update

* AutoModel

* ...

* lol

* ...

* add test

* update

* make fix-copies

---------

Co-authored-by: Dhruv Nair <dhruv.nair@gmail.com>

.github/workflows/nightly_tests.yml

@@ -417,7 +417,7 @@ jobs:
             additional_deps: ["peft"]
           - backend: "gguf"
             test_location: "gguf"
-            additional_deps: []
+            additional_deps: ["peft"]
           - backend: "torchao"
             test_location: "torchao"
             additional_deps: []

docs/source/en/api/pipelines/deepfloyd_if.md

@@ -14,6 +14,7 @@ specific language governing permissions and limitations under the License.
 
 <div class="flex flex-wrap space-x-1">
   <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
 </div>
 
 ## Overview

docs/source/en/api/pipelines/flux.md

@@ -14,6 +14,7 @@ specific language governing permissions and limitations under the License.
 
 <div class="flex flex-wrap space-x-1">
   <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
 </div>
 
 Flux is a series of text-to-image generation models based on diffusion transformers. To know more about Flux, check out the original [blog post](https://blackforestlabs.ai/announcing-black-forest-labs/) by the creators of Flux, Black Forest Labs.

docs/source/en/api/pipelines/kolors.md

@@ -14,6 +14,7 @@ specific language governing permissions and limitations under the License.
 
 <div class="flex flex-wrap space-x-1">
   <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
 </div>
 
 ![](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/kolors/kolors_header_collage.png)

docs/source/en/api/pipelines/ltx_video.md

@@ -16,6 +16,7 @@
 
 <div class="flex flex-wrap space-x-1">
   <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
 </div>
 
 [LTX Video](https://huggingface.co/Lightricks/LTX-Video) is the first DiT-based video generation model capable of generating high-quality videos in real-time. It produces 24 FPS videos at a 768x512 resolution faster than they can be watched. Trained on a large-scale dataset of diverse videos, the model generates high-resolution videos with realistic and varied content. We provide a model for both text-to-video as well as image + text-to-video usecases.

docs/source/en/api/pipelines/sana.md

@@ -16,6 +16,7 @@
 
 <div class="flex flex-wrap space-x-1">
   <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
 </div>
 
 [SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers](https://huggingface.co/papers/2410.10629) from NVIDIA and MIT HAN Lab, by Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, Song Han.

docs/source/en/api/pipelines/sana_sprint.md

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License. -->
 
-# SanaSprintPipeline
+# SANA-Sprint
 
 <div class="flex flex-wrap space-x-1">
   <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>

docs/source/en/api/pipelines/stable_diffusion/stable_diffusion_3.md

@@ -14,6 +14,7 @@ specific language governing permissions and limitations under the License.
 
 <div class="flex flex-wrap space-x-1">
   <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
 </div>
 
 Stable Diffusion 3 (SD3) was proposed in [Scaling Rectified Flow Transformers for High-Resolution Image Synthesis](https://arxiv.org/pdf/2403.03206.pdf) by Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Muller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach.

docs/source/en/api/pipelines/stable_diffusion/stable_diffusion_xl.md

@@ -14,6 +14,7 @@ specific language governing permissions and limitations under the License.
 
 <div class="flex flex-wrap space-x-1">
   <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
 </div>
 
 Stable Diffusion XL (SDXL) was proposed in [SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis](https://huggingface.co/papers/2307.01952) by Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach.

docs/source/en/optimization/memory.md

@@ -178,6 +178,9 @@ pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch
 # We can utilize the enable_group_offload method for Diffusers model implementations
 pipe.transformer.enable_group_offload(onload_device=onload_device, offload_device=offload_device, offload_type="leaf_level", use_stream=True)
 
+# Uncomment the following to also allow recording the current streams.
+# pipe.transformer.enable_group_offload(onload_device=onload_device, offload_device=offload_device, offload_type="leaf_level", use_stream=True, record_stream=True)
+
 # For any other model implementations, the apply_group_offloading function can be used
 apply_group_offloading(pipe.text_encoder, onload_device=onload_device, offload_type="block_level", num_blocks_per_group=2)
 apply_group_offloading(pipe.vae, onload_device=onload_device, offload_type="leaf_level")
@@ -205,6 +208,7 @@ Group offloading (for CUDA devices with support for asynchronous data transfer s
 - The `use_stream` parameter can be used with CUDA devices to enable prefetching layers for onload. It defaults to `False`. Layer prefetching allows overlapping computation and data transfer of model weights, which drastically reduces the overall execution time compared to other offloading methods. However, it can increase the CPU RAM usage significantly. Ensure that available CPU RAM that is at least twice the size of the model when setting `use_stream=True`. You can find more information about CUDA streams [here](https://pytorch.org/docs/stable/generated/torch.cuda.Stream.html)
 - If specifying `use_stream=True` on VAEs with tiling enabled, make sure to do a dummy forward pass (possibly with dummy inputs) before the actual inference to avoid device-mismatch errors. This may not work on all implementations. Please open an issue if you encounter any problems.
 - The parameter `low_cpu_mem_usage` can be set to `True` to reduce CPU memory usage when using streams for group offloading. This is useful when the CPU memory is the bottleneck, but it may counteract the benefits of using streams and increase the overall execution time. The CPU memory savings come from creating pinned-tensors on-the-fly instead of pre-pinning them. This parameter is better suited for using `leaf_level` offloading.
+- When using `use_stream=True`, users can additionally specify `record_stream=True` to get better speedups at the expense of slightly increased memory usage. Refer to the [official PyTorch docs](https://pytorch.org/docs/stable/generated/torch.Tensor.record_stream.html) to know more about this.
 
 For more information about available parameters and an explanation of how group offloading works, refer to [`~hooks.group_offloading.apply_group_offloading`].
 

docs/source/en/optimization/mps.md

@@ -12,6 +12,9 @@ specific language governing permissions and limitations under the License.
 
 # Metal Performance Shaders (MPS)
 
+> [!TIP]
+> Pipelines with a <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22"> badge indicate a model can take advantage of the MPS backend on Apple silicon devices for faster inference. Feel free to open a [Pull Request](https://github.com/huggingface/diffusers/compare) to add this badge to pipelines that are missing it.
+
 🤗 Diffusers is compatible with Apple silicon (M1/M2 chips) using the PyTorch [`mps`](https://pytorch.org/docs/stable/notes/mps.html) device, which uses the Metal framework to leverage the GPU on MacOS devices. You'll need to have:
 
 - macOS computer with Apple silicon (M1/M2) hardware
@@ -37,7 +40,7 @@ image
 
 <Tip warning={true}>
 
-Generating multiple prompts in a batch can [crash](https://github.com/huggingface/diffusers/issues/363) or fail to work reliably. We believe this is related to the [`mps`](https://github.com/pytorch/pytorch/issues/84039) backend in PyTorch. While this is being investigated, you should iterate instead of batching.
+The PyTorch [mps](https://pytorch.org/docs/stable/notes/mps.html) backend does not support NDArray sizes greater than `2**32`. Please open an [Issue](https://github.com/huggingface/diffusers/issues/new/choose) if you encounter this problem so we can investigate.
 
 </Tip>
 
@@ -59,6 +62,10 @@ If you're using **PyTorch 1.13**, you need to "prime" the pipeline with an addit
 
 ## Troubleshoot
 
+This section lists some common issues with using the `mps` backend and how to solve them.
+
+### Attention slicing
+
 M1/M2 performance is very sensitive to memory pressure. When this occurs, the system automatically swaps if it needs to which significantly degrades performance.
 
 To prevent this from happening, we recommend *attention slicing* to reduce memory pressure during inference and prevent swapping. This is especially relevant if your computer has less than 64GB of system RAM, or if you generate images at non-standard resolutions larger than 512×512 pixels. Call the [`~DiffusionPipeline.enable_attention_slicing`] function on your pipeline:
@@ -72,3 +79,7 @@ pipeline.enable_attention_slicing()
 ```
 
 Attention slicing performs the costly attention operation in multiple steps instead of all at once. It usually improves performance by ~20% in computers without universal memory, but we've observed *better performance* in most Apple silicon computers unless you have 64GB of RAM or more.
+
+### Batch inference
+
+Generating multiple prompts in a batch can crash or fail to work reliably. If this is the case, try iterating instead of batching.

docs/source/en/using-diffusers/loading.md

@@ -105,7 +105,7 @@ import torch
 
 pipe = HunyuanVideoPipeline.from_pretrained(
     "hunyuanvideo-community/HunyuanVideo",
-    torch_dtype={'transformer': torch.bfloat16, 'default': torch.float16},
+    torch_dtype={"transformer": torch.bfloat16, "default": torch.float16},
 )
 print(pipe.transformer.dtype, pipe.vae.dtype)  # (torch.bfloat16, torch.float16)
 ```

docs/source/en/using-diffusers/loading_adapters.md

@@ -194,6 +194,59 @@ Currently, [`~loaders.StableDiffusionLoraLoaderMixin.set_adapters`] only support
 
