/comfyanonymous/ComfyUI"
 IMPL_DIFFUSERS = "diffusers"
+IMPL_FLUX = "https://github.com/black-forest-labs/flux"
+IMPL_CHROMA = "https://huggingface.co/lodestones/Chroma"
+IMPL_LUMINA = "https://github.com/Alpha-VLLM/Lumina-Image-2.0"
+IMPL_HUNYUAN_IMAGE = "https://github.com/Tencent-Hunyuan/HunyuanImage-2.1"
 
 PRED_TYPE_EPSILON = "epsilon"
 PRED_TYPE_V = "v"
 
 
+@dataclass
+class ModelSpecMetadata:
+    """
+    ModelSpec 1.0.1 compliant metadata for safetensors models.
+    All fields correspond to modelspec.* keys in the final metadata.
+    """
+
+    # === MUST ===
+    architecture: str
+    implementation: str
+    title: str
+    resolution: str
+    sai_model_spec: str = "1.0.1"
+
+    # === SHOULD ===
+    description: str | None = None
+    author: str | None = None
+    date: str | None = None
+    hash_sha256: str | None = None
+
+    # === CAN ===
+    implementation_version: str | None = None
+    license: str | None = None
+    usage_hint: str | None = None
+    thumbnail: str | None = None
+    tags: str | None = None
+    merged_from: str | None = None
+    trigger_phrase: str | None = None
+    prediction_type: str | None = None
+    timestep_range: str | None = None
+    encoder_layer: str | None = None
+    preprocessor: str | None = None
+    is_negative_embedding: str | None = None
+    unet_dtype: str | None = None
+    vae_dtype: str | None = None
+
+    # === Additional metadata ===
+    additional_fields: dict[str, str] = field(default_factory=dict)
+
+    def to_metadata_dict(self) -> dict[str, str]:
+        """Convert dataclass to metadata dictionary with modelspec. prefixes."""
+        metadata = {}
+
+        # Add all non-None fields with modelspec prefix
+        for field_name, value in self.__dict__.items():
+            if field_name == "additional_fields":
+                # Handle additional fields separately
+                for key, val in value.items():
+                    if key.startswith("modelspec."):
+                        metadata[key] = val
+                    else:
+                        metadata[f"modelspec.{key}"] = val
+            elif value is not None:
+                metadata[f"modelspec.{field_name}"] = value
+
+        return metadata
+
+    @classmethod
+    def from_args(cls, args, **kwargs) -> "ModelSpecMetadata":
+        """Create ModelSpecMetadata from argparse Namespace, extracting metadata_* fields."""
+        metadata_fields = {}
+
+        # Extract all metadata_* attributes from args
+        for attr_name in dir(args):
+            if attr_name.startswith("metadata_") and not attr_name.startswith("metadata___"):
+                value = getattr(args, attr_name, None)
+                if value is not None:
+                    # Remove metadata_ prefix
+                    field_name = attr_name[9:]  # len("metadata_") = 9
+                    metadata_fields[field_name] = value
+
+        # Handle known standard fields
+        standard_fields = {
+            "author": metadata_fields.pop("author", None),
+            "description": metadata_fields.pop("description", None),
+            "license": metadata_fields.pop("license", None),
+            "tags": metadata_fields.pop("tags", None),
+        }
+
+        # Remove None values
+        standard_fields = {k: v for k, v in standard_fields.items() if v is not None}
+
+        # Merge with kwargs and remaining metadata fields
+        all_fields = {**standard_fields, **kwargs}
+        if metadata_fields:
+            all_fields["additional_fields"] = metadata_fields
+
+        return cls(**all_fields)
+
+
+def determine_architecture(
+    v2: bool, v_parameterization: bool, sdxl: bool, lora: bool, textual_inversion: bool, model_config: dict[str, str] | None = None
+) -> str:
+    """Determine model architecture string from parameters."""
+
+    model_config = model_config or {}
+
+    if sdxl:
+        arch = ARCH_SD_XL_V1_BASE
+    elif "sd3" in model_config:
+        arch = ARCH_SD3_M + "-" + model_config["sd3"]
+    elif "flux" in model_config:
+        flux_type = model_config["flux"]
+        if flux_type == "dev":
+            arch = ARCH_FLUX_1_DEV
+        elif flux_type == "schnell":
+            arch = ARCH_FLUX_1_SCHNELL
+        elif flux_type == "chroma":
+            arch = ARCH_FLUX_1_CHROMA
+        else:
+            arch = ARCH_FLUX_1_UNKNOWN
+    elif "lumina" in model_config:
+        lumina_type = model_config["lumina"]
+        if lumina_type == "lumina2":
+            arch = ARCH_LUMINA_2
+        else:
+            arch = ARCH_LUMINA_UNKNOWN
+    elif "hunyuan_image" in model_config:
+        hunyuan_image_type = model_config["hunyuan_image"]
+        if hunyuan_image_type == "2.1":
+            arch = ARCH_HUNYUAN_IMAGE_2_1
+        else:
+            arch = ARCH_HUNYUAN_IMAGE_UNKNOWN
+    elif v2:
+        arch = ARCH_SD_V2_768_V if v_parameterization else ARCH_SD_V2_512
+    else:
+        arch = ARCH_SD_V1
+
+    # Add adapter suffix
+    if lora:
+        arch += f"/{ADAPTER_LORA}"
+    elif textual_inversion:
+        arch += f"/{ADAPTER_TEXTUAL_INVERSION}"
+
+    return arch
+
+
+def determine_implementation(
+    lora: bool,
+    textual_inversion: bool,
+    sdxl: bool,
+    model_config: dict[str, str] | None = None,
+    is_stable_diffusion_ckpt: bool | None = None,
+) -> str:
+    """Determine implementation string from parameters."""
+
+    model_config = model_config or {}
+
+    if "flux" in model_config:
+        if model_config["flux"] == "chroma":
+            return IMPL_CHROMA
+        else:
+            return IMPL_FLUX
+    elif "lumina" in model_config:
+        return IMPL_LUMINA
+    elif (lora and sdxl) or textual_inversion or is_stable_diffusion_ckpt:
+        return IMPL_STABILITY_AI
+    else:
+        return IMPL_DIFFUSERS
+
+
+def get_implementation_version() -> str:
+    """Get the current implementation version as sd-scripts/{commit_hash}."""
+    try:
+        # Get the git commit hash
+        result = subprocess.run(
+            ["git", "rev-parse", "HEAD"],
+            capture_output=True,
+            text=True,
+            cwd=os.path.dirname(os.path.dirname(__file__)),  # Go up to sd-scripts root
+            timeout=5,
+        )
+
+        if result.returncode == 0:
+            commit_hash = result.stdout.strip()
+            return f"sd-scripts/{commit_hash}"
+        else:
+            logger.warning("Failed to get git commit hash, using fallback")
+            return "sd-scripts/unknown"
+
+    except (subprocess.TimeoutExpired, subprocess.SubprocessError, FileNotFoundError) as e:
+        logger.warning(f"Could not determine git commit: {e}")
+        return "sd-scripts/unknown"
+
+
+def file_to_data_url(file_path: str) -> str:
+    """Convert a file path to a data URL for embedding in metadata."""
+    if not os.path.exists(file_path):
+        raise FileNotFoundError(f"File not found: {file_path}")
+
+    # Get MIME type
+    mime_type, _ = mimetypes.guess_type(file_path)
+    if mime_type is None:
+        # Default to binary if we can't detect
+        mime_type = "application/octet-stream"
+
+    # Read file and encode as base64
+    with open(file_path, "rb") as f:
+        file_data = f.read()
+
+    encoded_data = base64.b64encode(file_data).decode("ascii")
+
+    return f"data:{mime_type};base64,{encoded_data}"
+
+
+def determine_resolution(
+    reso: Union[int, tuple[int, int]] | None = None,
+    sdxl: bool = False,
+    model_config: dict[str, str] | None = None,
+    v2: bool = False,
+    v_parameterization: bool = False,
+) -> str:
+    """Determine resolution string from parameters."""
+
+    model_config = model_config or {}
+
+    if reso is not None:
+        # Handle comma separated string
+        if isinstance(reso, str):
+            reso = tuple(map(int, reso.split(",")))
+        # Handle single int
+        if isinstance(reso, int):
+            reso = (reso, reso)
+        # Handle single-element tuple
+        if len(reso) == 1:
+            reso = (reso[0], reso[0])
+    else:
+        # Determine default resolution based on model type
+        if sdxl or "sd3" in model_config or "flux" in model_config or "lumina" in model_config:
+            reso = (1024, 1024)
+        elif v2 and v_parameterization:
+            reso = (768, 768)
+        else:
+            reso = (512, 512)
+
+    return f"{reso[0]}x{reso[1]}"
+
+
 def load_bytes_in_safetensors(tensors):
     bytes = safetensors.torch.save(tensors)
     b = BytesIO(bytes)
@@ -91,62 +364,42 @@ def update_hash_sha256(metadata: dict, state_dict: dict):
     raise NotImplementedError
 
 
-def build_metadata(
-    state_dict: Optional[dict],
+def build_metadata_dataclass(
+    state_dict: dict | None,
     v2: bool,
     v_parameterization: bool,
     sdxl: bool,
     lora: bool,
     textual_inversion: bool,
     timestamp: float,
-    title: Optional[str] = None,
-    reso: Optional[Union[int, Tuple[int, int]]] = None,
-    is_stable_diffusion_ckpt: Optional[bool] = None,
-    author: Optional[str] = None,
-    description: Optional[str] = None,
-    license: Optional[str] = None,
-    tags: Optional[str] = None,
-    merged_from: Optional[str] = None,
-    timesteps: Optional[Tuple[int, int]] = None,
-    clip_skip: Optional[int] = None,
-):
-    # if state_dict is None, hash is not calculated
+    title: str | None = None,
+    reso: int | tuple[int, int] | None = None,
+    is_stable_diffusion_ckpt: bool | None = None,
+    author: str | None = None,
+    description: str | None = None,
+    license: str | None = None,
+    tags: str | None = None,
+    merged_from: str | None = None,
+    timesteps: tuple[int, int] | None = None,
+    clip_skip: int | None = None,
+    model_config: dict | None = None,
+    optional_metadata: dict | None = None,
+) -> ModelSpecMetadata:
+    """
+    Build ModelSpec 1.0.1 compliant metadata dataclass.
 
-    metadata = {}
-    metadata.update(BASE_METADATA)
+    Args:
+        model_config: Dict containing model type info, e.g. {"flux": "dev"}, {"sd3": "large"}
+        optional_metadata: Dict of additional metadata fields to include
+    """
 
-    # TODO メモリを消費せずかつ正しいハッシュ計算の方法がわかったら実装する
-    # if state_dict is not None:
-    # hash = precalculate_safetensors_hashes(state_dict)
-    # metadata["modelspec.hash_sha256"] = hash
-
-    if sdxl:
-        arch = ARCH_SD_XL_V1_BASE
-    elif v2:
-        if v_parameterization:
-            arch = ARCH_SD_V2_768_V
-        else:
-            arch = ARCH_SD_V2_512
-    else:
-        arch = ARCH_SD_V1
-
-    if lora:
-        arch += f"/{ADAPTER_LORA}"
-    elif textual_inversion:
-        arch += f"/{ADAPTER_TEXTUAL_INVERSION}"
-
-    metadata["modelspec.architecture"] = arch
+    # Use helper functions for complex logic
+    architecture = determine_architecture(v2, v_parameterization, sdxl, lora, textual_inversion, model_config)
 
     if not lora and not textual_inversion and is_stable_diffusion_ckpt is None:
-        is_stable_diffusion_ckpt = True # default is stable diffusion ckpt if not lora and not textual_inversion
+        is_stable_diffusion_ckpt = True  # default is stable diffusion ckpt if not lora and not textual_inversion
 
-    if (lora and sdxl) or textual_inversion or is_stable_diffusion_ckpt:
-        # Stable Diffusion ckpt, TI, SDXL LoRA
-        impl = IMPL_STABILITY_AI
-    else:
-        # v1/v2 LoRA or Diffusers
-        impl = IMPL_DIFFUSERS
-    metadata["modelspec.implementation"] = impl
+    implementation = determine_implementation(lora, textual_inversion, sdxl, model_config, is_stable_diffusion_ckpt)
 
     if title is None:
         if lora:
@@ -156,97 +409,150 @@ def build_metadata(
         else:
             title = "Checkpoint"
         title += f"@{timestamp}"
-    metadata[MODELSPEC_TITLE] = title
-
-    if author is not None:
-        metadata["modelspec.author"] = author
-    else:
-        del metadata["modelspec.author"]
-
-    if description is not None:
-        metadata["modelspec.description"] = description
-    else:
-        del metadata["modelspec.description"]
-
-    if merged_from is not None:
-        metadata["modelspec.merged_from"] = merged_from
-    else:
-        del metadata["modelspec.merged_from"]
-
-    if license is not None:
-        metadata["modelspec.license"] = license
-    else:
-        del metadata["modelspec.license"]
-
-    if tags is not None:
-        metadata["modelspec.tags"] = tags
-    else:
-        del metadata["modelspec.tags"]
 
     # remove microsecond from time
     int_ts = int(timestamp)
-
     # time to iso-8601 compliant date
     date = datetime.datetime.fromtimestamp(int_ts).isoformat()
-    metadata["modelspec.date"] = date
 
-    if reso is not None:
-        # comma separated to tuple
-        if isinstance(reso, str):
-            reso = tuple(map(int, reso.split(",")))
-        if len(reso) == 1:
-            reso = (reso[0], reso[0])
-    else:
-        # resolution is defined in dataset, so use default
-        if sdxl:
-            reso = 1024
-        elif v2 and v_parameterization:
-            reso = 768
+    # Use helper function for resolution
+    resolution = determine_resolution(reso, sdxl, model_config, v2, v_parameterization)
+
+    # Handle prediction type - Flux models don't use prediction_type
+    model_config = model_config or {}
+    prediction_type = None
+    if "flux" not in model_config:
+        if v_parameterization:
+            prediction_type = PRED_TYPE_V
         else:
-            reso = 512
-    if isinstance(reso, int):
-        reso = (reso, reso)
-
-    metadata["modelspec.resolution"] = f"{reso[0]}x{reso[1]}"
-
-    if v_parameterization:
-        metadata["modelspec.prediction_type"] = PRED_TYPE_V
-    else:
-        metadata["modelspec.prediction_type"] = PRED_TYPE_EPSILON
+            prediction_type = PRED_TYPE_EPSILON
 
+    # Handle timesteps
+    timestep_range = None
     if timesteps is not None:
         if isinstance(timesteps, str) or isinstance(timesteps, int):
             timesteps = (timesteps, timesteps)
         if len(timesteps) == 1:
             timesteps = (timesteps[0], timesteps[0])
-        metadata["modelspec.timestep_range"] = f"{timesteps[0]},{timesteps[1]}"
-    else:
-        del metadata["modelspec.timestep_range"]
+        timestep_range = f"{timesteps[0]},{timesteps[1]}"
 
+    # Handle encoder layer (clip skip)
+    encoder_layer = None
     if clip_skip is not None:
-        metadata["modelspec.encoder_layer"] = f"{clip_skip}"
-    else:
-        del metadata["modelspec.encoder_layer"]
+        encoder_layer = f"{clip_skip}"
+
+    # TODO: Implement hash calculation when memory-efficient method is available
+    # hash_sha256 = None
+    # if state_dict is not None:
+    #     hash_sha256 = precalculate_safetensors_hashes(state_dict)
+
+    # Process thumbnail - convert file path to data URL if needed
+    processed_optional_metadata = optional_metadata.copy() if optional_metadata else {}
+    if "thumbnail" in processed_optional_metadata:
+        thumbnail_value = processed_optional_metadata["thumbnail"]
+        # Check if it's already a data URL or if it's a file path
+        if thumbnail_value and not thumbnail_value.startswith("data:"):
+            try:
+                processed_optional_metadata["thumbnail"] = file_to_data_url(thumbnail_value)
+                logger.info(f"Converted thumbnail file {thumbnail_value} to data URL")
+            except FileNotFoundError as e:
+                logger.warning(f"Thumbnail file not found, skipping: {e}")
+                del processed_optional_metadata["thumbnail"]
+            except Exception as e:
+                logger.warning(f"Failed to convert thumbnail to data URL: {e}")
+                del processed_optional_metadata["thumbnail"]
+
+    # Automatically set implementation version if not provided
+    if "implementation_version" not in processed_optional_metadata:
+        processed_optional_metadata["implementation_version"] = get_implementation_version()
+
+    # Create the dataclass
+    metadata = ModelSpecMetadata(
+        architecture=architecture,
+        implementation=implementation,
+        title=title,
+        description=description,
+        author=author,
+        date=date,
+        license=license,
+        tags=tags,
+        merged_from=merged_from,
+        resolution=resolution,
+        prediction_type=prediction_type,
+        timestep_range=timestep_range,
+        encoder_layer=encoder_layer,
+        additional_fields=processed_optional_metadata,
+    )
 
-    # # assert all values are filled
-    # assert all([v is not None for v in metadata.values()]), metadata
-    if not all([v is not None for v in metadata.values()]):
-        print(f"Internal error: some metadata values are None: {metadata}")
-    
     return metadata
 
 
+def build_metadata(
+    state_dict: dict | None,
+    v2: bool,
+    v_parameterization: bool,
+    sdxl: bool,
+    lora: bool,
+    textual_inversion: bool,
+    timestamp: float,
+    title: str | None = None,
+    reso: int | tuple[int, int] | None = None,
+    is_stable_diffusion_ckpt: bool | None = None,
+    author: str | None = None,
+    description: str | None = None,
+    license: str | None = None,
+    tags: str | None = None,
+    merged_from: str | None = None,
+    timesteps: tuple[int, int] | None = None,
+    clip_skip: int | None = None,
+    model_config: dict | None = None,
+    optional_metadata: dict | None = None,
+) -> dict[str, str]:
+    """
+    Build ModelSpec 1.0.1 compliant metadata for safetensors models.
+    Legacy function that returns dict - prefer build_metadata_dataclass for new code.
+
+    Args:
+        model_config: Dict containing model type info, e.g. {"flux": "dev"}, {"sd3": "large"}
+        optional_metadata: Dict of additional metadata fields to include
+    """
+    # Use the dataclass function and convert to dict
+    metadata_obj = build_metadata_dataclass(
+        state_dict=state_dict,
+        v2=v2,
+        v_parameterization=v_parameterization,
+        sdxl=sdxl,
+        lora=lora,
+        textual_inversion=textual_inversion,
+        timestamp=timestamp,
+        title=title,
+        reso=reso,
+        is_stable_diffusion_ckpt=is_stable_diffusion_ckpt,
+        author=author,
+        description=description,
+        license=license,
+        tags=tags,
+        merged_from=merged_from,
+        timesteps=timesteps,
+        clip_skip=clip_skip,
+        model_config=model_config,
+        optional_metadata=optional_metadata,
+    )
+
+    return metadata_obj.to_metadata_dict()
+
+
 # region utils
 
 
-def get_title(metadata: dict) -> Optional[str]:
+def get_title(metadata: dict) -> str | None:
     return metadata.get(MODELSPEC_TITLE, None)
 
 
 def load_metadata_from_safetensors(model: str) -> dict:
     if not model.endswith(".safetensors"):
         return {}
-    
+
     with safetensors.safe_open(model, framework="pt") as f:
         metadata = f.metadata()
     if metadata is None:
@@ -254,7 +560,7 @@ def load_metadata_from_safetensors(model: str) -> dict:
     return metadata
 
 
-def build_merged_from(models: List[str]) -> str:
+def build_merged_from(models: list[str]) -> str:
     def get_title(model: str):
         metadata = load_metadata_from_safetensors(model)
         title = metadata.get(MODELSPEC_TITLE, None)
@@ -266,6 +572,77 @@ def build_merged_from(models: List[str]) -> str:
     return ", ".join(titles)
 
 
+def add_model_spec_arguments(parser: argparse.ArgumentParser):
+    """Add all ModelSpec metadata arguments to the parser."""
+
+    parser.add_argument(
+        "--metadata_title",
+        type=str,
+        default=None,
+        help="title for model metadata (default is output_name) / メタデータに書き込まれるモデルタイトル、省略時はoutput_name",
+    )
+    parser.add_argument(
+        "--metadata_author",
+        type=str,
+        default=None,
+        help="author name for model metadata / メタデータに書き込まれるモデル作者名",
+    )
+    parser.add_argument(
+        "--metadata_description",
+        type=str,
+        default=None,
+        help="description for model metadata / メタデータに書き込まれるモデル説明",
+    )
+    parser.add_argument(
+        "--metadata_license",
+        type=str,
+        default=None,
+        help="license for model metadata / メタデータに書き込まれるモデルライセンス",
+    )
+    parser.add_argument(
+        "--metadata_tags",
+        type=str,
+        default=None,
+        help="tags for model metadata, separated by comma / メタデータに書き込まれるモデルタグ、カンマ区切り",
+    )
+    parser.add_argument(
+        "--metadata_usage_hint",
+        type=str,
+        default=None,
+        help="usage hint for model metadata / メタデータに書き込まれる使用方法のヒント",
+    )
+    parser.add_argument(
+        "--metadata_thumbnail",
+        type=str,
+        default=None,
+        help="thumbnail image as data URL or file path (will be converted to data URL) for model metadata / メタデータに書き込まれるサムネイル画像（データURLまたはファイルパス、ファイルパスの場合はデータURLに変換されます）",
+    )
+    parser.add_argument(
+        "--metadata_merged_from",
+        type=str,
+        default=None,
+        help="source models for merged model metadata / メタデータに書き込まれるマージ元モデル名",
+    )
+    parser.add_argument(
+        "--metadata_trigger_phrase",
+        type=str,
+        default=None,
+        help="trigger phrase for model metadata / メタデータに書き込まれるトリガーフレーズ",
+    )
+    parser.add_argument(
+        "--metadata_preprocessor",
+        type=str,
+        default=None,
+        help="preprocessor used for model metadata / メタデータに書き込まれる前処理手法",
+    )
+    parser.add_argument(
+        "--metadata_is_negative_embedding",
+        type=str,
+        default=None,
+        help="whether this is a negative embedding for model metadata / メタデータに書き込まれるネガティブ埋め込みかどうか",
+    )
+
+
 # endregion
 
