sd3 training

README.md

@@ -1,5 +1,30 @@
 This repository contains training, generation and utility scripts for Stable Diffusion.
 
+## SD3 training
+
+SD3 training is done with `sd3_train.py`. 
+
+`optimizer_type = "adafactor"` is recommended for 24GB VRAM GPUs. `cache_text_encoder_outputs_to_disk` and `cache_latents_to_disk` are necessary currently. 
+
+`clip_l`, `clip_g` and `t5xxl` can be specified if the checkpoint does not include them.  
+
+t5xxl doesn't seem to work with `fp16`, so use`bf16` or `fp32`. 
+
+There are `t5xxl_device` and `t5xxl_dtype` options for `t5xxl` device and dtype. 
+
+```toml
+learning_rate = 1e-5 # seems to be too high
+optimizer_type = "adafactor"
+optimizer_args = [ "scale_parameter=False", "relative_step=False", "warmup_init=False" ]
+cache_text_encoder_outputs = true
+cache_text_encoder_outputs_to_disk = true
+vae_batch_size = 1
+cache_latents = true
+cache_latents_to_disk = true
+```
+
+--- 
+
 [__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。
 

library/sai_model_spec.py

@@ -6,8 +6,10 @@ import os
 from typing import List, Optional, Tuple, Union
 import safetensors
 from library.utils import setup_logging
+
 setup_logging()
 import logging
+
 logger = logging.getLogger(__name__)
 
 r"""
@@ -55,11 +57,14 @@ 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-medium"
+ARCH_SD3_UNKNOWN = "stable-diffusion-3"
 
 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"
 
 PRED_TYPE_EPSILON = "epsilon"
@@ -113,7 +118,11 @@ def build_metadata(
     merged_from: Optional[str] = None,
     timesteps: Optional[Tuple[int, int]] = None,
     clip_skip: Optional[int] = None,
+    sd3: str = None,
 ):
+    """
+    sd3: only supports "m"
+    """
     # if state_dict is None, hash is not calculated
 
     metadata = {}
@@ -126,6 +135,11 @@ def build_metadata(
 
     if sdxl:
         arch = ARCH_SD_XL_V1_BASE
+    elif sd3 is not None:
+        if sd3 == "m":
+            arch = ARCH_SD3_M
+        else:
+            arch = ARCH_SD3_UNKNOWN
     elif v2:
         if v_parameterization:
             arch = ARCH_SD_V2_768_V
@@ -142,7 +156,7 @@ def build_metadata(
     metadata["modelspec.architecture"] = arch
 
     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
@@ -236,7 +250,7 @@ def build_metadata(
     # assert all([v is not None for v in metadata.values()]), metadata
     if not all([v is not None for v in metadata.values()]):
         logger.error(f"Internal error: some metadata values are None: {metadata}")
-    
+
     return metadata
 
 
@@ -250,7 +264,7 @@ def get_title(metadata: dict) -> Optional[str]:
 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:

library/sd3_models.py

@@ -1,11 +1,13 @@
-# some modules/classes are copied and modified from https://github.com/mcmonkey4eva/sd3-ref 
+# some modules/classes are copied and modified from https://github.com/mcmonkey4eva/sd3-ref
 # the original code is licensed under the MIT License
 
 # and some module/classes are contributed from KohakuBlueleaf. Thanks for the contribution!
 
+from ast import Tuple
 from functools import partial
 import math
-from typing import Dict, Optional
+from types import SimpleNamespace
+from typing import Dict, List, Optional, Union
 import einops
 import numpy as np
 import torch
@@ -106,6 +108,8 @@ class SD3Tokenizer:
         self.clip_l = SDTokenizer(tokenizer=clip_tokenizer)
         self.clip_g = SDXLClipGTokenizer(clip_tokenizer)
         self.t5xxl = T5XXLTokenizer() if t5xxl else None
+        # t5xxl has 99999999 max length, clip has 77
+        self.model_max_length = self.clip_l.max_length  # 77
 
     def tokenize_with_weights(self, text: str):
         return (
@@ -870,6 +874,10 @@ class MMDiT(nn.Module):
         self.final_layer = UnPatch(self.hidden_size, patch_size, self.out_channels)
         # self.initialize_weights()
 
+    @property
+    def model_type(self):
+        return "m"  # only support medium
+
     def enable_gradient_checkpointing(self):
         self.gradient_checkpointing = True
         for block in self.joint_blocks:
@@ -1013,6 +1021,10 @@ def create_mmdit_sd3_medium_configs(attn_mode: str):
 # endregion
 
 # region VAE
+# TODO support xformers
+
+VAE_SCALE_FACTOR = 1.5305
+VAE_SHIFT_FACTOR = 0.0609
 
 
 def Normalize(in_channels, num_groups=32, dtype=torch.float32, device=None):
@@ -1222,6 +1234,14 @@ class SDVAE(torch.nn.Module):
         self.encoder = VAEEncoder(dtype=dtype, device=device)
         self.decoder = VAEDecoder(dtype=dtype, device=device)
 
+    @property
+    def device(self):
+        return next(self.parameters()).device
+
+    @property
+    def dtype(self):
+        return next(self.parameters()).dtype
+
     @torch.autocast("cuda", dtype=torch.float16)
     def decode(self, latent):
         return self.decoder(latent)
@@ -1234,6 +1254,43 @@ class SDVAE(torch.nn.Module):
         std = torch.exp(0.5 * logvar)
         return mean + std * torch.randn_like(mean)
 
+    @staticmethod
+    def process_in(latent):
+        return (latent - VAE_SHIFT_FACTOR) * VAE_SCALE_FACTOR
+
+    @staticmethod
+    def process_out(latent):
+        return (latent / VAE_SCALE_FACTOR) + VAE_SHIFT_FACTOR
+
+
+class VAEOutput:
+    def __init__(self, latent):
+        self.latent = latent
+
+    @property
+    def latent_dist(self):
+        return self
+
+    def sample(self):
+        return self.latent
+
+
+class VAEWrapper:
+    def __init__(self, vae):
+        self.vae = vae
+
+    @property
+    def device(self):
+        return self.vae.device
+
+    @property
+    def dtype(self):
+        return self.vae.dtype
+
+    # latents = vae.encode(img_tensors).latent_dist.sample().to("cpu")
+    def encode(self, image):
+        return VAEOutput(self.vae.encode(image))
+
 
