sd3 training

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