Update train.py

train.py

@@ -13,11 +13,9 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-"""Fine-tuning script for Stable Diffusion XL for text2image."""
+"""Fine-tuning script for Stable Diffusion XL for text2image with support for LoRA."""
 
 import argparse
-import functools
-import gc
 import logging
 import math
 import os
@@ -26,7 +24,6 @@ import shutil
 from contextlib import nullcontext
 from pathlib import Path
 
-import accelerate
 import datasets
 import numpy as np
 import torch
@@ -35,45 +32,55 @@ import torch.utils.checkpoint
 import transformers
 from accelerate import Accelerator
 from accelerate.logging import get_logger
-from accelerate.utils import DistributedType, ProjectConfiguration, set_seed
-from datasets import concatenate_datasets, load_dataset
+from accelerate.utils import DistributedDataParallelKwargs, DistributedType, ProjectConfiguration, set_seed
+from datasets import load_dataset
 from huggingface_hub import create_repo, upload_folder
 from packaging import version
+from peft import LoraConfig, set_peft_model_state_dict
+from peft.utils import get_peft_model_state_dict
 from torchvision import transforms
 from torchvision.transforms.functional import crop
 from tqdm.auto import tqdm
 from transformers import AutoTokenizer, PretrainedConfig
 
 import diffusers
-from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionXLPipeline, UNet2DConditionModel
+from diffusers import (
+    AutoencoderKL,
+    DDPMScheduler,
+    StableDiffusionXLPipeline,
+    UNet2DConditionModel,
+)
+from diffusers.loaders import StableDiffusionLoraLoaderMixin
 from diffusers.optimization import get_scheduler
-from diffusers.training_utils import EMAModel, compute_snr
-from diffusers.utils import check_min_version, is_wandb_available
+from diffusers.training_utils import _set_state_dict_into_text_encoder, cast_training_params, compute_snr
+from diffusers.utils import (
+    check_min_version,
+    convert_state_dict_to_diffusers,
+    convert_unet_state_dict_to_peft,
+    is_wandb_available,
+)
 from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
 from diffusers.utils.import_utils import is_torch_npu_available, is_xformers_available
 from diffusers.utils.torch_utils import is_compiled_module
 
 
+if is_wandb_available():
+    import wandb
+
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
 check_min_version("0.36.0.dev0")
 
 logger = get_logger(__name__)
 if is_torch_npu_available():
-    import torch_npu
-
     torch.npu.config.allow_internal_format = False
 
-DATASET_NAME_MAPPING = {
-    "lambdalabs/naruto-blip-captions": ("image", "text"),
-}
-
 
 def save_model_card(
     repo_id: str,
     images: list = None,
-    validation_prompt: str = None,
     base_model: str = None,
     dataset_name: str = None,
+    train_text_encoder: bool = False,
     repo_folder: str = None,
     vae_path: str = None,
 ):
@@ -84,14 +91,15 @@ def save_model_card(
             img_str += f"![img_{i}](./image_{i}.png)\n"
 
     model_description = f"""
-# Text-to-image finetuning - {repo_id}
+# LoRA text2image fine-tuning - {repo_id}
 
-This pipeline was finetuned from **{base_model}** on the **{dataset_name}** dataset. Below are some example images generated with the finetuned pipeline using the following prompt: {validation_prompt}: \n
+These are LoRA adaption weights for {base_model}. The weights were fine-tuned on the {dataset_name} dataset. You can find some example images in the following. \n
 {img_str}
 
+LoRA for the text encoder was enabled: {train_text_encoder}.
+
 Special VAE used for training: {vae_path}.
 """
-
     model_card = load_or_create_model_card(
         repo_id_or_path=repo_id,
         from_training=True,
@@ -105,14 +113,56 @@ Special VAE used for training: {vae_path}.
         "stable-diffusion-xl",
         "stable-diffusion-xl-diffusers",
         "text-to-image",
-        "diffusers-training",
         "diffusers",
+        "diffusers-training",
+        "lora",
     ]
     model_card = populate_model_card(model_card, tags=tags)
 
     model_card.save(os.path.join(repo_folder, "README.md"))
 