 

library/sd3_train_utils.py

@@ -0,0 +1,945 @@
+import argparse
+import math
+import os
+import toml
+import json
+import time
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+from safetensors.torch import save_file
+from accelerate import Accelerator, PartialState
+from tqdm import tqdm
+from PIL import Image
+from transformers import CLIPTextModelWithProjection, T5EncoderModel
+
+from library.device_utils import init_ipex, clean_memory_on_device
+
+init_ipex()
+
+# from transformers import CLIPTokenizer
+# from library import model_util
+# , sdxl_model_util, train_util, sdxl_original_unet
+# from library.sdxl_lpw_stable_diffusion import SdxlStableDiffusionLongPromptWeightingPipeline
+from .utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+from library import sd3_models, sd3_utils, strategy_base, train_util
+
+
+def save_models(
+    ckpt_path: str,
+    mmdit: Optional[sd3_models.MMDiT],
+    vae: Optional[sd3_models.SDVAE],
+    clip_l: Optional[CLIPTextModelWithProjection],
+    clip_g: Optional[CLIPTextModelWithProjection],
+    t5xxl: Optional[T5EncoderModel],
+    sai_metadata: Optional[dict],
+    save_dtype: Optional[torch.dtype] = None,
+):
+    r"""
+    Save models to checkpoint file. Only supports unified checkpoint format.
+    """
+
+    state_dict = {}
+
+    def update_sd(prefix, sd):
+        for k, v in sd.items():
+            key = prefix + k
+            if save_dtype is not None:
+                v = v.detach().clone().to("cpu").to(save_dtype)
+            state_dict[key] = v
+
+    update_sd("model.diffusion_model.", mmdit.state_dict())
+    update_sd("first_stage_model.", vae.state_dict())
+
+    # do not support unified checkpoint format for now
+    # if clip_l is not None:
+    #     update_sd("text_encoders.clip_l.", clip_l.state_dict())
+    # if clip_g is not None:
+    #     update_sd("text_encoders.clip_g.", clip_g.state_dict())
+    # if t5xxl is not None:
+    #     update_sd("text_encoders.t5xxl.", t5xxl.state_dict())
+
+    save_file(state_dict, ckpt_path, metadata=sai_metadata)
+
+    if clip_l is not None:
+        clip_l_path = ckpt_path.replace(".safetensors", "_clip_l.safetensors")
+        save_file(clip_l.state_dict(), clip_l_path)
+    if clip_g is not None:
+        clip_g_path = ckpt_path.replace(".safetensors", "_clip_g.safetensors")
+        save_file(clip_g.state_dict(), clip_g_path)
+    if t5xxl is not None:
+        t5xxl_path = ckpt_path.replace(".safetensors", "_t5xxl.safetensors")
+        t5xxl_state_dict = t5xxl.state_dict()
+
+        # replace "shared.weight" with copy of it to avoid annoying shared tensor error on safetensors.save_file
+        shared_weight = t5xxl_state_dict["shared.weight"]
+        shared_weight_copy = shared_weight.detach().clone()
+        t5xxl_state_dict["shared.weight"] = shared_weight_copy
+
+        save_file(t5xxl_state_dict, t5xxl_path)
+
+
+def save_sd3_model_on_train_end(
+    args: argparse.Namespace,
+    save_dtype: torch.dtype,
+    epoch: int,
+    global_step: int,
+    clip_l: Optional[CLIPTextModelWithProjection],
+    clip_g: Optional[CLIPTextModelWithProjection],
+    t5xxl: Optional[T5EncoderModel],
+    mmdit: sd3_models.MMDiT,
+    vae: sd3_models.SDVAE,
+):
+    def sd_saver(ckpt_file, epoch_no, global_step):
+        sai_metadata = train_util.get_sai_model_spec(
+            None, args, False, False, False, is_stable_diffusion_ckpt=True, sd3=mmdit.model_type
+        )
+        save_models(ckpt_file, mmdit, vae, clip_l, clip_g, t5xxl, sai_metadata, save_dtype)
+
+    train_util.save_sd_model_on_train_end_common(args, True, True, epoch, global_step, sd_saver, None)
+
+
+# epochとstepの保存、メタデータにepoch/stepが含まれ引数が同じになるため、統合している
+# on_epoch_end: Trueならepoch終了時、Falseならstep経過時
+def save_sd3_model_on_epoch_end_or_stepwise(
+    args: argparse.Namespace,
+    on_epoch_end: bool,
+    accelerator,
+    save_dtype: torch.dtype,
+    epoch: int,
+    num_train_epochs: int,
+    global_step: int,
+    clip_l: Optional[CLIPTextModelWithProjection],
+    clip_g: Optional[CLIPTextModelWithProjection],
+    t5xxl: Optional[T5EncoderModel],
+    mmdit: sd3_models.MMDiT,
+    vae: sd3_models.SDVAE,
+):
+    def sd_saver(ckpt_file, epoch_no, global_step):
+        sai_metadata = train_util.get_sai_model_spec(
+            None, args, False, False, False, is_stable_diffusion_ckpt=True, sd3=mmdit.model_type
+        )
+        save_models(ckpt_file, mmdit, vae, clip_l, clip_g, t5xxl, sai_metadata, save_dtype)
+
+    train_util.save_sd_model_on_epoch_end_or_stepwise_common(
+        args,
+        on_epoch_end,
+        accelerator,
+        True,
+        True,
+        epoch,
+        num_train_epochs,
+        global_step,
+        sd_saver,
+        None,
+    )
+
+
+def add_sd3_training_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--clip_l",
+        type=str,
+        required=False,
+        help="CLIP-L model path. if not specified, use ckpt's state_dict / CLIP-Lモデルのパス。指定しない場合はckptのstate_dictを使用",
+    )
+    parser.add_argument(
+        "--clip_g",
+        type=str,
+        required=False,
+        help="CLIP-G model path. if not specified, use ckpt's state_dict / CLIP-Gモデルのパス。指定しない場合はckptのstate_dictを使用",
+    )
+    parser.add_argument(
+        "--t5xxl",
+        type=str,
+        required=False,
+        help="T5-XXL model path. if not specified, use ckpt's state_dict / T5-XXLモデルのパス。指定しない場合はckptのstate_dictを使用",
+    )
+    parser.add_argument(
+        "--save_clip",
+        action="store_true",
+        help="[DOES NOT WORK] unified checkpoint is not supported / 統合チェックポイントはまだサポートされていません",
+    )
+    parser.add_argument(
+        "--save_t5xxl",
+        action="store_true",
+        help="[DOES NOT WORK] unified checkpoint is not supported / 統合チェックポイントはまだサポートされていません",
+    )
+
+    parser.add_argument(
+        "--t5xxl_device",
+        type=str,
+        default=None,
+        help="[DOES NOT WORK] not supported yet. T5-XXL device. if not specified, use accelerator's device / T5-XXLデバイス。指定しない場合はacceleratorのデバイスを使用",
+    )
+    parser.add_argument(
+        "--t5xxl_dtype",
+        type=str,
+        default=None,
+        help="[DOES NOT WORK] not supported yet. T5-XXL dtype. if not specified, use default dtype (from mixed precision) / T5-XXL dtype。指定しない場合はデフォルトのdtype（mixed precisionから）を使用",
+    )
+
+    parser.add_argument(
+        "--t5xxl_max_token_length",
+        type=int,
+        default=256,
+        help="maximum token length for T5-XXL. 256 is the default value / T5-XXLの最大トークン長。デフォルトは256",
+    )
+    parser.add_argument(
+        "--apply_lg_attn_mask",
+        action="store_true",
+        help="apply attention mask (zero embs) to CLIP-L and G / CLIP-LとGにアテンションマスク（ゼロ埋め）を適用する",
+    )
+    parser.add_argument(
+        "--apply_t5_attn_mask",
+        action="store_true",
+        help="apply attention mask (zero embs) to T5-XXL / T5-XXLにアテンションマスク（ゼロ埋め）を適用する",
+    )
+    parser.add_argument(
+        "--clip_l_dropout_rate",
+        type=float,
+        default=0.0,
+        help="Dropout rate for CLIP-L encoder, default is 0.0 / CLIP-Lエンコーダのドロップアウト率、デフォルトは0.0",
+    )
+    parser.add_argument(
+        "--clip_g_dropout_rate",
+        type=float,
+        default=0.0,
+        help="Dropout rate for CLIP-G encoder, default is 0.0 / CLIP-Gエンコーダのドロップアウト率、デフォルトは0.0",
+    )
+    parser.add_argument(
+        "--t5_dropout_rate",
+        type=float,
+        default=0.0,
+        help="Dropout rate for T5 encoder, default is 0.0 / T5エンコーダのドロップアウト率、デフォルトは0.0",
+    )
+    parser.add_argument(
+        "--pos_emb_random_crop_rate",
+        type=float,
+        default=0.0,
+        help="Random crop rate for positional embeddings, default is 0.0. Only for SD3.5M"
+        " / 位置埋め込みのランダムクロップ率、デフォルトは0.0。SD3.5M以外では予期しない動作になります",
+    )
+    parser.add_argument(
+        "--enable_scaled_pos_embed",
+        action="store_true",
+        help="Scale position embeddings for each resolution during multi-resolution training. Only for SD3.5M"
+        " / 複数解像度学習時に解像度ごとに位置埋め込みをスケーリングする。SD3.5M以外では予期しない動作になります",
+    )
+
+    # Dependencies of Diffusers noise sampler has been removed for clarity in training
+
+    parser.add_argument(
+        "--training_shift",
+        type=float,
+        default=1.0,
+        help="Discrete flow shift for training timestep distribution adjustment, applied in addition to the weighting scheme, default is 1.0. /タイムステップ分布のための離散フローシフト、重み付けスキームの上に適用される、デフォルトは1.0。",
+    )
+
+
+def verify_sdxl_training_args(args: argparse.Namespace, supportTextEncoderCaching: bool = True):
+    assert not args.v2, "v2 cannot be enabled in SDXL training / SDXL学習ではv2を有効にすることはできません"
+    if args.v_parameterization:
+        logger.warning("v_parameterization will be unexpected / SDXL学習ではv_parameterizationは想定外の動作になります")
+
+    if args.clip_skip is not None:
+        logger.warning("clip_skip will be unexpected / SDXL学習ではclip_skipは動作しません")
+
+    # if args.multires_noise_iterations:
+    #     logger.info(
+    #         f"Warning: SDXL has been trained with noise_offset={DEFAULT_NOISE_OFFSET}, but noise_offset is disabled due to multires_noise_iterations / SDXLはnoise_offset={DEFAULT_NOISE_OFFSET}で学習されていますが、multires_noise_iterationsが有効になっているためnoise_offsetは無効になります"
+    #     )
+    # else:
+    #     if args.noise_offset is None:
+    #         args.noise_offset = DEFAULT_NOISE_OFFSET
+    #     elif args.noise_offset != DEFAULT_NOISE_OFFSET:
+    #         logger.info(
+    #             f"Warning: SDXL has been trained with noise_offset={DEFAULT_NOISE_OFFSET} / SDXLはnoise_offset={DEFAULT_NOISE_OFFSET}で学習されています"
+    #         )
+    #     logger.info(f"noise_offset is set to {args.noise_offset} / noise_offsetが{args.noise_offset}に設定されました")
+
+    assert (
+        not hasattr(args, "weighted_captions") or not args.weighted_captions
+    ), "weighted_captions cannot be enabled in SDXL training currently / SDXL学習では今のところweighted_captionsを有効にすることはできません"
+
+    if supportTextEncoderCaching:
+        if args.cache_text_encoder_outputs_to_disk and not args.cache_text_encoder_outputs:
+            args.cache_text_encoder_outputs = True
+            logger.warning(
+                "cache_text_encoder_outputs is enabled because cache_text_encoder_outputs_to_disk is enabled / "
+                + "cache_text_encoder_outputs_to_diskが有効になっているためcache_text_encoder_outputsが有効になりました"
+            )
+
+
+# temporary copied from sd3_minimal_inferece.py
+
+
+def get_all_sigmas(sampling: sd3_utils.ModelSamplingDiscreteFlow, steps):
+    start = sampling.timestep(sampling.sigma_max)
+    end = sampling.timestep(sampling.sigma_min)
+    timesteps = torch.linspace(start, end, steps)
+    sigs = []
+    for x in range(len(timesteps)):
+        ts = timesteps[x]
+        sigs.append(sampling.sigma(ts))
+    sigs += [0.0]
+    return torch.FloatTensor(sigs)
+
+
+def max_denoise(model_sampling, sigmas):
+    max_sigma = float(model_sampling.sigma_max)
+    sigma = float(sigmas[0])
+    return math.isclose(max_sigma, sigma, rel_tol=1e-05) or sigma > max_sigma
+
+
+def do_sample(
+    height: int,
+    width: int,
+    seed: int,
+    cond: Tuple[torch.Tensor, torch.Tensor],
+    neg_cond: Tuple[torch.Tensor, torch.Tensor],
+    mmdit: sd3_models.MMDiT,
+    steps: int,
+    guidance_scale: float,
+    dtype: torch.dtype,
+    device: str,
+):
+    latent = torch.zeros(1, 16, height // 8, width // 8, device=device)
+    latent = latent.to(dtype).to(device)
+
+    # noise = get_noise(seed, latent).to(device)
+    if seed is not None:
+        generator = torch.manual_seed(seed)
+    else:
+        generator = None
+    noise = (
+        torch.randn(latent.size(), dtype=torch.float32, layout=latent.layout, generator=generator, device="cpu")
+        .to(latent.dtype)
+        .to(device)
+    )
+
+    model_sampling = sd3_utils.ModelSamplingDiscreteFlow(shift=3.0)  # 3.0 is for SD3
+
+    sigmas = get_all_sigmas(model_sampling, steps).to(device)
+
+    noise_scaled = model_sampling.noise_scaling(sigmas[0], noise, latent, max_denoise(model_sampling, sigmas))
+
+    c_crossattn = torch.cat([cond[0], neg_cond[0]]).to(device).to(dtype)
+    y = torch.cat([cond[1], neg_cond[1]]).to(device).to(dtype)
+
+    x = noise_scaled.to(device).to(dtype)
+    # print(x.shape)
+
+    # with torch.no_grad():
+    for i in tqdm(range(len(sigmas) - 1)):
+        sigma_hat = sigmas[i]
+
+        timestep = model_sampling.timestep(sigma_hat).float()
+        timestep = torch.FloatTensor([timestep, timestep]).to(device)
+
+        x_c_nc = torch.cat([x, x], dim=0)
+        # print(x_c_nc.shape, timestep.shape, c_crossattn.shape, y.shape)
+
+        mmdit.prepare_block_swap_before_forward()
+        model_output = mmdit(x_c_nc, timestep, context=c_crossattn, y=y)
+        model_output = model_output.float()
+        batched = model_sampling.calculate_denoised(sigma_hat, model_output, x)
+
+        pos_out, neg_out = batched.chunk(2)
+        denoised = neg_out + (pos_out - neg_out) * guidance_scale
+        # print(denoised.shape)
+
+        # d = to_d(x, sigma_hat, denoised)
+        dims_to_append = x.ndim - sigma_hat.ndim
+        sigma_hat_dims = sigma_hat[(...,) + (None,) * dims_to_append]
+        # print(dims_to_append, x.shape, sigma_hat.shape, denoised.shape, sigma_hat_dims.shape)
+        """Converts a denoiser output to a Karras ODE derivative."""
+        d = (x - denoised) / sigma_hat_dims
+
+        dt = sigmas[i + 1] - sigma_hat
+
+        # Euler method
+        x = x + d * dt
+        x = x.to(dtype)
+
+    mmdit.prepare_block_swap_before_forward()
+    return x
+
+
+def sample_images(
+    accelerator: Accelerator,
+    args: argparse.Namespace,
+    epoch,
+    steps,
+    mmdit,
+    vae,
+    text_encoders,
+    sample_prompts_te_outputs,
+    prompt_replacement=None,
+):
+    if steps == 0:
+        if not args.sample_at_first:
+            return
+    else:
+        if args.sample_every_n_steps is None and args.sample_every_n_epochs is None:
+            return
+        if args.sample_every_n_epochs is not None:
+            # sample_every_n_steps は無視する
+            if epoch is None or epoch % args.sample_every_n_epochs != 0:
+                return
+        else:
+            if steps % args.sample_every_n_steps != 0 or epoch is not None:  # steps is not divisible or end of epoch
+                return
+
+    logger.info("")
+    logger.info(f"generating sample images at step / サンプル画像生成 ステップ: {steps}")
+    if not os.path.isfile(args.sample_prompts) and sample_prompts_te_outputs is None:
+        logger.error(f"No prompt file / プロンプトファイルがありません: {args.sample_prompts}")
+        return
+
+    distributed_state = PartialState()  # for multi gpu distributed inference. this is a singleton, so it's safe to use it here
+
+    # unwrap unet and text_encoder(s)
+    mmdit = accelerator.unwrap_model(mmdit)
+    text_encoders = None if text_encoders is None else [accelerator.unwrap_model(te) for te in text_encoders]
+    # print([(te.parameters().__next__().device if te is not None else None) for te in text_encoders])
+
+    prompts = train_util.load_prompts(args.sample_prompts)
+
+    save_dir = args.output_dir + "/sample"
+    os.makedirs(save_dir, exist_ok=True)
+
+    # save random state to restore later
+    rng_state = torch.get_rng_state()
+    cuda_rng_state = None
+    try:
+        cuda_rng_state = torch.cuda.get_rng_state() if torch.cuda.is_available() else None
+    except Exception:
+        pass
+
+    if distributed_state.num_processes <= 1:
+        # If only one device is available, just use the original prompt list. We don't need to care about the distribution of prompts.
+        with torch.no_grad(), accelerator.autocast():
+            for prompt_dict in prompts:
+                sample_image_inference(
+                    accelerator,
+                    args,
+                    mmdit,
+                    text_encoders,
+                    vae,
+                    save_dir,
+                    prompt_dict,
+                    epoch,
+                    steps,
+                    sample_prompts_te_outputs,
+                    prompt_replacement,
+                )
+    else:
+        # Creating list with N elements, where each element is a list of prompt_dicts, and N is the number of processes available (number of devices available)
+        # prompt_dicts are assigned to lists based on order of processes, to attempt to time the image creation time to match enum order. Probably only works when steps and sampler are identical.
+        per_process_prompts = []  # list of lists
+        for i in range(distributed_state.num_processes):
+            per_process_prompts.append(prompts[i :: distributed_state.num_processes])
+
+        with torch.no_grad():
+            with distributed_state.split_between_processes(per_process_prompts) as prompt_dict_lists:
+                for prompt_dict in prompt_dict_lists[0]:
+                    sample_image_inference(
+                        accelerator,
+                        args,
+                        mmdit,
+                        text_encoders,
+                        vae,
+                        save_dir,
+                        prompt_dict,
+                        epoch,
+                        steps,
+                        sample_prompts_te_outputs,
+                        prompt_replacement,
+                    )
+
+    torch.set_rng_state(rng_state)
+    if cuda_rng_state is not None:
+        torch.cuda.set_rng_state(cuda_rng_state)
+
+    clean_memory_on_device(accelerator.device)
+
+
+def sample_image_inference(
+    accelerator: Accelerator,
+    args: argparse.Namespace,
+    mmdit: sd3_models.MMDiT,
+    text_encoders: List[Union[CLIPTextModelWithProjection, T5EncoderModel]],
+    vae: sd3_models.SDVAE,
+    save_dir,
+    prompt_dict,
+    epoch,
+    steps,
+    sample_prompts_te_outputs,
+    prompt_replacement,
+):
+    assert isinstance(prompt_dict, dict)
+    negative_prompt = prompt_dict.get("negative_prompt")
+    sample_steps = prompt_dict.get("sample_steps", 30)
+    width = prompt_dict.get("width", 512)
+    height = prompt_dict.get("height", 512)
+    scale = prompt_dict.get("scale", 7.5)
+    seed = prompt_dict.get("seed")
+    # controlnet_image = prompt_dict.get("controlnet_image")
+    prompt: str = prompt_dict.get("prompt", "")
+    # sampler_name: str = prompt_dict.get("sample_sampler", args.sample_sampler)
+
+    if prompt_replacement is not None:
+        prompt = prompt.replace(prompt_replacement[0], prompt_replacement[1])
+        if negative_prompt is not None:
+            negative_prompt = negative_prompt.replace(prompt_replacement[0], prompt_replacement[1])
+
+    if seed is not None:
+        torch.manual_seed(seed)
+        torch.cuda.manual_seed(seed)
+    else:
+        # True random sample image generation
+        torch.seed()
+        torch.cuda.seed()
+
+    if negative_prompt is None:
+        negative_prompt = ""
+
+    height = max(64, height - height % 8)  # round to divisible by 8
+    width = max(64, width - width % 8)  # round to divisible by 8
+    logger.info(f"prompt: {prompt}")
+    logger.info(f"negative_prompt: {negative_prompt}")
+    logger.info(f"height: {height}")
+    logger.info(f"width: {width}")
+    logger.info(f"sample_steps: {sample_steps}")
+    logger.info(f"scale: {scale}")
+    # logger.info(f"sample_sampler: {sampler_name}")
+    if seed is not None:
+        logger.info(f"seed: {seed}")
+
+    # encode prompts
+    tokenize_strategy = strategy_base.TokenizeStrategy.get_strategy()
+    encoding_strategy = strategy_base.TextEncodingStrategy.get_strategy()
+
+    def encode_prompt(prpt):
+        text_encoder_conds = []
+        if sample_prompts_te_outputs and prpt in sample_prompts_te_outputs:
+            text_encoder_conds = sample_prompts_te_outputs[prpt]
+            print(f"Using cached text encoder outputs for prompt: {prpt}")
+        if text_encoders is not None:
+            print(f"Encoding prompt: {prpt}")
+            tokens_and_masks = tokenize_strategy.tokenize(prpt)
+            # strategy has apply_t5_attn_mask option
+            encoded_text_encoder_conds = encoding_strategy.encode_tokens(tokenize_strategy, text_encoders, tokens_and_masks)
+
+            # if text_encoder_conds is not cached, use encoded_text_encoder_conds
+            if len(text_encoder_conds) == 0:
+                text_encoder_conds = encoded_text_encoder_conds
+            else:
+                # if encoded_text_encoder_conds is not None, update cached text_encoder_conds
+                for i in range(len(encoded_text_encoder_conds)):
+                    if encoded_text_encoder_conds[i] is not None:
+                        text_encoder_conds[i] = encoded_text_encoder_conds[i]
+        return text_encoder_conds
+
+    lg_out, t5_out, pooled, l_attn_mask, g_attn_mask, t5_attn_mask = encode_prompt(prompt)
+    cond = encoding_strategy.concat_encodings(lg_out, t5_out, pooled)
+
+    # encode negative prompts
+    lg_out, t5_out, pooled, l_attn_mask, g_attn_mask, t5_attn_mask = encode_prompt(negative_prompt)
+    neg_cond = encoding_strategy.concat_encodings(lg_out, t5_out, pooled)
+
+    # sample image
+    clean_memory_on_device(accelerator.device)
+    with accelerator.autocast(), torch.no_grad():
+        # mmdit may be fp8, so we need weight_dtype here. vae is always in that dtype.
+        latents = do_sample(height, width, seed, cond, neg_cond, mmdit, sample_steps, scale, vae.dtype, accelerator.device)
+
+    # latent to image
+    clean_memory_on_device(accelerator.device)
+    org_vae_device = vae.device  # will be on cpu
+    vae.to(accelerator.device)
+    latents = vae.process_out(latents.to(vae.device, dtype=vae.dtype))
+    image = vae.decode(latents)
+    vae.to(org_vae_device)
+    clean_memory_on_device(accelerator.device)
+
+    image = image.float()
+    image = torch.clamp((image + 1.0) / 2.0, min=0.0, max=1.0)[0]
+    decoded_np = 255.0 * np.moveaxis(image.cpu().numpy(), 0, 2)
+    decoded_np = decoded_np.astype(np.uint8)
+
+    image = Image.fromarray(decoded_np)
+    # adding accelerator.wait_for_everyone() here should sync up and ensure that sample images are saved in the same order as the original prompt list
+    # but adding 'enum' to the filename should be enough
+
+    ts_str = time.strftime("%Y%m%d%H%M%S", time.localtime())
+    num_suffix = f"e{epoch:06d}" if epoch is not None else f"{steps:06d}"
+    seed_suffix = "" if seed is None else f"_{seed}"
+    i: int = prompt_dict["enum"]
+    img_filename = f"{'' if args.output_name is None else args.output_name + '_'}{num_suffix}_{i:02d}_{ts_str}{seed_suffix}.png"
+    image.save(os.path.join(save_dir, img_filename))
+
+    # send images to wandb if enabled
+    if "wandb" in [tracker.name for tracker in accelerator.trackers]:
+        wandb_tracker = accelerator.get_tracker("wandb")
+
+        import wandb
+
+        # not to commit images to avoid inconsistency between training and logging steps
+        wandb_tracker.log({f"sample_{i}": wandb.Image(image, caption=prompt)}, commit=False)  # positive prompt as a caption
+
+
+# region Diffusers
+
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.schedulers.scheduling_utils import SchedulerMixin
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.utils import BaseOutput
+
+
+@dataclass
+class FlowMatchEulerDiscreteSchedulerOutput(BaseOutput):
+    """
+    Output class for the scheduler's `step` function output.
+
+    Args:
+        prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+            Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+            denoising loop.
+    """
+
+    prev_sample: torch.FloatTensor
+
+
+class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
+    """
+    Euler scheduler.
+
+    This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+    methods the library implements for all schedulers such as loading and saving.
+
+    Args:
+        num_train_timesteps (`int`, defaults to 1000):
+            The number of diffusion steps to train the model.
+        timestep_spacing (`str`, defaults to `"linspace"`):
+            The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+            Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+        shift (`float`, defaults to 1.0):
+            The shift value for the timestep schedule.
+    """
+
+    _compatibles = []
+    order = 1
+
+    @register_to_config
+    def __init__(
+        self,
+        num_train_timesteps: int = 1000,
+        shift: float = 1.0,
+    ):
+        timesteps = np.linspace(1, num_train_timesteps, num_train_timesteps, dtype=np.float32)[::-1].copy()
+        timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32)
+
+        sigmas = timesteps / num_train_timesteps
+        sigmas = shift * sigmas / (1 + (shift - 1) * sigmas)
+
+        self.timesteps = sigmas * num_train_timesteps
+
+        self._step_index = None
+        self._begin_index = None
+
+        self.sigmas = sigmas.to("cpu")  # to avoid too much CPU/GPU communication
+        self.sigma_min = self.sigmas[-1].item()
+        self.sigma_max = self.sigmas[0].item()
+
+    @property
+    def step_index(self):
+        """
+        The index counter for current timestep. It will increase 1 after each scheduler step.
+        """
+        return self._step_index
+
+    @property
+    def begin_index(self):
+        """
+        The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+        """
+        return self._begin_index
+
+    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+    def set_begin_index(self, begin_index: int = 0):
+        """
+        Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+        Args:
+            begin_index (`int`):
+                The begin index for the scheduler.
+        """
+        self._begin_index = begin_index
+
+    def scale_noise(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[float, torch.FloatTensor],
+        noise: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        """
+        Forward process in flow-matching
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The input sample.
+            timestep (`int`, *optional*):
+                The current timestep in the diffusion chain.
+
+        Returns:
+            `torch.FloatTensor`:
+                A scaled input sample.
+        """
+        if self.step_index is None:
+            self._init_step_index(timestep)
+
+        sigma = self.sigmas[self.step_index]
+        sample = sigma * noise + (1.0 - sigma) * sample
+
+        return sample
+
+    def _sigma_to_t(self, sigma):
+        return sigma * self.config.num_train_timesteps
+
+    def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+        """
+        Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+        Args:
+            num_inference_steps (`int`):
+                The number of diffusion steps used when generating samples with a pre-trained model.
+            device (`str` or `torch.device`, *optional*):
+                The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        """
+        self.num_inference_steps = num_inference_steps
+
+        timesteps = np.linspace(self._sigma_to_t(self.sigma_max), self._sigma_to_t(self.sigma_min), num_inference_steps)
+
+        sigmas = timesteps / self.config.num_train_timesteps
+        sigmas = self.config.shift * sigmas / (1 + (self.config.shift - 1) * sigmas)
+        sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device)
+
+        timesteps = sigmas * self.config.num_train_timesteps
+        self.timesteps = timesteps.to(device=device)
+        self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])
+
+        self._step_index = None
+        self._begin_index = None
+
+    def index_for_timestep(self, timestep, schedule_timesteps=None):
+        if schedule_timesteps is None:
+            schedule_timesteps = self.timesteps
+
+        indices = (schedule_timesteps == timestep).nonzero()
+
+        # The sigma index that is taken for the **very** first `step`
+        # is always the second index (or the last index if there is only 1)
+        # This way we can ensure we don't accidentally skip a sigma in
+        # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+        pos = 1 if len(indices) > 1 else 0
+
+        return indices[pos].item()
+
+    def _init_step_index(self, timestep):
+        if self.begin_index is None:
+            if isinstance(timestep, torch.Tensor):
+                timestep = timestep.to(self.timesteps.device)
+            self._step_index = self.index_for_timestep(timestep)
+        else:
+            self._step_index = self._begin_index
+
+    def step(
+        self,
+        model_output: torch.FloatTensor,
+        timestep: Union[float, torch.FloatTensor],
+        sample: torch.FloatTensor,
+        s_churn: float = 0.0,
+        s_tmin: float = 0.0,
+        s_tmax: float = float("inf"),
+        s_noise: float = 1.0,
+        generator: Optional[torch.Generator] = None,
+        return_dict: bool = True,
+    ) -> Union[FlowMatchEulerDiscreteSchedulerOutput, Tuple]:
+        """
+        Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+        process from the learned model outputs (most often the predicted noise).
+
+        Args:
+            model_output (`torch.FloatTensor`):
+                The direct output from learned diffusion model.
+            timestep (`float`):
+                The current discrete timestep in the diffusion chain.
+            sample (`torch.FloatTensor`):
+                A current instance of a sample created by the diffusion process.
+            s_churn (`float`):
+            s_tmin  (`float`):
+            s_tmax  (`float`):
+            s_noise (`float`, defaults to 1.0):
+                Scaling factor for noise added to the sample.
+            generator (`torch.Generator`, *optional*):
+                A random number generator.
+            return_dict (`bool`):
+                Whether or not to return a [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or
+                tuple.
+
+        Returns:
+            [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or `tuple`:
+                If return_dict is `True`, [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] is
+                returned, otherwise a tuple is returned where the first element is the sample tensor.
+        """
+
+        if isinstance(timestep, int) or isinstance(timestep, torch.IntTensor) or isinstance(timestep, torch.LongTensor):
+            raise ValueError(
+                (
+                    "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to"
+                    " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass"
+                    " one of the `scheduler.timesteps` as a timestep."
+                ),
+            )
+
+        if self.step_index is None:
+            self._init_step_index(timestep)
+
+        # Upcast to avoid precision issues when computing prev_sample
+        sample = sample.to(torch.float32)
+
+        sigma = self.sigmas[self.step_index]
+
+        gamma = min(s_churn / (len(self.sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigma <= s_tmax else 0.0
+
+        noise = randn_tensor(model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator)
+
+        eps = noise * s_noise
+        sigma_hat = sigma * (gamma + 1)
+
+        if gamma > 0:
+            sample = sample + eps * (sigma_hat**2 - sigma**2) ** 0.5
+
+        # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
+        # NOTE: "original_sample" should not be an expected prediction_type but is left in for
+        # backwards compatibility
+
+        # if self.config.prediction_type == "vector_field":
+
+        denoised = sample - model_output * sigma
+        # 2. Convert to an ODE derivative
+        derivative = (sample - denoised) / sigma_hat
+
+        dt = self.sigmas[self.step_index + 1] - sigma_hat
+
+        prev_sample = sample + derivative * dt
+        # Cast sample back to model compatible dtype
+        prev_sample = prev_sample.to(model_output.dtype)
+
+        # upon completion increase step index by one
+        self._step_index += 1
+
+        if not return_dict:
+            return (prev_sample,)
+
+        return FlowMatchEulerDiscreteSchedulerOutput(prev_sample=prev_sample)
+
+    def __len__(self):
+        return self.config.num_train_timesteps
+
+
+def get_sigmas(noise_scheduler, timesteps, device, n_dim=4, dtype=torch.float32):
+    sigmas = noise_scheduler.sigmas.to(device=device, dtype=dtype)
+    schedule_timesteps = noise_scheduler.timesteps.to(device)
+    timesteps = timesteps.to(device)
+    step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
+
+    sigma = sigmas[step_indices].flatten()
+    while len(sigma.shape) < n_dim:
+        sigma = sigma.unsqueeze(-1)
+    return sigma
+
+
+def compute_density_for_timestep_sampling(
+    weighting_scheme: str, batch_size: int, logit_mean: float = None, logit_std: float = None, mode_scale: float = None
+):
+    """Compute the density for sampling the timesteps when doing SD3 training.
+
+    Courtesy: This was contributed by Rafie Walker in https://github.com/huggingface/diffusers/pull/8528.
+
+    SD3 paper reference: https://arxiv.org/abs/2403.03206v1.
+    """
+    if weighting_scheme == "logit_normal":
+        # See 3.1 in the SD3 paper ($rf/lognorm(0.00,1.00)$).
+        u = torch.normal(mean=logit_mean, std=logit_std, size=(batch_size,), device="cpu")
+        u = torch.nn.functional.sigmoid(u)
+    elif weighting_scheme == "mode":
+        u = torch.rand(size=(batch_size,), device="cpu")
+        u = 1 - u - mode_scale * (torch.cos(math.pi * u / 2) ** 2 - 1 + u)
+    else:
+        u = torch.rand(size=(batch_size,), device="cpu")
+    return u
+
+
+def compute_loss_weighting_for_sd3(weighting_scheme: str, sigmas=None):
+    """Computes loss weighting scheme for SD3 training.
+
+    Courtesy: This was contributed by Rafie Walker in https://github.com/huggingface/diffusers/pull/8528.
+
+    SD3 paper reference: https://arxiv.org/abs/2403.03206v1.
+    """
+    if weighting_scheme == "sigma_sqrt":
+        weighting = (sigmas**-2.0).float()
+    elif weighting_scheme == "cosmap":
+        bot = 1 - 2 * sigmas + 2 * sigmas**2
+        weighting = 2 / (math.pi * bot)
+    else:
+        weighting = torch.ones_like(sigmas)
+    return weighting
+
+
+# endregion
+
+
+def get_noisy_model_input_and_timesteps(args, latents, noise, device, dtype) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    bsz = latents.shape[0]
+
+    # Sample a random timestep for each image
+    # for weighting schemes where we sample timesteps non-uniformly
+    u = compute_density_for_timestep_sampling(
+        weighting_scheme=args.weighting_scheme,
+        batch_size=bsz,
+        logit_mean=args.logit_mean,
+        logit_std=args.logit_std,
+        mode_scale=args.mode_scale,
+    )
+    t_min = args.min_timestep if args.min_timestep is not None else 0
+    t_max = args.max_timestep if args.max_timestep is not None else 1000
+    shift = args.training_shift
+
+    # weighting shift, value >1 will shift distribution to noisy side (focus more on overall structure), value <1 will shift towards less-noisy side (focus more on details)
+    u = (u * shift) / (1 + (shift - 1) * u)
+
+    indices = (u * (t_max - t_min) + t_min).long()
+    timesteps = indices.to(device=device, dtype=dtype)
+
+    # sigmas according to flowmatching
+    sigmas = timesteps / 1000
+    sigmas = sigmas.view(-1, 1, 1, 1)
+    noisy_model_input = sigmas * noise + (1.0 - sigmas) * latents
+
+    return noisy_model_input, timesteps, sigmas