 </Tip>
 
+### Hotswapping LoRA adapters
+
+A common use case when serving multiple adapters is to load one adapter first, generate images, load another adapter, generate more images, load another adapter, etc. This workflow normally requires calling [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`], [`~loaders.StableDiffusionLoraLoaderMixin.set_adapters`], and possibly [`~loaders.peft.PeftAdapterMixin.delete_adapters`] to save memory. Moreover, if the model is compiled using `torch.compile`, performing these steps requires recompilation, which takes time.
+
+To better support this common workflow, you can "hotswap" a LoRA adapter, to avoid accumulating memory and in some cases, recompilation. It requires an adapter to already be loaded, and the new adapter weights are swapped in-place for the existing adapter.
+
+Pass `hotswap=True` when loading a LoRA adapter to enable this feature. It is important to indicate the name of the existing adapter, (`default_0` is the default adapter name), to be swapped. If you loaded the first adapter with a different name, use that name instead.
+
+```python
+pipe = ...
+# load adapter 1 as normal
+pipeline.load_lora_weights(file_name_adapter_1)
+# generate some images with adapter 1
+...
+# now hot swap the 2nd adapter
+pipeline.load_lora_weights(file_name_adapter_2, hotswap=True, adapter_name="default_0")
+# generate images with adapter 2
+```
+
+
+<Tip warning={true}>
+
+Hotswapping is not currently supported for LoRA adapters that target the text encoder.
+
+</Tip>
+
+For compiled models, it is often (though not always if the second adapter targets identical LoRA ranks and scales) necessary to call [`~loaders.lora_base.LoraBaseMixin.enable_lora_hotswap`] to avoid recompilation. Use [`~loaders.lora_base.LoraBaseMixin.enable_lora_hotswap`] _before_ loading the first adapter, and `torch.compile` should be called _after_ loading the first adapter.
+
+```python
+pipe = ...
+# call this extra method
+pipe.enable_lora_hotswap(target_rank=max_rank)
+# now load adapter 1
+pipe.load_lora_weights(file_name_adapter_1)
+# now compile the unet of the pipeline
+pipe.unet = torch.compile(pipeline.unet, ...)
+# generate some images with adapter 1
+...
+# now hot swap adapter 2
+pipeline.load_lora_weights(file_name_adapter_2, hotswap=True, adapter_name="default_0")
+# generate images with adapter 2
+```
+
+The `target_rank=max_rank` argument is important for setting the maximum rank among all LoRA adapters that will be loaded. If you have one adapter with rank 8 and another with rank 16, pass `target_rank=16`. You should use a higher value if in doubt. By default, this value is 128.
+
+However, there can be situations where recompilation is unavoidable. For example, if the hotswapped adapter targets more layers than the initial adapter, then recompilation is triggered. Try to load the adapter that targets the most layers first. Refer to the PEFT docs on [hotswapping](https://huggingface.co/docs/peft/main/en/package_reference/hotswap#peft.utils.hotswap.hotswap_adapter) for more details about the limitations of this feature.
+
+<Tip>
+
+Move your code inside the `with torch._dynamo.config.patch(error_on_recompile=True)` context manager to detect if a model was recompiled. If you detect recompilation despite following all the steps above, please open an issue with [Diffusers](https://github.com/huggingface/diffusers/issues) with a reproducible example.
+
+</Tip>
+
 ### Kohya and TheLastBen
 
 Other popular LoRA trainers from the community include those by [Kohya](https://github.com/kohya-ss/sd-scripts/) and [TheLastBen](https://github.com/TheLastBen/fast-stable-diffusion). These trainers create different LoRA checkpoints than those trained by 🤗 Diffusers, but they can still be loaded in the same way.

examples/advanced_diffusion_training/requirements.txt

@@ -1,7 +1,8 @@
-accelerate>=0.16.0
+accelerate>=0.31.0
 torchvision
-transformers>=4.25.1
+transformers>=4.41.2
 ftfy
 tensorboard
 Jinja2
-peft==0.7.0
+peft>=0.11.1
+sentencepiece

examples/advanced_diffusion_training/train_dreambooth_lora_flux_advanced.py

@@ -24,7 +24,7 @@ import re
 import shutil
 from contextlib import nullcontext
 from pathlib import Path
-from typing import List, Optional, Union
+from typing import List, Optional
 
 import numpy as np
 import torch
@@ -228,10 +228,20 @@ def log_validation(
 
     # run inference
     generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
-    autocast_ctx = nullcontext()
+    autocast_ctx = torch.autocast(accelerator.device.type) if not is_final_validation else nullcontext()
 
-    with autocast_ctx:
-        images = [pipeline(**pipeline_args, generator=generator).images[0] for _ in range(args.num_validation_images)]
+    # pre-calculate  prompt embeds, pooled prompt embeds, text ids because t5 does not support autocast
+    with torch.no_grad():
+        prompt_embeds, pooled_prompt_embeds, text_ids = pipeline.encode_prompt(
+            pipeline_args["prompt"], prompt_2=pipeline_args["prompt"]
+        )
+    images = []
+    for _ in range(args.num_validation_images):
+        with autocast_ctx:
+            image = pipeline(
+                prompt_embeds=prompt_embeds, pooled_prompt_embeds=pooled_prompt_embeds, generator=generator
+            ).images[0]
+            images.append(image)
 
     for tracker in accelerator.trackers:
         phase_name = "test" if is_final_validation else "validation"
@@ -657,6 +667,7 @@ def parse_args(input_args=None):
     parser.add_argument(
         "--adam_weight_decay_text_encoder", type=float, default=1e-03, help="Weight decay to use for text_encoder"
     )
+
     parser.add_argument(
         "--lora_layers",
         type=str,
@@ -666,6 +677,7 @@ def parse_args(input_args=None):
             'E.g. - "to_k,to_q,to_v,to_out.0" will result in lora training of attention layers only. For more examples refer to https://github.com/huggingface/diffusers/blob/main/examples/advanced_diffusion_training/README_flux.md'
         ),
     )
+
     parser.add_argument(
         "--adam_epsilon",
         type=float,
@@ -738,6 +750,15 @@ def parse_args(input_args=None):
             " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
         ),
     )
+    parser.add_argument(
+        "--upcast_before_saving",
+        action="store_true",
+        default=False,
+        help=(
+            "Whether to upcast the trained transformer layers to float32 before saving (at the end of training). "
+            "Defaults to precision dtype used for training to save memory"
+        ),
+    )
     parser.add_argument(
         "--prior_generation_precision",
         type=str,
@@ -818,9 +839,9 @@ class TokenEmbeddingsHandler:
         idx = 0
         for tokenizer, text_encoder in zip(self.tokenizers, self.text_encoders):
             assert isinstance(inserting_toks, list), "inserting_toks should be a list of strings."
-            assert all(
-                isinstance(tok, str) for tok in inserting_toks
-            ), "All elements in inserting_toks should be strings."
+            assert all(isinstance(tok, str) for tok in inserting_toks), (
+                "All elements in inserting_toks should be strings."
+            )
 
             self.inserting_toks = inserting_toks
             special_tokens_dict = {"additional_special_tokens": self.inserting_toks}
@@ -1147,7 +1168,7 @@ def tokenize_prompt(tokenizer, prompt, max_sequence_length, add_special_tokens=F
     return text_input_ids
 
 
-def _get_t5_prompt_embeds(
+def _encode_prompt_with_t5(
     text_encoder,
     tokenizer,
     max_sequence_length=512,
@@ -1176,7 +1197,10 @@ def _get_t5_prompt_embeds(
 
     prompt_embeds = text_encoder(text_input_ids.to(device))[0]
 
-    dtype = text_encoder.dtype
+    if hasattr(text_encoder, "module"):
+        dtype = text_encoder.module.dtype
+    else:
+        dtype = text_encoder.dtype
     prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
 
     _, seq_len, _ = prompt_embeds.shape
@@ -1188,7 +1212,7 @@ def _get_t5_prompt_embeds(
     return prompt_embeds
 
 
-def _get_clip_prompt_embeds(
+def _encode_prompt_with_clip(
     text_encoder,
     tokenizer,
     prompt: str,
@@ -1217,9 +1241,13 @@ def _get_clip_prompt_embeds(
 
     prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=False)
 
+    if hasattr(text_encoder, "module"):
+        dtype = text_encoder.module.dtype
+    else:
+        dtype = text_encoder.dtype
     # Use pooled output of CLIPTextModel
     prompt_embeds = prompt_embeds.pooler_output
-    prompt_embeds = prompt_embeds.to(dtype=text_encoder.dtype, device=device)
+    prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
 
     # duplicate text embeddings for each generation per prompt, using mps friendly method
     prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
@@ -1238,136 +1266,35 @@ def encode_prompt(
     text_input_ids_list=None,
 ):
     prompt = [prompt] if isinstance(prompt, str) else prompt
-    batch_size = len(prompt)
-    dtype = text_encoders[0].dtype
+    if hasattr(text_encoders[0], "module"):
+        dtype = text_encoders[0].module.dtype
+    else:
+        dtype = text_encoders[0].dtype
 
-    pooled_prompt_embeds = _get_clip_prompt_embeds(
+    pooled_prompt_embeds = _encode_prompt_with_clip(
         text_encoder=text_encoders[0],
         tokenizer=tokenizers[0],
         prompt=prompt,
         device=device if device is not None else text_encoders[0].device,
         num_images_per_prompt=num_images_per_prompt,
-        text_input_ids=text_input_ids_list[0] if text_input_ids_list is not None else None,
+        text_input_ids=text_input_ids_list[0] if text_input_ids_list else None,
     )
 
-    prompt_embeds = _get_t5_prompt_embeds(
+    prompt_embeds = _encode_prompt_with_t5(
         text_encoder=text_encoders[1],
         tokenizer=tokenizers[1],
         max_sequence_length=max_sequence_length,
         prompt=prompt,
         num_images_per_prompt=num_images_per_prompt,
         device=device if device is not None else text_encoders[1].device,
-        text_input_ids=text_input_ids_list[1] if text_input_ids_list is not None else None,
+        text_input_ids=text_input_ids_list[1] if text_input_ids_list else None,
     )
 