 
+def log_validation(
+    pipeline,
+    args,
+    accelerator,
+    epoch,
+    is_final_validation=False,
+):
+    logger.info(
+        f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
+        f" {args.validation_prompt}."
+    )
+    pipeline = pipeline.to(accelerator.device)
+    pipeline.set_progress_bar_config(disable=True)
+
+    # run inference
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
+    pipeline_args = {"prompt": args.validation_prompt}
+    if torch.backends.mps.is_available():
+        autocast_ctx = nullcontext()
+    else:
+        autocast_ctx = torch.autocast(accelerator.device.type)
+
+    with autocast_ctx:
+        images = [pipeline(**pipeline_args, generator=generator).images[0] for _ in range(args.num_validation_images)]
+
+    for tracker in accelerator.trackers:
+        phase_name = "test" if is_final_validation else "validation"
+        if tracker.name == "tensorboard":
+            np_images = np.stack([np.asarray(img) for img in images])
+            tracker.writer.add_images(phase_name, np_images, epoch, dataformats="NHWC")
+        if tracker.name == "wandb":
+            tracker.log(
+                {
+                    phase_name: [
+                        wandb.Image(image, caption=f"{i}: {args.validation_prompt}") for i, image in enumerate(images)
+                    ]
+                }
+            )
+    return images
+
+
 def import_model_class_from_model_name_or_path(
     pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
 ):
@@ -226,16 +276,10 @@ def parse_args(input_args=None):
             "value if set."
         ),
     )
-    parser.add_argument(
-        "--proportion_empty_prompts",
-        type=float,
-        default=0,
-        help="Proportion of image prompts to be replaced with empty strings. Defaults to 0 (no prompt replacement).",
-    )
     parser.add_argument(
         "--output_dir",
         type=str,
-        default="sdxl-model-finetuned",
+        default="sd-model-finetuned-lora",
         help="The output directory where the model predictions and checkpoints will be written.",
     )
     parser.add_argument(
@@ -268,6 +312,11 @@ def parse_args(input_args=None):
         action="store_true",
         help="whether to randomly flip images horizontally",
     )
+    parser.add_argument(
+        "--train_text_encoder",
+        action="store_true",
+        help="Whether to train the text encoder. If set, the text encoder should be float32 precision.",
+    )
     parser.add_argument(
         "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
     )
@@ -338,55 +387,6 @@ def parse_args(input_args=None):
     parser.add_argument(
         "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
     )
-    parser.add_argument(
-        "--timestep_bias_strategy",
-        type=str,
-        default="none",
-        choices=["earlier", "later", "range", "none"],
-        help=(
-            "The timestep bias strategy, which may help direct the model toward learning low or high frequency details."
-            " Choices: ['earlier', 'later', 'range', 'none']."
-            " The default is 'none', which means no bias is applied, and training proceeds normally."
-            " The value of 'later' will increase the frequency of the model's final training timesteps."
-        ),
-    )
-    parser.add_argument(
-        "--timestep_bias_multiplier",
-        type=float,
-        default=1.0,
-        help=(
-            "The multiplier for the bias. Defaults to 1.0, which means no bias is applied."
-            " A value of 2.0 will double the weight of the bias, and a value of 0.5 will halve it."
-        ),
-    )
-    parser.add_argument(
-        "--timestep_bias_begin",
-        type=int,
-        default=0,
-        help=(
-            "When using `--timestep_bias_strategy=range`, the beginning (inclusive) timestep to bias."
-            " Defaults to zero, which equates to having no specific bias."
-        ),
-    )
-    parser.add_argument(
-        "--timestep_bias_end",
-        type=int,
-        default=1000,
-        help=(
-            "When using `--timestep_bias_strategy=range`, the final timestep (inclusive) to bias."
-            " Defaults to 1000, which is the number of timesteps that Stable Diffusion is trained on."
-        ),
-    )
-    parser.add_argument(
-        "--timestep_bias_portion",
-        type=float,
-        default=0.25,
-        help=(
-            "The portion of timesteps to bias. Defaults to 0.25, which 25% of timesteps will be biased."
-            " A value of 0.5 will bias one half of the timesteps. The value provided for `--timestep_bias_strategy` determines"
-            " whether the biased portions are in the earlier or later timesteps."
-        ),
-    )
     parser.add_argument(
         "--snr_gamma",
         type=float,
@@ -394,7 +394,6 @@ def parse_args(input_args=None):
         help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
         "More details here: https://huggingface.co/papers/2303.09556.",
     )
-    parser.add_argument("--use_ema", action="store_true", help="Whether to use EMA model.")
     parser.add_argument(
         "--allow_tf32",
         action="store_true",
@@ -464,12 +463,23 @@ def parse_args(input_args=None):
     )
     parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
     parser.add_argument(
-        "--enable_npu_flash_attention", action="store_true", help="Whether or not to use npu flash attention."
+        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
     )
     parser.add_argument(
-        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
+        "--enable_npu_flash_attention", action="store_true", help="Whether or not to use npu flash attention."
     )
     parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.")
+    parser.add_argument(
+        "--rank",
+        type=int,
+        default=4,
+        help=("The dimension of the LoRA update matrices."),
+    )
+    parser.add_argument(
+        "--debug_loss",
+        action="store_true",
+        help="debug loss for each image, if filenames are available in the dataset",
+    )
     parser.add_argument(
         "--image_interpolation_mode",
         type=str,
@@ -492,109 +502,53 @@ def parse_args(input_args=None):
     # Sanity checks
     if args.dataset_name is None and args.train_data_dir is None:
         raise ValueError("Need either a dataset name or a training folder.")
-    if args.proportion_empty_prompts < 0 or args.proportion_empty_prompts > 1:
-        raise ValueError("`--proportion_empty_prompts` must be in the range [0, 1].")
 
     return args
 
 
-# Adapted from pipelines.StableDiffusionXLPipeline.encode_prompt
-def encode_prompt(batch, text_encoders, tokenizers, proportion_empty_prompts, caption_column, is_train=True):
-    prompt_embeds_list = []
-    prompt_batch = batch[caption_column]
-
-    captions = []
-    for caption in prompt_batch:
-        if random.random() < proportion_empty_prompts:
-            captions.append("")
-        elif isinstance(caption, str):
-            captions.append(caption)
-        elif isinstance(caption, (list, np.ndarray)):
-            # take a random caption if there are multiple
-            captions.append(random.choice(caption) if is_train else caption[0])
-
-    with torch.no_grad():
-        for tokenizer, text_encoder in zip(tokenizers, text_encoders):
-            text_inputs = tokenizer(
-                captions,
-                padding="max_length",
-                max_length=tokenizer.model_max_length,
-                truncation=True,
-                return_tensors="pt",
-            )
-            text_input_ids = text_inputs.input_ids
-            prompt_embeds = text_encoder(
-                text_input_ids.to(text_encoder.device),
-                output_hidden_states=True,
-                return_dict=False,
-            )
-
-            # We are only ALWAYS interested in the pooled output of the final text encoder
-            pooled_prompt_embeds = prompt_embeds[0]
-            prompt_embeds = prompt_embeds[-1][-2]
-            bs_embed, seq_len, _ = prompt_embeds.shape
-            prompt_embeds = prompt_embeds.view(bs_embed, seq_len, -1)
-            prompt_embeds_list.append(prompt_embeds)
-
-    prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
-    pooled_prompt_embeds = pooled_prompt_embeds.view(bs_embed, -1)
-    return {"prompt_embeds": prompt_embeds.cpu(), "pooled_prompt_embeds": pooled_prompt_embeds.cpu()}
-
-
-def compute_vae_encodings(batch, vae):
-    images = batch.pop("pixel_values")
-    pixel_values = torch.stack(list(images))
-    pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
-    pixel_values = pixel_values.to(vae.device, dtype=vae.dtype)
+DATASET_NAME_MAPPING = {
+    "lambdalabs/naruto-blip-captions": ("image", "text"),
+}
 