library/sd3_utils.py

@@ -0,0 +1,302 @@
+from dataclasses import dataclass
+import math
+import re
+from typing import Dict, List, Optional, Union
+import torch
+import safetensors
+from safetensors.torch import load_file
+from accelerate import init_empty_weights
+from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPConfig, CLIPTextConfig
+
+from .utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+from library import sd3_models
+
+# TODO move some of functions to model_util.py
+from library import sdxl_model_util
+
+# region models
+
+# TODO remove dependency on flux_utils
+from library.safetensors_utils import load_safetensors
+from library.flux_utils import load_t5xxl as flux_utils_load_t5xxl
+
+
+def analyze_state_dict_state(state_dict: Dict, prefix: str = ""):
+    logger.info(f"Analyzing state dict state...")
+
+    # analyze configs
+    patch_size = state_dict[f"{prefix}x_embedder.proj.weight"].shape[2]
+    depth = state_dict[f"{prefix}x_embedder.proj.weight"].shape[0] // 64
+    num_patches = state_dict[f"{prefix}pos_embed"].shape[1]
+    pos_embed_max_size = round(math.sqrt(num_patches))
+    adm_in_channels = state_dict[f"{prefix}y_embedder.mlp.0.weight"].shape[1]
+    context_shape = state_dict[f"{prefix}context_embedder.weight"].shape
+    qk_norm = "rms" if f"{prefix}joint_blocks.0.context_block.attn.ln_k.weight" in state_dict.keys() else None
+
+    #  x_block_self_attn_layers.append(int(key.split(".x_block.attn2.ln_k.weight")[0].split(".")[-1]))
+    x_block_self_attn_layers = []
+    re_attn = re.compile(r"\.(\d+)\.x_block\.attn2\.ln_k\.weight")
+    for key in list(state_dict.keys()):
+        m = re_attn.search(key)
+        if m:
+            x_block_self_attn_layers.append(int(m.group(1)))
+
+    context_embedder_in_features = context_shape[1]
+    context_embedder_out_features = context_shape[0]
+
+    # only supports 3-5-large, medium or 3-medium. This is added after `stable-diffusion-3-`.
+    if qk_norm is not None:
+        if len(x_block_self_attn_layers) == 0:
+            model_type = "5-large"
+        else:
+            model_type = "5-medium"
+    else:
+        model_type = "medium"
+
+    params = sd3_models.SD3Params(
+        patch_size=patch_size,
+        depth=depth,
+        num_patches=num_patches,
+        pos_embed_max_size=pos_embed_max_size,
+        adm_in_channels=adm_in_channels,
+        qk_norm=qk_norm,
+        x_block_self_attn_layers=x_block_self_attn_layers,
+        context_embedder_in_features=context_embedder_in_features,
+        context_embedder_out_features=context_embedder_out_features,
+        model_type=model_type,
+    )
+    logger.info(f"Analyzed state dict state: {params}")
+    return params
+
+
+def load_mmdit(
+    state_dict: Dict, dtype: Optional[Union[str, torch.dtype]], device: Union[str, torch.device], attn_mode: str = "torch"
+) -> sd3_models.MMDiT:
+    mmdit_sd = {}
+
+    mmdit_prefix = "model.diffusion_model."
+    for k in list(state_dict.keys()):
+        if k.startswith(mmdit_prefix):
+            mmdit_sd[k[len(mmdit_prefix) :]] = state_dict.pop(k)
+
+    # load MMDiT
+    logger.info("Building MMDit")
+    params = analyze_state_dict_state(mmdit_sd)
+    with init_empty_weights():
+        mmdit = sd3_models.create_sd3_mmdit(params, attn_mode)
+
+    logger.info("Loading state dict...")
+    info = mmdit.load_state_dict(mmdit_sd, strict=False, assign=True)
+    logger.info(f"Loaded MMDiT: {info}")
+    return mmdit
+
+
+def load_clip_l(
+    clip_l_path: Optional[str],
+    dtype: Optional[Union[str, torch.dtype]],
+    device: Union[str, torch.device],
+    disable_mmap: bool = False,
+    state_dict: Optional[Dict] = None,
+):
+    clip_l_sd = None
+    if clip_l_path is None:
+        if "text_encoders.clip_l.transformer.text_model.embeddings.position_embedding.weight" in state_dict:
+            # found clip_l: remove prefix "text_encoders.clip_l."
+            logger.info("clip_l is included in the checkpoint")
+            clip_l_sd = {}
+            prefix = "text_encoders.clip_l."
+            for k in list(state_dict.keys()):
+                if k.startswith(prefix):
+                    clip_l_sd[k[len(prefix) :]] = state_dict.pop(k)
+        elif clip_l_path is None:
+            logger.info("clip_l is not included in the checkpoint and clip_l_path is not provided")
+            return None
+
+    # load clip_l
+    logger.info("Building CLIP-L")
+    config = CLIPTextConfig(
+        vocab_size=49408,
+        hidden_size=768,
+        intermediate_size=3072,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        max_position_embeddings=77,
+        hidden_act="quick_gelu",
+        layer_norm_eps=1e-05,
+        dropout=0.0,
+        attention_dropout=0.0,
+        initializer_range=0.02,
+        initializer_factor=1.0,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        model_type="clip_text_model",
+        projection_dim=768,
+        # torch_dtype="float32",
+        # transformers_version="4.25.0.dev0",
+    )
+    with init_empty_weights():
+        clip = CLIPTextModelWithProjection(config)
+
+    if clip_l_sd is None:
+        logger.info(f"Loading state dict from {clip_l_path}")
+        clip_l_sd = load_safetensors(clip_l_path, device=str(device), disable_mmap=disable_mmap, dtype=dtype)
+
+    if "text_projection.weight" not in clip_l_sd:
+        logger.info("Adding text_projection.weight to clip_l_sd")
+        clip_l_sd["text_projection.weight"] = torch.eye(768, dtype=dtype, device=device)
+
+    info = clip.load_state_dict(clip_l_sd, strict=False, assign=True)
+    logger.info(f"Loaded CLIP-L: {info}")
+    return clip
+
+
+def load_clip_g(
+    clip_g_path: Optional[str],
+    dtype: Optional[Union[str, torch.dtype]],
+    device: Union[str, torch.device],
+    disable_mmap: bool = False,
+    state_dict: Optional[Dict] = None,
+):
+    clip_g_sd = None
+    if state_dict is not None:
+        if "text_encoders.clip_g.transformer.text_model.embeddings.position_embedding.weight" in state_dict:
+            # found clip_g: remove prefix "text_encoders.clip_g."
+            logger.info("clip_g is included in the checkpoint")
+            clip_g_sd = {}
+            prefix = "text_encoders.clip_g."
+            for k in list(state_dict.keys()):
+                if k.startswith(prefix):
+                    clip_g_sd[k[len(prefix) :]] = state_dict.pop(k)
+        elif clip_g_path is None:
+            logger.info("clip_g is not included in the checkpoint and clip_g_path is not provided")
+            return None
+
+    # load clip_g
+    logger.info("Building CLIP-G")
+    config = CLIPTextConfig(
+        vocab_size=49408,
+        hidden_size=1280,
+        intermediate_size=5120,
+        num_hidden_layers=32,
+        num_attention_heads=20,
+        max_position_embeddings=77,
+        hidden_act="gelu",
+        layer_norm_eps=1e-05,
+        dropout=0.0,
+        attention_dropout=0.0,
+        initializer_range=0.02,
+        initializer_factor=1.0,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        model_type="clip_text_model",
+        projection_dim=1280,
+        # torch_dtype="float32",
+        # transformers_version="4.25.0.dev0",
+    )
+    with init_empty_weights():
+        clip = CLIPTextModelWithProjection(config)
+
+    if clip_g_sd is None:
+        logger.info(f"Loading state dict from {clip_g_path}")
+        clip_g_sd = load_safetensors(clip_g_path, device=str(device), disable_mmap=disable_mmap, dtype=dtype)
+    info = clip.load_state_dict(clip_g_sd, strict=False, assign=True)
+    logger.info(f"Loaded CLIP-G: {info}")
+    return clip
+
+
+def load_t5xxl(
+    t5xxl_path: Optional[str],
+    dtype: Optional[Union[str, torch.dtype]],
+    device: Union[str, torch.device],
+    disable_mmap: bool = False,
+    state_dict: Optional[Dict] = None,
+):
+    t5xxl_sd = None
+    if state_dict is not None:
+        if "text_encoders.t5xxl.transformer.encoder.block.0.layer.0.SelfAttention.k.weight" in state_dict:
+            # found t5xxl: remove prefix "text_encoders.t5xxl."
+            logger.info("t5xxl is included in the checkpoint")
+            t5xxl_sd = {}
+            prefix = "text_encoders.t5xxl."
+            for k in list(state_dict.keys()):
+                if k.startswith(prefix):
+                    t5xxl_sd[k[len(prefix) :]] = state_dict.pop(k)
+        elif t5xxl_path is None:
+            logger.info("t5xxl is not included in the checkpoint and t5xxl_path is not provided")
+            return None
+
+    return flux_utils_load_t5xxl(t5xxl_path, dtype, device, disable_mmap, state_dict=t5xxl_sd)
+
+
+def load_vae(
+    vae_path: Optional[str],
+    vae_dtype: Optional[Union[str, torch.dtype]],
+    device: Optional[Union[str, torch.device]],
+    disable_mmap: bool = False,
+    state_dict: Optional[Dict] = None,
+):
+    vae_sd = {}
+    if vae_path:
+        logger.info(f"Loading VAE from {vae_path}...")
+        vae_sd = load_safetensors(vae_path, device, disable_mmap, dtype=vae_dtype)
+    else:
+        # remove prefix "first_stage_model."
+        vae_sd = {}
+        vae_prefix = "first_stage_model."
+        for k in list(state_dict.keys()):
+            if k.startswith(vae_prefix):
+                vae_sd[k[len(vae_prefix) :]] = state_dict.pop(k)
+
+    logger.info("Building VAE")
+    vae = sd3_models.SDVAE(vae_dtype, device)
+    logger.info("Loading state dict...")
+    info = vae.load_state_dict(vae_sd)
+    logger.info(f"Loaded VAE: {info}")
+    vae.to(device=device, dtype=vae_dtype)  # make sure it's in the right device and dtype
+    return vae
+
+
+# endregion
+
+
+class ModelSamplingDiscreteFlow:
+    """Helper for sampler scheduling (ie timestep/sigma calculations) for Discrete Flow models"""
+
+    def __init__(self, shift=1.0):
+        self.shift = shift
+        timesteps = 1000
+        self.sigmas = self.sigma(torch.arange(1, timesteps + 1, 1))
+
+    @property
+    def sigma_min(self):
+        return self.sigmas[0]
+
+    @property
+    def sigma_max(self):
+        return self.sigmas[-1]
+
+    def timestep(self, sigma):
+        return sigma * 1000
+
+    def sigma(self, timestep: torch.Tensor):
+        timestep = timestep / 1000.0
+        if self.shift == 1.0:
+            return timestep
+        return self.shift * timestep / (1 + (self.shift - 1) * timestep)
+
+    def calculate_denoised(self, sigma, model_output, model_input):
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (model_output.ndim - 1))
+        return model_input - model_output * sigma
+
+    def noise_scaling(self, sigma, noise, latent_image, max_denoise=False):
+        # assert max_denoise is False, "max_denoise not implemented"
+        # max_denoise is always True, I'm not sure why it's there
+        return sigma * noise + (1.0 - sigma) * latent_image

library/sdxl_lpw_stable_diffusion.py

@@ -13,12 +13,20 @@ from tqdm import tqdm
 from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
 
 from diffusers import SchedulerMixin, StableDiffusionPipeline
-from diffusers.models import AutoencoderKL, UNet2DConditionModel
-from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker
+from diffusers.models import AutoencoderKL
+from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
 from diffusers.utils import logging
 from PIL import Image
 