-    text_ids = torch.zeros(batch_size, prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
-    text_ids = text_ids.repeat(num_images_per_prompt, 1, 1)
+    text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
 
     return prompt_embeds, pooled_prompt_embeds, text_ids
 
 
-# CustomFlowMatchEulerDiscreteScheduler was taken from ostris ai-toolkit trainer:
-# https://github.com/ostris/ai-toolkit/blob/9ee1ef2a0a2a9a02b92d114a95f21312e5906e54/toolkit/samplers/custom_flowmatch_sampler.py#L95
-class CustomFlowMatchEulerDiscreteScheduler(FlowMatchEulerDiscreteScheduler):
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        with torch.no_grad():
-            # create weights for timesteps
-            num_timesteps = 1000
-
-            # generate the multiplier based on cosmap loss weighing
-            # this is only used on linear timesteps for now
-
-            # cosine map weighing is higher in the middle and lower at the ends
-            # bot = 1 - 2 * self.sigmas + 2 * self.sigmas ** 2
-            # cosmap_weighing = 2 / (math.pi * bot)
-
-            # sigma sqrt weighing is significantly higher at the end and lower at the beginning
-            sigma_sqrt_weighing = (self.sigmas**-2.0).float()
-            # clip at 1e4 (1e6 is too high)
-            sigma_sqrt_weighing = torch.clamp(sigma_sqrt_weighing, max=1e4)
-            # bring to a mean of 1
-            sigma_sqrt_weighing = sigma_sqrt_weighing / sigma_sqrt_weighing.mean()
-
-            # Create linear timesteps from 1000 to 0
-            timesteps = torch.linspace(1000, 0, num_timesteps, device="cpu")
-
-            self.linear_timesteps = timesteps
-            # self.linear_timesteps_weights = cosmap_weighing
-            self.linear_timesteps_weights = sigma_sqrt_weighing
-
-            # self.sigmas = self.get_sigmas(timesteps, n_dim=1, dtype=torch.float32, device='cpu')
-            pass
-
-    def get_weights_for_timesteps(self, timesteps: torch.Tensor) -> torch.Tensor:
-        # Get the indices of the timesteps
-        step_indices = [(self.timesteps == t).nonzero().item() for t in timesteps]
-
-        # Get the weights for the timesteps
-        weights = self.linear_timesteps_weights[step_indices].flatten()
-
-        return weights
-
-    def get_sigmas(self, timesteps: torch.Tensor, n_dim, dtype, device) -> torch.Tensor:
-        sigmas = self.sigmas.to(device=device, dtype=dtype)
-        schedule_timesteps = self.timesteps.to(device)
-        timesteps = timesteps.to(device)
-        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
-
-        sigma = sigmas[step_indices].flatten()
-        while len(sigma.shape) < n_dim:
-            sigma = sigma.unsqueeze(-1)
-
-        return sigma
-
-    def add_noise(
-        self,
-        original_samples: torch.Tensor,
-        noise: torch.Tensor,
-        timesteps: torch.Tensor,
-    ) -> torch.Tensor:
-        ## ref https://github.com/huggingface/diffusers/blob/fbe29c62984c33c6cf9cf7ad120a992fe6d20854/examples/dreambooth/train_dreambooth_sd3.py#L1578
-        ## Add noise according to flow matching.
-        ## zt = (1 - texp) * x + texp * z1
-
-        # sigmas = get_sigmas(timesteps, n_dim=model_input.ndim, dtype=model_input.dtype)
-        # noisy_model_input = (1.0 - sigmas) * model_input + sigmas * noise
-
-        # timestep needs to be in [0, 1], we store them in [0, 1000]
-        # noisy_sample = (1 - timestep) * latent + timestep * noise
-        t_01 = (timesteps / 1000).to(original_samples.device)
-        noisy_model_input = (1 - t_01) * original_samples + t_01 * noise
-
-        # n_dim = original_samples.ndim
-        # sigmas = self.get_sigmas(timesteps, n_dim, original_samples.dtype, original_samples.device)
-        # noisy_model_input = (1.0 - sigmas) * original_samples + sigmas * noise
-        return noisy_model_input
-
-    def scale_model_input(self, sample: torch.Tensor, timestep: Union[float, torch.Tensor]) -> torch.Tensor:
-        return sample
-
-    def set_train_timesteps(self, num_timesteps, device, linear=False):
-        if linear:
-            timesteps = torch.linspace(1000, 0, num_timesteps, device=device)
-            self.timesteps = timesteps
-            return timesteps
-        else:
-            # distribute them closer to center. Inference distributes them as a bias toward first
-            # Generate values from 0 to 1
-            t = torch.sigmoid(torch.randn((num_timesteps,), device=device))
-
-            # Scale and reverse the values to go from 1000 to 0
-            timesteps = (1 - t) * 1000
-
-            # Sort the timesteps in descending order
-            timesteps, _ = torch.sort(timesteps, descending=True)
-
-            self.timesteps = timesteps.to(device=device)
-
-            return timesteps
-
-
 def main(args):
     if args.report_to == "wandb" and args.hub_token is not None:
         raise ValueError(
@@ -1499,7 +1426,7 @@ def main(args):
     )
 
     # Load scheduler and models
-    noise_scheduler = CustomFlowMatchEulerDiscreteScheduler.from_pretrained(
+    noise_scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
         args.pretrained_model_name_or_path, subfolder="scheduler"
     )
     noise_scheduler_copy = copy.deepcopy(noise_scheduler)
@@ -1619,7 +1546,6 @@ def main(args):
         target_modules=target_modules,
     )
     transformer.add_adapter(transformer_lora_config)
-
     if args.train_text_encoder:
         text_lora_config = LoraConfig(
             r=args.rank,
@@ -1679,7 +1605,7 @@ def main(args):
         lora_state_dict = FluxPipeline.lora_state_dict(input_dir)
 
         transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
         }
         transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
         incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -1727,7 +1653,6 @@ def main(args):
         cast_training_params(models, dtype=torch.float32)
 
     transformer_lora_parameters = list(filter(lambda p: p.requires_grad, transformer.parameters()))
-
     if args.train_text_encoder:
         text_lora_parameters_one = list(filter(lambda p: p.requires_grad, text_encoder_one.parameters()))
     # if we use textual inversion, we freeze all parameters except for the token embeddings
@@ -1737,7 +1662,8 @@ def main(args):
         for name, param in text_encoder_one.named_parameters():
             if "token_embedding" in name:
                 # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
-                param.data = param.to(dtype=torch.float32)
+                if args.mixed_precision == "fp16":
+                    param.data = param.to(dtype=torch.float32)
                 param.requires_grad = True
                 text_lora_parameters_one.append(param)
             else:
@@ -1747,7 +1673,8 @@ def main(args):
             for name, param in text_encoder_two.named_parameters():
                 if "shared" in name:
                     # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
-                    param.data = param.to(dtype=torch.float32)
+                    if args.mixed_precision == "fp16":
+                        param.data = param.to(dtype=torch.float32)
                     param.requires_grad = True
                     text_lora_parameters_two.append(param)
                 else:
@@ -1828,6 +1755,7 @@ def main(args):
             optimizer_class = bnb.optim.AdamW8bit
         else:
             optimizer_class = torch.optim.AdamW
+
         optimizer = optimizer_class(
             params_to_optimize,
             betas=(args.adam_beta1, args.adam_beta2),
@@ -2021,6 +1949,7 @@ def main(args):
                 lr_scheduler,
             )
         else:
+            print("I SHOULD BE HERE")
             transformer, text_encoder_one, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
                 transformer, text_encoder_one, optimizer, train_dataloader, lr_scheduler
             )
@@ -2125,7 +2054,7 @@ def main(args):
                 if args.train_text_encoder:
                     text_encoder_one.train()
                     # set top parameter requires_grad = True for gradient checkpointing works
-                    accelerator.unwrap_model(text_encoder_one).text_model.embeddings.requires_grad_(True)
+                    unwrap_model(text_encoder_one).text_model.embeddings.requires_grad_(True)
                 elif args.train_text_encoder_ti:  # textual inversion / pivotal tuning
                     text_encoder_one.train()
                 if args.enable_t5_ti:
@@ -2137,6 +2066,11 @@ def main(args):
                 pivoted_tr = True
 
         for step, batch in enumerate(train_dataloader):
+            models_to_accumulate = [transformer]
+            if not freeze_text_encoder:
+                models_to_accumulate.extend([text_encoder_one])
+                if args.enable_t5_ti:
+                    models_to_accumulate.extend([text_encoder_two])
             if pivoted_te:
                 # stopping optimization of text_encoder params
                 optimizer.param_groups[te_idx]["lr"] = 0.0
@@ -2145,7 +2079,7 @@ def main(args):
                 logger.info(f"PIVOT TRANSFORMER {epoch}")
                 optimizer.param_groups[0]["lr"] = 0.0
 
-            with accelerator.accumulate(transformer):
+            with accelerator.accumulate(models_to_accumulate):
                 prompts = batch["prompts"]
 
                 # encode batch prompts when custom prompts are provided for each image -
@@ -2189,7 +2123,7 @@ def main(args):
                 model_input = (model_input - vae_config_shift_factor) * vae_config_scaling_factor
                 model_input = model_input.to(dtype=weight_dtype)
 
-                vae_scale_factor = 2 ** (len(vae_config_block_out_channels))
+                vae_scale_factor = 2 ** (len(vae_config_block_out_channels) - 1)
 
                 latent_image_ids = FluxPipeline._prepare_latent_image_ids(
                     model_input.shape[0],
@@ -2228,7 +2162,7 @@ def main(args):
                 )
 