-    with torch.no_grad():
-        model_input = vae.encode(pixel_values).latent_dist.sample()
-    model_input = model_input * vae.config.scaling_factor
 
-    # There might have slightly performance improvement
-    # by changing model_input.cpu() to accelerator.gather(model_input)
-    return {"model_input": model_input.cpu()}
+def tokenize_prompt(tokenizer, prompt):
+    text_inputs = tokenizer(
+        prompt,
+        padding="max_length",
+        max_length=tokenizer.model_max_length,
+        truncation=True,
+        return_tensors="pt",
+    )
+    text_input_ids = text_inputs.input_ids
+    return text_input_ids
 
 
-def generate_timestep_weights(args, num_timesteps):
-    weights = torch.ones(num_timesteps)
+# Adapted from pipelines.StableDiffusionXLPipeline.encode_prompt
+def encode_prompt(text_encoders, tokenizers, prompt, text_input_ids_list=None):
+    prompt_embeds_list = []
 
-    # Determine the indices to bias
-    num_to_bias = int(args.timestep_bias_portion * num_timesteps)
+    for i, text_encoder in enumerate(text_encoders):
+        if tokenizers is not None:
+            tokenizer = tokenizers[i]
+            text_input_ids = tokenize_prompt(tokenizer, prompt)
+        else:
+            assert text_input_ids_list is not None
+            text_input_ids = text_input_ids_list[i]
 
-    if args.timestep_bias_strategy == "later":
-        bias_indices = slice(-num_to_bias, None)
-    elif args.timestep_bias_strategy == "earlier":
-        bias_indices = slice(0, num_to_bias)
-    elif args.timestep_bias_strategy == "range":
-        # Out of the possible 1000 timesteps, we might want to focus on eg. 200-500.
-        range_begin = args.timestep_bias_begin
-        range_end = args.timestep_bias_end
-        if range_begin < 0:
-            raise ValueError(
-                "When using the range strategy for timestep bias, you must provide a beginning timestep greater or equal to zero."
-            )
-        if range_end > num_timesteps:
-            raise ValueError(
-                "When using the range strategy for timestep bias, you must provide an ending timestep smaller than the number of timesteps."
-            )
-        bias_indices = slice(range_begin, range_end)
-    else:  # 'none' or any other string
-        return weights
-    if args.timestep_bias_multiplier <= 0:
-        return ValueError(
-            "The parameter --timestep_bias_multiplier is not intended to be used to disable the training of specific timesteps."
-            " If it was intended to disable timestep bias, use `--timestep_bias_strategy none` instead."
-            " A timestep bias multiplier less than or equal to 0 is not allowed."
+        prompt_embeds = text_encoder(
+            text_input_ids.to(text_encoder.device), output_hidden_states=True, return_dict=False
         )
 
-    # Apply the bias
-    weights[bias_indices] *= args.timestep_bias_multiplier
-
-    # Normalize
-    weights /= weights.sum()
+        # We are only ALWAYS interested in the pooled output of the final text encoder
+        pooled_prompt_embeds = prompt_embeds[0]
+        prompt_embeds = prompt_embeds[-1][-2]
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.view(bs_embed, seq_len, -1)
+        prompt_embeds_list.append(prompt_embeds)
 
-    return weights
+    prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+    pooled_prompt_embeds = pooled_prompt_embeds.view(bs_embed, -1)
+    return prompt_embeds, pooled_prompt_embeds
 
 
 def main(args):
@@ -606,30 +560,22 @@ def main(args):
 
     logging_dir = Path(args.output_dir, args.logging_dir)
 
-    accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
-
     if torch.backends.mps.is_available() and args.mixed_precision == "bf16":
         # due to pytorch#99272, MPS does not yet support bfloat16.
         raise ValueError(
             "Mixed precision training with bfloat16 is not supported on MPS. Please use fp16 (recommended) or fp32 instead."
         )
 
+    accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
+    kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
     accelerator = Accelerator(
         gradient_accumulation_steps=args.gradient_accumulation_steps,
         mixed_precision=args.mixed_precision,
         log_with=args.report_to,
         project_config=accelerator_project_config,
+        kwargs_handlers=[kwargs],
     )
 
-    # Disable AMP for MPS.
-    if torch.backends.mps.is_available():
-        accelerator.native_amp = False
-
-    if args.report_to == "wandb":
-        if not is_wandb_available():
-            raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
-        import wandb
-
     # Make one log on every process with the configuration for debugging.
     logging.basicConfig(
         format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
@@ -684,7 +630,6 @@ def main(args):
 
     # Load scheduler and models
     noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
-    # Check for terminal SNR in combination with SNR Gamma
     text_encoder_one = text_encoder_cls_one.from_pretrained(
         args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
     )
@@ -706,14 +651,13 @@ def main(args):
         args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
     )
 
-    # Freeze vae and text encoders.
+    # We only train the additional adapter LoRA layers
     vae.requires_grad_(False)
     text_encoder_one.requires_grad_(False)
     text_encoder_two.requires_grad_(False)
-    # Set unet as trainable.
-    unet.train()
+    unet.requires_grad_(False)
 