-from library import sdxl_model_util, sdxl_train_util, train_util
+from library import (
+    sdxl_model_util,
+    sdxl_train_util,
+    strategy_base,
+    strategy_sdxl,
+    train_util,
+    sdxl_original_unet,
+    sdxl_original_control_net,
+)
 
 
 try:
@@ -537,7 +545,7 @@ class SdxlStableDiffusionLongPromptWeightingPipeline:
         vae: AutoencoderKL,
         text_encoder: List[CLIPTextModel],
         tokenizer: List[CLIPTokenizer],
-        unet: UNet2DConditionModel,
+        unet: Union[sdxl_original_unet.SdxlUNet2DConditionModel, sdxl_original_control_net.SdxlControlledUNet],
         scheduler: SchedulerMixin,
         # clip_skip: int,
         safety_checker: StableDiffusionSafetyChecker,
@@ -594,74 +602,6 @@ class SdxlStableDiffusionLongPromptWeightingPipeline:
                 return torch.device(module._hf_hook.execution_device)
         return self.device
 
-    def _encode_prompt(
-        self,
-        prompt,
-        device,
-        num_images_per_prompt,
-        do_classifier_free_guidance,
-        negative_prompt,
-        max_embeddings_multiples,
-        is_sdxl_text_encoder2,
-    ):
-        r"""
-        Encodes the prompt into text encoder hidden states.
-
-        Args:
-            prompt (`str` or `list(int)`):
-                prompt to be encoded
-            device: (`torch.device`):
-                torch device
-            num_images_per_prompt (`int`):
-                number of images that should be generated per prompt
-            do_classifier_free_guidance (`bool`):
-                whether to use classifier free guidance or not
-            negative_prompt (`str` or `List[str]`):
-                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
-                if `guidance_scale` is less than `1`).
-            max_embeddings_multiples (`int`, *optional*, defaults to `3`):
-                The max multiple length of prompt embeddings compared to the max output length of text encoder.
-        """
-        batch_size = len(prompt) if isinstance(prompt, list) else 1
-
-        if negative_prompt is None:
-            negative_prompt = [""] * batch_size
-        elif isinstance(negative_prompt, str):
-            negative_prompt = [negative_prompt] * batch_size
-        if batch_size != len(negative_prompt):
-            raise ValueError(
-                f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
-                f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
-                " the batch size of `prompt`."
-            )
-
-        text_embeddings, text_pool, uncond_embeddings, uncond_pool = get_weighted_text_embeddings(
-            pipe=self,
-            prompt=prompt,
-            uncond_prompt=negative_prompt if do_classifier_free_guidance else None,
-            max_embeddings_multiples=max_embeddings_multiples,
-            clip_skip=self.clip_skip,
-            is_sdxl_text_encoder2=is_sdxl_text_encoder2,
-        )
-        bs_embed, seq_len, _ = text_embeddings.shape
-        text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)  # ??
-        text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
-        if text_pool is not None:
-            text_pool = text_pool.repeat(1, num_images_per_prompt)
-            text_pool = text_pool.view(bs_embed * num_images_per_prompt, -1)
-
-        if do_classifier_free_guidance:
-            bs_embed, seq_len, _ = uncond_embeddings.shape
-            uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
-            uncond_embeddings = uncond_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
-            if uncond_pool is not None:
-                uncond_pool = uncond_pool.repeat(1, num_images_per_prompt)
-                uncond_pool = uncond_pool.view(bs_embed * num_images_per_prompt, -1)
-
-            return text_embeddings, text_pool, uncond_embeddings, uncond_pool
-
-        return text_embeddings, text_pool, None, None
-
     def check_inputs(self, prompt, height, width, strength, callback_steps):
         if not isinstance(prompt, str) and not isinstance(prompt, list):
             raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
@@ -792,7 +732,7 @@ class SdxlStableDiffusionLongPromptWeightingPipeline:
         max_embeddings_multiples: Optional[int] = 3,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
-        controlnet=None,
+        controlnet: sdxl_original_control_net.SdxlControlNet = None,
         controlnet_image=None,
         callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
         is_cancelled_callback: Optional[Callable[[], bool]] = None,
@@ -896,34 +836,30 @@ class SdxlStableDiffusionLongPromptWeightingPipeline:
         do_classifier_free_guidance = guidance_scale > 1.0
 
         # 3. Encode input prompt
-        # 実装を簡単にするためにtokenzer/text encoderを切り替えて二回呼び出す
-        # To simplify the implementation, switch the tokenzer/text encoder and call it twice
-        text_embeddings_list = []
-        text_pool = None
-        uncond_embeddings_list = []
-        uncond_pool = None
-        for i in range(len(self.tokenizers)):
-            self.tokenizer = self.tokenizers[i]
-            self.text_encoder = self.text_encoders[i]
+        tokenize_strategy: strategy_sdxl.SdxlTokenizeStrategy = strategy_base.TokenizeStrategy.get_strategy()
+        encoding_strategy: strategy_sdxl.SdxlTextEncodingStrategy = strategy_base.TextEncodingStrategy.get_strategy()
 
-            text_embeddings, tp1, uncond_embeddings, up1 = self._encode_prompt(
-                prompt,
-                device,
-                num_images_per_prompt,
-                do_classifier_free_guidance,
-                negative_prompt,
-                max_embeddings_multiples,
-                is_sdxl_text_encoder2=i == 1,
+        text_input_ids, text_weights = tokenize_strategy.tokenize_with_weights(prompt)
+        hidden_states_1, hidden_states_2, text_pool = encoding_strategy.encode_tokens_with_weights(
+            tokenize_strategy, self.text_encoders, text_input_ids, text_weights
+        )
+        text_embeddings = torch.cat([hidden_states_1, hidden_states_2], dim=-1)
+
+        if do_classifier_free_guidance:
+            input_ids, weights = tokenize_strategy.tokenize_with_weights(negative_prompt or "")
+            hidden_states_1, hidden_states_2, uncond_pool = encoding_strategy.encode_tokens_with_weights(
+                tokenize_strategy, self.text_encoders, input_ids, weights
             )
-            text_embeddings_list.append(text_embeddings)
-            uncond_embeddings_list.append(uncond_embeddings)
+            uncond_embeddings = torch.cat([hidden_states_1, hidden_states_2], dim=-1)
+        else:
+            uncond_embeddings = None
+            uncond_pool = None
 
-            if tp1 is not None:
-                text_pool = tp1
-            if up1 is not None:
-                uncond_pool = up1
-
-        dtype = self.unet.dtype
+        unet_dtype = self.unet.dtype
+        dtype = unet_dtype
+        if hasattr(dtype, "itemsize") and dtype.itemsize == 1:  # fp8
+            dtype = torch.float16
+            self.unet.to(dtype)
 
         # 4. Preprocess image and mask
         if isinstance(image, PIL.Image.Image):
@@ -966,23 +902,23 @@ class SdxlStableDiffusionLongPromptWeightingPipeline:
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
         # create size embs and concat embeddings for SDXL
-        orig_size = torch.tensor([height, width]).repeat(batch_size * num_images_per_prompt, 1).to(dtype)
+        orig_size = torch.tensor([height, width]).repeat(batch_size * num_images_per_prompt, 1).to(device, dtype)
         crop_size = torch.zeros_like(orig_size)
         target_size = orig_size
-        embs = sdxl_train_util.get_size_embeddings(orig_size, crop_size, target_size, device).to(dtype)
+        embs = sdxl_train_util.get_size_embeddings(orig_size, crop_size, target_size, device).to(device, dtype)
 
         # make conditionings
+        text_pool = text_pool.to(device, dtype)
         if do_classifier_free_guidance:
-            text_embeddings = torch.cat(text_embeddings_list, dim=2)
-            uncond_embeddings = torch.cat(uncond_embeddings_list, dim=2)
-            text_embedding = torch.cat([uncond_embeddings, text_embeddings]).to(dtype)
+            text_embedding = torch.cat([uncond_embeddings, text_embeddings]).to(device, dtype)
 
-            cond_vector = torch.cat([text_pool, embs], dim=1)
-            uncond_vector = torch.cat([uncond_pool, embs], dim=1)
-            vector_embedding = torch.cat([uncond_vector, cond_vector]).to(dtype)
+            uncond_pool = uncond_pool.to(device, dtype)
+            cond_vector = torch.cat([text_pool, embs], dim=1).to(dtype)
+            uncond_vector = torch.cat([uncond_pool, embs], dim=1).to(dtype)
+            vector_embedding = torch.cat([uncond_vector, cond_vector])
         else:
-            text_embedding = torch.cat(text_embeddings_list, dim=2).to(dtype)
-            vector_embedding = torch.cat([text_pool, embs], dim=1).to(dtype)
+            text_embedding = text_embeddings.to(device, dtype)
+            vector_embedding = torch.cat([text_pool, embs], dim=1)
 
         # 8. Denoising loop
         for i, t in enumerate(self.progress_bar(timesteps)):
@@ -990,22 +926,14 @@ class SdxlStableDiffusionLongPromptWeightingPipeline:
             latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
             latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
 
-            unet_additional_args = {}
-            if controlnet is not None:
-                down_block_res_samples, mid_block_res_sample = controlnet(
-                    latent_model_input,
-                    t,
-                    encoder_hidden_states=text_embeddings,
-                    controlnet_cond=controlnet_image,
-                    conditioning_scale=1.0,
-                    guess_mode=False,
-                    return_dict=False,
-                )
-                unet_additional_args["down_block_additional_residuals"] = down_block_res_samples
-                unet_additional_args["mid_block_additional_residual"] = mid_block_res_sample
+            # FIXME SD1 ControlNet is not working
 
             # predict the noise residual
-            noise_pred = self.unet(latent_model_input, t, text_embedding, vector_embedding)
+            if controlnet is not None:
+                input_resi_add, mid_add = controlnet(latent_model_input, t, text_embedding, vector_embedding, controlnet_image)
+                noise_pred = self.unet(latent_model_input, t, text_embedding, vector_embedding, input_resi_add, mid_add)
+            else:
+                noise_pred = self.unet(latent_model_input, t, text_embedding, vector_embedding)
             noise_pred = noise_pred.to(dtype)  # U-Net changes dtype in LoRA training
 
             # perform guidance
@@ -1028,6 +956,7 @@ class SdxlStableDiffusionLongPromptWeightingPipeline:
                 if is_cancelled_callback is not None and is_cancelled_callback():
                     return None
 
+        self.unet.to(unet_dtype)
         return latents
 
     def latents_to_image(self, latents):

library/sdxl_model_util.py

@@ -1,4 +1,5 @@
 import torch
+import safetensors
 from accelerate import init_empty_weights
 from accelerate.utils.modeling import set_module_tensor_to_device
 from safetensors.torch import load_file, save_file
@@ -7,7 +8,12 @@ from typing import List
 from diffusers import AutoencoderKL, EulerDiscreteScheduler, UNet2DConditionModel
 from library import model_util
 from library import sdxl_original_unet
+from library.utils import setup_logging
 
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
 
 VAE_SCALE_FACTOR = 0.13025
 MODEL_VERSION_SDXL_BASE_V1_0 = "sdxl_base_v1-0"
@@ -131,7 +137,7 @@ def convert_sdxl_text_encoder_2_checkpoint(checkpoint, max_length):
 
     # temporary workaround for text_projection.weight.weight for Playground-v2
     if "text_projection.weight.weight" in new_sd:
-        print(f"convert_sdxl_text_encoder_2_checkpoint: convert text_projection.weight.weight to text_projection.weight")
+        logger.info("convert_sdxl_text_encoder_2_checkpoint: convert text_projection.weight.weight to text_projection.weight")
         new_sd["text_projection.weight"] = new_sd["text_projection.weight.weight"]
         del new_sd["text_projection.weight.weight"]
 
@@ -160,17 +166,20 @@ def _load_state_dict_on_device(model, state_dict, device, dtype=None):
     raise RuntimeError("Error(s) in loading state_dict for {}:\n\t{}".format(model.__class__.__name__, "\n\t".join(error_msgs)))
 
 
-def load_models_from_sdxl_checkpoint(model_version, ckpt_path, map_location, dtype=None):
+def load_models_from_sdxl_checkpoint(model_version, ckpt_path, map_location, dtype=None, disable_mmap=False):
     # model_version is reserved for future use
     # dtype is used for full_fp16/bf16 integration. Text Encoder will remain fp32, because it runs on CPU when caching
 
     # Load the state dict
     if model_util.is_safetensors(ckpt_path):
         checkpoint = None
-        try:
-            state_dict = load_file(ckpt_path, device=map_location)
-        except:
-            state_dict = load_file(ckpt_path)  # prevent device invalid Error
+        if disable_mmap:
+            state_dict = safetensors.torch.load(open(ckpt_path, "rb").read())
+        else:
+            try:
+                state_dict = load_file(ckpt_path, device=map_location)
+            except:
+                state_dict = load_file(ckpt_path)  # prevent device invalid Error
         epoch = None
         global_step = None
     else:
@@ -186,20 +195,20 @@ def load_models_from_sdxl_checkpoint(model_version, ckpt_path, map_location, dty
         checkpoint = None
 
     # U-Net
-    print("building U-Net")
+    logger.info("building U-Net")
     with init_empty_weights():
         unet = sdxl_original_unet.SdxlUNet2DConditionModel()
 
-    print("loading U-Net from checkpoint")
+    logger.info("loading U-Net from checkpoint")
     unet_sd = {}
     for k in list(state_dict.keys()):
         if k.startswith("model.diffusion_model."):
             unet_sd[k.replace("model.diffusion_model.", "")] = state_dict.pop(k)
     info = _load_state_dict_on_device(unet, unet_sd, device=map_location, dtype=dtype)
-    print("U-Net: ", info)
+    logger.info(f"U-Net: {info}")
 
     # Text Encoders
-    print("building text encoders")
+    logger.info("building text encoders")
 
     # Text Encoder 1 is same to Stability AI's SDXL
     text_model1_cfg = CLIPTextConfig(
@@ -252,7 +261,7 @@ def load_models_from_sdxl_checkpoint(model_version, ckpt_path, map_location, dty
     with init_empty_weights():
         text_model2 = CLIPTextModelWithProjection(text_model2_cfg)
 
-    print("loading text encoders from checkpoint")
+    logger.info("loading text encoders from checkpoint")
     te1_sd = {}
     te2_sd = {}
     for k in list(state_dict.keys()):
@@ -266,22 +275,22 @@ def load_models_from_sdxl_checkpoint(model_version, ckpt_path, map_location, dty
         te1_sd.pop("text_model.embeddings.position_ids")
 
     info1 = _load_state_dict_on_device(text_model1, te1_sd, device=map_location)  # remain fp32
-    print("text encoder 1:", info1)
+    logger.info(f"text encoder 1: {info1}")
 
     converted_sd, logit_scale = convert_sdxl_text_encoder_2_checkpoint(te2_sd, max_length=77)
     info2 = _load_state_dict_on_device(text_model2, converted_sd, device=map_location)  # remain fp32
-    print("text encoder 2:", info2)
+    logger.info(f"text encoder 2: {info2}")
 
     # prepare vae
-    print("building VAE")
+    logger.info("building VAE")
     vae_config = model_util.create_vae_diffusers_config()
     with init_empty_weights():
         vae = AutoencoderKL(**vae_config)
 
-    print("loading VAE from checkpoint")
+    logger.info("loading VAE from checkpoint")
     converted_vae_checkpoint = model_util.convert_ldm_vae_checkpoint(state_dict, vae_config)
     info = _load_state_dict_on_device(vae, converted_vae_checkpoint, device=map_location, dtype=dtype)
-    print("VAE:", info)
+    logger.info(f"VAE: {info}")
 
     ckpt_info = (epoch, global_step) if epoch is not None else None
     return text_model1, text_model2, vae, unet, logit_scale, ckpt_info