                 # handle guidance
-                if transformer.config.guidance_embeds:
+                if unwrap_model(transformer).config.guidance_embeds:
                     guidance = torch.tensor([args.guidance_scale], device=accelerator.device)
                     guidance = guidance.expand(model_input.shape[0])
                 else:
@@ -2288,16 +2222,26 @@ def main(args):
                 accelerator.backward(loss)
                 if accelerator.sync_gradients:
                     if not freeze_text_encoder:
-                        if args.train_text_encoder:
+                        if args.train_text_encoder:  # text encoder tuning
                             params_to_clip = itertools.chain(transformer.parameters(), text_encoder_one.parameters())
                         elif pure_textual_inversion:
-                            params_to_clip = itertools.chain(
-                                text_encoder_one.parameters(), text_encoder_two.parameters()
-                            )
+                            if args.enable_t5_ti:
+                                params_to_clip = itertools.chain(
+                                    text_encoder_one.parameters(), text_encoder_two.parameters()
+                                )
+                            else:
+                                params_to_clip = itertools.chain(text_encoder_one.parameters())
                         else:
-                            params_to_clip = itertools.chain(
-                                transformer.parameters(), text_encoder_one.parameters(), text_encoder_two.parameters()
-                            )
+                            if args.enable_t5_ti:
+                                params_to_clip = itertools.chain(
+                                    transformer.parameters(),
+                                    text_encoder_one.parameters(),
+                                    text_encoder_two.parameters(),
+                                )
+                            else:
+                                params_to_clip = itertools.chain(
+                                    transformer.parameters(), text_encoder_one.parameters()
+                                )
                     else:
                         params_to_clip = itertools.chain(transformer.parameters())
                     accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
@@ -2339,6 +2283,10 @@ def main(args):
 
                         save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
                         accelerator.save_state(save_path)
+                        if args.train_text_encoder_ti:
+                            embedding_handler.save_embeddings(
+                                f"{args.output_dir}/{Path(args.output_dir).name}_emb_checkpoint_{global_step}.safetensors"
+                            )
                         logger.info(f"Saved state to {save_path}")
 
             logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
@@ -2351,14 +2299,16 @@ def main(args):
         if accelerator.is_main_process:
             if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
                 # create pipeline
-                if freeze_text_encoder:
+                if freeze_text_encoder:  # no text encoder one, two optimizations
                     text_encoder_one, text_encoder_two = load_text_encoders(text_encoder_cls_one, text_encoder_cls_two)
+                    text_encoder_one.to(weight_dtype)
+                    text_encoder_two.to(weight_dtype)
                 pipeline = FluxPipeline.from_pretrained(
                     args.pretrained_model_name_or_path,
                     vae=vae,
-                    text_encoder=accelerator.unwrap_model(text_encoder_one),
-                    text_encoder_2=accelerator.unwrap_model(text_encoder_two),
-                    transformer=accelerator.unwrap_model(transformer),
+                    text_encoder=unwrap_model(text_encoder_one),
+                    text_encoder_2=unwrap_model(text_encoder_two),
+                    transformer=unwrap_model(transformer),
                     revision=args.revision,
                     variant=args.variant,
                     torch_dtype=weight_dtype,
@@ -2372,21 +2322,21 @@ def main(args):
                     epoch=epoch,
                     torch_dtype=weight_dtype,
                 )
-                images = None
-                del pipeline
-
                 if freeze_text_encoder:
                     del text_encoder_one, text_encoder_two
                     free_memory()
-                elif args.train_text_encoder:
-                    del text_encoder_two
-                    free_memory()
+
+                images = None
+                del pipeline
 
     # Save the lora layers
     accelerator.wait_for_everyone()
     if accelerator.is_main_process:
         transformer = unwrap_model(transformer)
-        transformer = transformer.to(weight_dtype)
+        if args.upcast_before_saving:
+            transformer.to(torch.float32)
+        else:
+            transformer = transformer.to(weight_dtype)
         transformer_lora_layers = get_peft_model_state_dict(transformer)
 
         if args.train_text_encoder:
@@ -2428,8 +2378,8 @@ def main(args):
                 accelerator=accelerator,
                 pipeline_args=pipeline_args,
                 epoch=epoch,
-                torch_dtype=weight_dtype,
                 is_final_validation=True,
+                torch_dtype=weight_dtype,
             )
 
         save_model_card(
@@ -2452,6 +2402,7 @@ def main(args):
                 commit_message="End of training",
                 ignore_patterns=["step_*", "epoch_*"],
             )
+
         images = None
         del pipeline
 

examples/advanced_diffusion_training/train_dreambooth_lora_sd15_advanced.py

@@ -200,7 +200,8 @@ Special VAE used for training: {vae_path}.
         "diffusers",
         "diffusers-training",
         lora,
-        "template:sd-lora" "stable-diffusion",
+        "template:sd-lora",
+        "stable-diffusion",
         "stable-diffusion-diffusers",
     ]
     model_card = populate_model_card(model_card, tags=tags)
@@ -724,9 +725,9 @@ class TokenEmbeddingsHandler:
         idx = 0
         for tokenizer, text_encoder in zip(self.tokenizers, self.text_encoders):
             assert isinstance(inserting_toks, list), "inserting_toks should be a list of strings."
-            assert all(
-                isinstance(tok, str) for tok in inserting_toks
-            ), "All elements in inserting_toks should be strings."
+            assert all(isinstance(tok, str) for tok in inserting_toks), (
+                "All elements in inserting_toks should be strings."
+            )
 
             self.inserting_toks = inserting_toks
             special_tokens_dict = {"additional_special_tokens": self.inserting_toks}
@@ -746,9 +747,9 @@ class TokenEmbeddingsHandler:
                 .to(dtype=self.dtype)
                 * std_token_embedding
             )
-            self.embeddings_settings[
-                f"original_embeddings_{idx}"
-            ] = text_encoder.text_model.embeddings.token_embedding.weight.data.clone()
+            self.embeddings_settings[f"original_embeddings_{idx}"] = (
+                text_encoder.text_model.embeddings.token_embedding.weight.data.clone()
+            )
             self.embeddings_settings[f"std_token_embedding_{idx}"] = std_token_embedding
 
             inu = torch.ones((len(tokenizer),), dtype=torch.bool)
@@ -1322,7 +1323,7 @@ def main(args):
 
         lora_state_dict, network_alphas = StableDiffusionPipeline.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 = {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:

examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py

@@ -890,9 +890,9 @@ class TokenEmbeddingsHandler:
         idx = 0
         for tokenizer, text_encoder in zip(self.tokenizers, self.text_encoders):
             assert isinstance(inserting_toks, list), "inserting_toks should be a list of strings."
-            assert all(
-                isinstance(tok, str) for tok in inserting_toks
-            ), "All elements in inserting_toks should be strings."
+            assert all(isinstance(tok, str) for tok in inserting_toks), (
+                "All elements in inserting_toks should be strings."
+            )
 
             self.inserting_toks = inserting_toks
             special_tokens_dict = {"additional_special_tokens": self.inserting_toks}
@@ -912,9 +912,9 @@ class TokenEmbeddingsHandler:
                 .to(dtype=self.dtype)
                 * std_token_embedding
             )
-            self.embeddings_settings[
-                f"original_embeddings_{idx}"
-            ] = text_encoder.text_model.embeddings.token_embedding.weight.data.clone()
+            self.embeddings_settings[f"original_embeddings_{idx}"] = (
+                text_encoder.text_model.embeddings.token_embedding.weight.data.clone()
+            )
             self.embeddings_settings[f"std_token_embedding_{idx}"] = std_token_embedding
 
             inu = torch.ones((len(tokenizer),), dtype=torch.bool)
@@ -1647,7 +1647,7 @@ def main(args):
 
         lora_state_dict, network_alphas = 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 = {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:

examples/amused/train_amused.py

@@ -720,7 +720,7 @@ def main(args):
     # Train!
     logger.info("***** Running training *****")
     logger.info(f"  Num training steps = {args.max_train_steps}")
-    logger.info(f"  Instantaneous batch size per device = { args.train_batch_size}")
+    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}")
     logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
 

examples/cogvideo/train_cogvideox_image_to_video_lora.py

@@ -1138,7 +1138,7 @@ def main(args):
         lora_state_dict = CogVideoXImageToVideoPipeline.lora_state_dict(input_dir)
 
         transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
         }
         transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
         incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")

examples/cogvideo/train_cogvideox_lora.py

@@ -1159,7 +1159,7 @@ def main(args):
         lora_state_dict = CogVideoXPipeline.lora_state_dict(input_dir)
 
         transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
         }
         transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
         incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")

examples/community/adaptive_mask_inpainting.py

@@ -1103,7 +1103,7 @@ class AdaptiveMaskInpaintPipeline(
                     f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                     f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                     f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                     " `pipeline.unet` or your `default_mask_image` or `image` input."
                 )
         elif num_channels_unet != 4:

examples/community/hd_painter.py

@@ -686,7 +686,7 @@ class StableDiffusionHDPainterPipeline(StableDiffusionInpaintPipeline):
                     f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                     f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                     f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                     " `pipeline.unet` or your `mask_image` or `image` input."
                 )
         elif num_channels_unet != 4:

examples/community/img2img_inpainting.py

@@ -362,7 +362,7 @@ class ImageToImageInpaintingPipeline(DiffusionPipeline):
                 f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                 f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                 f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                 " `pipeline.unet` or your `mask_image` or `image` input."
             )
 

examples/community/llm_grounded_diffusion.py

@@ -1120,7 +1120,7 @@ class LLMGroundedDiffusionPipeline(
 
             if verbose:
                 logger.info(
-                    f"time index {index}, loss: {loss.item()/loss_scale:.3f} (de-scaled with scale {loss_scale:.1f}), loss threshold: {loss_threshold:.3f}"
+                    f"time index {index}, loss: {loss.item() / loss_scale:.3f} (de-scaled with scale {loss_scale:.1f}), loss threshold: {loss_threshold:.3f}"
                 )
 
             try:
@@ -1184,7 +1184,7 @@ class LLMGroundedDiffusionPipeline(
 
                     if verbose:
                         logger.info(
-                            f"time index {index}, loss: {loss.item()/loss_scale:.3f}, loss threshold: {loss_threshold:.3f}, iteration: {iteration}"
+                            f"time index {index}, loss: {loss.item() / loss_scale:.3f}, loss threshold: {loss_threshold:.3f}, iteration: {iteration}"
                         )
 
             finally:

examples/community/mod_controlnet_tile_sr_sdxl.py

@@ -701,7 +701,7 @@ class StableDiffusionXLControlNetTileSRPipeline(
             raise ValueError("`max_tile_size` cannot be None.")
         elif not isinstance(max_tile_size, int) or max_tile_size not in (1024, 1280):
             raise ValueError(
-                f"`max_tile_size` has to be in 1024 or 1280 but is {max_tile_size} of type" f" {type(max_tile_size)}."
+                f"`max_tile_size` has to be in 1024 or 1280 but is {max_tile_size} of type {type(max_tile_size)}."
             )
         if tile_gaussian_sigma is None:
             raise ValueError("`tile_gaussian_sigma` cannot be None.")