 # endregion
 
@@ -1370,15 +1427,39 @@ class CLIPTextModel(torch.nn.Module):
 
 
 class ClipTokenWeightEncoder:
-    def encode_token_weights(self, token_weight_pairs):
-        tokens = list(map(lambda a: a[0], token_weight_pairs[0]))
-        out, pooled = self([tokens])
-        if pooled is not None:
-            first_pooled = pooled[0:1].cpu()
+    # def encode_token_weights(self, token_weight_pairs):
+    #     tokens = list(map(lambda a: a[0], token_weight_pairs[0]))
+    #     out, pooled = self([tokens])
+    #     if pooled is not None:
+    #         first_pooled = pooled[0:1]
+    #     else:
+    #         first_pooled = pooled
+    #     output = [out[0:1]]
+    #     return torch.cat(output, dim=-2), first_pooled
+
+    # fix to support batched inputs
+    # : Union[List[Tuple[torch.Tensor, torch.Tensor]], List[List[Tuple[torch.Tensor, torch.Tensor]]]]
+    def encode_token_weights(self, list_of_token_weight_pairs):
+        has_batch = isinstance(list_of_token_weight_pairs[0][0], list)
+
+        if has_batch:
+            list_of_tokens = []
+            for pairs in list_of_token_weight_pairs:
+                tokens = [a[0] for a in pairs[0]]  # I'm not sure why this is [0]
+                list_of_tokens.append(tokens)
         else:
-            first_pooled = pooled
-        output = [out[0:1]]
-        return torch.cat(output, dim=-2).cpu(), first_pooled
+            list_of_tokens = [[a[0] for a in list_of_token_weight_pairs[0]]]
+
+        out, pooled = self(list_of_tokens)
+        if has_batch:
+            return out, pooled
+        else:
+            if pooled is not None:
+                first_pooled = pooled[0:1]
+            else:
+                first_pooled = pooled
+            output = [out[0:1]]
+            return torch.cat(output, dim=-2), first_pooled
 
 
 class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
@@ -1694,6 +1775,7 @@ class T5Stack(torch.nn.Module):
         x = self.embed_tokens(input_ids)
         past_bias = None
         for i, l in enumerate(self.block):
+            # uncomment to debug layerwise output: fp16 may cause issues
             # print(i, x.mean(), x.std())
             x, past_bias = l(x, past_bias)
             if i == intermediate_output:

library/sd3_train_utils.py

@@ -0,0 +1,544 @@
+import argparse
+import math
+import os
+from typing import Optional, Tuple
+
+import torch
+from safetensors.torch import save_file
+
+from library import sd3_models, sd3_utils, train_util
+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__)
+
+from .sdxl_train_util import match_mixed_precision
+
+
+def load_target_model(args, accelerator, attn_mode, weight_dtype, t5xxl_device, t5xxl_dtype) -> Tuple[
+    sd3_models.MMDiT,
+    Optional[sd3_models.SDClipModel],
+    Optional[sd3_models.SDXLClipG],
+    Optional[sd3_models.T5XXLModel],
+    sd3_models.SDVAE,
+]:
+    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:
+            logger.info(f"loading model for process {accelerator.state.local_process_index}/{accelerator.state.num_processes}")
+
+            mmdit, clip_l, clip_g, t5xxl, vae = sd3_utils.load_models(
+                args.pretrained_model_name_or_path,
+                args.clip_l,
+                args.clip_g,
+                args.t5xxl,
+                args.vae,
+                attn_mode,
+                accelerator.device if args.lowram else "cpu",
+                weight_dtype,
+                args.disable_mmap_load_safetensors,
+                t5xxl_device,
+                t5xxl_dtype,
+            )
+
+            # work on low-ram device
+            if args.lowram:
+                if clip_l is not None:
+                    clip_l.to(accelerator.device)
+                if clip_g is not None:
+                    clip_g.to(accelerator.device)
+                if t5xxl is not None:
+                    t5xxl.to(accelerator.device)
+                vae.to(accelerator.device)
+                mmdit.to(accelerator.device)
+
+            clean_memory_on_device(accelerator.device)
+        accelerator.wait_for_everyone()
+
+    return mmdit, clip_l, clip_g, t5xxl, vae
+
+
+def save_models(
+    ckpt_path: str,
+    mmdit: sd3_models.MMDiT,
+    vae: sd3_models.SDVAE,
+    clip_l: sd3_models.SDClipModel,
+    clip_g: sd3_models.SDXLClipG,
+    t5xxl: Optional[sd3_models.T5XXLModel],
+    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())
+
+    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)
+
+
+def save_sd3_model_on_train_end(
+    args: argparse.Namespace,
+    save_dtype: torch.dtype,
+    epoch: int,
+    global_step: int,
+    clip_l: sd3_models.SDClipModel,
+    clip_g: sd3_models.SDXLClipG,
+    t5xxl: Optional[sd3_models.T5XXLModel],
+    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: sd3_models.SDClipModel,
+    clip_g: sd3_models.SDXLClipG,
+    t5xxl: Optional[sd3_models.T5XXLModel],
+    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(
+        "--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(
+        "--disable_mmap_load_safetensors",
+        action="store_true",
+        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",
+    )
+
+    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="save CLIP models to checkpoint / CLIPモデルをチェックポイントに保存する"
+    )
+    parser.add_argument(
+        "--save_t5xxl", action="store_true", help="save T5-XXL model to checkpoint / T5-XXLモデルをチェックポイントに保存する"
+    )
+
+    parser.add_argument(
+        "--t5xxl_device",
+        type=str,
+        default=None,
+        help="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="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",
+    )
+
+    # copy from Diffusers
+    parser.add_argument(
+        "--weighting_scheme",
+        type=str,
+        default="logit_normal",
+        choices=["sigma_sqrt", "logit_normal", "mode", "cosmap"],
+    )
+    parser.add_argument(
+        "--logit_mean", type=float, default=0.0, help="mean to use when using the `'logit_normal'` weighting scheme."
+    )
+    parser.add_argument("--logit_std", type=float, default=1.0, help="std to use when using the `'logit_normal'` weighting scheme.")
+    parser.add_argument(
+        "--mode_scale",
+        type=float,
+        default=1.29,
+        help="Scale of mode weighting scheme. Only effective when using the `'mode'` as the `weighting_scheme`.",
+    )
+
+
+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"
+            )
+
+
+def sample_images(*args, **kwargs):
+    return train_util.sample_images_common(SdxlStableDiffusionLongPromptWeightingPipeline, *args, **kwargs)
+
+
+# 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
+
+
+# endregion