library/sdxl_original_control_net.py

@@ -0,0 +1,272 @@
+# some parts are modified from Diffusers library (Apache License 2.0)
+
+import math
+from types import SimpleNamespace
+from typing import Any, Optional
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import functional as F
+from einops import rearrange
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+from library import sdxl_original_unet
+from library.sdxl_model_util import convert_sdxl_unet_state_dict_to_diffusers, convert_diffusers_unet_state_dict_to_sdxl
+
+
+class ControlNetConditioningEmbedding(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+        dims = [16, 32, 96, 256]
+
+        self.conv_in = nn.Conv2d(3, dims[0], kernel_size=3, padding=1)
+        self.blocks = nn.ModuleList([])
+
+        for i in range(len(dims) - 1):
+            channel_in = dims[i]
+            channel_out = dims[i + 1]
+            self.blocks.append(nn.Conv2d(channel_in, channel_in, kernel_size=3, padding=1))
+            self.blocks.append(nn.Conv2d(channel_in, channel_out, kernel_size=3, padding=1, stride=2))
+
+        self.conv_out = nn.Conv2d(dims[-1], 320, kernel_size=3, padding=1)
+        nn.init.zeros_(self.conv_out.weight)  # zero module weight
+        nn.init.zeros_(self.conv_out.bias)  # zero module bias
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        x = F.silu(x)
+        for block in self.blocks:
+            x = block(x)
+            x = F.silu(x)
+        x = self.conv_out(x)
+        return x
+
+
+class SdxlControlNet(sdxl_original_unet.SdxlUNet2DConditionModel):
+    def __init__(self, multiplier: Optional[float] = None, **kwargs):
+        super().__init__(**kwargs)
+        self.multiplier = multiplier
+
+        # remove unet layers
+        self.output_blocks = nn.ModuleList([])
+        del self.out
+
+        self.controlnet_cond_embedding = ControlNetConditioningEmbedding()
+
+        dims = [320, 320, 320, 320, 640, 640, 640, 1280, 1280]
+        self.controlnet_down_blocks = nn.ModuleList([])
+        for dim in dims:
+            self.controlnet_down_blocks.append(nn.Conv2d(dim, dim, kernel_size=1))
+            nn.init.zeros_(self.controlnet_down_blocks[-1].weight)  # zero module weight
+            nn.init.zeros_(self.controlnet_down_blocks[-1].bias)  # zero module bias
+
+        self.controlnet_mid_block = nn.Conv2d(1280, 1280, kernel_size=1)
+        nn.init.zeros_(self.controlnet_mid_block.weight)  # zero module weight
+        nn.init.zeros_(self.controlnet_mid_block.bias)  # zero module bias
+
+    def init_from_unet(self, unet: sdxl_original_unet.SdxlUNet2DConditionModel):
+        unet_sd = unet.state_dict()
+        unet_sd = {k: v for k, v in unet_sd.items() if not k.startswith("out")}
+        sd = super().state_dict()
+        sd.update(unet_sd)
+        info = super().load_state_dict(sd, strict=True, assign=True)
+        return info
+
+    def load_state_dict(self, state_dict: dict, strict: bool = True, assign: bool = True) -> Any:
+        # convert state_dict to SAI format
+        unet_sd = {}
+        for k in list(state_dict.keys()):
+            if not k.startswith("controlnet_"):
+                unet_sd[k] = state_dict.pop(k)
+        unet_sd = convert_diffusers_unet_state_dict_to_sdxl(unet_sd)
+        state_dict.update(unet_sd)
+        super().load_state_dict(state_dict, strict=strict, assign=assign)
+
+    def state_dict(self, destination=None, prefix="", keep_vars=False):
+        # convert state_dict to Diffusers format
+        state_dict = super().state_dict(destination, prefix, keep_vars)
+        control_net_sd = {}
+        for k in list(state_dict.keys()):
+            if k.startswith("controlnet_"):
+                control_net_sd[k] = state_dict.pop(k)
+        state_dict = convert_sdxl_unet_state_dict_to_diffusers(state_dict)
+        state_dict.update(control_net_sd)
+        return state_dict
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        timesteps: Optional[torch.Tensor] = None,
+        context: Optional[torch.Tensor] = None,
+        y: Optional[torch.Tensor] = None,
+        cond_image: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        # broadcast timesteps to batch dimension
+        timesteps = timesteps.expand(x.shape[0])
+
+        t_emb = sdxl_original_unet.get_timestep_embedding(timesteps, self.model_channels, downscale_freq_shift=0)
+        t_emb = t_emb.to(x.dtype)
+        emb = self.time_embed(t_emb)
+
+        assert x.shape[0] == y.shape[0], f"batch size mismatch: {x.shape[0]} != {y.shape[0]}"
+        assert x.dtype == y.dtype, f"dtype mismatch: {x.dtype} != {y.dtype}"
+        emb = emb + self.label_emb(y)
+
+        def call_module(module, h, emb, context):
+            x = h
+            for layer in module:
+                if isinstance(layer, sdxl_original_unet.ResnetBlock2D):
+                    x = layer(x, emb)
+                elif isinstance(layer, sdxl_original_unet.Transformer2DModel):
+                    x = layer(x, context)
+                else:
+                    x = layer(x)
+            return x
+
+        h = x
+        multiplier = self.multiplier if self.multiplier is not None else 1.0
+        hs = []
+        for i, module in enumerate(self.input_blocks):
+            h = call_module(module, h, emb, context)
+            if i == 0:
+                h = self.controlnet_cond_embedding(cond_image) + h
+            hs.append(self.controlnet_down_blocks[i](h) * multiplier)
+
+        h = call_module(self.middle_block, h, emb, context)
+        h = self.controlnet_mid_block(h) * multiplier
+
+        return hs, h
+
+
+class SdxlControlledUNet(sdxl_original_unet.SdxlUNet2DConditionModel):
+    """
+    This class is for training purpose only.
+    """
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+    def forward(self, x, timesteps=None, context=None, y=None, input_resi_add=None, mid_add=None, **kwargs):
+        # broadcast timesteps to batch dimension
+        timesteps = timesteps.expand(x.shape[0])
+
+        hs = []
+        t_emb = sdxl_original_unet.get_timestep_embedding(timesteps, self.model_channels, downscale_freq_shift=0)
+        t_emb = t_emb.to(x.dtype)
+        emb = self.time_embed(t_emb)
+
+        assert x.shape[0] == y.shape[0], f"batch size mismatch: {x.shape[0]} != {y.shape[0]}"
+        assert x.dtype == y.dtype, f"dtype mismatch: {x.dtype} != {y.dtype}"
+        emb = emb + self.label_emb(y)
+
+        def call_module(module, h, emb, context):
+            x = h
+            for layer in module:
+                if isinstance(layer, sdxl_original_unet.ResnetBlock2D):
+                    x = layer(x, emb)
+                elif isinstance(layer, sdxl_original_unet.Transformer2DModel):
+                    x = layer(x, context)
+                else:
+                    x = layer(x)
+            return x
+
+        h = x
+        for module in self.input_blocks:
+            h = call_module(module, h, emb, context)
+            hs.append(h)
+
+        h = call_module(self.middle_block, h, emb, context)
+        h = h + mid_add
+
+        for module in self.output_blocks:
+            resi = hs.pop() + input_resi_add.pop()
+            h = torch.cat([h, resi], dim=1)
+            h = call_module(module, h, emb, context)
+
+        h = h.type(x.dtype)
+        h = call_module(self.out, h, emb, context)
+
+        return h
+
+
+if __name__ == "__main__":
+    import time
+
+    logger.info("create unet")
+    unet = SdxlControlledUNet()
+    unet.to("cuda", torch.bfloat16)
+    unet.set_use_sdpa(True)
+    unet.set_gradient_checkpointing(True)
+    unet.train()
+
+    logger.info("create control_net")
+    control_net = SdxlControlNet()
+    control_net.to("cuda")
+    control_net.set_use_sdpa(True)
+    control_net.set_gradient_checkpointing(True)
+    control_net.train()
+
+    logger.info("Initialize control_net from unet")
+    control_net.init_from_unet(unet)
+
+    unet.requires_grad_(False)
+    control_net.requires_grad_(True)
+
+    # 使用メモリ量確認用の疑似学習ループ
+    logger.info("preparing optimizer")
+
+    # optimizer = torch.optim.SGD(unet.parameters(), lr=1e-3, nesterov=True, momentum=0.9) # not working
+
+    import bitsandbytes
+
+    optimizer = bitsandbytes.adam.Adam8bit(control_net.parameters(), lr=1e-3)  # not working
+    # optimizer = bitsandbytes.optim.RMSprop8bit(unet.parameters(), lr=1e-3)  # working at 23.5 GB with torch2
+    # optimizer=bitsandbytes.optim.Adagrad8bit(unet.parameters(), lr=1e-3)  # working at 23.5 GB with torch2
+
+    # import transformers
+    # optimizer = transformers.optimization.Adafactor(unet.parameters(), relative_step=True)  # working at 22.2GB with torch2
+
+    scaler = torch.cuda.amp.GradScaler(enabled=True)
+
+    logger.info("start training")
+    steps = 10
+    batch_size = 1
+
+    for step in range(steps):
+        logger.info(f"step {step}")
+        if step == 1:
+            time_start = time.perf_counter()
+
+        x = torch.randn(batch_size, 4, 128, 128).cuda()  # 1024x1024
+        t = torch.randint(low=0, high=1000, size=(batch_size,), device="cuda")
+        txt = torch.randn(batch_size, 77, 2048).cuda()
+        vector = torch.randn(batch_size, sdxl_original_unet.ADM_IN_CHANNELS).cuda()
+        cond_img = torch.rand(batch_size, 3, 1024, 1024).cuda()
+
+        with torch.cuda.amp.autocast(enabled=True, dtype=torch.bfloat16):
+            input_resi_add, mid_add = control_net(x, t, txt, vector, cond_img)
+            output = unet(x, t, txt, vector, input_resi_add, mid_add)
+            target = torch.randn_like(output)
+            loss = torch.nn.functional.mse_loss(output, target)
+
+        scaler.scale(loss).backward()
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad(set_to_none=True)
+
+    time_end = time.perf_counter()
+    logger.info(f"elapsed time: {time_end - time_start} [sec] for last {steps - 1} steps")
+
+    logger.info("finish training")
+    sd = control_net.state_dict()
+
+    from safetensors.torch import save_file
+
+    save_file(sd, r"E:\Work\SD\Tmp\sdxl\ctrl\control_net.safetensors")

library/sdxl_original_unet.py

@@ -30,7 +30,12 @@ import torch.utils.checkpoint
 from torch import nn
 from torch.nn import functional as F
 from einops import rearrange
+from library.utils import setup_logging
 
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
 
 IN_CHANNELS: int = 4
 OUT_CHANNELS: int = 4
@@ -332,7 +337,7 @@ class ResnetBlock2D(nn.Module):
 
     def forward(self, x, emb):
         if self.training and self.gradient_checkpointing:
-            # print("ResnetBlock2D: gradient_checkpointing")
+            # logger.info("ResnetBlock2D: gradient_checkpointing")
 
             def create_custom_forward(func):
                 def custom_forward(*inputs):
@@ -366,7 +371,7 @@ class Downsample2D(nn.Module):
 
     def forward(self, hidden_states):
         if self.training and self.gradient_checkpointing:
-            # print("Downsample2D: gradient_checkpointing")
+            # logger.info("Downsample2D: gradient_checkpointing")
 
             def create_custom_forward(func):
                 def custom_forward(*inputs):
@@ -653,7 +658,7 @@ class BasicTransformerBlock(nn.Module):
 
     def forward(self, hidden_states, context=None, timestep=None):
         if self.training and self.gradient_checkpointing:
-            # print("BasicTransformerBlock: checkpointing")
+            # logger.info("BasicTransformerBlock: checkpointing")
 
             def create_custom_forward(func):
                 def custom_forward(*inputs):
@@ -796,7 +801,7 @@ class Upsample2D(nn.Module):
 
     def forward(self, hidden_states, output_size=None):
         if self.training and self.gradient_checkpointing:
-            # print("Upsample2D: gradient_checkpointing")
+            # logger.info("Upsample2D: gradient_checkpointing")
 
             def create_custom_forward(func):
                 def custom_forward(*inputs):
@@ -1046,7 +1051,7 @@ class SdxlUNet2DConditionModel(nn.Module):
         for block in blocks:
             for module in block:
                 if hasattr(module, "set_use_memory_efficient_attention"):
-                    # print(module.__class__.__name__)
+                    # logger.info(module.__class__.__name__)
                     module.set_use_memory_efficient_attention(xformers, mem_eff)
 
     def set_use_sdpa(self, sdpa: bool) -> None:
@@ -1061,7 +1066,7 @@ class SdxlUNet2DConditionModel(nn.Module):
         for block in blocks:
             for module in block.modules():
                 if hasattr(module, "gradient_checkpointing"):
-                    # print(module.__class__.__name__, module.gradient_checkpointing, "->", value)
+                    # logger.info(f{module.__class__.__name__} {module.gradient_checkpointing} -> {value}")
                     module.gradient_checkpointing = value
 
     # endregion
@@ -1071,7 +1076,7 @@ class SdxlUNet2DConditionModel(nn.Module):
         timesteps = timesteps.expand(x.shape[0])
 
         hs = []
-        t_emb = get_timestep_embedding(timesteps, self.model_channels)  # , repeat_only=False)
+        t_emb = get_timestep_embedding(timesteps, self.model_channels, downscale_freq_shift=0)  # , repeat_only=False)
         t_emb = t_emb.to(x.dtype)
         emb = self.time_embed(t_emb)
 
@@ -1083,7 +1088,7 @@ class SdxlUNet2DConditionModel(nn.Module):
         def call_module(module, h, emb, context):
             x = h
             for layer in module:
-                # print(layer.__class__.__name__, x.dtype, emb.dtype, context.dtype if context is not None else None)
+                # logger.info(layer.__class__.__name__, x.dtype, emb.dtype, context.dtype if context is not None else None)
                 if isinstance(layer, ResnetBlock2D):
                     x = layer(x, emb)
                 elif isinstance(layer, Transformer2DModel):
@@ -1129,20 +1134,20 @@ class InferSdxlUNet2DConditionModel:
     # call original model's methods
     def __getattr__(self, name):
         return getattr(self.delegate, name)
-    
+
     def __call__(self, *args, **kwargs):
         return self.delegate(*args, **kwargs)
 
     def set_deep_shrink(self, ds_depth_1, ds_timesteps_1=650, ds_depth_2=None, ds_timesteps_2=None, ds_ratio=0.5):
         if ds_depth_1 is None:
-            print("Deep Shrink is disabled.")
+            logger.info("Deep Shrink is disabled.")
             self.ds_depth_1 = None
             self.ds_timesteps_1 = None
             self.ds_depth_2 = None
             self.ds_timesteps_2 = None
             self.ds_ratio = None
         else:
-            print(
+            logger.info(
                 f"Deep Shrink is enabled: [depth={ds_depth_1}/{ds_depth_2}, timesteps={ds_timesteps_1}/{ds_timesteps_2}, ratio={ds_ratio}]"
             )
             self.ds_depth_1 = ds_depth_1
@@ -1151,9 +1156,9 @@ class InferSdxlUNet2DConditionModel:
             self.ds_timesteps_2 = ds_timesteps_2 if ds_timesteps_2 is not None else 1000
             self.ds_ratio = ds_ratio
 
-    def forward(self, x, timesteps=None, context=None, y=None, **kwargs):
+    def forward(self, x, timesteps=None, context=None, y=None, input_resi_add=None, mid_add=None, **kwargs):
         r"""
-        current implementation is a copy of `SdxlUNet2DConditionModel.forward()` with Deep Shrink.
+        current implementation is a copy of `SdxlUNet2DConditionModel.forward()` with Deep Shrink and ControlNet.
         """
         _self = self.delegate
 
@@ -1161,7 +1166,7 @@ class InferSdxlUNet2DConditionModel:
         timesteps = timesteps.expand(x.shape[0])
 
         hs = []
-        t_emb = get_timestep_embedding(timesteps, _self.model_channels)  # , repeat_only=False)
+        t_emb = get_timestep_embedding(timesteps, _self.model_channels, downscale_freq_shift=0)  # , repeat_only=False)
         t_emb = t_emb.to(x.dtype)
         emb = _self.time_embed(t_emb)
 
@@ -1204,6 +1209,8 @@ class InferSdxlUNet2DConditionModel:
             hs.append(h)
 
         h = call_module(_self.middle_block, h, emb, context)
+        if mid_add is not None:
+            h = h + mid_add
 
         for module in _self.output_blocks:
             # Deep Shrink
@@ -1212,7 +1219,11 @@ class InferSdxlUNet2DConditionModel:
                     # print("upsample", h.shape, hs[-1].shape)
                     h = resize_like(h, hs[-1])
 
-            h = torch.cat([h, hs.pop()], dim=1)
+            resi = hs.pop()
+            if input_resi_add is not None:
+                resi = resi + input_resi_add.pop()
+
+            h = torch.cat([h, resi], dim=1)
             h = call_module(module, h, emb, context)
 
         # Deep Shrink: in case of depth 0
@@ -1229,7 +1240,7 @@ class InferSdxlUNet2DConditionModel:
 if __name__ == "__main__":
     import time
 
-    print("create unet")
+    logger.info("create unet")
     unet = SdxlUNet2DConditionModel()
 
     unet.to("cuda")
@@ -1238,7 +1249,7 @@ if __name__ == "__main__":
     unet.train()
 
     # 使用メモリ量確認用の疑似学習ループ
-    print("preparing optimizer")
+    logger.info("preparing optimizer")
 
     # optimizer = torch.optim.SGD(unet.parameters(), lr=1e-3, nesterov=True, momentum=0.9) # not working
 
@@ -1253,12 +1264,12 @@ if __name__ == "__main__":
 
     scaler = torch.cuda.amp.GradScaler(enabled=True)
 
-    print("start training")
+    logger.info("start training")
     steps = 10
     batch_size = 1
 
     for step in range(steps):
-        print(f"step {step}")
+        logger.info(f"step {step}")
         if step == 1:
             time_start = time.perf_counter()
 
@@ -1278,4 +1289,4 @@ if __name__ == "__main__":
         optimizer.zero_grad(set_to_none=True)
 
     time_end = time.perf_counter()
-    print(f"elapsed time: {time_end - time_start} [sec] for last {steps - 1} steps")
+    logger.info(f"elapsed time: {time_end - time_start} [sec] for last {steps - 1} steps")

library/sdxl_train_util.py

@@ -1,14 +1,23 @@
 import argparse
-import gc
 import math
 import os
 from typing import Optional
+
 import torch
+from library.device_utils import init_ipex, clean_memory_on_device
+
+init_ipex()
+
 from accelerate import init_empty_weights
 from tqdm import tqdm
 from transformers import CLIPTokenizer
 from library import model_util, sdxl_model_util, train_util, sdxl_original_unet
-from library.sdxl_lpw_stable_diffusion import SdxlStableDiffusionLongPromptWeightingPipeline
+from .utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
 
 TOKENIZER1_PATH = "openai/clip-vit-large-patch14"
 TOKENIZER2_PATH = "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"
@@ -17,11 +26,10 @@ TOKENIZER2_PATH = "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"
 
 
 def load_target_model(args, accelerator, model_version: str, weight_dtype):
-    # load models for each process
     model_dtype = match_mixed_precision(args, weight_dtype)  # prepare fp16/bf16
     for pi in range(accelerator.state.num_processes):
         if pi == accelerator.state.local_process_index:
-            print(f"loading model for process {accelerator.state.local_process_index}/{accelerator.state.num_processes}")
+            logger.info(f"loading model for process {accelerator.state.local_process_index}/{accelerator.state.num_processes}")
 
             (
                 load_stable_diffusion_format,
@@ -38,6 +46,7 @@ def load_target_model(args, accelerator, model_version: str, weight_dtype):
                 weight_dtype,
                 accelerator.device if args.lowram else "cpu",
                 model_dtype,
+                args.disable_mmap_load_safetensors,
             )
 
             # work on low-ram device
@@ -47,22 +56,21 @@ def load_target_model(args, accelerator, model_version: str, weight_dtype):
                 unet.to(accelerator.device)
                 vae.to(accelerator.device)
 
-            gc.collect()
-            torch.cuda.empty_cache()
+            clean_memory_on_device(accelerator.device)
         accelerator.wait_for_everyone()
 
     return load_stable_diffusion_format, text_encoder1, text_encoder2, vae, unet, logit_scale, ckpt_info
 
 
 def _load_target_model(
-    name_or_path: str, vae_path: Optional[str], model_version: str, weight_dtype, device="cpu", model_dtype=None
+    name_or_path: str, vae_path: Optional[str], model_version: str, weight_dtype, device="cpu", model_dtype=None, disable_mmap=False
 ):
     # model_dtype only work with full fp16/bf16
     name_or_path = os.readlink(name_or_path) if os.path.islink(name_or_path) else name_or_path
     load_stable_diffusion_format = os.path.isfile(name_or_path)  # determine SD or Diffusers
 
     if load_stable_diffusion_format:
-        print(f"load StableDiffusion checkpoint: {name_or_path}")
+        logger.info(f"load StableDiffusion checkpoint: {name_or_path}")
         (
             text_encoder1,
             text_encoder2,
@@ -70,13 +78,13 @@ def _load_target_model(
             unet,
             logit_scale,
             ckpt_info,
-        ) = sdxl_model_util.load_models_from_sdxl_checkpoint(model_version, name_or_path, device, model_dtype)
+        ) = sdxl_model_util.load_models_from_sdxl_checkpoint(model_version, name_or_path, device, model_dtype, disable_mmap)
     else:
         # Diffusers model is loaded to CPU
         from diffusers import StableDiffusionXLPipeline
 
         variant = "fp16" if weight_dtype == torch.float16 else None
-        print(f"load Diffusers pretrained models: {name_or_path}, variant={variant}")
+        logger.info(f"load Diffusers pretrained models: {name_or_path}, variant={variant}")
         try:
             try:
                 pipe = StableDiffusionXLPipeline.from_pretrained(
@@ -84,12 +92,12 @@ def _load_target_model(
                 )
             except EnvironmentError as ex:
                 if variant is not None:
-                    print("try to load fp32 model")
+                    logger.info("try to load fp32 model")
                     pipe = StableDiffusionXLPipeline.from_pretrained(name_or_path, variant=None, tokenizer=None)
                 else:
                     raise ex
         except EnvironmentError as ex:
-            print(
+            logger.error(
                 f"model is not found as a file or in Hugging Face, perhaps file name is wrong? / 指定したモデル名のファイル、またはHugging Faceのモデルが見つかりません。ファイル名が誤っているかもしれません: {name_or_path}"
             )
             raise ex
@@ -112,7 +120,7 @@ def _load_target_model(
         with init_empty_weights():
             unet = sdxl_original_unet.SdxlUNet2DConditionModel()  # overwrite unet
         sdxl_model_util._load_state_dict_on_device(unet, state_dict, device=device, dtype=model_dtype)
-        print("U-Net converted to original U-Net")
+        logger.info("U-Net converted to original U-Net")
 
         logit_scale = None
         ckpt_info = None
@@ -120,13 +128,13 @@ def _load_target_model(
     # VAEを読み込む
     if vae_path is not None:
         vae = model_util.load_vae(vae_path, weight_dtype)
-        print("additional VAE loaded")
+        logger.info("additional VAE loaded")
 
     return load_stable_diffusion_format, text_encoder1, text_encoder2, vae, unet, logit_scale, ckpt_info
 
 
 def load_tokenizers(args: argparse.Namespace):
-    print("prepare tokenizers")
+    logger.info("prepare tokenizers")
 
     original_paths = [TOKENIZER1_PATH, TOKENIZER2_PATH]
     tokeniers = []
@@ -135,14 +143,14 @@ def load_tokenizers(args: argparse.Namespace):
         if args.tokenizer_cache_dir:
             local_tokenizer_path = os.path.join(args.tokenizer_cache_dir, original_path.replace("/", "_"))
             if os.path.exists(local_tokenizer_path):
-                print(f"load tokenizer from cache: {local_tokenizer_path}")
+                logger.info(f"load tokenizer from cache: {local_tokenizer_path}")
                 tokenizer = CLIPTokenizer.from_pretrained(local_tokenizer_path)
 
         if tokenizer is None:
             tokenizer = CLIPTokenizer.from_pretrained(original_path)
 
         if args.tokenizer_cache_dir and not os.path.exists(local_tokenizer_path):
-            print(f"save Tokenizer to cache: {local_tokenizer_path}")
+            logger.info(f"save Tokenizer to cache: {local_tokenizer_path}")
             tokenizer.save_pretrained(local_tokenizer_path)
 
         if i == 1:
@@ -151,7 +159,7 @@ def load_tokenizers(args: argparse.Namespace):
         tokeniers.append(tokenizer)
 
     if hasattr(args, "max_token_length") and args.max_token_length is not None:
-        print(f"update token length: {args.max_token_length}")
+        logger.info(f"update token length: {args.max_token_length}")
 
     return tokeniers
 
@@ -318,50 +326,58 @@ def save_sd_model_on_epoch_end_or_stepwise(
     )
 
 
-def add_sdxl_training_arguments(parser: argparse.ArgumentParser):
+def add_sdxl_training_arguments(parser: argparse.ArgumentParser, support_text_encoder_caching: bool = True):
+    if support_text_encoder_caching:
+        parser.add_argument(
+            "--cache_text_encoder_outputs",
+            action="store_true",
+            help="cache text encoder outputs / text encoderの出力をキャッシュする",
+        )
+        parser.add_argument(
+            "--cache_text_encoder_outputs_to_disk",
+            action="store_true",
+            help="cache text encoder outputs to disk / text encoderの出力をディスクにキャッシュする",
+        )
     parser.add_argument(
-        "--cache_text_encoder_outputs", action="store_true", help="cache text encoder outputs / text encoderの出力をキャッシュする"
-    )
-    parser.add_argument(
-        "--cache_text_encoder_outputs_to_disk",
+        "--disable_mmap_load_safetensors",
         action="store_true",
-        help="cache text encoder outputs to disk / text encoderの出力をディスクにキャッシュする",
+        help="disable mmap load for safetensors. Speed up model loading in WSL environment / safetensorsのmmapロードを無効にする。WSL環境等でモデル読み込みを高速化できる",
     )
 
 
-def verify_sdxl_training_args(args: argparse.Namespace, supportTextEncoderCaching: bool = True):
+def verify_sdxl_training_args(args: argparse.Namespace, support_text_encoder_caching: bool = True):
     assert not args.v2, "v2 cannot be enabled in SDXL training / SDXL学習ではv2を有効にすることはできません"
-    if args.v_parameterization:
-        print("v_parameterization will be unexpected / SDXL学習ではv_parameterizationは想定外の動作になります")
 
     if args.clip_skip is not None:
-        print("clip_skip will be unexpected / SDXL学習ではclip_skipは動作しません")
+        logger.warning("clip_skip will be unexpected / SDXL学習ではclip_skipは動作しません")
 