examples/community/pipeline_flux_differential_img2img.py

@@ -488,7 +488,7 @@ class FluxDifferentialImg2ImgPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         if padding_mask_crop is not None:
             if not isinstance(image, PIL.Image.Image):
                 raise ValueError(
-                    f"The image should be a PIL image when inpainting mask crop, but is of type" f" {type(image)}."
+                    f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}."
                 )
             if not isinstance(mask_image, PIL.Image.Image):
                 raise ValueError(
@@ -496,7 +496,7 @@ class FluxDifferentialImg2ImgPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
                     f" {type(mask_image)}."
                 )
             if output_type != "pil":
-                raise ValueError(f"The output type should be PIL when inpainting mask crop, but is" f" {output_type}.")
+                raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.")
 
         if max_sequence_length is not None and max_sequence_length > 512:
             raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}")

examples/community/pipeline_prompt2prompt.py

@@ -907,12 +907,12 @@ def create_controller(
 
     # reweight
     if edit_type == "reweight":
-        assert (
-            equalizer_words is not None and equalizer_strengths is not None
-        ), "To use reweight edit, please specify equalizer_words and equalizer_strengths."
-        assert len(equalizer_words) == len(
-            equalizer_strengths
-        ), "equalizer_words and equalizer_strengths must be of same length."
+        assert equalizer_words is not None and equalizer_strengths is not None, (
+            "To use reweight edit, please specify equalizer_words and equalizer_strengths."
+        )
+        assert len(equalizer_words) == len(equalizer_strengths), (
+            "equalizer_words and equalizer_strengths must be of same length."
+        )
         equalizer = get_equalizer(prompts[1], equalizer_words, equalizer_strengths, tokenizer=tokenizer)
         return AttentionReweight(
             prompts,

examples/community/pipeline_sdxl_style_aligned.py

@@ -1738,7 +1738,7 @@ class StyleAlignedSDXLPipeline(
                         f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                         f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                         f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                        f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                        f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                         " `pipeline.unet` or your `mask_image` or `image` input."
                     )
             elif num_channels_unet != 4:

examples/community/pipeline_stable_diffusion_upscale_ldm3d.py

@@ -689,7 +689,7 @@ class StableDiffusionUpscaleLDM3DPipeline(
                 f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                 f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                 f" `num_channels_image`: {num_channels_image} "
-                f" = {num_channels_latents+num_channels_image}. Please verify the config of"
+                f" = {num_channels_latents + num_channels_image}. Please verify the config of"
                 " `pipeline.unet` or your `image` input."
             )
 

examples/community/pipeline_stable_diffusion_xl_attentive_eraser.py

@@ -1028,7 +1028,7 @@ class StableDiffusionXL_AE_Pipeline(
         if padding_mask_crop is not None:
             if not isinstance(image, PIL.Image.Image):
                 raise ValueError(
-                    f"The image should be a PIL image when inpainting mask crop, but is of type" f" {type(image)}."
+                    f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}."
                 )
             if not isinstance(mask_image, PIL.Image.Image):
                 raise ValueError(
@@ -1036,7 +1036,7 @@ class StableDiffusionXL_AE_Pipeline(
                     f" {type(mask_image)}."
                 )
             if output_type != "pil":
-                raise ValueError(f"The output type should be PIL when inpainting mask crop, but is" f" {output_type}.")
+                raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.")
 
         if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
             raise ValueError(
@@ -2050,7 +2050,7 @@ class StableDiffusionXL_AE_Pipeline(
                     f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                     f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                     f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                     " `pipeline.unet` or your `mask_image` or `image` input."
                 )
         elif num_channels_unet != 4:

examples/community/pipeline_stable_diffusion_xl_controlnet_adapter_inpaint.py

@@ -1578,7 +1578,7 @@ class StableDiffusionXLControlNetAdapterInpaintPipeline(
                     f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                     f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                     f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
-                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                     " `pipeline.unet` or your `mask_image` or `image` input."
                 )
         elif num_channels_unet != 4:

examples/community/scheduling_ufogen.py

@@ -288,8 +288,7 @@ class UFOGenScheduler(SchedulerMixin, ConfigMixin):
 
             if timesteps[0] >= self.config.num_train_timesteps:
                 raise ValueError(
-                    f"`timesteps` must start before `self.config.train_timesteps`:"
-                    f" {self.config.num_train_timesteps}."
+                    f"`timesteps` must start before `self.config.train_timesteps`: {self.config.num_train_timesteps}."
                 )
 
             timesteps = np.array(timesteps, dtype=np.int64)

examples/consistency_distillation/train_lcm_distill_lora_sd_wds.py

@@ -89,7 +89,7 @@ def get_module_kohya_state_dict(module, prefix: str, dtype: torch.dtype, adapter
 
         # Set alpha parameter
         if "lora_down" in kohya_key:
-            alpha_key = f'{kohya_key.split(".")[0]}.alpha'
+            alpha_key = f"{kohya_key.split('.')[0]}.alpha"
             kohya_ss_state_dict[alpha_key] = torch.tensor(module.peft_config[adapter_name].lora_alpha).to(dtype)
 
     return kohya_ss_state_dict

examples/consistency_distillation/train_lcm_distill_lora_sdxl.py

@@ -901,7 +901,7 @@ def main(args):
             unet_ = accelerator.unwrap_model(unet)
             lora_state_dict, _ = StableDiffusionXLPipeline.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.")
+                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")

examples/consistency_distillation/train_lcm_distill_lora_sdxl_wds.py

@@ -95,7 +95,7 @@ def get_module_kohya_state_dict(module, prefix: str, dtype: torch.dtype, adapter
 
         # Set alpha parameter
         if "lora_down" in kohya_key:
-            alpha_key = f'{kohya_key.split(".")[0]}.alpha'
+            alpha_key = f"{kohya_key.split('.')[0]}.alpha"
             kohya_ss_state_dict[alpha_key] = torch.tensor(module.peft_config[adapter_name].lora_alpha).to(dtype)
 
     return kohya_ss_state_dict

examples/controlnet/train_controlnet.py

@@ -927,17 +927,22 @@ def main(args):
     )
 
     # Scheduler and math around the number of training steps.
-    overrode_max_train_steps = False
-    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    # Check the PR https://github.com/huggingface/diffusers/pull/8312 for detailed explanation.
+    num_warmup_steps_for_scheduler = args.lr_warmup_steps * accelerator.num_processes
     if args.max_train_steps is None:
-        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
-        overrode_max_train_steps = True
+        len_train_dataloader_after_sharding = math.ceil(len(train_dataloader) / accelerator.num_processes)
+        num_update_steps_per_epoch = math.ceil(len_train_dataloader_after_sharding / args.gradient_accumulation_steps)
+        num_training_steps_for_scheduler = (
+            args.num_train_epochs * num_update_steps_per_epoch * accelerator.num_processes
+        )
+    else:
+        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes
 
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
         num_cycles=args.lr_num_cycles,
         power=args.lr_power,
     )
@@ -962,8 +967,14 @@ def main(args):
 
     # 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)
-    if overrode_max_train_steps:
+    if args.max_train_steps is None:
         args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        if num_training_steps_for_scheduler != args.max_train_steps * accelerator.num_processes:
+            logger.warning(
+                f"The length of the 'train_dataloader' after 'accelerator.prepare' ({len(train_dataloader)}) does not match "
+                f"the expected length ({len_train_dataloader_after_sharding}) when the learning rate scheduler was created. "
+                f"This inconsistency may result in the learning rate scheduler not functioning properly."
+            )
     # Afterwards we recalculate our number of training epochs
     args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
 

examples/controlnet/train_controlnet_sd3.py

@@ -17,6 +17,7 @@ import argparse
 import contextlib
 import copy
 import functools
+import gc
 import logging
 import math
 import os
@@ -52,6 +53,7 @@ from diffusers.optimization import get_scheduler
 from diffusers.training_utils import compute_density_for_timestep_sampling, compute_loss_weighting_for_sd3, free_memory
 from diffusers.utils import check_min_version, is_wandb_available, make_image_grid
 from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
+from diffusers.utils.testing_utils import backend_empty_cache
 from diffusers.utils.torch_utils import is_compiled_module
 
 
@@ -74,8 +76,9 @@ def log_validation(controlnet, args, accelerator, weight_dtype, step, is_final_v
 
     pipeline = StableDiffusion3ControlNetPipeline.from_pretrained(
         args.pretrained_model_name_or_path,
-        controlnet=controlnet,
+        controlnet=None,
         safety_checker=None,
+        transformer=None,
         revision=args.revision,
         variant=args.variant,
         torch_dtype=weight_dtype,
@@ -102,18 +105,55 @@ def log_validation(controlnet, args, accelerator, weight_dtype, step, is_final_v
             "number of `args.validation_image` and `args.validation_prompt` should be checked in `parse_args`"
         )
 