library/sd3_utils.py

@@ -1,30 +1,226 @@
 import math
-from typing import Dict
+from typing import Dict, Optional, Union
 import torch
+import safetensors
+from safetensors.torch import load_file
+from accelerate import init_empty_weights
+
+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
+
+
+def load_models(
+    ckpt_path: str,
+    clip_l_path: str,
+    clip_g_path: str,
+    t5xxl_path: str,
+    vae_path: str,
+    attn_mode: str,
+    device: Union[str, torch.device],
+    weight_dtype: torch.dtype,
+    disable_mmap: bool = False,
+    t5xxl_device: Optional[str] = None,
+    t5xxl_dtype: Optional[str] = None,
+):
+    def load_state_dict(path: str, dvc: Union[str, torch.device] = device):
+        if disable_mmap:
+            return safetensors.torch.load(open(path, "rb").read())
+        else:
+            try:
+                return load_file(path, device=dvc)
+            except:
+                return load_file(path)  # prevent device invalid Error
+
+    t5xxl_device = t5xxl_device or device
+
+    logger.info(f"Loading SD3 models from {ckpt_path}...")
+    state_dict = load_state_dict(ckpt_path)
+
+    # load clip_l
+    clip_l_sd = None
+    if clip_l_path:
+        logger.info(f"Loading clip_l from {clip_l_path}...")
+        clip_l_sd = load_state_dict(clip_l_path)
+        for key in list(clip_l_sd.keys()):
+            clip_l_sd["transformer." + key] = clip_l_sd.pop(key)
+    else:
+        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)
+
+    # load clip_g
+    clip_g_sd = None
+    if clip_g_path:
+        logger.info(f"Loading clip_g from {clip_g_path}...")
+        clip_g_sd = load_state_dict(clip_g_path)
+        for key in list(clip_g_sd.keys()):
+            clip_g_sd["transformer." + key] = clip_g_sd.pop(key)
+    else:
+        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)
+
+    # load t5xxl
+    t5xxl_sd = None
+    if t5xxl_path:
+        logger.info(f"Loading t5xxl from {t5xxl_path}...")
+        t5xxl_sd = load_state_dict(t5xxl_path, t5xxl_device)
+        for key in list(t5xxl_sd.keys()):
+            t5xxl_sd["transformer." + key] = t5xxl_sd.pop(key)
+    else:
+        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)
+
+    # MMDiT and VAE
+    vae_sd = {}
+    if vae_path:
+        logger.info(f"Loading VAE from {vae_path}...")
+        vae_sd = load_state_dict(vae_path)
+    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)
+
+    mmdit_prefix = "model.diffusion_model."
+    for k in list(state_dict.keys()):
+        if k.startswith(mmdit_prefix):
+            state_dict[k[len(mmdit_prefix) :]] = state_dict.pop(k)
+        else:
+            state_dict.pop(k)  # remove other keys
+
+    # load MMDiT
+    logger.info("Building MMDit")
+    with init_empty_weights():
+        mmdit = sd3_models.create_mmdit_sd3_medium_configs(attn_mode)
+
+    logger.info("Loading state dict...")
+    info = sdxl_model_util._load_state_dict_on_device(mmdit, state_dict, device, weight_dtype)
+    logger.info(f"Loaded MMDiT: {info}")
+
+    # load ClipG and ClipL
+    if clip_l_sd is None:
+        clip_l = None
+    else:
+        logger.info("Building ClipL")
+        clip_l = sd3_models.create_clip_l(device, weight_dtype, clip_l_sd)
+        logger.info("Loading state dict...")
+        info = clip_l.load_state_dict(clip_l_sd)
+        logger.info(f"Loaded ClipL: {info}")
+        clip_l.set_attn_mode(attn_mode)
+
+    if clip_g_sd is None:
+        clip_g = None
+    else:
+        logger.info("Building ClipG")
+        clip_g = sd3_models.create_clip_g(device, weight_dtype, clip_g_sd)
+        logger.info("Loading state dict...")
+        info = clip_g.load_state_dict(clip_g_sd)
+        logger.info(f"Loaded ClipG: {info}")
+        clip_g.set_attn_mode(attn_mode)
+
+    # load T5XXL
+    if t5xxl_sd is None:
+        t5xxl = None
+    else:
+        logger.info("Building T5XXL")
+        t5xxl = sd3_models.create_t5xxl(t5xxl_device, t5xxl_dtype, t5xxl_sd)
+        logger.info("Loading state dict...")
+        info = t5xxl.load_state_dict(t5xxl_sd)
+        logger.info(f"Loaded T5XXL: {info}")
+        t5xxl.set_attn_mode(attn_mode)
+
+    # load VAE
+    logger.info("Building VAE")
+    vae = sd3_models.SDVAE()
+    logger.info("Loading state dict...")
+    info = vae.load_state_dict(vae_sd)
+    logger.info(f"Loaded VAE: {info}")
+
+    return mmdit, clip_l, clip_g, t5xxl, vae
+
+
+# endregion
+# region utils
+
 
 def get_cond(
     prompt: str,
     tokenizer: sd3_models.SD3Tokenizer,
     clip_l: sd3_models.SDClipModel,
     clip_g: sd3_models.SDXLClipG,
-    t5xxl: sd3_models.T5XXLModel,
+    t5xxl: Optional[sd3_models.T5XXLModel] = None,
+    device: Optional[torch.device] = None,
+    dtype: Optional[torch.dtype] = None,
 ):
     l_tokens, g_tokens, t5_tokens = tokenizer.tokenize_with_weights(prompt)
+    return get_cond_from_tokens(l_tokens, g_tokens, t5_tokens, clip_l, clip_g, t5xxl, device=device, dtype=dtype)
+
+
+def get_cond_from_tokens(
+    l_tokens,
+    g_tokens,
+    t5_tokens,
+    clip_l: sd3_models.SDClipModel,
+    clip_g: sd3_models.SDXLClipG,
+    t5xxl: Optional[sd3_models.T5XXLModel] = None,
+    device: Optional[torch.device] = None,
+    dtype: Optional[torch.dtype] = None,
+):
     l_out, l_pooled = clip_l.encode_token_weights(l_tokens)
     g_out, g_pooled = clip_g.encode_token_weights(g_tokens)
     lg_out = torch.cat([l_out, g_out], dim=-1)
     lg_out = torch.nn.functional.pad(lg_out, (0, 4096 - lg_out.shape[-1]))
+    if device is not None:
+        lg_out = lg_out.to(device=device)
+        l_pooled = l_pooled.to(device=device)
+        g_pooled = g_pooled.to(device=device)
+    if dtype is not None:
+        lg_out = lg_out.to(dtype=dtype)
+        l_pooled = l_pooled.to(dtype=dtype)
+        g_pooled = g_pooled.to(dtype=dtype)
 
+    # t5xxl may be in another device (eg. cpu)
     if t5_tokens is None:
-        t5_out = torch.zeros((lg_out.shape[0], 77, 4096), device=lg_out.device)
+        t5_out = torch.zeros((lg_out.shape[0], 77, 4096), device=lg_out.device, dtype=lg_out.dtype)
     else:
-        t5_out, t5_pooled = t5xxl.encode_token_weights(t5_tokens)  # t5_out is [1, 77, 4096], t5_pooled is None
-        t5_out = t5_out.to(lg_out.dtype)
+        t5_out, _ = t5xxl.encode_token_weights(t5_tokens)  # t5_out is [1, 77, 4096], t5_pooled is None
+        if device is not None:
+            t5_out = t5_out.to(device=device)
+        if dtype is not None:
+            t5_out = t5_out.to(dtype=dtype)
 
-    return torch.cat([lg_out, t5_out], dim=-2), torch.cat((l_pooled, g_pooled), dim=-1)
+    # return torch.cat([lg_out, t5_out], dim=-2), torch.cat((l_pooled, g_pooled), dim=-1)
+    return lg_out, t5_out, torch.cat((l_pooled, g_pooled), dim=-1)
 