-    # For mixed precision training we cast all non-trainable weights to half-precision
+    # For mixed precision training we cast all non-trainable weights (vae, non-lora text_encoder and non-lora unet) to half-precision
     # as these weights are only used for inference, keeping weights in full precision is not required.
     weight_dtype = torch.float32
     if accelerator.mixed_precision == "fp16":
@@ -723,22 +667,22 @@ def main(args):
 
     # Move unet, vae and text_encoder to device and cast to weight_dtype
     # The VAE is in float32 to avoid NaN losses.
-    vae.to(accelerator.device, dtype=torch.float32)
+    unet.to(accelerator.device, dtype=weight_dtype)
+
+    if args.pretrained_vae_model_name_or_path is None:
+        vae.to(accelerator.device, dtype=torch.float32)
+    else:
+        vae.to(accelerator.device, dtype=weight_dtype)
     text_encoder_one.to(accelerator.device, dtype=weight_dtype)
     text_encoder_two.to(accelerator.device, dtype=weight_dtype)
 
-    # Create EMA for the unet.
-    if args.use_ema:
-        ema_unet = UNet2DConditionModel.from_pretrained(
-            args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
-        )
-        ema_unet = EMAModel(ema_unet.parameters(), model_cls=UNet2DConditionModel, model_config=ema_unet.config)
     if args.enable_npu_flash_attention:
         if is_torch_npu_available():
             logger.info("npu flash attention enabled.")
             unet.enable_npu_flash_attention()
         else:
             raise ValueError("npu flash attention requires torch_npu extensions and is supported only on npu devices.")
+
     if args.enable_xformers_memory_efficient_attention:
         if is_xformers_available():
             import xformers
@@ -752,44 +696,122 @@ def main(args):
         else:
             raise ValueError("xformers is not available. Make sure it is installed correctly")
 
-    # `accelerate` 0.16.0 will have better support for customized saving
-    if version.parse(accelerate.__version__) >= version.parse("0.16.0"):
-        # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
-        def save_model_hook(models, weights, output_dir):
-            if accelerator.is_main_process:
-                if args.use_ema:
-                    ema_unet.save_pretrained(os.path.join(output_dir, "unet_ema"))
+    # now we will add new LoRA weights to the attention layers
+    # Set correct lora layers
+    unet_lora_config = LoraConfig(
+        r=args.rank,
+        lora_alpha=args.rank,
+        init_lora_weights="gaussian",
+        target_modules=["to_k", "to_q", "to_v", "to_out.0"],
+    )
+
+    unet.add_adapter(unet_lora_config)
+
+    # The text encoder comes from 🤗 transformers, we will also attach adapters to it.
+    if args.train_text_encoder:
+        # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
+        text_lora_config = LoraConfig(
+            r=args.rank,
+            lora_alpha=args.rank,
+            init_lora_weights="gaussian",
+            target_modules=["q_proj", "k_proj", "v_proj", "out_proj"],
+        )
+        text_encoder_one.add_adapter(text_lora_config)
+        text_encoder_two.add_adapter(text_lora_config)
 
-                for i, model in enumerate(models):
-                    model.save_pretrained(os.path.join(output_dir, "unet"))
+    def unwrap_model(model):
+        model = accelerator.unwrap_model(model)
+        model = model._orig_mod if is_compiled_module(model) else model
+        return model
 
-                    # make sure to pop weight so that corresponding model is not saved again
-                    if weights:
-                        weights.pop()
+    # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
+    def save_model_hook(models, weights, output_dir):
+        if accelerator.is_main_process:
+            # there are only two options here. Either are just the unet attn processor layers
+            # or there are the unet and text encoder attn layers
+            unet_lora_layers_to_save = None
+            text_encoder_one_lora_layers_to_save = None
+            text_encoder_two_lora_layers_to_save = None
+
+            for model in models:
+                if isinstance(unwrap_model(model), type(unwrap_model(unet))):
+                    unet_lora_layers_to_save = convert_state_dict_to_diffusers(get_peft_model_state_dict(model))
+                elif isinstance(unwrap_model(model), type(unwrap_model(text_encoder_one))):
+                    text_encoder_one_lora_layers_to_save = convert_state_dict_to_diffusers(
+                        get_peft_model_state_dict(model)
+                    )
+                elif isinstance(unwrap_model(model), type(unwrap_model(text_encoder_two))):
+                    text_encoder_two_lora_layers_to_save = convert_state_dict_to_diffusers(
+                        get_peft_model_state_dict(model)
+                    )
+                else:
+                    raise ValueError(f"unexpected save model: {model.__class__}")
 
-        def load_model_hook(models, input_dir):
-            if args.use_ema:
-                load_model = EMAModel.from_pretrained(os.path.join(input_dir, "unet_ema"), UNet2DConditionModel)
-                ema_unet.load_state_dict(load_model.state_dict())
-                ema_unet.to(accelerator.device)
-                del load_model
+                # make sure to pop weight so that corresponding model is not saved again
+                if weights:
+                    weights.pop()
 
-            for _ in range(len(models)):
-                # pop models so that they are not loaded again
-                model = models.pop()
+            StableDiffusionXLPipeline.save_lora_weights(
+                output_dir,
+                unet_lora_layers=unet_lora_layers_to_save,
+                text_encoder_lora_layers=text_encoder_one_lora_layers_to_save,
+                text_encoder_2_lora_layers=text_encoder_two_lora_layers_to_save,
+            )
+
+    def load_model_hook(models, input_dir):
+        unet_ = None
+        text_encoder_one_ = None
+        text_encoder_two_ = None
 
-                # load diffusers style into model
-                load_model = UNet2DConditionModel.from_pretrained(input_dir, subfolder="unet")
-                model.register_to_config(**load_model.config)
+        while len(models) > 0:
+            model = models.pop()
 