+    with torch.no_grad():
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = pipeline.encode_prompt(
+            validation_prompts,
+            prompt_2=None,
+            prompt_3=None,
+        )
+
+    del pipeline
+    gc.collect()
+    backend_empty_cache(accelerator.device.type)
+
+    pipeline = StableDiffusion3ControlNetPipeline.from_pretrained(
+        args.pretrained_model_name_or_path,
+        controlnet=controlnet,
+        safety_checker=None,
+        text_encoder=None,
+        text_encoder_2=None,
+        text_encoder_3=None,
+        revision=args.revision,
+        variant=args.variant,
+        torch_dtype=weight_dtype,
+    )
+    pipeline.enable_model_cpu_offload(device=accelerator.device.type)
+    pipeline.set_progress_bar_config(disable=True)
+
     image_logs = []
     inference_ctx = contextlib.nullcontext() if is_final_validation else torch.autocast(accelerator.device.type)
 
-    for validation_prompt, validation_image in zip(validation_prompts, validation_images):
+    for i, validation_image in enumerate(validation_images):
         validation_image = Image.open(validation_image).convert("RGB")
+        validation_prompt = validation_prompts[i]
 
         images = []
 
         for _ in range(args.num_validation_images):
             with inference_ctx:
                 image = pipeline(
-                    validation_prompt, control_image=validation_image, num_inference_steps=20, generator=generator
+                    prompt_embeds=prompt_embeds[i].unsqueeze(0),
+                    negative_prompt_embeds=negative_prompt_embeds[i].unsqueeze(0),
+                    pooled_prompt_embeds=pooled_prompt_embeds[i].unsqueeze(0),
+                    negative_pooled_prompt_embeds=negative_pooled_prompt_embeds[i].unsqueeze(0),
+                    control_image=validation_image,
+                    num_inference_steps=20,
+                    generator=generator,
                 ).images[0]
 
             images.append(image)
@@ -655,6 +695,7 @@ def make_train_dataset(args, tokenizer_one, tokenizer_two, tokenizer_three, acce
             dataset = load_dataset(
                 args.train_data_dir,
                 cache_dir=args.cache_dir,
+                trust_remote_code=True,
             )
         # See more about loading custom images at
         # https://huggingface.co/docs/datasets/v2.0.0/en/dataset_script

examples/custom_diffusion/retrieve.py

@@ -50,9 +50,11 @@ def retrieve(class_prompt, class_data_dir, num_class_images):
     total = 0
     pbar = tqdm(desc="downloading real regularization images", total=num_class_images)
 
-    with open(f"{class_data_dir}/caption.txt", "w") as f1, open(f"{class_data_dir}/urls.txt", "w") as f2, open(
-        f"{class_data_dir}/images.txt", "w"
-    ) as f3:
+    with (
+        open(f"{class_data_dir}/caption.txt", "w") as f1,
+        open(f"{class_data_dir}/urls.txt", "w") as f2,
+        open(f"{class_data_dir}/images.txt", "w") as f3,
+    ):
         while total < num_class_images:
             images = class_images[count]
             count += 1

examples/custom_diffusion/train_custom_diffusion.py

@@ -731,18 +731,18 @@ def main(args):
             if not class_images_dir.exists():
                 class_images_dir.mkdir(parents=True, exist_ok=True)
             if args.real_prior:
-                assert (
-                    class_images_dir / "images"
-                ).exists(), f"Please run: python retrieve.py --class_prompt \"{concept['class_prompt']}\" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}"
-                assert (
-                    len(list((class_images_dir / "images").iterdir())) == args.num_class_images
-                ), f"Please run: python retrieve.py --class_prompt \"{concept['class_prompt']}\" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}"
-                assert (
-                    class_images_dir / "caption.txt"
-                ).exists(), f"Please run: python retrieve.py --class_prompt \"{concept['class_prompt']}\" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}"
-                assert (
-                    class_images_dir / "images.txt"
-                ).exists(), f"Please run: python retrieve.py --class_prompt \"{concept['class_prompt']}\" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}"
+                assert (class_images_dir / "images").exists(), (
+                    f'Please run: python retrieve.py --class_prompt "{concept["class_prompt"]}" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}'
+                )
+                assert len(list((class_images_dir / "images").iterdir())) == args.num_class_images, (
+                    f'Please run: python retrieve.py --class_prompt "{concept["class_prompt"]}" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}'
+                )
+                assert (class_images_dir / "caption.txt").exists(), (
+                    f'Please run: python retrieve.py --class_prompt "{concept["class_prompt"]}" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}'
+                )
+                assert (class_images_dir / "images.txt").exists(), (
+                    f'Please run: python retrieve.py --class_prompt "{concept["class_prompt"]}" --class_data_dir {class_images_dir} --num_class_images {args.num_class_images}'
+                )
                 concept["class_prompt"] = os.path.join(class_images_dir, "caption.txt")
                 concept["class_data_dir"] = os.path.join(class_images_dir, "images.txt")
                 args.concepts_list[i] = concept

examples/dreambooth/train_dreambooth.py

@@ -1014,7 +1014,7 @@ def main(args):
 
     if args.train_text_encoder and unwrap_model(text_encoder).dtype != torch.float32:
         raise ValueError(
-            f"Text encoder loaded as datatype {unwrap_model(text_encoder).dtype}." f" {low_precision_error_string}"
+            f"Text encoder loaded as datatype {unwrap_model(text_encoder).dtype}. {low_precision_error_string}"
         )
 
     # Enable TF32 for faster training on Ampere GPUs,

examples/dreambooth/train_dreambooth_flux.py

@@ -895,7 +895,10 @@ def _encode_prompt_with_t5(
 
     prompt_embeds = text_encoder(text_input_ids.to(device))[0]
 
-    dtype = text_encoder.dtype
+    if hasattr(text_encoder, "module"):
+        dtype = text_encoder.module.dtype
+    else:
+        dtype = text_encoder.dtype
     prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
 
     _, seq_len, _ = prompt_embeds.shape
@@ -936,9 +939,13 @@ def _encode_prompt_with_clip(
 
     prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=False)
 
+    if hasattr(text_encoder, "module"):
+        dtype = text_encoder.module.dtype
+    else:
+        dtype = text_encoder.dtype
     # Use pooled output of CLIPTextModel
     prompt_embeds = prompt_embeds.pooler_output
-    prompt_embeds = prompt_embeds.to(dtype=text_encoder.dtype, device=device)
+    prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
 
     # duplicate text embeddings for each generation per prompt, using mps friendly method
     prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
@@ -958,7 +965,12 @@ def encode_prompt(
 ):
     prompt = [prompt] if isinstance(prompt, str) else prompt
     batch_size = len(prompt)
-    dtype = text_encoders[0].dtype
+
+    if hasattr(text_encoders[0], "module"):
+        dtype = text_encoders[0].module.dtype
+    else:
+        dtype = text_encoders[0].dtype
+
     device = device if device is not None else text_encoders[1].device
     pooled_prompt_embeds = _encode_prompt_with_clip(
         text_encoder=text_encoders[0],
@@ -1590,7 +1602,7 @@ def main(args):
                 )
 
                 # handle guidance
-                if accelerator.unwrap_model(transformer).config.guidance_embeds:
+                if unwrap_model(transformer).config.guidance_embeds:
                     guidance = torch.tensor([args.guidance_scale], device=accelerator.device)
                     guidance = guidance.expand(model_input.shape[0])
                 else:
@@ -1716,9 +1728,9 @@ def main(args):
                 pipeline = FluxPipeline.from_pretrained(
                     args.pretrained_model_name_or_path,
                     vae=vae,
-                    text_encoder=accelerator.unwrap_model(text_encoder_one, keep_fp32_wrapper=False),
-                    text_encoder_2=accelerator.unwrap_model(text_encoder_two, keep_fp32_wrapper=False),
-                    transformer=accelerator.unwrap_model(transformer, keep_fp32_wrapper=False),
+                    text_encoder=unwrap_model(text_encoder_one, keep_fp32_wrapper=False),
+                    text_encoder_2=unwrap_model(text_encoder_two, keep_fp32_wrapper=False),
+                    transformer=unwrap_model(transformer, keep_fp32_wrapper=False),
                     revision=args.revision,
                     variant=args.variant,
                     torch_dtype=weight_dtype,

examples/dreambooth/train_dreambooth_lora.py

@@ -982,7 +982,7 @@ def main(args):
 
         lora_state_dict, network_alphas = 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 = {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")
 

examples/dreambooth/train_dreambooth_lora_flux.py

@@ -177,16 +177,25 @@ def log_validation(
         f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
         f" {args.validation_prompt}."
     )
-    pipeline = pipeline.to(accelerator.device)
+    pipeline = pipeline.to(accelerator.device, dtype=torch_dtype)
     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
-    # autocast_ctx = torch.autocast(accelerator.device.type) if not is_final_validation else nullcontext()
-    autocast_ctx = nullcontext()
+    autocast_ctx = torch.autocast(accelerator.device.type) if not is_final_validation else nullcontext()
 
-    with autocast_ctx:
-        images = [pipeline(**pipeline_args, generator=generator).images[0] for _ in range(args.num_validation_images)]
+    # pre-calculate  prompt embeds, pooled prompt embeds, text ids because t5 does not support autocast
+    with torch.no_grad():
+        prompt_embeds, pooled_prompt_embeds, text_ids = pipeline.encode_prompt(
+            pipeline_args["prompt"], prompt_2=pipeline_args["prompt"]
+        )
+    images = []
+    for _ in range(args.num_validation_images):
+        with autocast_ctx:
+            image = pipeline(
+                prompt_embeds=prompt_embeds, pooled_prompt_embeds=pooled_prompt_embeds, generator=generator
+            ).images[0]
+            images.append(image)
 
     for tracker in accelerator.trackers:
         phase_name = "test" if is_final_validation else "validation"
@@ -203,8 +212,7 @@ def log_validation(
             )
 
     del pipeline
-    if torch.cuda.is_available():
-        torch.cuda.empty_cache()
+    free_memory()
 
     return images
 
@@ -932,7 +940,10 @@ def _encode_prompt_with_t5(
 
     prompt_embeds = text_encoder(text_input_ids.to(device))[0]
 