 
 # used if other sd3 models is available
@@ -111,3 +307,6 @@ class ModelSamplingDiscreteFlow:
         # 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
+
+
+# endregion

library/train_util.py

@@ -58,7 +58,7 @@ from diffusers import (
     KDPM2AncestralDiscreteScheduler,
     AutoencoderKL,
 )
-from library import custom_train_functions
+from library import custom_train_functions, sd3_utils
 from library.original_unet import UNet2DConditionModel
 from huggingface_hub import hf_hub_download
 import numpy as np
@@ -135,6 +135,7 @@ IMAGE_TRANSFORMS = transforms.Compose(
 )
 
 TEXT_ENCODER_OUTPUTS_CACHE_SUFFIX = "_te_outputs.npz"
+TEXT_ENCODER_OUTPUTS_CACHE_SUFFIX_SD3 = "_sd3_te.npz"
 
 
 class ImageInfo:
@@ -985,7 +986,7 @@ class BaseDataset(torch.utils.data.Dataset):
             ]
         )
 
-    def cache_latents(self, vae, vae_batch_size=1, cache_to_disk=False, is_main_process=True):
+    def cache_latents(self, vae, vae_batch_size=1, cache_to_disk=False, is_main_process=True, file_suffix=".npz"):
         # マルチGPUには対応していないので、そちらはtools/cache_latents.pyを使うこと
         logger.info("caching latents.")
 
@@ -1006,7 +1007,7 @@ class BaseDataset(torch.utils.data.Dataset):
 
             # check disk cache exists and size of latents
             if cache_to_disk:
-                info.latents_npz = os.path.splitext(info.absolute_path)[0] + ".npz"
+                info.latents_npz = os.path.splitext(info.absolute_path)[0] + file_suffix
                 if not is_main_process:  # store to info only
                     continue
 
@@ -1040,14 +1041,43 @@ class BaseDataset(torch.utils.data.Dataset):
         for batch in tqdm(batches, smoothing=1, total=len(batches)):
             cache_batch_latents(vae, cache_to_disk, batch, subset.flip_aug, subset.alpha_mask, subset.random_crop)
 
-    # weight_dtypeを指定するとText Encoderそのもの、およひ出力がweight_dtypeになる
-    # SDXLでのみ有効だが、datasetのメソッドとする必要があるので、sdxl_train_util.pyではなくこちらに実装する
-    # SD1/2に対応するにはv2のフラグを持つ必要があるので後回し
+    # if weight_dtype is specified, Text Encoder itself and output will be converted to the dtype
+    # this method is only for SDXL, but it should be implemented here because it needs to be a method of dataset
+    # to support SD1/2, it needs a flag for v2, but it is postponed
     def cache_text_encoder_outputs(
-        self, tokenizers, text_encoders, device, weight_dtype, cache_to_disk=False, is_main_process=True
+        self, tokenizers, text_encoders, device, output_dtype, cache_to_disk=False, is_main_process=True
     ):
         assert len(tokenizers) == 2, "only support SDXL"
+        return self.cache_text_encoder_outputs_common(
+            tokenizers, text_encoders, [device, device], output_dtype, [output_dtype], cache_to_disk, is_main_process
+        )
 
+    # same as above, but for SD3
+    def cache_text_encoder_outputs_sd3(
+        self, tokenizer, text_encoders, devices, output_dtype, te_dtypes, cache_to_disk=False, is_main_process=True
+    ):
+        return self.cache_text_encoder_outputs_common(
+            [tokenizer],
+            text_encoders,
+            devices,
+            output_dtype,
+            te_dtypes,
+            cache_to_disk,
+            is_main_process,
+            TEXT_ENCODER_OUTPUTS_CACHE_SUFFIX_SD3,
+        )
+
+    def cache_text_encoder_outputs_common(
+        self,
+        tokenizers,
+        text_encoders,
+        devices,
+        output_dtype,
+        te_dtypes,
+        cache_to_disk=False,
+        is_main_process=True,
+        file_suffix=TEXT_ENCODER_OUTPUTS_CACHE_SUFFIX,
+    ):
         # latentsのキャッシュと同様に、ディスクへのキャッシュに対応する
         # またマルチGPUには対応していないので、そちらはtools/cache_latents.pyを使うこと
         logger.info("caching text encoder outputs.")
@@ -1058,13 +1088,14 @@ class BaseDataset(torch.utils.data.Dataset):
         for info in tqdm(image_infos):
             # subset = self.image_to_subset[info.image_key]
             if cache_to_disk:
-                te_out_npz = os.path.splitext(info.absolute_path)[0] + TEXT_ENCODER_OUTPUTS_CACHE_SUFFIX
+                te_out_npz = os.path.splitext(info.absolute_path)[0] + file_suffix
                 info.text_encoder_outputs_npz = te_out_npz
 
                 if not is_main_process:  # store to info only
                     continue
 
                 if os.path.exists(te_out_npz):
+                    # TODO check varidity of cache here
                     continue
 
             image_infos_to_cache.append(info)
@@ -1073,18 +1104,23 @@ class BaseDataset(torch.utils.data.Dataset):
             return
 
         # prepare tokenizers and text encoders
-        for text_encoder in text_encoders:
+        for text_encoder, device, te_dtype in zip(text_encoders, devices, te_dtypes):
             text_encoder.to(device)
-            if weight_dtype is not None:
-                text_encoder.to(dtype=weight_dtype)
+            if te_dtype is not None:
+                text_encoder.to(dtype=te_dtype)
 
         # create batch
+        is_sd3 = len(tokenizers) == 1
         batch = []
         batches = []
         for info in image_infos_to_cache:
-            input_ids1 = self.get_input_ids(info.caption, tokenizers[0])
-            input_ids2 = self.get_input_ids(info.caption, tokenizers[1])
-            batch.append((info, input_ids1, input_ids2))
+            if not is_sd3:
+                input_ids1 = self.get_input_ids(info.caption, tokenizers[0])
+                input_ids2 = self.get_input_ids(info.caption, tokenizers[1])
+                batch.append((info, input_ids1, input_ids2))
+            else:
+                l_tokens, g_tokens, t5_tokens = tokenizers[0].tokenize_with_weights(info.caption)
+                batch.append((info, l_tokens, g_tokens, t5_tokens))
 
             if len(batch) >= self.batch_size:
                 batches.append(batch)
@@ -1095,13 +1131,32 @@ class BaseDataset(torch.utils.data.Dataset):
 