-                model.load_state_dict(load_model.state_dict())
-                del load_model
+            if isinstance(model, type(unwrap_model(unet))):
+                unet_ = model
+            elif isinstance(model, type(unwrap_model(text_encoder_one))):
+                text_encoder_one_ = model
+            elif isinstance(model, type(unwrap_model(text_encoder_two))):
+                text_encoder_two_ = model
+            else:
+                raise ValueError(f"unexpected save model: {model.__class__}")
+
+        lora_state_dict, _ = StableDiffusionLoraLoaderMixin.lora_state_dict(input_dir)
+        unet_state_dict = {f"{k.replace('unet.', '')}": v for k, v in lora_state_dict.items() if k.startswith("unet.")}
+        unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
+        incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")
+        if incompatible_keys is not None:
+            # check only for unexpected keys
+            unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
+            if unexpected_keys:
+                logger.warning(
+                    f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
+                    f" {unexpected_keys}. "
+                )
 
-        accelerator.register_save_state_pre_hook(save_model_hook)
-        accelerator.register_load_state_pre_hook(load_model_hook)
+        if args.train_text_encoder:
+            _set_state_dict_into_text_encoder(lora_state_dict, prefix="text_encoder.", text_encoder=text_encoder_one_)
+
+            _set_state_dict_into_text_encoder(
+                lora_state_dict, prefix="text_encoder_2.", text_encoder=text_encoder_two_
+            )
+
+        # Make sure the trainable params are in float32. This is again needed since the base models
+        # are in `weight_dtype`. More details:
+        # https://github.com/huggingface/diffusers/pull/6514#discussion_r1449796804
+        if args.mixed_precision == "fp16":
+            models = [unet_]
+            if args.train_text_encoder:
+                models.extend([text_encoder_one_, text_encoder_two_])
+            cast_training_params(models, dtype=torch.float32)
+
+    accelerator.register_save_state_pre_hook(save_model_hook)
+    accelerator.register_load_state_pre_hook(load_model_hook)
 
     if args.gradient_checkpointing:
         unet.enable_gradient_checkpointing()
+        if args.train_text_encoder:
+            text_encoder_one.gradient_checkpointing_enable()
+            text_encoder_two.gradient_checkpointing_enable()
 
     # Enable TF32 for faster training on Ampere GPUs,
     # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
@@ -801,6 +823,13 @@ def main(args):
             args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
         )
 
+    # Make sure the trainable params are in float32.
+    if args.mixed_precision == "fp16":
+        models = [unet]
+        if args.train_text_encoder:
+            models.extend([text_encoder_one, text_encoder_two])
+        cast_training_params(models, dtype=torch.float32)
+
     # Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs
     if args.use_8bit_adam:
         try:
@@ -815,7 +844,13 @@ def main(args):
         optimizer_class = torch.optim.AdamW
 
     # Optimizer creation
-    params_to_optimize = unet.parameters()
+    params_to_optimize = list(filter(lambda p: p.requires_grad, unet.parameters()))
+    if args.train_text_encoder:
+        params_to_optimize = (
+            params_to_optimize
+            + list(filter(lambda p: p.requires_grad, text_encoder_one.parameters()))
+            + list(filter(lambda p: p.requires_grad, text_encoder_two.parameters()))
+        )
     optimizer = optimizer_class(
         params_to_optimize,
         lr=args.learning_rate,
@@ -870,13 +905,39 @@ def main(args):
             )
 
     # Preprocessing the datasets.
+    # We need to tokenize input captions and transform the images.
+    def tokenize_captions(examples, is_train=True):
+        captions = []
+        for caption in examples[caption_column]:
+            if isinstance(caption, str):
+                captions.append(caption)
+            elif isinstance(caption, (list, np.ndarray)):
+                # take a random caption if there are multiple
+                captions.append(random.choice(caption) if is_train else caption[0])
+            else:
+                raise ValueError(
+                    f"Caption column `{caption_column}` should contain either strings or lists of strings."
+                )
+        tokens_one = tokenize_prompt(tokenizer_one, captions)
+        tokens_two = tokenize_prompt(tokenizer_two, captions)
+        return tokens_one, tokens_two
+
+    # Get the specified interpolation method from the args
     interpolation = getattr(transforms.InterpolationMode, args.image_interpolation_mode.upper(), None)
+
+    # Raise an error if the interpolation method is invalid
     if interpolation is None:
-        raise ValueError(f"Unsupported interpolation mode {interpolation=}.")
-    train_resize = transforms.Resize(args.resolution, interpolation=interpolation)
+        raise ValueError(f"Unsupported interpolation mode {args.image_interpolation_mode}.")
+    # Preprocessing the datasets.
+    train_resize = transforms.Resize(args.resolution, interpolation=interpolation)  # Use dynamic interpolation method
     train_crop = transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution)
     train_flip = transforms.RandomHorizontalFlip(p=1.0)
-    train_transforms = transforms.Compose([transforms.ToTensor(), transforms.Normalize([0.5], [0.5])])
+    train_transforms = transforms.Compose(
+        [
+            transforms.ToTensor(),
+            transforms.Normalize([0.5], [0.5]),
+        ]
+    )
 
     def preprocess_train(examples):
         images = [image.convert("RGB") for image in examples[image_column]]
@@ -905,73 +966,44 @@ def main(args):
         examples["original_sizes"] = original_sizes
         examples["crop_top_lefts"] = crop_top_lefts
         examples["pixel_values"] = all_images
+        tokens_one, tokens_two = tokenize_captions(examples)
+        examples["input_ids_one"] = tokens_one
+        examples["input_ids_two"] = tokens_two
+        if args.debug_loss:
+            fnames = [os.path.basename(image.filename) for image in examples[image_column] if image.filename]
+            if fnames:
+                examples["filenames"] = fnames
         return examples
 