-    dtype = text_encoder.dtype
+    if hasattr(text_encoder, "module"):
+        dtype = text_encoder.module.dtype
+    else:
+        dtype = text_encoder.dtype
     prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
 
     _, seq_len, _ = prompt_embeds.shape
@@ -973,9 +984,13 @@ def _encode_prompt_with_clip(
 
     prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=False)
 
+    if hasattr(text_encoder, "module"):
+        dtype = text_encoder.module.dtype
+    else:
+        dtype = text_encoder.dtype
     # Use pooled output of CLIPTextModel
     prompt_embeds = prompt_embeds.pooler_output
-    prompt_embeds = prompt_embeds.to(dtype=text_encoder.dtype, device=device)
+    prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
 
     # duplicate text embeddings for each generation per prompt, using mps friendly method
     prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
@@ -994,7 +1009,11 @@ def encode_prompt(
     text_input_ids_list=None,
 ):
     prompt = [prompt] if isinstance(prompt, str) else prompt
-    dtype = text_encoders[0].dtype
+
+    if hasattr(text_encoders[0], "module"):
+        dtype = text_encoders[0].module.dtype
+    else:
+        dtype = text_encoders[0].dtype
 
     pooled_prompt_embeds = _encode_prompt_with_clip(
         text_encoder=text_encoders[0],
@@ -1275,7 +1294,7 @@ def main(args):
         lora_state_dict = FluxPipeline.lora_state_dict(input_dir)
 
         transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
         }
         transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
         incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
@@ -1619,7 +1638,7 @@ def main(args):
         if args.train_text_encoder:
             text_encoder_one.train()
             # set top parameter requires_grad = True for gradient checkpointing works
-            accelerator.unwrap_model(text_encoder_one).text_model.embeddings.requires_grad_(True)
+            unwrap_model(text_encoder_one).text_model.embeddings.requires_grad_(True)
 
         for step, batch in enumerate(train_dataloader):
             models_to_accumulate = [transformer]
@@ -1710,7 +1729,7 @@ def main(args):
                 )
 
                 # handle guidance
-                if accelerator.unwrap_model(transformer).config.guidance_embeds:
+                if unwrap_model(transformer).config.guidance_embeds:
                     guidance = torch.tensor([args.guidance_scale], device=accelerator.device)
                     guidance = guidance.expand(model_input.shape[0])
                 else:
@@ -1828,9 +1847,9 @@ def main(args):
                 pipeline = FluxPipeline.from_pretrained(
                     args.pretrained_model_name_or_path,
                     vae=vae,
-                    text_encoder=accelerator.unwrap_model(text_encoder_one),
-                    text_encoder_2=accelerator.unwrap_model(text_encoder_two),
-                    transformer=accelerator.unwrap_model(transformer),
+                    text_encoder=unwrap_model(text_encoder_one),
+                    text_encoder_2=unwrap_model(text_encoder_two),
+                    transformer=unwrap_model(transformer),
                     revision=args.revision,
                     variant=args.variant,
                     torch_dtype=weight_dtype,

examples/dreambooth/train_dreambooth_lora_lumina2.py

@@ -1053,7 +1053,7 @@ def main(args):
         lora_state_dict = Lumina2Text2ImgPipeline.lora_state_dict(input_dir)
 
         transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
         }
         transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
         incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")

examples/dreambooth/train_dreambooth_lora_sana.py

@@ -1064,7 +1064,7 @@ def main(args):
         lora_state_dict = SanaPipeline.lora_state_dict(input_dir)
 
         transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
         }
         transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
         incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")

examples/dreambooth/train_dreambooth_lora_sd3.py

@@ -1355,7 +1355,7 @@ def main(args):
         lora_state_dict = StableDiffusion3Pipeline.lora_state_dict(input_dir)
 
         transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
         }
         transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
         incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")

examples/dreambooth/train_dreambooth_lora_sdxl.py

@@ -118,7 +118,7 @@ def save_model_card(
             )
 
     model_description = f"""
-# {'SDXL' if 'playground' not in base_model else 'Playground'} LoRA DreamBooth - {repo_id}
+# {"SDXL" if "playground" not in base_model else "Playground"} LoRA DreamBooth - {repo_id}
 
 <Gallery />
 
@@ -669,6 +669,16 @@ def parse_args(input_args=None):
         ),
     )
 
+    parser.add_argument(
+        "--image_interpolation_mode",
+        type=str,
+        default="lanczos",
+        choices=[
+            f.lower() for f in dir(transforms.InterpolationMode) if not f.startswith("__") and not f.endswith("__")
+        ],
+        help="The image interpolation method to use for resizing images.",
+    )
+
     if input_args is not None:
         args = parser.parse_args(input_args)
     else:
@@ -790,7 +800,12 @@ class DreamBoothDataset(Dataset):
         self.original_sizes = []
         self.crop_top_lefts = []
         self.pixel_values = []
-        train_resize = transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR)
+
+        interpolation = getattr(transforms.InterpolationMode, args.image_interpolation_mode.upper(), None)
+        if interpolation is None:
+            raise ValueError(f"Unsupported interpolation mode {interpolation=}.")
+        train_resize = transforms.Resize(size, interpolation=interpolation)
+
         train_crop = transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size)
         train_flip = transforms.RandomHorizontalFlip(p=1.0)
         train_transforms = transforms.Compose(
@@ -1271,7 +1286,7 @@ def main(args):
 
         lora_state_dict, network_alphas = 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 = {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:

examples/flux-control/train_control_lora_flux.py

@@ -91,9 +91,9 @@ def log_validation(flux_transformer, args, accelerator, weight_dtype, step, is_f
             torch_dtype=weight_dtype,
         )
         pipeline.load_lora_weights(args.output_dir)
-        assert (
-            pipeline.transformer.config.in_channels == initial_channels * 2
-        ), f"{pipeline.transformer.config.in_channels=}"
+        assert pipeline.transformer.config.in_channels == initial_channels * 2, (
+            f"{pipeline.transformer.config.in_channels=}"
+        )
 
     pipeline.to(accelerator.device)
     pipeline.set_progress_bar_config(disable=True)
@@ -954,7 +954,7 @@ def main(args):
 
             lora_state_dict = FluxControlPipeline.lora_state_dict(input_dir)
             transformer_lora_state_dict = {
-                f'{k.replace("transformer.", "")}': v
+                f"{k.replace('transformer.', '')}": v
                 for k, v in lora_state_dict.items()
                 if k.startswith("transformer.") and "lora" in k
             }

examples/model_search/pipeline_easy.py

@@ -1081,9 +1081,9 @@ class AutoConfig:
                     f"textual_inversion_path: {search_word} -> {textual_inversion_path.model_status.site_url}"
                 )
 
-                pretrained_model_name_or_paths[
-                    pretrained_model_name_or_paths.index(search_word)
-                ] = textual_inversion_path.model_path
+                pretrained_model_name_or_paths[pretrained_model_name_or_paths.index(search_word)] = (
+                    textual_inversion_path.model_path
+                )
 
         self.load_textual_inversion(
             pretrained_model_name_or_paths, token=tokens, tokenizer=tokenizer, text_encoder=text_encoder, **kwargs

examples/research_projects/anytext/anytext.py

@@ -187,9 +187,9 @@ def get_clip_token_for_string(tokenizer, string):
         return_tensors="pt",
     )
     tokens = batch_encoding["input_ids"]
-    assert (
-        torch.count_nonzero(tokens - 49407) == 2
-    ), f"String '{string}' maps to more than a single token. Please use another string"
+    assert torch.count_nonzero(tokens - 49407) == 2, (
+        f"String '{string}' maps to more than a single token. Please use another string"
+    )
     return tokens[0, 1]
 
 

examples/research_projects/anytext/ocr_recog/RecSVTR.py

@@ -312,9 +312,9 @@ class PatchEmbed(nn.Module):
 
     def forward(self, x):
         B, C, H, W = x.shape
-        assert (
-            H == self.img_size[0] and W == self.img_size[1]
-        ), f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+        assert H == self.img_size[0] and W == self.img_size[1], (
+            f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+        )
         x = self.proj(x).flatten(2).permute(0, 2, 1)
         return x
 

examples/research_projects/colossalai/train_dreambooth_colossalai.py

@@ -619,7 +619,7 @@ def main(args):
 
             optimizer.step()
             lr_scheduler.step()
-            logger.info(f"max GPU_mem cost is {torch.cuda.max_memory_allocated()/2**20} MB", ranks=[0])
+            logger.info(f"max GPU_mem cost is {torch.cuda.max_memory_allocated() / 2**20} MB", ranks=[0])
             # Checks if the accelerator has performed an optimization step behind the scenes
             progress_bar.update(1)
             global_step += 1

examples/research_projects/controlnet/train_controlnet_webdataset.py

@@ -803,21 +803,20 @@ def parse_args(input_args=None):
         "--control_type",
         type=str,
         default="canny",
-        help=("The type of controlnet conditioning image to use. One of `canny`, `depth`" " Defaults to `canny`."),
+        help=("The type of controlnet conditioning image to use. One of `canny`, `depth` Defaults to `canny`."),
     )
     parser.add_argument(
         "--transformer_layers_per_block",
         type=str,
         default=None,
-        help=("The number of layers per block in the transformer. If None, defaults to" " `args.transformer_layers`."),
+        help=("The number of layers per block in the transformer. If None, defaults to `args.transformer_layers`."),
     )
     parser.add_argument(
         "--old_style_controlnet",
         action="store_true",
         default=False,
         help=(
-            "Use the old style controlnet, which is a single transformer layer with"
-            " a single head. Defaults to False."
+            "Use the old style controlnet, which is a single transformer layer with a single head. Defaults to False."
         ),
     )
 

examples/research_projects/diffusion_dpo/train_diffusion_dpo.py

@@ -86,7 +86,7 @@ def import_model_class_from_model_name_or_path(pretrained_model_name_or_path: st
 
 
 def log_validation(args, unet, accelerator, weight_dtype, epoch, is_final_validation=False):
-    logger.info(f"Running validation... \n Generating images with prompts:\n" f" {VALIDATION_PROMPTS}.")
+    logger.info(f"Running validation... \n Generating images with prompts:\n {VALIDATION_PROMPTS}.")
 