         # iterate batches: call text encoder and cache outputs for memory or disk
         logger.info("caching text encoder outputs...")
-        for batch in tqdm(batches):
-            infos, input_ids1, input_ids2 = zip(*batch)
-            input_ids1 = torch.stack(input_ids1, dim=0)
-            input_ids2 = torch.stack(input_ids2, dim=0)
-            cache_batch_text_encoder_outputs(
-                infos, tokenizers, text_encoders, self.max_token_length, cache_to_disk, input_ids1, input_ids2, weight_dtype
-            )
+        if not is_sd3:
+            for batch in tqdm(batches):
+                infos, input_ids1, input_ids2 = zip(*batch)
+                input_ids1 = torch.stack(input_ids1, dim=0)
+                input_ids2 = torch.stack(input_ids2, dim=0)
+                cache_batch_text_encoder_outputs(
+                    infos, tokenizers, text_encoders, self.max_token_length, cache_to_disk, input_ids1, input_ids2, output_dtype
+                )
+        else:
+            for batch in tqdm(batches):
+                infos, l_tokens, g_tokens, t5_tokens = zip(*batch)
+
+                # stack tokens
+                # l_tokens = [tokens[0] for tokens in l_tokens]
+                # g_tokens = [tokens[0] for tokens in g_tokens]
+                # t5_tokens = [tokens[0] for tokens in t5_tokens]
+
+                cache_batch_text_encoder_outputs_sd3(
+                    infos,
+                    tokenizers[0],
+                    text_encoders,
+                    self.max_token_length,
+                    cache_to_disk,
+                    (l_tokens, g_tokens, t5_tokens),
+                    output_dtype,
+                )
 
     def get_image_size(self, image_path):
         return imagesize.get(image_path)
@@ -1332,6 +1387,7 @@ class BaseDataset(torch.utils.data.Dataset):
                     captions.append(caption)
 
                 if not self.token_padding_disabled:  # this option might be omitted in future
+                    # TODO get_input_ids must support SD3
                     if self.XTI_layers:
                         token_caption = self.get_input_ids(caption_layer, self.tokenizers[0])
                     else:
@@ -2140,10 +2196,10 @@ class DatasetGroup(torch.utils.data.ConcatDataset):
         for dataset in self.datasets:
             dataset.enable_XTI(*args, **kwargs)
 
-    def cache_latents(self, vae, vae_batch_size=1, cache_to_disk=False, is_main_process=True):
+    def cache_latents(self, vae, vae_batch_size=1, cache_to_disk=False, is_main_process=True, file_suffix=".npz"):
         for i, dataset in enumerate(self.datasets):
             logger.info(f"[Dataset {i}]")
-            dataset.cache_latents(vae, vae_batch_size, cache_to_disk, is_main_process)
+            dataset.cache_latents(vae, vae_batch_size, cache_to_disk, is_main_process, file_suffix)
 
     def cache_text_encoder_outputs(
         self, tokenizers, text_encoders, device, weight_dtype, cache_to_disk=False, is_main_process=True
@@ -2152,6 +2208,15 @@ class DatasetGroup(torch.utils.data.ConcatDataset):
             logger.info(f"[Dataset {i}]")
             dataset.cache_text_encoder_outputs(tokenizers, text_encoders, device, weight_dtype, cache_to_disk, is_main_process)
 
+    def cache_text_encoder_outputs_sd3(
+        self, tokenizer, text_encoders, device, output_dtype, te_dtypes, cache_to_disk=False, is_main_process=True
+    ):
+        for i, dataset in enumerate(self.datasets):
+            logger.info(f"[Dataset {i}]")
+            dataset.cache_text_encoder_outputs_sd3(
+                tokenizer, text_encoders, device, output_dtype, te_dtypes, cache_to_disk, is_main_process
+            )
+
     def set_caching_mode(self, caching_mode):
         for dataset in self.datasets:
             dataset.set_caching_mode(caching_mode)
@@ -2585,6 +2650,30 @@ def cache_batch_text_encoder_outputs(
             info.text_encoder_pool2 = pool2
 
 
+def cache_batch_text_encoder_outputs_sd3(
+    image_infos, tokenizer, text_encoders, max_token_length, cache_to_disk, input_ids, output_dtype
+):
+    # make input_ids for each text encoder
+    l_tokens, g_tokens, t5_tokens = input_ids
+    
+    clip_l, clip_g, t5xxl = text_encoders
+    with torch.no_grad():
+        b_lg_out, b_t5_out, b_pool = sd3_utils.get_cond_from_tokens(
+            l_tokens, g_tokens, t5_tokens, clip_l, clip_g, t5xxl, "cpu", output_dtype
+        )
+        b_lg_out = b_lg_out.detach()
+        b_t5_out = b_t5_out.detach()
+        b_pool = b_pool.detach()
+
+    for info, lg_out, t5_out, pool in zip(image_infos, b_lg_out, b_t5_out, b_pool):
+        if cache_to_disk:
+            save_text_encoder_outputs_to_disk(info.text_encoder_outputs_npz, lg_out, t5_out, pool)
+        else:
+            info.text_encoder_outputs1 = lg_out
+            info.text_encoder_outputs2 = t5_out
+            info.text_encoder_pool2 = pool
+
+
 def save_text_encoder_outputs_to_disk(npz_path, hidden_state1, hidden_state2, pool2):
     np.savez(
         npz_path,
@@ -2907,6 +2996,7 @@ def get_sai_model_spec(
     lora: bool,
     textual_inversion: bool,
     is_stable_diffusion_ckpt: Optional[bool] = None,  # None for TI and LoRA
+    sd3: str = None,
 ):
     timestamp = time.time()
 
@@ -2940,6 +3030,7 @@ def get_sai_model_spec(
         tags=args.metadata_tags,
         timesteps=timesteps,
         clip_skip=args.clip_skip,  # None or int
+        sd3=sd3,
     )
     return metadata
 

sd3_minimal_inference.py

@@ -320,8 +320,11 @@ if __name__ == "__main__":
     # prepare embeddings
     logger.info("Encoding prompts...")
     # embeds, pooled_embed
-    cond = sd3_utils.get_cond(args.prompt, tokenizer, clip_l, clip_g, t5xxl)
-    neg_cond = sd3_utils.get_cond(args.negative_prompt, tokenizer, clip_l, clip_g, t5xxl)
+    lg_out, t5_out, pooled = sd3_utils.get_cond(args.prompt, tokenizer, clip_l, clip_g, t5xxl)
+    cond = torch.cat([lg_out, t5_out], dim=-2), pooled
+
+    lg_out, t5_out, pooled = sd3_utils.get_cond(args.negative_prompt, tokenizer, clip_l, clip_g, t5xxl)
+    neg_cond = torch.cat([lg_out, t5_out], dim=-2), pooled
 