     with accelerator.main_process_first():
         if args.max_train_samples is not None:
             dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples))
         # Set the training transforms
-        train_dataset = dataset["train"].with_transform(preprocess_train)
-
-    # Let's first compute all the embeddings so that we can free up the text encoders
-    # from memory. We will pre-compute the VAE encodings too.
-    text_encoders = [text_encoder_one, text_encoder_two]
-    tokenizers = [tokenizer_one, tokenizer_two]
-    compute_embeddings_fn = functools.partial(
-        encode_prompt,
-        text_encoders=text_encoders,
-        tokenizers=tokenizers,
-        proportion_empty_prompts=args.proportion_empty_prompts,
-        caption_column=args.caption_column,
-    )
-    compute_vae_encodings_fn = functools.partial(compute_vae_encodings, vae=vae)
-    with accelerator.main_process_first():
-        from datasets.fingerprint import Hasher
-
-        # fingerprint used by the cache for the other processes to load the result
-        # details: https://github.com/huggingface/diffusers/pull/4038#discussion_r1266078401
-        new_fingerprint = Hasher.hash(args)
-        new_fingerprint_for_vae = Hasher.hash((vae_path, args))
-        train_dataset_with_embeddings = train_dataset.map(
-            compute_embeddings_fn, batched=True, new_fingerprint=new_fingerprint
-        )
-        train_dataset_with_vae = train_dataset.map(
-            compute_vae_encodings_fn,
-            batched=True,
-            batch_size=args.train_batch_size,
-            new_fingerprint=new_fingerprint_for_vae,
-        )
-        precomputed_dataset = concatenate_datasets(
-            [train_dataset_with_embeddings, train_dataset_with_vae.remove_columns(["image", "text"])], axis=1
-        )
-        precomputed_dataset = precomputed_dataset.with_transform(preprocess_train)
-
-    del compute_vae_encodings_fn, compute_embeddings_fn, text_encoder_one, text_encoder_two
-    del text_encoders, tokenizers, vae
-    gc.collect()
-    if is_torch_npu_available():
-        torch_npu.npu.empty_cache()
-    elif torch.cuda.is_available():
-        torch.cuda.empty_cache()
+        train_dataset = dataset["train"].with_transform(preprocess_train, output_all_columns=True)
 
     def collate_fn(examples):
-        model_input = torch.stack([torch.tensor(example["model_input"]) for example in examples])
+        pixel_values = torch.stack([example["pixel_values"] for example in examples])
+        pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
         original_sizes = [example["original_sizes"] for example in examples]
         crop_top_lefts = [example["crop_top_lefts"] for example in examples]
-        prompt_embeds = torch.stack([torch.tensor(example["prompt_embeds"]) for example in examples])
-        pooled_prompt_embeds = torch.stack([torch.tensor(example["pooled_prompt_embeds"]) for example in examples])
-
-        return {
-            "model_input": model_input,
-            "prompt_embeds": prompt_embeds,
-            "pooled_prompt_embeds": pooled_prompt_embeds,
+        input_ids_one = torch.stack([example["input_ids_one"] for example in examples])
+        input_ids_two = torch.stack([example["input_ids_two"] for example in examples])
+        result = {
+            "pixel_values": pixel_values,
+            "input_ids_one": input_ids_one,
+            "input_ids_two": input_ids_two,
             "original_sizes": original_sizes,
             "crop_top_lefts": crop_top_lefts,
         }
 
+        filenames = [example["filenames"] for example in examples if "filenames" in example]
+        if filenames:
+            result["filenames"] = filenames
+        return result
+
     # DataLoaders creation:
     train_dataloader = torch.utils.data.DataLoader(
-        precomputed_dataset,
+        train_dataset,
         shuffle=True,
         collate_fn=collate_fn,
         batch_size=args.train_batch_size,
@@ -993,12 +1025,14 @@ def main(args):
     )
 
     # Prepare everything with our `accelerator`.
-    unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-        unet, optimizer, train_dataloader, lr_scheduler
-    )
-
-    if args.use_ema:
-        ema_unet.to(accelerator.device)
+    if args.train_text_encoder:
+        unet, text_encoder_one, text_encoder_two, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+            unet, text_encoder_one, text_encoder_two, optimizer, train_dataloader, lr_scheduler
+        )
+    else:
+        unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+            unet, optimizer, train_dataloader, lr_scheduler
+        )
 
     # We need to recalculate our total training steps as the size of the training dataloader may have changed.
     num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
@@ -1010,24 +1044,13 @@ def main(args):
     # We need to initialize the trackers we use, and also store our configuration.
     # The trackers initializes automatically on the main process.
     if accelerator.is_main_process:
-        accelerator.init_trackers("text2image-fine-tune-sdxl", config=vars(args))
-
-    # Function for unwrapping if torch.compile() was used in accelerate.
-    def unwrap_model(model):
-        model = accelerator.unwrap_model(model)
-        model = model._orig_mod if is_compiled_module(model) else model
-        return model
-
-    if torch.backends.mps.is_available() or "playground" in args.pretrained_model_name_or_path:
-        autocast_ctx = nullcontext()
-    else:
-        autocast_ctx = torch.autocast(accelerator.device.type)
+        accelerator.init_trackers("text2image-fine-tune", config=vars(args))
 