     # if args.multires_noise_iterations:
-    #     print(
+    #     logger.info(
     #         f"Warning: SDXL has been trained with noise_offset={DEFAULT_NOISE_OFFSET}, but noise_offset is disabled due to multires_noise_iterations / SDXLはnoise_offset={DEFAULT_NOISE_OFFSET}で学習されていますが、multires_noise_iterationsが有効になっているためnoise_offsetは無効になります"
     #     )
     # else:
     #     if args.noise_offset is None:
     #         args.noise_offset = DEFAULT_NOISE_OFFSET
     #     elif args.noise_offset != DEFAULT_NOISE_OFFSET:
-    #         print(
+    #         logger.info(
     #             f"Warning: SDXL has been trained with noise_offset={DEFAULT_NOISE_OFFSET} / SDXLはnoise_offset={DEFAULT_NOISE_OFFSET}で学習されています"
     #         )
-    #     print(f"noise_offset is set to {args.noise_offset} / noise_offsetが{args.noise_offset}に設定されました")
+    #     logger.info(f"noise_offset is set to {args.noise_offset} / noise_offsetが{args.noise_offset}に設定されました")
 
-    assert (
-        not hasattr(args, "weighted_captions") or not args.weighted_captions
-    ), "weighted_captions cannot be enabled in SDXL training currently / SDXL学習では今のところweighted_captionsを有効にすることはできません"
+    # assert (
+    #     not hasattr(args, "weighted_captions") or not args.weighted_captions
+    # ), "weighted_captions cannot be enabled in SDXL training currently / SDXL学習では今のところweighted_captionsを有効にすることはできません"
 
-    if supportTextEncoderCaching:
+    if support_text_encoder_caching:
         if args.cache_text_encoder_outputs_to_disk and not args.cache_text_encoder_outputs:
             args.cache_text_encoder_outputs = True
-            print(
+            logger.warning(
                 "cache_text_encoder_outputs is enabled because cache_text_encoder_outputs_to_disk is enabled / "
                 + "cache_text_encoder_outputs_to_diskが有効になっているためcache_text_encoder_outputsが有効になりました"
             )
 
 
 def sample_images(*args, **kwargs):
+    from library.sdxl_lpw_stable_diffusion import SdxlStableDiffusionLongPromptWeightingPipeline
+
     return train_util.sample_images_common(SdxlStableDiffusionLongPromptWeightingPipeline, *args, **kwargs)

library/slicing_vae.py

@@ -26,7 +26,10 @@ from diffusers.models.modeling_utils import ModelMixin
 from diffusers.models.unet_2d_blocks import UNetMidBlock2D, get_down_block, get_up_block
 from diffusers.models.vae import DecoderOutput, DiagonalGaussianDistribution
 from diffusers.models.autoencoder_kl import AutoencoderKLOutput
-
+from .utils import setup_logging
+setup_logging()
+import logging
+logger = logging.getLogger(__name__)
 
 def slice_h(x, num_slices):
     # slice with pad 1 both sides: to eliminate side effect of padding of conv2d
@@ -89,7 +92,7 @@ def resblock_forward(_self, num_slices, input_tensor, temb, **kwargs):
     #     sliced_tensor = torch.chunk(x, num_div, dim=1)
     #     sliced_weight = torch.chunk(norm.weight, num_div, dim=0)
     #     sliced_bias = torch.chunk(norm.bias, num_div, dim=0)
-    #     print(sliced_tensor[0].shape, num_div, sliced_weight[0].shape, sliced_bias[0].shape)
+    #     logger.info(sliced_tensor[0].shape, num_div, sliced_weight[0].shape, sliced_bias[0].shape)
     #     normed_tensor = []
     #     for i in range(num_div):
     #         n = torch.group_norm(sliced_tensor[i], norm.num_groups, sliced_weight[i], sliced_bias[i], norm.eps)
@@ -243,7 +246,7 @@ class SlicingEncoder(nn.Module):
 
         self.num_slices = num_slices
         div = num_slices / (2 ** (len(self.down_blocks) - 1))  # 深い層はそこまで分割しなくていいので適宜減らす
-        # print(f"initial divisor: {div}")
+        # logger.info(f"initial divisor: {div}")
         if div >= 2:
             div = int(div)
             for resnet in self.mid_block.resnets:
@@ -253,11 +256,11 @@ class SlicingEncoder(nn.Module):
         for i, down_block in enumerate(self.down_blocks[::-1]):
             if div >= 2:
                 div = int(div)
-                # print(f"down block: {i} divisor: {div}")
+                # logger.info(f"down block: {i} divisor: {div}")
                 for resnet in down_block.resnets:
                     resnet.forward = wrapper(resblock_forward, resnet, div)
                 if down_block.downsamplers is not None:
-                    # print("has downsample")
+                    # logger.info("has downsample")
                     for downsample in down_block.downsamplers:
                         downsample.forward = wrapper(self.downsample_forward, downsample, div * 2)
             div *= 2
@@ -307,7 +310,7 @@ class SlicingEncoder(nn.Module):
     def downsample_forward(self, _self, num_slices, hidden_states):
         assert hidden_states.shape[1] == _self.channels
         assert _self.use_conv and _self.padding == 0
-        print("downsample forward", num_slices, hidden_states.shape)
+        logger.info(f"downsample forward {num_slices} {hidden_states.shape}")
 
         org_device = hidden_states.device
         cpu_device = torch.device("cpu")
@@ -350,7 +353,7 @@ class SlicingEncoder(nn.Module):
                 hidden_states = torch.cat([hidden_states, x], dim=2)
 
         hidden_states = hidden_states.to(org_device)
-        # print("downsample forward done", hidden_states.shape)
+        # logger.info(f"downsample forward done {hidden_states.shape}")
         return hidden_states
 
 
@@ -426,7 +429,7 @@ class SlicingDecoder(nn.Module):
 
         self.num_slices = num_slices
         div = num_slices / (2 ** (len(self.up_blocks) - 1))
-        print(f"initial divisor: {div}")
+        logger.info(f"initial divisor: {div}")
         if div >= 2:
             div = int(div)
             for resnet in self.mid_block.resnets:
@@ -436,11 +439,11 @@ class SlicingDecoder(nn.Module):
         for i, up_block in enumerate(self.up_blocks):
             if div >= 2:
                 div = int(div)
-                # print(f"up block: {i} divisor: {div}")
+                # logger.info(f"up block: {i} divisor: {div}")
                 for resnet in up_block.resnets:
                     resnet.forward = wrapper(resblock_forward, resnet, div)
                 if up_block.upsamplers is not None:
-                    # print("has upsample")
+                    # logger.info("has upsample")
                     for upsample in up_block.upsamplers:
                         upsample.forward = wrapper(self.upsample_forward, upsample, div * 2)
             div *= 2
@@ -528,7 +531,7 @@ class SlicingDecoder(nn.Module):
             del x
 
         hidden_states = torch.cat(sliced, dim=2)
-        # print("us hidden_states", hidden_states.shape)
+        # logger.info(f"us hidden_states {hidden_states.shape}")
         del sliced
 
         hidden_states = hidden_states.to(org_device)

library/strategy_base.py

@@ -0,0 +1,637 @@
+# base class for platform strategies. this file defines the interface for strategies
+
+import os
+import re
+from typing import Any, List, Optional, Tuple, Union, Callable
+
+import numpy as np
+import torch
+from transformers import CLIPTokenizer, CLIPTextModel, CLIPTextModelWithProjection
+
+
+# TODO remove circular import by moving ImageInfo to a separate file
+# from library.train_util import ImageInfo
+
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class TokenizeStrategy:
+    _strategy = None  # strategy instance: actual strategy class
+
+    _re_attention = re.compile(
+        r"""\\\(|
+\\\)|
+\\\[|
+\\]|
+\\\\|
+\\|
+\(|
+\[|
+:([+-]?[.\d]+)\)|
+\)|
+]|
+[^\\()\[\]:]+|
+:
+""",
+        re.X,
+    )
+
+    @classmethod
+    def set_strategy(cls, strategy):
+        if cls._strategy is not None:
+            raise RuntimeError(f"Internal error. {cls.__name__} strategy is already set")
+        cls._strategy = strategy
+
+    @classmethod
+    def get_strategy(cls) -> Optional["TokenizeStrategy"]:
+        return cls._strategy
+
+    def _load_tokenizer(
+        self, model_class: Any, model_id: str, subfolder: Optional[str] = None, tokenizer_cache_dir: Optional[str] = None
+    ) -> Any:
+        tokenizer = None
+        if tokenizer_cache_dir:
+            local_tokenizer_path = os.path.join(tokenizer_cache_dir, model_id.replace("/", "_"))
+            if os.path.exists(local_tokenizer_path):
+                logger.info(f"load tokenizer from cache: {local_tokenizer_path}")
+                tokenizer = model_class.from_pretrained(local_tokenizer_path)  # same for v1 and v2
+
+        if tokenizer is None:
+            tokenizer = model_class.from_pretrained(model_id, subfolder=subfolder)
+
+        if tokenizer_cache_dir and not os.path.exists(local_tokenizer_path):
+            logger.info(f"save Tokenizer to cache: {local_tokenizer_path}")
+            tokenizer.save_pretrained(local_tokenizer_path)
+
+        return tokenizer
+
+    def tokenize(self, text: Union[str, List[str]]) -> List[torch.Tensor]:
+        raise NotImplementedError
+
+    def tokenize_with_weights(self, text: Union[str, List[str]]) -> Tuple[List[torch.Tensor], List[torch.Tensor]]:
+        """
+        returns: [tokens1, tokens2, ...], [weights1, weights2, ...]
+        """
+        raise NotImplementedError
+
+    def _get_weighted_input_ids(
+        self, tokenizer: CLIPTokenizer, text: str, max_length: Optional[int] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        max_length includes starting and ending tokens.
+        """
+
+        def parse_prompt_attention(text):
+            """
+            Parses a string with attention tokens and returns a list of pairs: text and its associated weight.
+            Accepted tokens are:
+            (abc) - increases attention to abc by a multiplier of 1.1
+            (abc:3.12) - increases attention to abc by a multiplier of 3.12
+            [abc] - decreases attention to abc by a multiplier of 1.1
+            \( - literal character '('
+            \[ - literal character '['
+            \) - literal character ')'
+            \] - literal character ']'
+            \\ - literal character '\'
+            anything else - just text
+            >>> parse_prompt_attention('normal text')
+            [['normal text', 1.0]]
+            >>> parse_prompt_attention('an (important) word')
+            [['an ', 1.0], ['important', 1.1], [' word', 1.0]]
+            >>> parse_prompt_attention('(unbalanced')
+            [['unbalanced', 1.1]]
+            >>> parse_prompt_attention('\(literal\]')
+            [['(literal]', 1.0]]
+            >>> parse_prompt_attention('(unnecessary)(parens)')
+            [['unnecessaryparens', 1.1]]
+            >>> parse_prompt_attention('a (((house:1.3)) [on] a (hill:0.5), sun, (((sky))).')
+            [['a ', 1.0],
+            ['house', 1.5730000000000004],
+            [' ', 1.1],
+            ['on', 1.0],
+            [' a ', 1.1],
+            ['hill', 0.55],
+            [', sun, ', 1.1],
+            ['sky', 1.4641000000000006],
+            ['.', 1.1]]
+            """
+
+            res = []
+            round_brackets = []
+            square_brackets = []
+
+            round_bracket_multiplier = 1.1
+            square_bracket_multiplier = 1 / 1.1
+
+            def multiply_range(start_position, multiplier):
+                for p in range(start_position, len(res)):
+                    res[p][1] *= multiplier
+
+            for m in TokenizeStrategy._re_attention.finditer(text):
+                text = m.group(0)
+                weight = m.group(1)
+
+                if text.startswith("\\"):
+                    res.append([text[1:], 1.0])
+                elif text == "(":
+                    round_brackets.append(len(res))
+                elif text == "[":
+                    square_brackets.append(len(res))
+                elif weight is not None and len(round_brackets) > 0:
+                    multiply_range(round_brackets.pop(), float(weight))
+                elif text == ")" and len(round_brackets) > 0:
+                    multiply_range(round_brackets.pop(), round_bracket_multiplier)
+                elif text == "]" and len(square_brackets) > 0:
+                    multiply_range(square_brackets.pop(), square_bracket_multiplier)
+                else:
+                    res.append([text, 1.0])
+
+            for pos in round_brackets:
+                multiply_range(pos, round_bracket_multiplier)
+
+            for pos in square_brackets:
+                multiply_range(pos, square_bracket_multiplier)
+
+            if len(res) == 0:
+                res = [["", 1.0]]
+
+            # merge runs of identical weights
+            i = 0
+            while i + 1 < len(res):
+                if res[i][1] == res[i + 1][1]:
+                    res[i][0] += res[i + 1][0]
+                    res.pop(i + 1)
+                else:
+                    i += 1
+
+            return res
+
+        def get_prompts_with_weights(text: str, max_length: int):
+            r"""
+            Tokenize a list of prompts and return its tokens with weights of each token. max_length does not include starting and ending token.
+
+            No padding, starting or ending token is included.
+            """
+            truncated = False
+
+            texts_and_weights = parse_prompt_attention(text)
+            tokens = []
+            weights = []
+            for word, weight in texts_and_weights:
+                # tokenize and discard the starting and the ending token
+                token = tokenizer(word).input_ids[1:-1]
+                tokens += token
+                # copy the weight by length of token
+                weights += [weight] * len(token)
+                # stop if the text is too long (longer than truncation limit)
+                if len(tokens) > max_length:
+                    truncated = True
+                    break
+            # truncate
+            if len(tokens) > max_length:
+                truncated = True
+                tokens = tokens[:max_length]
+                weights = weights[:max_length]
+            if truncated:
+                logger.warning("Prompt was truncated. Try to shorten the prompt or increase max_embeddings_multiples")
+            return tokens, weights
+
+        def pad_tokens_and_weights(tokens, weights, max_length, bos, eos, pad):
+            r"""
+            Pad the tokens (with starting and ending tokens) and weights (with 1.0) to max_length.
+            """
+            tokens = [bos] + tokens + [eos] + [pad] * (max_length - 2 - len(tokens))
+            weights = [1.0] + weights + [1.0] * (max_length - 1 - len(weights))
+            return tokens, weights
+
+        if max_length is None:
+            max_length = tokenizer.model_max_length
+
+        tokens, weights = get_prompts_with_weights(text, max_length - 2)
+        tokens, weights = pad_tokens_and_weights(
+            tokens, weights, max_length, tokenizer.bos_token_id, tokenizer.eos_token_id, tokenizer.pad_token_id
+        )
+        return torch.tensor(tokens).unsqueeze(0), torch.tensor(weights).unsqueeze(0)
+
+    def _get_input_ids(
+        self, tokenizer: CLIPTokenizer, text: str, max_length: Optional[int] = None, weighted: bool = False
+    ) -> torch.Tensor:
+        """
+        for SD1.5/2.0/SDXL
+        TODO support batch input
+        """
+        if max_length is None:
+            max_length = tokenizer.model_max_length - 2
+
+        if weighted:
+            input_ids, weights = self._get_weighted_input_ids(tokenizer, text, max_length)
+        else:
+            input_ids = tokenizer(text, padding="max_length", truncation=True, max_length=max_length, return_tensors="pt").input_ids
+
+        if max_length > tokenizer.model_max_length:
+            input_ids = input_ids.squeeze(0)
+            iids_list = []
+            if tokenizer.pad_token_id == tokenizer.eos_token_id:
+                # v1
+                # 77以上の時は "<BOS> .... <EOS> <EOS> <EOS>" でトータル227とかになっているので、"<BOS>...<EOS>"の三連に変換する
+                # 1111氏のやつは , で区切る、とかしているようだが　とりあえず単純に
+                for i in range(1, max_length - tokenizer.model_max_length + 2, tokenizer.model_max_length - 2):  # (1, 152, 75)
+                    ids_chunk = (
+                        input_ids[0].unsqueeze(0),
+                        input_ids[i : i + tokenizer.model_max_length - 2],
+                        input_ids[-1].unsqueeze(0),
+                    )
+                    ids_chunk = torch.cat(ids_chunk)
+                    iids_list.append(ids_chunk)
+            else:
+                # v2 or SDXL
+                # 77以上の時は "<BOS> .... <EOS> <PAD> <PAD>..." でトータル227とかになっているので、"<BOS>...<EOS> <PAD> <PAD> ..."の三連に変換する
+                for i in range(1, max_length - tokenizer.model_max_length + 2, tokenizer.model_max_length - 2):
+                    ids_chunk = (
+                        input_ids[0].unsqueeze(0),  # BOS
+                        input_ids[i : i + tokenizer.model_max_length - 2],
+                        input_ids[-1].unsqueeze(0),
+                    )  # PAD or EOS
+                    ids_chunk = torch.cat(ids_chunk)
+
+                    # 末尾が <EOS> <PAD> または <PAD> <PAD> の場合は、何もしなくてよい
+                    # 末尾が x <PAD/EOS> の場合は末尾を <EOS> に変える（x <EOS> なら結果的に変化なし）
+                    if ids_chunk[-2] != tokenizer.eos_token_id and ids_chunk[-2] != tokenizer.pad_token_id:
+                        ids_chunk[-1] = tokenizer.eos_token_id
+                    # 先頭が <BOS> <PAD> ... の場合は <BOS> <EOS> <PAD> ... に変える
+                    if ids_chunk[1] == tokenizer.pad_token_id:
+                        ids_chunk[1] = tokenizer.eos_token_id
+
+                    iids_list.append(ids_chunk)
+
+            input_ids = torch.stack(iids_list)  # 3,77
+
+            if weighted:
+                weights = weights.squeeze(0)
+                new_weights = torch.ones(input_ids.shape)
+                for i in range(1, max_length - tokenizer.model_max_length + 2, tokenizer.model_max_length - 2):
+                    b = i // (tokenizer.model_max_length - 2)
+                    new_weights[b, 1 : 1 + tokenizer.model_max_length - 2] = weights[i : i + tokenizer.model_max_length - 2]
+                weights = new_weights
+
+        if weighted:
+            return input_ids, weights
+        return input_ids
+
+
+class TextEncodingStrategy:
+    _strategy = None  # strategy instance: actual strategy class
+
+    @classmethod
+    def set_strategy(cls, strategy):
+        if cls._strategy is not None:
+            raise RuntimeError(f"Internal error. {cls.__name__} strategy is already set")
+        cls._strategy = strategy
+
+    @classmethod
+    def get_strategy(cls) -> Optional["TextEncodingStrategy"]:
+        return cls._strategy
+
+    def encode_tokens(
+        self, tokenize_strategy: TokenizeStrategy, models: List[Any], tokens: List[torch.Tensor]
+    ) -> List[torch.Tensor]:
+        """
+        Encode tokens into embeddings and outputs.
+        :param tokens: list of token tensors for each TextModel
+        :return: list of output embeddings for each architecture
+        """
+        raise NotImplementedError
+
+    def encode_tokens_with_weights(
+        self, tokenize_strategy: TokenizeStrategy, models: List[Any], tokens: List[torch.Tensor], weights: List[torch.Tensor]
+    ) -> List[torch.Tensor]:
+        """
+        Encode tokens into embeddings and outputs.
+        :param tokens: list of token tensors for each TextModel
+        :param weights: list of weight tensors for each TextModel
+        :return: list of output embeddings for each architecture
+        """
+        raise NotImplementedError
+
+
+class TextEncoderOutputsCachingStrategy:
+    _strategy = None  # strategy instance: actual strategy class
+
+    def __init__(
+        self,
+        cache_to_disk: bool,
+        batch_size: Optional[int],
+        skip_disk_cache_validity_check: bool,
+        is_partial: bool = False,
+        is_weighted: bool = False,
+    ) -> None:
+        self._cache_to_disk = cache_to_disk
+        self._batch_size = batch_size
+        self.skip_disk_cache_validity_check = skip_disk_cache_validity_check
+        self._is_partial = is_partial
+        self._is_weighted = is_weighted
+
+    @classmethod
+    def set_strategy(cls, strategy):
+        if cls._strategy is not None:
+            raise RuntimeError(f"Internal error. {cls.__name__} strategy is already set")
+        cls._strategy = strategy
+
+    @classmethod
+    def get_strategy(cls) -> Optional["TextEncoderOutputsCachingStrategy"]:
+        return cls._strategy
+
+    @property
+    def cache_to_disk(self):
+        return self._cache_to_disk
+
+    @property
+    def batch_size(self):
+        return self._batch_size
+
+    @property
+    def is_partial(self):
+        return self._is_partial
+
+    @property
+    def is_weighted(self):
+        return self._is_weighted
+
+    def get_outputs_npz_path(self, image_abs_path: str) -> str:
+        raise NotImplementedError
+
+    def load_outputs_npz(self, npz_path: str) -> List[np.ndarray]:
+        raise NotImplementedError
+
+    def is_disk_cached_outputs_expected(self, npz_path: str) -> bool:
+        raise NotImplementedError
+
+    def cache_batch_outputs(
+        self, tokenize_strategy: TokenizeStrategy, models: List[Any], text_encoding_strategy: TextEncodingStrategy, batch: List
+    ):
+        raise NotImplementedError
+
+
+class LatentsCachingStrategy:
+    # TODO commonize utillity functions to this class, such as npz handling etc.
+
+    _strategy = None  # strategy instance: actual strategy class
+
+    def __init__(self, cache_to_disk: bool, batch_size: int, skip_disk_cache_validity_check: bool) -> None:
+        self._cache_to_disk = cache_to_disk
+        self._batch_size = batch_size
+        self.skip_disk_cache_validity_check = skip_disk_cache_validity_check
+
+    @classmethod
+    def set_strategy(cls, strategy):
+        if cls._strategy is not None:
+            raise RuntimeError(f"Internal error. {cls.__name__} strategy is already set")
+        cls._strategy = strategy
+
+    @classmethod
+    def get_strategy(cls) -> Optional["LatentsCachingStrategy"]:
+        return cls._strategy
+
+    @property
+    def cache_to_disk(self):
+        return self._cache_to_disk
+
+    @property
+    def batch_size(self):
+        return self._batch_size
+
+    @property
+    def cache_suffix(self):
+        raise NotImplementedError
+
+    def get_image_size_from_disk_cache_path(self, absolute_path: str, npz_path: str) -> Tuple[Optional[int], Optional[int]]:
+        w, h = os.path.splitext(npz_path)[0].split("_")[-2].split("x")
+        return int(w), int(h)
+
+    def get_latents_npz_path(self, absolute_path: str, image_size: Tuple[int, int]) -> str:
+        raise NotImplementedError
+
+    def is_disk_cached_latents_expected(
+        self, bucket_reso: Tuple[int, int], npz_path: str, flip_aug: bool, alpha_mask: bool
+    ) -> bool:
+        raise NotImplementedError
+
+    def cache_batch_latents(self, model: Any, batch: List, flip_aug: bool, alpha_mask: bool, random_crop: bool):
+        raise NotImplementedError
+
+    def _default_is_disk_cached_latents_expected(
+        self,
+        latents_stride: int,
+        bucket_reso: Tuple[int, int],
+        npz_path: str,
+        flip_aug: bool,
+        apply_alpha_mask: bool,
+        multi_resolution: bool = False,
+    ) -> bool:
+        """
+        Args:
+            latents_stride: stride of latents
+            bucket_reso: resolution of the bucket
+            npz_path: path to the npz file
+            flip_aug: whether to flip images
+            apply_alpha_mask: whether to apply alpha mask
+            multi_resolution: whether to use multi-resolution latents
+
+        Returns:
+            bool
+        """
+        if not self.cache_to_disk:
+            return False
+        if not os.path.exists(npz_path):
+            return False
+        if self.skip_disk_cache_validity_check:
+            return True
+
+        expected_latents_size = (bucket_reso[1] // latents_stride, bucket_reso[0] // latents_stride)  # bucket_reso is (W, H)
+
+        # e.g. "_32x64", HxW
+        key_reso_suffix = f"_{expected_latents_size[0]}x{expected_latents_size[1]}" if multi_resolution else ""
+
+        try:
+            npz = np.load(npz_path)
+            if "latents" + key_reso_suffix not in npz:
+                return False
+            if flip_aug and "latents_flipped" + key_reso_suffix not in npz:
+                return False
+            if apply_alpha_mask and "alpha_mask" + key_reso_suffix not in npz:
+                return False
+        except Exception as e:
+            logger.error(f"Error loading file: {npz_path}")
+            raise e
+
+        return True
+
+    # TODO remove circular dependency for ImageInfo
+    def _default_cache_batch_latents(
+        self,
+        encode_by_vae: Callable,
+        vae_device: torch.device,
+        vae_dtype: torch.dtype,
+        image_infos: List,
+        flip_aug: bool,
+        apply_alpha_mask: bool,
+        random_crop: bool,
+        multi_resolution: bool = False,
+    ):
+        """
+        Default implementation for cache_batch_latents. Image loading, VAE, flipping, alpha mask handling are common.
+
+        Args:
+            encode_by_vae: function to encode images by VAE
+            vae_device: device to use for VAE
+            vae_dtype: dtype to use for VAE
+            image_infos: list of ImageInfo
+            flip_aug: whether to flip images
+            apply_alpha_mask: whether to apply alpha mask
+            random_crop: whether to random crop images
+            multi_resolution: whether to use multi-resolution latents
+        
+        Returns: 
+            None
+        """
+        from library import train_util  # import here to avoid circular import
+
+        img_tensor, alpha_masks, original_sizes, crop_ltrbs = train_util.load_images_and_masks_for_caching(
+            image_infos, apply_alpha_mask, random_crop
+        )
+        img_tensor = img_tensor.to(device=vae_device, dtype=vae_dtype)
+
+        with torch.no_grad():
+            latents_tensors = encode_by_vae(img_tensor).to("cpu")
+        if flip_aug:
+            img_tensor = torch.flip(img_tensor, dims=[3])
+            with torch.no_grad():
+                flipped_latents = encode_by_vae(img_tensor).to("cpu")
+        else:
+            flipped_latents = [None] * len(latents_tensors)
+
+        # for info, latents, flipped_latent, alpha_mask in zip(image_infos, latents_tensors, flipped_latents, alpha_masks):
+        for i in range(len(image_infos)):
+            info = image_infos[i]
+            latents = latents_tensors[i]
+            flipped_latent = flipped_latents[i]
+            alpha_mask = alpha_masks[i]
+            original_size = original_sizes[i]
+            crop_ltrb = crop_ltrbs[i]
+
+            latents_size = latents.shape[1:3]  # H, W
+            key_reso_suffix = f"_{latents_size[0]}x{latents_size[1]}" if multi_resolution else ""  # e.g. "_32x64", HxW
+
+            if self.cache_to_disk:
+                self.save_latents_to_disk(
+                    info.latents_npz, latents, original_size, crop_ltrb, flipped_latent, alpha_mask, key_reso_suffix
+                )
+            else:
+                info.latents_original_size = original_size
+                info.latents_crop_ltrb = crop_ltrb
+                info.latents = latents
+                if flip_aug:
+                    info.latents_flipped = flipped_latent
+                info.alpha_mask = alpha_mask
+
+    def load_latents_from_disk(
+        self, npz_path: str, bucket_reso: Tuple[int, int]
+    ) -> Tuple[Optional[np.ndarray], Optional[List[int]], Optional[List[int]], Optional[np.ndarray], Optional[np.ndarray]]:
+        """
+        for SD/SDXL
+
+        Args:
+            npz_path (str): Path to the npz file.
+            bucket_reso (Tuple[int, int]): The resolution of the bucket.
+        
+        Returns:
+            Tuple[
+                Optional[np.ndarray], 
+                Optional[List[int]], 
+                Optional[List[int]], 
+                Optional[np.ndarray], 
+                Optional[np.ndarray]
+            ]: Latent np tensors, original size, crop (left top, right bottom), flipped latents, alpha mask
+        """
+        return self._default_load_latents_from_disk(None, npz_path, bucket_reso)
+
+    def _default_load_latents_from_disk(
+        self, latents_stride: Optional[int], npz_path: str, bucket_reso: Tuple[int, int]
+    ) -> Tuple[Optional[np.ndarray], Optional[List[int]], Optional[List[int]], Optional[np.ndarray], Optional[np.ndarray]]:
+        """
+        Args:
+            latents_stride (Optional[int]): Stride for latents. If None, load all latents.
+            npz_path (str): Path to the npz file.
+            bucket_reso (Tuple[int, int]): The resolution of the bucket.
+       
+        Returns:
+            Tuple[
+                Optional[np.ndarray], 
+                Optional[List[int]], 
+                Optional[List[int]], 
+                Optional[np.ndarray], 
+                Optional[np.ndarray]
+            ]: Latent np tensors, original size, crop (left top, right bottom), flipped latents, alpha mask
+        """
+        if latents_stride is None:
+            key_reso_suffix = ""
+        else:
+            latents_size = (bucket_reso[1] // latents_stride, bucket_reso[0] // latents_stride)  # bucket_reso is (W, H)
+            key_reso_suffix = f"_{latents_size[0]}x{latents_size[1]}"  # e.g. "_32x64", HxW
+
+        npz = np.load(npz_path)
+        if "latents" + key_reso_suffix not in npz:
+            raise ValueError(f"latents{key_reso_suffix} not found in {npz_path}")
+
+        latents = npz["latents" + key_reso_suffix]
+        original_size = npz["original_size" + key_reso_suffix].tolist()
+        crop_ltrb = npz["crop_ltrb" + key_reso_suffix].tolist()
+        flipped_latents = npz["latents_flipped" + key_reso_suffix] if "latents_flipped" + key_reso_suffix in npz else None
+        alpha_mask = npz["alpha_mask" + key_reso_suffix] if "alpha_mask" + key_reso_suffix in npz else None
+        return latents, original_size, crop_ltrb, flipped_latents, alpha_mask
+
+    def save_latents_to_disk(
+        self,
+        npz_path,
+        latents_tensor,
+        original_size,
+        crop_ltrb,
+        flipped_latents_tensor=None,
+        alpha_mask=None,
+        key_reso_suffix="",
+    ):
+        """
+        Args:
+            npz_path (str): Path to the npz file.
+            latents_tensor (torch.Tensor): Latent tensor
+            original_size (List[int]): Original size of the image
+            crop_ltrb (List[int]): Crop left top right bottom
+            flipped_latents_tensor (Optional[torch.Tensor]): Flipped latent tensor
+            alpha_mask (Optional[torch.Tensor]): Alpha mask
+            key_reso_suffix (str): Key resolution suffix
+
+        Returns:
+            None
+        """
+        kwargs = {}
+
+        if os.path.exists(npz_path):
+            # load existing npz and update it
+            npz = np.load(npz_path)
+            for key in npz.files:
+                kwargs[key] = npz[key]
+
+        # TODO float() is needed if vae is in bfloat16. Remove it if vae is float16.
+        kwargs["latents" + key_reso_suffix] = latents_tensor.float().cpu().numpy()
+        kwargs["original_size" + key_reso_suffix] = np.array(original_size)
+        kwargs["crop_ltrb" + key_reso_suffix] = np.array(crop_ltrb)
+        if flipped_latents_tensor is not None:
+            kwargs["latents_flipped" + key_reso_suffix] = flipped_latents_tensor.float().cpu().numpy()
+        if alpha_mask is not None:
+            kwargs["alpha_mask" + key_reso_suffix] = alpha_mask.float().cpu().numpy()
+        np.savez(npz_path, **kwargs)