     # generate image
     logger.info("Generating image...")

sd3_train.py

@@ -0,0 +1,907 @@
+# training with captions
+
+import argparse
+import copy
+import math
+import os
+from multiprocessing import Value
+from typing import List
+import toml
+
+from tqdm import tqdm
+
+import torch
+from library.device_utils import init_ipex, clean_memory_on_device
+
+
+init_ipex()
+
+from accelerate.utils import set_seed
+from diffusers import DDPMScheduler
+from library import deepspeed_utils, sd3_models, sd3_train_utils, sd3_utils
+
+# , sdxl_model_util
+
+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,
+)
+import library.custom_train_functions as custom_train_functions
+
+# from library.custom_train_functions import (
+#     apply_snr_weight,
+#     prepare_scheduler_for_custom_training,
+#     scale_v_prediction_loss_like_noise_prediction,
+#     add_v_prediction_like_loss,
+#     apply_debiased_estimation,
+#     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)
+
+    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"
+    assert (
+        not args.train_text_encoder or not args.cache_text_encoder_outputs
+    ), "cache_text_encoder_outputs is not supported when training text encoder / text encoderを学習するときはcache_text_encoder_outputsはサポートされていません"
+
+    # if args.block_lr:
+    #     block_lrs = [float(lr) for lr in args.block_lr.split(",")]
+    #     assert (
+    #         len(block_lrs) == UNET_NUM_BLOCKS_FOR_BLOCK_LR
+    #     ), f"block_lr must have {UNET_NUM_BLOCKS_FOR_BLOCK_LR} values / block_lrは{UNET_NUM_BLOCKS_FOR_BLOCK_LR}個の値を指定してください"
+    # else:
+    #     block_lrs = None
+
+    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
+
+    # load tokenizer
+    sd3_tokenizer = sd3_models.SD3Tokenizer()
+
+    # 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, tokenizer=[sd3_tokenizer])
+        train_dataset_group = config_util.generate_dataset_group_by_blueprint(blueprint.dataset_group)
+    else:
+        train_dataset_group = train_util.load_arbitrary_dataset(args, [sd3_tokenizer])
+
+    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(8)  # TODO これでいいか確認
+
+    if args.debug_dataset:
+        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)
+    vae_dtype = weight_dtype  # torch.float32 if args.no_half_vae else weight_dtype # SD3 VAE works with fp16
+
+    t5xxl_dtype = weight_dtype
+    if args.t5xxl_dtype is not None:
+        if args.t5xxl_dtype == "fp16":
+            t5xxl_dtype = torch.float16
+        elif args.t5xxl_dtype == "bf16":
+            t5xxl_dtype = torch.bfloat16
+        elif args.t5xxl_dtype == "fp32" or args.t5xxl_dtype == "float":
+            t5xxl_dtype = torch.float32
+        else:
+            raise ValueError(f"unexpected t5xxl_dtype: {args.t5xxl_dtype}")
+    t5xxl_device = accelerator.device if args.t5xxl_device is None else args.t5xxl_device
+
+    # モデルを読み込む
+    attn_mode = "xformers" if args.xformers else "torch"
+
+    assert (
+        attn_mode == "torch"
+    ), f"attn_mode {attn_mode} is not supported. Please use `--sdpa` instead of `--xformers`. / attn_mode {attn_mode} はサポートされていません。`--xformers`の代わりに`--sdpa`を使ってください。"
+
+    mmdit, clip_l, clip_g, t5xxl, vae = sd3_train_utils.load_target_model(
+        args, accelerator, attn_mode, weight_dtype, t5xxl_device, t5xxl_dtype
+    )
+    assert clip_l is not None, "clip_l is required / clip_lは必須です"
+    assert clip_g is not None, "clip_g is required / clip_gは必須です"
+    # logit_scale = logit_scale.to(accelerator.device, dtype=weight_dtype)
+
+    # 学習を準備する
+    if cache_latents:
+        vae.to(accelerator.device, dtype=vae_dtype)
+        vae.requires_grad_(False)
+        vae.eval()
+        vae_wrapper = sd3_models.VAEWrapper(vae)  # make SD/SDXL compatible
+        with torch.no_grad():
+            train_dataset_group.cache_latents(
+                vae_wrapper, args.vae_batch_size, args.cache_latents_to_disk, accelerator.is_main_process, file_suffix="_sd3.npz"
+            )
+        vae.to("cpu")
+        clean_memory_on_device(accelerator.device)
+
+        accelerator.wait_for_everyone()
+
+    # 学習を準備する：モデルを適切な状態にする
+    if args.gradient_checkpointing:
+        mmdit.enable_gradient_checkpointing()
+    train_mmdit = args.learning_rate != 0
+    train_clip_l = False
+    train_clip_g = False
+    train_t5xxl = False
+
+    # if args.train_text_encoder:
+    #     # TODO each option for two text encoders?
+    #     accelerator.print("enable text encoder training")
+    #     if args.gradient_checkpointing:
+    #         text_encoder1.gradient_checkpointing_enable()
+    #         text_encoder2.gradient_checkpointing_enable()
+    #     lr_te1 = args.learning_rate_te1 if args.learning_rate_te1 is not None else args.learning_rate  # 0 means not train
+    #     lr_te2 = args.learning_rate_te2 if args.learning_rate_te2 is not None else args.learning_rate  # 0 means not train
+    #     train_clip_l = lr_te1 != 0
+    #     train_clip_g = lr_te2 != 0
+
+    #     # caching one text encoder output is not supported
+    #     if not train_clip_l:
+    #         text_encoder1.to(weight_dtype)
+    #     if not train_clip_g:
+    #         text_encoder2.to(weight_dtype)
+    #     text_encoder1.requires_grad_(train_clip_l)
+    #     text_encoder2.requires_grad_(train_clip_g)
+    #     text_encoder1.train(train_clip_l)
+    #     text_encoder2.train(train_clip_g)
+    # else:
+    clip_l.to(weight_dtype)
+    clip_g.to(weight_dtype)
+    clip_l.requires_grad_(False)
+    clip_g.requires_grad_(False)
+    clip_l.eval()
+    clip_g.eval()
+    if t5xxl is not None:
+        t5xxl.to(t5xxl_dtype)
+        t5xxl.requires_grad_(False)
+        t5xxl.eval()
+
+    # TextEncoderの出力をキャッシュする
+    if args.cache_text_encoder_outputs:
+        # Text Encodes are eval and no grad
+
+        with torch.no_grad(), accelerator.autocast():
+            train_dataset_group.cache_text_encoder_outputs_sd3(
+                sd3_tokenizer,
+                (clip_l, clip_g, t5xxl),
+                (accelerator.device, accelerator.device, t5xxl_device),
+                None,
+                (None, None, None),
+                args.cache_text_encoder_outputs_to_disk,
+                accelerator.is_main_process,
+            )
+        accelerator.wait_for_everyone()
+
+    if not cache_latents:
+        vae.requires_grad_(False)
+        vae.eval()
+        vae.to(accelerator.device, dtype=vae_dtype)
+
+    mmdit.requires_grad_(train_mmdit)
+    if not train_mmdit:
+        mmdit.to(accelerator.device, dtype=weight_dtype)  # because of unet is not prepared
+
+    training_models = []
+    params_to_optimize = []
+    # if train_unet:
+    training_models.append(mmdit)
+    # if block_lrs is None:
+    params_to_optimize.append({"params": list(mmdit.parameters()), "lr": args.learning_rate})
+    # else:
+    #     params_to_optimize.extend(get_block_params_to_optimize(mmdit, block_lrs))
+
+    # if train_clip_l:
+    #     training_models.append(text_encoder1)
+    #     params_to_optimize.append({"params": list(text_encoder1.parameters()), "lr": args.learning_rate_te1 or args.learning_rate})
+    # if train_clip_g:
+    #     training_models.append(text_encoder2)
+    #     params_to_optimize.append({"params": list(text_encoder2.parameters()), "lr": args.learning_rate_te2 or args.learning_rate})
+
+    # 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"train mmdit: {train_mmdit}")  # , text_encoder1: {train_clip_l}, text_encoder2: {train_clip_g}")
+    accelerator.print(f"number of models: {len(training_models)}")
+    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.fused_optimizer_groups:
+        # 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 group of parameters.
+        # This balances memory usage and management complexity.
+
+        # calculate total number of parameters
+        n_total_params = sum(len(params["params"]) for params in params_to_optimize)
+        params_per_group = math.ceil(n_total_params / args.fused_optimizer_groups)
+
+        # split params into groups, keeping the learning rate the same for all params in a group
+        # this will increase the number of groups if the learning rate is different for different params (e.g. U-Net and text encoders)
+        grouped_params = []
+        param_group = []
+        param_group_lr = -1
+        for group in params_to_optimize:
+            lr = group["lr"]
+            for p in group["params"]:
+                # if the learning rate is different for different params, start a new group
+                if lr != param_group_lr:
+                    if param_group:
+                        grouped_params.append({"params": param_group, "lr": param_group_lr})
+                        param_group = []
+                    param_group_lr = lr
+
+                param_group.append(p)
+
+                # if the group has enough parameters, start a new group
+                if len(param_group) == params_per_group:
+                    grouped_params.append({"params": param_group, "lr": param_group_lr})
+                    param_group = []
+                    param_group_lr = -1
+
+        if param_group:
+            grouped_params.append({"params": param_group, "lr": param_group_lr})
+
+        # 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 fused optimizer groups")
+
+    else:
+        _, _, optimizer = train_util.get_optimizer(args, trainable_params=params_to_optimize)
+
+    # dataloaderを準備する
+    # 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.fused_optimizer_groups:
+        # 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.")
+        mmdit.to(weight_dtype)
+        clip_l.to(weight_dtype)
+        clip_g.to(weight_dtype)
+        if t5xxl is not None:
+            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.")
+        mmdit.to(weight_dtype)
+        clip_l.to(weight_dtype)
+        clip_g.to(weight_dtype)
+        if t5xxl is not None:
+            t5xxl.to(weight_dtype)
+
+    # TODO check if this is necessary. SD3 uses pool for clip_l and clip_g
+    # # freeze last layer and final_layer_norm in te1 since we use the output of the penultimate layer
+    # if train_clip_l:
+    #     text_encoder1.text_model.encoder.layers[-1].requires_grad_(False)
+    #     text_encoder1.text_model.final_layer_norm.requires_grad_(False)
+
+    if args.deepspeed:
+        ds_model = deepspeed_utils.prepare_deepspeed_model(
+            args,
+            mmdit=mmdit,
+            # mmdie=mmdit if train_mmdit else None,
+            # text_encoder1=text_encoder1 if train_clip_l else None,
+            # text_encoder2=text_encoder2 if train_clip_g else None,
+        )
+        # 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がなんかよろしくやってくれるらしい
+        if train_mmdit:
+            mmdit = accelerator.prepare(mmdit)
+        # if train_clip_l:
+        #     text_encoder1 = accelerator.prepare(text_encoder1)
+        # if train_clip_g:
+        #     text_encoder2 = accelerator.prepare(text_encoder2)
+        optimizer, train_dataloader, lr_scheduler = accelerator.prepare(optimizer, train_dataloader, lr_scheduler)
+
+    # TextEncoderの出力をキャッシュするときには、すでに出力を取得済みなのでCPUへ移動する
+    if args.cache_text_encoder_outputs:
+        # move Text Encoders for sampling images. Text Encoder doesn't work on CPU with fp16
+        clip_l.to("cpu", dtype=torch.float32)
+        clip_g.to("cpu", dtype=torch.float32)
+        if t5xxl is not None:
+            t5xxl.to("cpu", dtype=torch.float32)
+        clean_memory_on_device(accelerator.device)
+    else:
+        # make sure Text Encoders are on GPU
+        # TODO support CPU for text encoders
+        clip_l.to(accelerator.device)
+        clip_g.to(accelerator.device)
+        if t5xxl is not None:
+            t5xxl.to(accelerator.device)
+
+    # TODO cache sample prompt's embeddings to free text encoder's memory
+    if args.cache_text_encoder_outputs:
+        if not args.save_t5xxl:
+            t5xxl = None  # free memory
+    clean_memory_on_device(accelerator.device)
+
+    # 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 in optimizer.param_groups:
+            for parameter in param_group["params"]:
+                if parameter.requires_grad:
+
+                    def __grad_hook(tensor: torch.Tensor, param_group=param_group):
+                        if accelerator.sync_gradients and args.max_grad_norm != 0.0:
+                            accelerator.clip_grad_norm_(tensor, args.max_grad_norm)
+                        optimizer.step_param(tensor, param_group)
+                        tensor.grad = None
+
+                    parameter.register_post_accumulate_grad_hook(__grad_hook)
+
+    elif args.fused_optimizer_groups:
+        # 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 optimizer_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(optimizer_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 = DDPMScheduler(
+    #     beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000, clip_sample=False
+    # )
+
+    noise_scheduler = sd3_train_utils.FlowMatchEulerDiscreteScheduler(num_train_timesteps=1000, shift=3.0)
+    noise_scheduler_copy = copy.deepcopy(noise_scheduler)
+
+    # prepare_scheduler_for_custom_training(noise_scheduler, accelerator.device)
+    # if args.zero_terminal_snr:
+    #     custom_train_functions.fix_noise_scheduler_betas_for_zero_terminal_snr(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,
+        )
+
+    # # For --sample_at_first
+    # sd3_train_utils.sample_images(
+    #     accelerator, args, 0, global_step, accelerator.device, vae, [tokenizer1, tokenizer2], [text_encoder1, text_encoder2], mmdit
+    # )
+
+    # following function will be moved to sd3_train_utils
+
+    def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
+        sigmas = noise_scheduler_copy.sigmas.to(device=accelerator.device, dtype=dtype)
+        schedule_timesteps = noise_scheduler_copy.timesteps.to(accelerator.device)
+        timesteps = timesteps.to(accelerator.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
+
+    loss_recorder = train_util.LossRecorder()
+    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.fused_optimizer_groups:
+                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).to(dtype=weight_dtype)
+                else:
+                    with torch.no_grad():
+                        # encode images to latents. images are [-1, 1]
+                        latents = vae.encode(batch["images"].to(vae_dtype)).to(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)
+                # latents = latents * sdxl_model_util.VAE_SCALE_FACTOR
+                latents = sd3_models.SDVAE.process_in(latents)
+
+                if "text_encoder_outputs1_list" not in batch or batch["text_encoder_outputs1_list"] is None:
+                    # not cached, get text encoder outputs
+                    # XXX This does not work yet
+                    input_ids_clip_l, input_ids_clip_g, input_ids_t5xxl = batch["input_ids"]
+                    with torch.set_grad_enabled(args.train_text_encoder):
+                        # TODO support weighted captions
+                        # TODO support length > 75
+                        input_ids_clip_l = input_ids_clip_l.to(accelerator.device)