     # Train!
     total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
 
     logger.info("***** Running training *****")
-    logger.info(f"  Num examples = {len(precomputed_dataset)}")
+    logger.info(f"  Num examples = {len(train_dataset)}")
     logger.info(f"  Num Epochs = {args.num_train_epochs}")
     logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
     logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
@@ -1073,11 +1096,25 @@ def main(args):
     )
 
     for epoch in range(first_epoch, args.num_train_epochs):
+        unet.train()
+        if args.train_text_encoder:
+            text_encoder_one.train()
+            text_encoder_two.train()
         train_loss = 0.0
         for step, batch in enumerate(train_dataloader):
             with accelerator.accumulate(unet):
+                # Convert images to latent space
+                if args.pretrained_vae_model_name_or_path is not None:
+                    pixel_values = batch["pixel_values"].to(dtype=weight_dtype)
+                else:
+                    pixel_values = batch["pixel_values"]
+
+                model_input = vae.encode(pixel_values).latent_dist.sample()
+                model_input = model_input * vae.config.scaling_factor
+                if args.pretrained_vae_model_name_or_path is None:
+                    model_input = model_input.to(weight_dtype)
+
                 # Sample noise that we'll add to the latents
-                model_input = batch["model_input"].to(accelerator.device)
                 noise = torch.randn_like(model_input)
                 if args.noise_offset:
                     # https://www.crosslabs.org//blog/diffusion-with-offset-noise
@@ -1086,29 +1123,23 @@ def main(args):
                     )
 
                 bsz = model_input.shape[0]
-                if args.timestep_bias_strategy == "none":
-                    # Sample a random timestep for each image without bias.
-                    timesteps = torch.randint(
-                        0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
-                    )
-                else:
-                    # Sample a random timestep for each image, potentially biased by the timestep weights.
-                    # Biasing the timestep weights allows us to spend less time training irrelevant timesteps.
-                    weights = generate_timestep_weights(args, noise_scheduler.config.num_train_timesteps).to(
-                        model_input.device
-                    )
-                    timesteps = torch.multinomial(weights, bsz, replacement=True).long()
+                # Sample a random timestep for each image
+                timesteps = torch.randint(
+                    0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
+                )
+                timesteps = timesteps.long()
 
                 # Add noise to the model input according to the noise magnitude at each timestep
                 # (this is the forward diffusion process)
-                noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps).to(dtype=weight_dtype)
+                noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps)
 
                 # time ids
                 def compute_time_ids(original_size, crops_coords_top_left):
                     # Adapted from pipeline.StableDiffusionXLPipeline._get_add_time_ids
                     target_size = (args.resolution, args.resolution)
                     add_time_ids = list(original_size + crops_coords_top_left + target_size)
-                    add_time_ids = torch.tensor([add_time_ids], device=accelerator.device, dtype=weight_dtype)
+                    add_time_ids = torch.tensor([add_time_ids])
+                    add_time_ids = add_time_ids.to(accelerator.device, dtype=weight_dtype)
                     return add_time_ids
 
                 add_time_ids = torch.cat(
@@ -1117,8 +1148,12 @@ def main(args):
 
                 # Predict the noise residual
                 unet_added_conditions = {"time_ids": add_time_ids}
-                prompt_embeds = batch["prompt_embeds"].to(accelerator.device, dtype=weight_dtype)
-                pooled_prompt_embeds = batch["pooled_prompt_embeds"].to(accelerator.device)
+                prompt_embeds, pooled_prompt_embeds = encode_prompt(
+                    text_encoders=[text_encoder_one, text_encoder_two],
+                    tokenizers=None,
+                    prompt=None,
+                    text_input_ids_list=[batch["input_ids_one"], batch["input_ids_two"]],
+                )
                 unet_added_conditions.update({"text_embeds": pooled_prompt_embeds})
                 model_pred = unet(
                     noisy_model_input,
@@ -1137,11 +1172,6 @@ def main(args):
                     target = noise
                 elif noise_scheduler.config.prediction_type == "v_prediction":
                     target = noise_scheduler.get_velocity(model_input, noise, timesteps)
-                elif noise_scheduler.config.prediction_type == "sample":
-                    # We set the target to latents here, but the model_pred will return the noise sample prediction.
-                    target = model_input
-                    # We will have to subtract the noise residual from the prediction to get the target sample.
-                    model_pred = model_pred - noise
                 else:
                     raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
 
@@ -1163,7 +1193,9 @@ def main(args):
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                     loss = loss.mean()
-
+                if args.debug_loss and "filenames" in batch:
+                    for fname in batch["filenames"]:
+                        accelerator.log({"loss_for_" + fname: loss}, step=global_step)
                 # Gather the losses across all processes for logging (if we use distributed training).
                 avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean()
                 train_loss += avg_loss.item() / args.gradient_accumulation_steps
@@ -1171,16 +1203,13 @@ def main(args):
                 # Backpropagate
                 accelerator.backward(loss)
                 if accelerator.sync_gradients:
-                    params_to_clip = unet.parameters()
-                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
+                    accelerator.clip_grad_norm_(params_to_optimize, args.max_grad_norm)
                 optimizer.step()
                 lr_scheduler.step()
                 optimizer.zero_grad()
 
             # Checks if the accelerator has performed an optimization step behind the scenes
             if accelerator.sync_gradients:
-                if args.use_ema:
-                    ema_unet.step(unet.parameters())
                 progress_bar.update(1)
                 global_step += 1
                 accelerator.log({"train_loss": train_loss}, step=global_step)
@@ -1221,137 +1250,80 @@ def main(args):
 
         if accelerator.is_main_process:
             if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
-                logger.info(
-                    f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
-                    f" {args.validation_prompt}."
-                )
-                if args.use_ema:
-                    # Store the UNet parameters temporarily and load the EMA parameters to perform inference.
-                    ema_unet.store(unet.parameters())
-                    ema_unet.copy_to(unet.parameters())
-
                 # create pipeline
-                vae = AutoencoderKL.from_pretrained(
-                    vae_path,
-                    subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
-                    revision=args.revision,
-                    variant=args.variant,
-                )
                 pipeline = StableDiffusionXLPipeline.from_pretrained(
                     args.pretrained_model_name_or_path,
                     vae=vae,
-                    unet=accelerator.unwrap_model(unet),
+                    text_encoder=unwrap_model(text_encoder_one),
+                    text_encoder_2=unwrap_model(text_encoder_two),
+                    unet=unwrap_model(unet),
                     revision=args.revision,
                     variant=args.variant,
                     torch_dtype=weight_dtype,
                 )
-                if args.prediction_type is not None:
-                    scheduler_args = {"prediction_type": args.prediction_type}
-                    pipeline.scheduler = pipeline.scheduler.from_config(pipeline.scheduler.config, **scheduler_args)
 