library/strategy_flux.py

@@ -0,0 +1,271 @@
+import os
+import glob
+from typing import Any, List, Optional, Tuple, Union
+import torch
+import numpy as np
+from transformers import CLIPTokenizer, T5TokenizerFast
+
+from library import flux_utils, train_util
+from library.strategy_base import LatentsCachingStrategy, TextEncodingStrategy, TokenizeStrategy, TextEncoderOutputsCachingStrategy
+
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+CLIP_L_TOKENIZER_ID = "openai/clip-vit-large-patch14"
+T5_XXL_TOKENIZER_ID = "google/t5-v1_1-xxl"
+
+
+class FluxTokenizeStrategy(TokenizeStrategy):
+    def __init__(self, t5xxl_max_length: int = 512, tokenizer_cache_dir: Optional[str] = None) -> None:
+        self.t5xxl_max_length = t5xxl_max_length
+        self.clip_l = self._load_tokenizer(CLIPTokenizer, CLIP_L_TOKENIZER_ID, tokenizer_cache_dir=tokenizer_cache_dir)
+        self.t5xxl = self._load_tokenizer(T5TokenizerFast, T5_XXL_TOKENIZER_ID, tokenizer_cache_dir=tokenizer_cache_dir)
+
+    def tokenize(self, text: Union[str, List[str]]) -> List[torch.Tensor]:
+        text = [text] if isinstance(text, str) else text
+
+        l_tokens = self.clip_l(text, max_length=77, padding="max_length", truncation=True, return_tensors="pt")
+        t5_tokens = self.t5xxl(text, max_length=self.t5xxl_max_length, padding="max_length", truncation=True, return_tensors="pt")
+
+        t5_attn_mask = t5_tokens["attention_mask"]
+        l_tokens = l_tokens["input_ids"]
+        t5_tokens = t5_tokens["input_ids"]
+
+        return [l_tokens, t5_tokens, t5_attn_mask]
+
+
+class FluxTextEncodingStrategy(TextEncodingStrategy):
+    def __init__(self, apply_t5_attn_mask: Optional[bool] = None) -> None:
+        """
+        Args:
+            apply_t5_attn_mask: Default value for apply_t5_attn_mask.
+        """
+        self.apply_t5_attn_mask = apply_t5_attn_mask
+
+    def encode_tokens(
+        self,
+        tokenize_strategy: TokenizeStrategy,
+        models: List[Any],
+        tokens: List[torch.Tensor],
+        apply_t5_attn_mask: Optional[bool] = None,
+    ) -> List[torch.Tensor]:
+        # supports single model inference
+
+        if apply_t5_attn_mask is None:
+            apply_t5_attn_mask = self.apply_t5_attn_mask
+
+        clip_l, t5xxl = models if len(models) == 2 else (models[0], None)
+        l_tokens, t5_tokens = tokens[:2]
+        t5_attn_mask = tokens[2] if len(tokens) > 2 else None
+
+        # clip_l is None when using T5 only
+        if clip_l is not None and l_tokens is not None:
+            l_pooled = clip_l(l_tokens.to(clip_l.device))["pooler_output"]
+        else:
+            l_pooled = None
+
+        # t5xxl is None when using CLIP only
+        if t5xxl is not None and t5_tokens is not None:
+            # t5_out is [b, max length, 4096]
+            attention_mask = None if not apply_t5_attn_mask else t5_attn_mask.to(t5xxl.device)
+            t5_out, _ = t5xxl(t5_tokens.to(t5xxl.device), attention_mask, return_dict=False, output_hidden_states=True)
+            # if zero_pad_t5_output:
+            #     t5_out = t5_out * t5_attn_mask.to(t5_out.device).unsqueeze(-1)
+            txt_ids = torch.zeros(t5_out.shape[0], t5_out.shape[1], 3, device=t5_out.device)
+        else:
+            t5_out = None
+            txt_ids = None
+            t5_attn_mask = None  # caption may be dropped/shuffled, so t5_attn_mask should not be used to make sure the mask is same as the cached one
+
+        return [l_pooled, t5_out, txt_ids, t5_attn_mask]  # returns t5_attn_mask for attention mask in transformer
+
+
+class FluxTextEncoderOutputsCachingStrategy(TextEncoderOutputsCachingStrategy):
+    FLUX_TEXT_ENCODER_OUTPUTS_NPZ_SUFFIX = "_flux_te.npz"
+
+    def __init__(
+        self,
+        cache_to_disk: bool,
+        batch_size: int,
+        skip_disk_cache_validity_check: bool,
+        is_partial: bool = False,
+        apply_t5_attn_mask: bool = False,
+    ) -> None:
+        super().__init__(cache_to_disk, batch_size, skip_disk_cache_validity_check, is_partial)
+        self.apply_t5_attn_mask = apply_t5_attn_mask
+
+        self.warn_fp8_weights = False
+
+    def get_outputs_npz_path(self, image_abs_path: str) -> str:
+        return os.path.splitext(image_abs_path)[0] + FluxTextEncoderOutputsCachingStrategy.FLUX_TEXT_ENCODER_OUTPUTS_NPZ_SUFFIX
+
+    def is_disk_cached_outputs_expected(self, npz_path: str):
+        if not self.cache_to_disk:
+            return False
+        if not os.path.exists(npz_path):
+            return False
+        if self.skip_disk_cache_validity_check:
+            return True
+
+        try:
+            npz = np.load(npz_path)
+            if "l_pooled" not in npz:
+                return False
+            if "t5_out" not in npz:
+                return False
+            if "txt_ids" not in npz:
+                return False
+            if "t5_attn_mask" not in npz:
+                return False
+            if "apply_t5_attn_mask" not in npz:
+                return False
+            npz_apply_t5_attn_mask = npz["apply_t5_attn_mask"]
+            if npz_apply_t5_attn_mask != self.apply_t5_attn_mask:
+                return False
+        except Exception as e:
+            logger.error(f"Error loading file: {npz_path}")
+            raise e
+
+        return True
+
+    def load_outputs_npz(self, npz_path: str) -> List[np.ndarray]:
+        data = np.load(npz_path)
+        l_pooled = data["l_pooled"]
+        t5_out = data["t5_out"]
+        txt_ids = data["txt_ids"]
+        t5_attn_mask = data["t5_attn_mask"]
+        # apply_t5_attn_mask should be same as self.apply_t5_attn_mask
+        return [l_pooled, t5_out, txt_ids, t5_attn_mask]
+
+    def cache_batch_outputs(
+        self, tokenize_strategy: TokenizeStrategy, models: List[Any], text_encoding_strategy: TextEncodingStrategy, infos: List
+    ):
+        if not self.warn_fp8_weights:
+            if flux_utils.get_t5xxl_actual_dtype(models[1]) == torch.float8_e4m3fn:
+                logger.warning(
+                    "T5 model is using fp8 weights for caching. This may affect the quality of the cached outputs."
+                    " / T5モデルはfp8の重みを使用しています。これはキャッシュの品質に影響を与える可能性があります。"
+                )
+            self.warn_fp8_weights = True
+
+        flux_text_encoding_strategy: FluxTextEncodingStrategy = text_encoding_strategy
+        captions = [info.caption for info in infos]
+
+        tokens_and_masks = tokenize_strategy.tokenize(captions)
+        with torch.no_grad():
+            # attn_mask is applied in text_encoding_strategy.encode_tokens if apply_t5_attn_mask is True
+            l_pooled, t5_out, txt_ids, _ = flux_text_encoding_strategy.encode_tokens(tokenize_strategy, models, tokens_and_masks)
+
+        if l_pooled.dtype == torch.bfloat16:
+            l_pooled = l_pooled.float()
+        if t5_out.dtype == torch.bfloat16:
+            t5_out = t5_out.float()
+        if txt_ids.dtype == torch.bfloat16:
+            txt_ids = txt_ids.float()
+
+        l_pooled = l_pooled.cpu().numpy()
+        t5_out = t5_out.cpu().numpy()
+        txt_ids = txt_ids.cpu().numpy()
+        t5_attn_mask = tokens_and_masks[2].cpu().numpy()
+
+        for i, info in enumerate(infos):
+            l_pooled_i = l_pooled[i]
+            t5_out_i = t5_out[i]
+            txt_ids_i = txt_ids[i]
+            t5_attn_mask_i = t5_attn_mask[i]
+            apply_t5_attn_mask_i = self.apply_t5_attn_mask
+
+            if self.cache_to_disk:
+                np.savez(
+                    info.text_encoder_outputs_npz,
+                    l_pooled=l_pooled_i,
+                    t5_out=t5_out_i,
+                    txt_ids=txt_ids_i,
+                    t5_attn_mask=t5_attn_mask_i,
+                    apply_t5_attn_mask=apply_t5_attn_mask_i,
+                )
+            else:
+                # it's fine that attn mask is not None. it's overwritten before calling the model if necessary
+                info.text_encoder_outputs = (l_pooled_i, t5_out_i, txt_ids_i, t5_attn_mask_i)
+
+
+class FluxLatentsCachingStrategy(LatentsCachingStrategy):
+    FLUX_LATENTS_NPZ_SUFFIX = "_flux.npz"
+
+    def __init__(self, cache_to_disk: bool, batch_size: int, skip_disk_cache_validity_check: bool) -> None:
+        super().__init__(cache_to_disk, batch_size, skip_disk_cache_validity_check)
+
+    @property
+    def cache_suffix(self) -> str:
+        return FluxLatentsCachingStrategy.FLUX_LATENTS_NPZ_SUFFIX
+
+    def get_latents_npz_path(self, absolute_path: str, image_size: Tuple[int, int]) -> str:
+        return (
+            os.path.splitext(absolute_path)[0]
+            + f"_{image_size[0]:04d}x{image_size[1]:04d}"
+            + FluxLatentsCachingStrategy.FLUX_LATENTS_NPZ_SUFFIX
+        )
+
+    def is_disk_cached_latents_expected(self, bucket_reso: Tuple[int, int], npz_path: str, flip_aug: bool, alpha_mask: bool):
+        return self._default_is_disk_cached_latents_expected(8, bucket_reso, npz_path, flip_aug, alpha_mask, multi_resolution=True)
+
+    def load_latents_from_disk(
+        self, npz_path: str, bucket_reso: Tuple[int, int]
+    ) -> Tuple[Optional[np.ndarray], Optional[List[int]], Optional[List[int]], Optional[np.ndarray], Optional[np.ndarray]]:
+        return self._default_load_latents_from_disk(8, npz_path, bucket_reso)  # support multi-resolution
+
+    # TODO remove circular dependency for ImageInfo
+    def cache_batch_latents(self, vae, image_infos: List, flip_aug: bool, alpha_mask: bool, random_crop: bool):
+        encode_by_vae = lambda img_tensor: vae.encode(img_tensor).to("cpu")
+        vae_device = vae.device
+        vae_dtype = vae.dtype
+
+        self._default_cache_batch_latents(
+            encode_by_vae, vae_device, vae_dtype, image_infos, flip_aug, alpha_mask, random_crop, multi_resolution=True
+        )
+
+        if not train_util.HIGH_VRAM:
+            train_util.clean_memory_on_device(vae.device)
+
+
+if __name__ == "__main__":
+    # test code for FluxTokenizeStrategy
+    # tokenizer = sd3_models.SD3Tokenizer()
+    strategy = FluxTokenizeStrategy(256)
+    text = "hello world"
+
+    l_tokens, g_tokens, t5_tokens = strategy.tokenize(text)
+    # print(l_tokens.shape)
+    print(l_tokens)
+    print(g_tokens)
+    print(t5_tokens)
+
+    texts = ["hello world", "the quick brown fox jumps over the lazy dog"]
+    l_tokens_2 = strategy.clip_l(texts, max_length=77, padding="max_length", truncation=True, return_tensors="pt")
+    g_tokens_2 = strategy.clip_g(texts, max_length=77, padding="max_length", truncation=True, return_tensors="pt")
+    t5_tokens_2 = strategy.t5xxl(
+        texts, max_length=strategy.t5xxl_max_length, padding="max_length", truncation=True, return_tensors="pt"
+    )
+    print(l_tokens_2)
+    print(g_tokens_2)
+    print(t5_tokens_2)
+
+    # compare
+    print(torch.allclose(l_tokens, l_tokens_2["input_ids"][0]))
+    print(torch.allclose(g_tokens, g_tokens_2["input_ids"][0]))
+    print(torch.allclose(t5_tokens, t5_tokens_2["input_ids"][0]))
+
+    text = ",".join(["hello world! this is long text"] * 50)
+    l_tokens, g_tokens, t5_tokens = strategy.tokenize(text)
+    print(l_tokens)
+    print(g_tokens)
+    print(t5_tokens)
+
+    print(f"model max length l: {strategy.clip_l.model_max_length}")
+    print(f"model max length g: {strategy.clip_g.model_max_length}")
+    print(f"model max length t5: {strategy.t5xxl.model_max_length}")