+                        input_ids_clip_g = input_ids_clip_g.to(accelerator.device)
+                        input_ids_t5xxl = input_ids_t5xxl.to(accelerator.device)
+
+                        # get text encoder outputs: outputs are concatenated
+                        context, pool = sd3_utils.get_cond_from_tokens(
+                            input_ids_clip_l, input_ids_clip_g, input_ids_t5xxl, clip_l, clip_g, t5xxl
+                        )
+                else:
+                    # encoder_hidden_states1 = batch["text_encoder_outputs1_list"].to(accelerator.device).to(weight_dtype)
+                    # encoder_hidden_states2 = batch["text_encoder_outputs2_list"].to(accelerator.device).to(weight_dtype)
+                    # pool2 = batch["text_encoder_pool2_list"].to(accelerator.device).to(weight_dtype)
+                    # TODO this reuses SDXL keys, it should be fixed
+                    lg_out = batch["text_encoder_outputs1_list"]
+                    t5_out = batch["text_encoder_outputs2_list"]
+                    pool = batch["text_encoder_pool2_list"]
+                    context = torch.cat([lg_out, t5_out], dim=-2)
+
+                # 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]
+
+                # 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 * noise_scheduler_copy.config.num_train_timesteps).long()
+                timesteps = noise_scheduler_copy.timesteps[indices].to(device=accelerator.device)
+
+                # Add noise according to flow matching.
+                sigmas = get_sigmas(timesteps, n_dim=latents.ndim, dtype=weight_dtype)
+                noisy_model_input = sigmas * noise + (1.0 - sigmas) * latents
+
+                # call model
+                with accelerator.autocast():
+                    model_pred = mmdit(noisy_model_input, timesteps, context=context, y=pool)
+
+                # Follow: Section 5 of https://arxiv.org/abs/2206.00364.
+                # Preconditioning of the model outputs.
+                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)
+
+                # flow matching loss
+                target = latents
+
+                # Compute regular loss. TODO simplify this
+                loss = torch.mean(
+                    (weighting.float() * (model_pred.float() - target.float()) ** 2).reshape(target.shape[0], -1),
+                    1,
+                )
+                loss = loss.mean()
+
+                accelerator.backward(loss)
+
+                if not (args.fused_backward_pass or args.fused_optimizer_groups):
+                    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.fused_optimizer_groups:
+                        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
+
+                # sdxl_train_util.sample_images(
+                #     accelerator,
+                #     args,
+                #     None,
+                #     global_step,
+                #     accelerator.device,
+                #     vae,
+                #     [tokenizer1, tokenizer2],
+                #     [text_encoder1, text_encoder2],
+                #     mmdit,
+                # )
+
+                # 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:
+                        sd3_train_utils.save_sd3_model_on_epoch_end_or_stepwise(
+                            args,
+                            False,
+                            accelerator,
+                            save_dtype,
+                            epoch,
+                            num_train_epochs,
+                            global_step,
+                            clip_l if args.save_clip else None,
+                            clip_g if args.save_clip else None,
+                            t5xxl if args.save_t5xxl else None,
+                            mmdit,
+                            vae,
+                        )
+
+            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:
+                logs = {"loss": current_loss}
+                train_util.append_lr_to_logs(logs, lr_scheduler, args.optimizer_type, including_unet=train_mmdit)
+
+                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 args.logging_dir is not None:
+            logs = {"loss/epoch": loss_recorder.moving_average}
+            accelerator.log(logs, step=epoch + 1)
+
+        accelerator.wait_for_everyone()
+
+        if args.save_every_n_epochs is not None:
+            if accelerator.is_main_process:
+                sd3_train_utils.save_sd3_model_on_epoch_end_or_stepwise(
+                    args,
+                    True,
+                    accelerator,
+                    save_dtype,
+                    epoch,
+                    num_train_epochs,
+                    global_step,
+                    clip_l if args.save_clip else None,
+                    clip_g if args.save_clip else None,
+                    t5xxl if args.save_t5xxl else None,
+                    mmdit,
+                    vae,
+                )
+
+        # sdxl_train_util.sample_images(
+        #     accelerator,
+        #     args,
+        #     epoch + 1,
+        #     global_step,
+        #     accelerator.device,
+        #     vae,
+        #     [tokenizer1, tokenizer2],
+        #     [text_encoder1, text_encoder2],
+        #     mmdit,
+        # )
+
+    is_main_process = accelerator.is_main_process
+    # if is_main_process:
+    mmdit = accelerator.unwrap_model(mmdit)
+    clip_l = accelerator.unwrap_model(clip_l)
+    clip_g = accelerator.unwrap_model(clip_g)
+    if t5xxl is not None:
+        t5xxl = accelerator.unwrap_model(t5xxl)
+
+    accelerator.end_training()
+
+    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:
+        sd3_train_utils.save_sd3_model_on_train_end(
+            args,
+            save_dtype,
+            epoch,
+            global_step,
+            clip_l if args.save_clip else None,
+            clip_g if args.save_clip else None,
+            t5xxl if args.save_t5xxl else None,
+            mmdit,
+            vae,
+        )
+        logger.info("model saved.")
+
+
+def setup_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser()
+
+    add_logging_arguments(parser)
+    train_util.add_sd_models_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)
+    custom_train_functions.add_custom_train_arguments(parser)
+    sd3_train_utils.add_sd3_training_arguments(parser)
+
+    # TE training is disabled temporarily
+
+    # parser.add_argument(
+    #     "--learning_rate_te1",
+    #     type=float,
+    #     default=None,
+    #     help="learning rate for text encoder 1 (ViT-L) / text encoder 1 (ViT-L)の学習率",
+    # )
+    # parser.add_argument(
+    #     "--learning_rate_te2",
+    #     type=float,
+    #     default=None,
+    #     help="learning rate for text encoder 2 (BiG-G) / text encoder 2 (BiG-G)の学習率",
+    # )
+
+    # 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(
+    #     "--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",
+    # )
+    # parser.add_argument(
+    #     "--block_lr",
+    #     type=str,
+    #     default=None,
+    #     help=f"learning rates for each block of U-Net, comma-separated, {UNET_NUM_BLOCKS_FOR_BLOCK_LR} values / "
+    #     + f"U-Netの各ブロックの学習率、カンマ区切り、{UNET_NUM_BLOCKS_FOR_BLOCK_LR}個の値",
+    # )
+    parser.add_argument(
+        "--fused_optimizer_groups",
+        type=int,
+        default=None,
+        help="number of optimizers for fused backward pass and optimizer step / fused backward passとoptimizer stepのためのoptimizer数",
+    )
+    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)