-                pipeline = pipeline.to(accelerator.device)
-                pipeline.set_progress_bar_config(disable=True)
-
-                # run inference
-                generator = (
-                    torch.Generator(device=accelerator.device).manual_seed(args.seed)
-                    if args.seed is not None
-                    else None
-                )
-                pipeline_args = {"prompt": args.validation_prompt}
-
-                with autocast_ctx:
-                    images = [
-                        pipeline(**pipeline_args, generator=generator, num_inference_steps=25).images[0]
-                        for _ in range(args.num_validation_images)
-                    ]
-
-                for tracker in accelerator.trackers:
-                    if tracker.name == "tensorboard":
-                        np_images = np.stack([np.asarray(img) for img in images])
-                        tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
-                    if tracker.name == "wandb":
-                        tracker.log(
-                            {
-                                "validation": [
-                                    wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
-                                    for i, image in enumerate(images)
-                                ]
-                            }
-                        )
+                images = log_validation(pipeline, args, accelerator, epoch)
 
                 del pipeline
-                if is_torch_npu_available():
-                    torch_npu.npu.empty_cache()
-                elif torch.cuda.is_available():
-                    torch.cuda.empty_cache()
-
-                if args.use_ema:
-                    # Switch back to the original UNet parameters.
-                    ema_unet.restore(unet.parameters())
+                torch.cuda.empty_cache()
 
+    # Save the lora layers
     accelerator.wait_for_everyone()
     if accelerator.is_main_process:
         unet = unwrap_model(unet)
-        if args.use_ema:
-            ema_unet.copy_to(unet.parameters())
+        unet_lora_state_dict = convert_state_dict_to_diffusers(get_peft_model_state_dict(unet))
 
-        # Serialize pipeline.
-        vae = AutoencoderKL.from_pretrained(
-            vae_path,
-            subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
-            revision=args.revision,
-            variant=args.variant,
-            torch_dtype=weight_dtype,
+        if args.train_text_encoder:
+            text_encoder_one = unwrap_model(text_encoder_one)
+            text_encoder_two = unwrap_model(text_encoder_two)
+
+            text_encoder_lora_layers = convert_state_dict_to_diffusers(get_peft_model_state_dict(text_encoder_one))
+            text_encoder_2_lora_layers = convert_state_dict_to_diffusers(get_peft_model_state_dict(text_encoder_two))
+        else:
+            text_encoder_lora_layers = None
+            text_encoder_2_lora_layers = None
+
+        StableDiffusionXLPipeline.save_lora_weights(
+            save_directory=args.output_dir,
+            unet_lora_layers=unet_lora_state_dict,
+            text_encoder_lora_layers=text_encoder_lora_layers,
+            text_encoder_2_lora_layers=text_encoder_2_lora_layers,
         )
+
+        del unet
+        del text_encoder_one
+        del text_encoder_two
+        del text_encoder_lora_layers
+        del text_encoder_2_lora_layers
+        torch.cuda.empty_cache()
+
+        # Final inference
+        # Make sure vae.dtype is consistent with the unet.dtype
+        if args.mixed_precision == "fp16":
+            vae.to(weight_dtype)
+        # Load previous pipeline
         pipeline = StableDiffusionXLPipeline.from_pretrained(
             args.pretrained_model_name_or_path,
-            unet=unet,
             vae=vae,
             revision=args.revision,
             variant=args.variant,
             torch_dtype=weight_dtype,
         )
-        if args.prediction_type is not None:
-            scheduler_args = {"prediction_type": args.prediction_type}
-            pipeline.scheduler = pipeline.scheduler.from_config(pipeline.scheduler.config, **scheduler_args)
-        pipeline.save_pretrained(args.output_dir)
+
+        # load attention processors
+        pipeline.load_lora_weights(args.output_dir)
 
         # run inference
-        images = []
         if args.validation_prompt and args.num_validation_images > 0:
-            pipeline = pipeline.to(accelerator.device)
-            generator = (
-                torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
-            )
-
-            with autocast_ctx:
-                images = [
-                    pipeline(args.validation_prompt, num_inference_steps=25, generator=generator).images[0]
-                    for _ in range(args.num_validation_images)
-                ]
-
-            for tracker in accelerator.trackers:
-                if tracker.name == "tensorboard":
-                    np_images = np.stack([np.asarray(img) for img in images])
-                    tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC")
-                if tracker.name == "wandb":
-                    tracker.log(
-                        {
-                            "test": [
-                                wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
-                                for i, image in enumerate(images)
-                            ]
-                        }
-                    )
+            images = log_validation(pipeline, args, accelerator, epoch, is_final_validation=True)
 
         if args.push_to_hub:
             save_model_card(
-                repo_id=repo_id,
+                repo_id,
                 images=images,
-                validation_prompt=args.validation_prompt,
                 base_model=args.pretrained_model_name_or_path,
                 dataset_name=args.dataset_name,
+                train_text_encoder=args.train_text_encoder,
                 repo_folder=args.output_dir,
                 vae_path=args.pretrained_vae_model_name_or_path,
             )