library/strategy_hunyuan_image.py

@@ -0,0 +1,218 @@
+import os
+from typing import Any, List, Optional, Tuple, Union
+import torch
+import numpy as np
+from transformers import AutoTokenizer, Qwen2Tokenizer
+
+from library import hunyuan_image_text_encoder, hunyuan_image_vae, train_util
+from library.strategy_base import LatentsCachingStrategy, TextEncodingStrategy, TokenizeStrategy, TextEncoderOutputsCachingStrategy
+
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class HunyuanImageTokenizeStrategy(TokenizeStrategy):
+    def __init__(self, tokenizer_cache_dir: Optional[str] = None) -> None:
+        self.vlm_tokenizer = self._load_tokenizer(
+            Qwen2Tokenizer, hunyuan_image_text_encoder.QWEN_2_5_VL_IMAGE_ID, tokenizer_cache_dir=tokenizer_cache_dir
+        )
+        self.byt5_tokenizer = self._load_tokenizer(
+            AutoTokenizer, hunyuan_image_text_encoder.BYT5_TOKENIZER_PATH, subfolder="", tokenizer_cache_dir=tokenizer_cache_dir
+        )
+
+    def tokenize(self, text: Union[str, List[str]]) -> List[torch.Tensor]:
+        text = [text] if isinstance(text, str) else text
+
+        vlm_tokens, vlm_mask = hunyuan_image_text_encoder.get_qwen_tokens(self.vlm_tokenizer, text)
+
+        # byt5_tokens, byt5_mask = hunyuan_image_text_encoder.get_byt5_text_tokens(self.byt5_tokenizer, text)
+        byt5_tokens = []
+        byt5_mask = []
+        for t in text:
+            tokens, mask = hunyuan_image_text_encoder.get_byt5_text_tokens(self.byt5_tokenizer, t)
+            if tokens is None:
+                tokens = torch.zeros((1, 1), dtype=torch.long)
+                mask = torch.zeros((1, 1), dtype=torch.long)
+            byt5_tokens.append(tokens)
+            byt5_mask.append(mask)
+        max_len = max([m.shape[1] for m in byt5_mask])
+        byt5_tokens = torch.cat([torch.nn.functional.pad(t, (0, max_len - t.shape[1]), value=0) for t in byt5_tokens], dim=0)
+        byt5_mask = torch.cat([torch.nn.functional.pad(m, (0, max_len - m.shape[1]), value=0) for m in byt5_mask], dim=0)
+
+        return [vlm_tokens, vlm_mask, byt5_tokens, byt5_mask]
+
+
+class HunyuanImageTextEncodingStrategy(TextEncodingStrategy):
+    def __init__(self) -> None:
+        pass
+
+    def encode_tokens(
+        self, tokenize_strategy: TokenizeStrategy, models: List[Any], tokens: List[torch.Tensor]
+    ) -> List[torch.Tensor]:
+        vlm_tokens, vlm_mask, byt5_tokens, byt5_mask = tokens
+
+        qwen2vlm, byt5 = models
+
+        # autocast and no_grad are handled in hunyuan_image_text_encoder
+        vlm_embed, vlm_mask = hunyuan_image_text_encoder.get_qwen_prompt_embeds_from_tokens(qwen2vlm, vlm_tokens, vlm_mask)
+
+        # ocr_mask, byt5_embed, byt5_mask = hunyuan_image_text_encoder.get_byt5_prompt_embeds_from_tokens(
+        #     byt5, byt5_tokens, byt5_mask
+        # )
+        ocr_mask, byt5_embed, byt5_updated_mask = [], [], []
+        for i in range(byt5_tokens.shape[0]):
+            ocr_m, byt5_e, byt5_m = hunyuan_image_text_encoder.get_byt5_prompt_embeds_from_tokens(
+                byt5, byt5_tokens[i : i + 1], byt5_mask[i : i + 1]
+            )
+            ocr_mask.append(torch.zeros((1,), dtype=torch.long) + (1 if ocr_m[0] else 0))  # 1 or 0
+            byt5_embed.append(byt5_e)
+            byt5_updated_mask.append(byt5_m)
+
+        ocr_mask = torch.cat(ocr_mask, dim=0).to(torch.bool)  # [B]
+        byt5_embed = torch.cat(byt5_embed, dim=0)
+        byt5_updated_mask = torch.cat(byt5_updated_mask, dim=0)
+
+        return [vlm_embed, vlm_mask, byt5_embed, byt5_updated_mask, ocr_mask]
+
+
+class HunyuanImageTextEncoderOutputsCachingStrategy(TextEncoderOutputsCachingStrategy):
+    HUNYUAN_IMAGE_TEXT_ENCODER_OUTPUTS_NPZ_SUFFIX = "_hi_te.npz"
+
+    def __init__(
+        self, cache_to_disk: bool, batch_size: int, skip_disk_cache_validity_check: bool, is_partial: bool = False
+    ) -> None:
+        super().__init__(cache_to_disk, batch_size, skip_disk_cache_validity_check, is_partial)
+
+    def get_outputs_npz_path(self, image_abs_path: str) -> str:
+        return (
+            os.path.splitext(image_abs_path)[0]
+            + HunyuanImageTextEncoderOutputsCachingStrategy.HUNYUAN_IMAGE_TEXT_ENCODER_OUTPUTS_NPZ_SUFFIX
+        )
+
+    def is_disk_cached_outputs_expected(self, npz_path: str):
+        if not self.cache_to_disk:
+            return False
+        if not os.path.exists(npz_path):
+            return False
+        if self.skip_disk_cache_validity_check:
+            return True
+
+        try:
+            npz = np.load(npz_path)
+            if "vlm_embed" not in npz:
+                return False
+            if "vlm_mask" not in npz:
+                return False
+            if "byt5_embed" not in npz:
+                return False
+            if "byt5_mask" not in npz:
+                return False
+            if "ocr_mask" not in npz:
+                return False
+        except Exception as e:
+            logger.error(f"Error loading file: {npz_path}")
+            raise e
+
+        return True
+
+    def load_outputs_npz(self, npz_path: str) -> List[np.ndarray]:
+        data = np.load(npz_path)
+        vln_embed = data["vlm_embed"]
+        vlm_mask = data["vlm_mask"]
+        byt5_embed = data["byt5_embed"]
+        byt5_mask = data["byt5_mask"]
+        ocr_mask = data["ocr_mask"]
+        return [vln_embed, vlm_mask, byt5_embed, byt5_mask, ocr_mask]
+
+    def cache_batch_outputs(
+        self, tokenize_strategy: TokenizeStrategy, models: List[Any], text_encoding_strategy: TextEncodingStrategy, infos: List
+    ):
+        huyuan_image_text_encoding_strategy: HunyuanImageTextEncodingStrategy = text_encoding_strategy
+        captions = [info.caption for info in infos]
+
+        tokens_and_masks = tokenize_strategy.tokenize(captions)
+        with torch.no_grad():
+            vlm_embed, vlm_mask, byt5_embed, byt5_mask, ocr_mask = huyuan_image_text_encoding_strategy.encode_tokens(
+                tokenize_strategy, models, tokens_and_masks
+            )
+
+        if vlm_embed.dtype == torch.bfloat16:
+            vlm_embed = vlm_embed.float()
+        if byt5_embed.dtype == torch.bfloat16:
+            byt5_embed = byt5_embed.float()
+
+        vlm_embed = vlm_embed.cpu().numpy()
+        vlm_mask = vlm_mask.cpu().numpy()
+        byt5_embed = byt5_embed.cpu().numpy()
+        byt5_mask = byt5_mask.cpu().numpy()
+        ocr_mask = ocr_mask.cpu().numpy()
+
+        for i, info in enumerate(infos):
+            vlm_embed_i = vlm_embed[i]
+            vlm_mask_i = vlm_mask[i]
+            byt5_embed_i = byt5_embed[i]
+            byt5_mask_i = byt5_mask[i]
+            ocr_mask_i = ocr_mask[i]
+
+            if self.cache_to_disk:
+                np.savez(
+                    info.text_encoder_outputs_npz,
+                    vlm_embed=vlm_embed_i,
+                    vlm_mask=vlm_mask_i,
+                    byt5_embed=byt5_embed_i,
+                    byt5_mask=byt5_mask_i,
+                    ocr_mask=ocr_mask_i,
+                )
+            else:
+                info.text_encoder_outputs = (vlm_embed_i, vlm_mask_i, byt5_embed_i, byt5_mask_i, ocr_mask_i)
+
+
+class HunyuanImageLatentsCachingStrategy(LatentsCachingStrategy):
+    HUNYUAN_IMAGE_LATENTS_NPZ_SUFFIX = "_hi.npz"
+
+    def __init__(self, cache_to_disk: bool, batch_size: int, skip_disk_cache_validity_check: bool) -> None:
+        super().__init__(cache_to_disk, batch_size, skip_disk_cache_validity_check)
+
+    @property
+    def cache_suffix(self) -> str:
+        return HunyuanImageLatentsCachingStrategy.HUNYUAN_IMAGE_LATENTS_NPZ_SUFFIX
+
+    def get_latents_npz_path(self, absolute_path: str, image_size: Tuple[int, int]) -> str:
+        return (
+            os.path.splitext(absolute_path)[0]
+            + f"_{image_size[0]:04d}x{image_size[1]:04d}"
+            + HunyuanImageLatentsCachingStrategy.HUNYUAN_IMAGE_LATENTS_NPZ_SUFFIX
+        )
+
+    def is_disk_cached_latents_expected(self, bucket_reso: Tuple[int, int], npz_path: str, flip_aug: bool, alpha_mask: bool):
+        return self._default_is_disk_cached_latents_expected(32, bucket_reso, npz_path, flip_aug, alpha_mask, multi_resolution=True)
+
+    def load_latents_from_disk(
+        self, npz_path: str, bucket_reso: Tuple[int, int]
+    ) -> Tuple[Optional[np.ndarray], Optional[List[int]], Optional[List[int]], Optional[np.ndarray], Optional[np.ndarray]]:
+        return self._default_load_latents_from_disk(32, npz_path, bucket_reso)  # support multi-resolution
+
+    # TODO remove circular dependency for ImageInfo
+    def cache_batch_latents(
+        self, vae: hunyuan_image_vae.HunyuanVAE2D, image_infos: List, flip_aug: bool, alpha_mask: bool, random_crop: bool
+    ):
+        # encode_by_vae = lambda img_tensor: vae.encode(img_tensor).sample()
+        def encode_by_vae(img_tensor):
+            # no_grad is handled in _default_cache_batch_latents
+            nonlocal vae
+            with torch.autocast(device_type=vae.device.type, dtype=vae.dtype):
+                return vae.encode(img_tensor).sample()
+
+        vae_device = vae.device
+        vae_dtype = vae.dtype
+
+        self._default_cache_batch_latents(
+            encode_by_vae, vae_device, vae_dtype, image_infos, flip_aug, alpha_mask, random_crop, multi_resolution=True
+        )
+
+        if not train_util.HIGH_VRAM:
+            train_util.clean_memory_on_device(vae.device)

library/strategy_lumina.py

@@ -0,0 +1,375 @@
+import glob
+import os
+from typing import Any, List, Optional, Tuple, Union
+
+import torch
+from transformers import AutoTokenizer, AutoModel, Gemma2Model, GemmaTokenizerFast
+from library import train_util
+from library.strategy_base import (
+    LatentsCachingStrategy,
+    TokenizeStrategy,
+    TextEncodingStrategy,
+    TextEncoderOutputsCachingStrategy,
+)
+import numpy as np
+from library.utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+GEMMA_ID = "google/gemma-2-2b"
+
+
+class LuminaTokenizeStrategy(TokenizeStrategy):
+    def __init__(
+        self, system_prompt:str, max_length: Optional[int], tokenizer_cache_dir: Optional[str] = None
+    ) -> None:
+        self.tokenizer: GemmaTokenizerFast = AutoTokenizer.from_pretrained(
+            GEMMA_ID, cache_dir=tokenizer_cache_dir
+        )
+        self.tokenizer.padding_side = "right"
+
+        if system_prompt is None:
+            system_prompt = ""
+        system_prompt_special_token = "<Prompt Start>"
+        system_prompt = f"{system_prompt} {system_prompt_special_token} " if system_prompt else ""
+        self.system_prompt = system_prompt
+
+        if max_length is None:
+            self.max_length = 256
+        else:
+            self.max_length = max_length
+
+    def tokenize(
+        self, text: Union[str, List[str]], is_negative: bool = False
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+            text (Union[str, List[str]]): Text to tokenize
+
+        Returns:
+            Tuple[torch.Tensor, torch.Tensor]:
+                token input ids, attention_masks
+        """
+        text = [text] if isinstance(text, str) else text
+        
+        # In training, we always add system prompt (is_negative=False)
+        if not is_negative:
+            # Add system prompt to the beginning of each text
+            text = [self.system_prompt + t for t in text]
+
+        encodings = self.tokenizer(
+            text,
+            max_length=self.max_length,
+            return_tensors="pt",
+            padding="max_length",
+            truncation=True,
+            pad_to_multiple_of=8,
+        )
+        return (encodings.input_ids, encodings.attention_mask)
+
+    def tokenize_with_weights(
+        self, text: str | List[str]
+    ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor]]:
+        """
+        Args:
+            text (Union[str, List[str]]): Text to tokenize
+
+        Returns:
+            Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor]]:
+                token input ids, attention_masks, weights
+        """
+        # Gemma doesn't support weighted prompts, return uniform weights
+        tokens, attention_masks = self.tokenize(text)
+        weights = [torch.ones_like(t) for t in tokens]
+        return tokens, attention_masks, weights
+
+
+class LuminaTextEncodingStrategy(TextEncodingStrategy):
+    def __init__(self) -> None:
+        super().__init__()
+
+    def encode_tokens(
+        self,
+        tokenize_strategy: TokenizeStrategy,
+        models: List[Any],
+        tokens: Tuple[torch.Tensor, torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Args:
+            tokenize_strategy (LuminaTokenizeStrategy): Tokenize strategy
+            models (List[Any]): Text encoders
+            tokens (Tuple[torch.Tensor, torch.Tensor]): tokens, attention_masks
+
+        Returns:
+            Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+                hidden_states, input_ids, attention_masks
+        """
+        text_encoder = models[0]
+        # Check model or torch dynamo OptimizedModule
+        assert isinstance(text_encoder, Gemma2Model) or isinstance(text_encoder._orig_mod, Gemma2Model), f"text encoder is not Gemma2Model {text_encoder.__class__.__name__}"
+        input_ids, attention_masks = tokens
+
+        outputs = text_encoder(
+            input_ids=input_ids.to(text_encoder.device),
+            attention_mask=attention_masks.to(text_encoder.device),
+            output_hidden_states=True,
+            return_dict=True,
+        )
+
+        return outputs.hidden_states[-2], input_ids, attention_masks
+
+    def encode_tokens_with_weights(
+        self,
+        tokenize_strategy: TokenizeStrategy,
+        models: List[Any],
+        tokens: Tuple[torch.Tensor, torch.Tensor],
+        weights: List[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Args:
+            tokenize_strategy (LuminaTokenizeStrategy): Tokenize strategy
+            models (List[Any]): Text encoders
+            tokens (Tuple[torch.Tensor, torch.Tensor]): tokens, attention_masks
+            weights_list (List[torch.Tensor]): Currently unused
+
+        Returns:
+            Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+                hidden_states, input_ids, attention_masks
+        """
+        # For simplicity, use uniform weighting
+        return self.encode_tokens(tokenize_strategy, models, tokens)
+
+
+class LuminaTextEncoderOutputsCachingStrategy(TextEncoderOutputsCachingStrategy):
+    LUMINA_TEXT_ENCODER_OUTPUTS_NPZ_SUFFIX = "_lumina_te.npz"
+
+    def __init__(
+        self,
+        cache_to_disk: bool,
+        batch_size: int,
+        skip_disk_cache_validity_check: bool,
+        is_partial: bool = False,
+    ) -> None:
+        super().__init__(
+            cache_to_disk,
+            batch_size,
+            skip_disk_cache_validity_check,
+            is_partial,
+        )
+
+    def get_outputs_npz_path(self, image_abs_path: str) -> str:
+        return (
+            os.path.splitext(image_abs_path)[0]
+            + LuminaTextEncoderOutputsCachingStrategy.LUMINA_TEXT_ENCODER_OUTPUTS_NPZ_SUFFIX
+        )
+
+    def is_disk_cached_outputs_expected(self, npz_path: str) -> bool:
+        """
+        Args:
+            npz_path (str): Path to the npz file.
+
+        Returns:
+            bool: True if the npz file is expected to be cached.
+        """
+        if not self.cache_to_disk:
+            return False
+        if not os.path.exists(npz_path):
+            return False
+        if self.skip_disk_cache_validity_check:
+            return True
+
+        try:
+            npz = np.load(npz_path)
+            if "hidden_state" not in npz:
+                return False
+            if "attention_mask" not in npz:
+                return False
+            if "input_ids" not in npz:
+                return False
+        except Exception as e:
+            logger.error(f"Error loading file: {npz_path}")
+            raise e
+
+        return True
+
+    def load_outputs_npz(self, npz_path: str) -> List[np.ndarray]:
+        """
+        Load outputs from a npz file
+
+        Returns:
+            List[np.ndarray]: hidden_state, input_ids, attention_mask
+        """
+        data = np.load(npz_path)
+        hidden_state = data["hidden_state"]
+        attention_mask = data["attention_mask"]
+        input_ids = data["input_ids"]
+        return [hidden_state, input_ids, attention_mask]
+
+    @torch.no_grad()
+    def cache_batch_outputs(
+        self,
+        tokenize_strategy: TokenizeStrategy,
+        models: List[Any],
+        text_encoding_strategy: TextEncodingStrategy,
+        batch: List[train_util.ImageInfo],
+    ) -> None:
+        """
+        Args:
+            tokenize_strategy (LuminaTokenizeStrategy): Tokenize strategy
+            models (List[Any]): Text encoders
+            text_encoding_strategy (LuminaTextEncodingStrategy):
+            infos (List): List of ImageInfo
+
+        Returns:
+            None
+        """
+        assert isinstance(text_encoding_strategy, LuminaTextEncodingStrategy)
+        assert isinstance(tokenize_strategy, LuminaTokenizeStrategy)
+
+        captions = [info.caption for info in batch]
+
+        if self.is_weighted:
+            tokens, attention_masks, weights_list = (
+                tokenize_strategy.tokenize_with_weights(captions)
+            )
+            hidden_state, input_ids, attention_masks = (
+                text_encoding_strategy.encode_tokens_with_weights(
+                    tokenize_strategy,
+                    models,
+                    (tokens, attention_masks),
+                    weights_list,
+                )
+            )
+        else:
+            tokens = tokenize_strategy.tokenize(captions)
+            hidden_state, input_ids, attention_masks = (
+                text_encoding_strategy.encode_tokens(
+                    tokenize_strategy, models, tokens
+                )
+            )
+
+        if hidden_state.dtype != torch.float32:
+            hidden_state = hidden_state.float()
+
+        hidden_state = hidden_state.cpu().numpy()
+        attention_mask = attention_masks.cpu().numpy() # (B, S)
+        input_ids = input_ids.cpu().numpy() # (B, S) 
+
+
+        for i, info in enumerate(batch):
+            hidden_state_i = hidden_state[i]
+            attention_mask_i = attention_mask[i]
+            input_ids_i = input_ids[i]
+
+            if self.cache_to_disk:
+                assert info.text_encoder_outputs_npz is not None, f"Text encoder cache outputs to disk not found for image {info.image_key}"
+                np.savez(
+                    info.text_encoder_outputs_npz,
+                    hidden_state=hidden_state_i,
+                    attention_mask=attention_mask_i,
+                    input_ids=input_ids_i,
+                )
+            else:
+                info.text_encoder_outputs = [
+                    hidden_state_i,
+                    input_ids_i,
+                    attention_mask_i,
+                ]
+
+
+class LuminaLatentsCachingStrategy(LatentsCachingStrategy):
+    LUMINA_LATENTS_NPZ_SUFFIX = "_lumina.npz"
+
+    def __init__(
+        self, cache_to_disk: bool, batch_size: int, skip_disk_cache_validity_check: bool
+    ) -> None:
+        super().__init__(cache_to_disk, batch_size, skip_disk_cache_validity_check)
+
+    @property
+    def cache_suffix(self) -> str:
+        return LuminaLatentsCachingStrategy.LUMINA_LATENTS_NPZ_SUFFIX
+
+    def get_latents_npz_path(
+        self, absolute_path: str, image_size: Tuple[int, int]
+    ) -> str:
+        return (
+            os.path.splitext(absolute_path)[0]
+            + f"_{image_size[0]:04d}x{image_size[1]:04d}"
+            + LuminaLatentsCachingStrategy.LUMINA_LATENTS_NPZ_SUFFIX
+        )
+
+    def is_disk_cached_latents_expected(
+        self,
+        bucket_reso: Tuple[int, int],
+        npz_path: str,
+        flip_aug: bool,
+        alpha_mask: bool,
+    ) -> bool:
+        """
+        Args:
+            bucket_reso (Tuple[int, int]): The resolution of the bucket.
+            npz_path (str): Path to the npz file.
+            flip_aug (bool): Whether to flip the image.
+            alpha_mask (bool): Whether to apply
+        """
+        return self._default_is_disk_cached_latents_expected(
+            8, bucket_reso, npz_path, flip_aug, alpha_mask, multi_resolution=True
+        )
+
+    def load_latents_from_disk(
+        self, npz_path: str, bucket_reso: Tuple[int, int]
+    ) -> Tuple[
+        Optional[np.ndarray],
+        Optional[List[int]],
+        Optional[List[int]],
+        Optional[np.ndarray],
+        Optional[np.ndarray],
+    ]:
+        """
+        Args:
+            npz_path (str): Path to the npz file.
+            bucket_reso (Tuple[int, int]): The resolution of the bucket.
+
+        Returns:
+            Tuple[
+                Optional[np.ndarray],
+                Optional[List[int]],
+                Optional[List[int]],
+                Optional[np.ndarray],
+                Optional[np.ndarray],
+            ]: Tuple of latent tensors, attention_mask, input_ids, latents, latents_unet
+        """
+        return self._default_load_latents_from_disk(
+            8, npz_path, bucket_reso
+        )  # support multi-resolution
+
+    # TODO remove circular dependency for ImageInfo
+    def cache_batch_latents(
+        self,
+        model,
+        batch: List,
+        flip_aug: bool,
+        alpha_mask: bool,
+        random_crop: bool,
+    ):
+        encode_by_vae = lambda img_tensor: model.encode(img_tensor).to("cpu")
+        vae_device = model.device
+        vae_dtype = model.dtype
+
+        self._default_cache_batch_latents(
+            encode_by_vae,
+            vae_device,
+            vae_dtype,
+            batch,
+            flip_aug,
+            alpha_mask,
+            random_crop,
+            multi_resolution=True,
+        )
+
+        if not train_util.HIGH_VRAM:
+            train_util.clean_memory_on_device(model.device)