     # create pipeline
     pipeline = DiffusionPipeline.from_pretrained(

examples/research_projects/diffusion_dpo/train_diffusion_dpo_sdxl.py

@@ -91,7 +91,7 @@ def import_model_class_from_model_name_or_path(
 
 
 def log_validation(args, unet, vae, accelerator, weight_dtype, epoch, is_final_validation=False):
-    logger.info(f"Running validation... \n Generating images with prompts:\n" f" {VALIDATION_PROMPTS}.")
+    logger.info(f"Running validation... \n Generating images with prompts:\n {VALIDATION_PROMPTS}.")
 
     if is_final_validation:
         if args.mixed_precision == "fp16":

examples/research_projects/diffusion_orpo/train_diffusion_orpo_sdxl_lora.py

@@ -91,7 +91,7 @@ def import_model_class_from_model_name_or_path(
 
 
 def log_validation(args, unet, vae, accelerator, weight_dtype, epoch, is_final_validation=False):
-    logger.info(f"Running validation... \n Generating images with prompts:\n" f" {VALIDATION_PROMPTS}.")
+    logger.info(f"Running validation... \n Generating images with prompts:\n {VALIDATION_PROMPTS}.")
 
     if is_final_validation:
         if args.mixed_precision == "fp16":
@@ -683,7 +683,7 @@ def main(args):
 
         lora_state_dict, network_alphas = StableDiffusionXLLoraLoaderMixin.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 = {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:

examples/research_projects/diffusion_orpo/train_diffusion_orpo_sdxl_lora_wds.py

@@ -89,7 +89,7 @@ def import_model_class_from_model_name_or_path(
 
 
 def log_validation(args, unet, vae, accelerator, weight_dtype, epoch, is_final_validation=False):
-    logger.info(f"Running validation... \n Generating images with prompts:\n" f" {VALIDATION_PROMPTS}.")
+    logger.info(f"Running validation... \n Generating images with prompts:\n {VALIDATION_PROMPTS}.")
 
     if is_final_validation:
         if args.mixed_precision == "fp16":
@@ -790,7 +790,7 @@ def main(args):
 
         lora_state_dict, network_alphas = StableDiffusionXLLoraLoaderMixin.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 = {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:

examples/research_projects/flux_lora_quantization/train_dreambooth_lora_flux_miniature.py

@@ -783,7 +783,7 @@ def main(args):
         lora_state_dict = FluxPipeline.lora_state_dict(input_dir)
 
         transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
         }
         transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
         incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")

examples/research_projects/gligen/demo.ipynb

@@ -26,8 +26,7 @@
     "%load_ext autoreload\n",
     "%autoreload 2\n",
     "\n",
-    "import torch\n",
-    "from diffusers import StableDiffusionGLIGENTextImagePipeline, StableDiffusionGLIGENPipeline"
+    "from diffusers import StableDiffusionGLIGENPipeline"
    ]
   },
   {
@@ -36,28 +35,25 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "import os\n",
+    "from transformers import CLIPTextModel, CLIPTokenizer\n",
+    "\n",
     "import diffusers\n",
     "from diffusers import (\n",
     "    AutoencoderKL,\n",
     "    DDPMScheduler,\n",
-    "    UNet2DConditionModel,\n",
-    "    UniPCMultistepScheduler,\n",
     "    EulerDiscreteScheduler,\n",
+    "    UNet2DConditionModel,\n",
     ")\n",
-    "from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer\n",
+    "\n",
+    "\n",
     "# pretrained_model_name_or_path = 'masterful/gligen-1-4-generation-text-box'\n",
     "\n",
-    "pretrained_model_name_or_path = '/root/data/zhizhonghuang/checkpoints/models--masterful--gligen-1-4-generation-text-box/snapshots/d2820dc1e9ba6ca082051ce79cfd3eb468ae2c83'\n",
+    "pretrained_model_name_or_path = \"/root/data/zhizhonghuang/checkpoints/models--masterful--gligen-1-4-generation-text-box/snapshots/d2820dc1e9ba6ca082051ce79cfd3eb468ae2c83\"\n",
     "\n",
     "tokenizer = CLIPTokenizer.from_pretrained(pretrained_model_name_or_path, subfolder=\"tokenizer\")\n",
     "noise_scheduler = DDPMScheduler.from_pretrained(pretrained_model_name_or_path, subfolder=\"scheduler\")\n",
-    "text_encoder = CLIPTextModel.from_pretrained(\n",
-    "    pretrained_model_name_or_path, subfolder=\"text_encoder\"\n",
-    ")\n",
-    "vae = AutoencoderKL.from_pretrained(\n",
-    "    pretrained_model_name_or_path, subfolder=\"vae\"\n",
-    ")\n",
+    "text_encoder = CLIPTextModel.from_pretrained(pretrained_model_name_or_path, subfolder=\"text_encoder\")\n",
+    "vae = AutoencoderKL.from_pretrained(pretrained_model_name_or_path, subfolder=\"vae\")\n",
     "# unet = UNet2DConditionModel.from_pretrained(\n",
     "#     pretrained_model_name_or_path, subfolder=\"unet\"\n",
     "# )\n",
@@ -71,9 +67,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "unet = UNet2DConditionModel.from_pretrained(\n",
-    "    '/root/data/zhizhonghuang/ckpt/GLIGEN_Text_Retrain_COCO'\n",
-    ")"
+    "unet = UNet2DConditionModel.from_pretrained(\"/root/data/zhizhonghuang/ckpt/GLIGEN_Text_Retrain_COCO\")"
    ]
   },
   {
@@ -108,6 +102,9 @@
    "metadata": {},
    "outputs": [],
    "source": [
+    "import numpy as np\n",
+    "\n",
+    "\n",
     "# prompt = 'A realistic image of landscape scene depicting a green car parking on the left of a blue truck, with a red air balloon and a bird in the sky'\n",
     "# gen_boxes = [('a green car', [21, 281, 211, 159]), ('a blue truck', [269, 283, 209, 160]), ('a red air balloon', [66, 8, 145, 135]), ('a bird', [296, 42, 143, 100])]\n",
     "\n",
@@ -117,10 +114,8 @@
     "# prompt = 'A realistic scene of three skiers standing in a line on the snow near a palm tree'\n",
     "# gen_boxes = [('a skier', [5, 152, 139, 168]), ('a skier', [278, 192, 121, 158]), ('a skier', [148, 173, 124, 155]), ('a palm tree', [404, 105, 103, 251])]\n",
     "\n",
-    "prompt = 'An oil painting of a pink dolphin jumping on the left of a steam boat on the sea'\n",
-    "gen_boxes = [('a steam boat', [232, 225, 257, 149]), ('a jumping pink dolphin', [21, 249, 189, 123])]\n",
-    "\n",
-    "import numpy as np\n",
+    "prompt = \"An oil painting of a pink dolphin jumping on the left of a steam boat on the sea\"\n",
+    "gen_boxes = [(\"a steam boat\", [232, 225, 257, 149]), (\"a jumping pink dolphin\", [21, 249, 189, 123])]\n",
     "\n",
     "boxes = np.array([x[1] for x in gen_boxes])\n",
     "boxes = boxes / 512\n",
@@ -166,7 +161,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "diffusers.utils.make_image_grid(images, 4, len(images)//4)"
+    "diffusers.utils.make_image_grid(images, 4, len(images) // 4)"
    ]
   },
   {
@@ -179,7 +174,7 @@
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "densecaption",
+   "display_name": "Python 3 (ipykernel)",
    "language": "python",
    "name": "python3"
   },
@@ -197,5 +192,5 @@
   }
  },
  "nbformat": 4,
- "nbformat_minor": 2
+ "nbformat_minor": 4
 }

examples/research_projects/instructpix2pix_lora/train_instruct_pix2pix_lora.py

@@ -15,8 +15,8 @@
 # limitations under the License.
 
 """
-    Script to fine-tune Stable Diffusion for LORA InstructPix2Pix.
-    Base code referred from: https://github.com/huggingface/diffusers/blob/main/examples/instruct_pix2pix/train_instruct_pix2pix.py
+Script to fine-tune Stable Diffusion for LORA InstructPix2Pix.
+Base code referred from: https://github.com/huggingface/diffusers/blob/main/examples/instruct_pix2pix/train_instruct_pix2pix.py
 """
 
 import argparse

examples/research_projects/multi_subject_dreambooth/train_multi_subject_dreambooth.py

@@ -763,9 +763,9 @@ def main(args):
         # Parse instance and class inputs, and double check that lengths match
         instance_data_dir = args.instance_data_dir.split(",")
         instance_prompt = args.instance_prompt.split(",")
-        assert all(
-            x == len(instance_data_dir) for x in [len(instance_data_dir), len(instance_prompt)]
-        ), "Instance data dir and prompt inputs are not of the same length."
+        assert all(x == len(instance_data_dir) for x in [len(instance_data_dir), len(instance_prompt)]), (
+            "Instance data dir and prompt inputs are not of the same length."
+        )
 
         if args.with_prior_preservation:
             class_data_dir = args.class_data_dir.split(",")
@@ -788,9 +788,9 @@ def main(args):
                     negative_validation_prompts.append(None)
             args.validation_negative_prompt = negative_validation_prompts
 
-            assert num_of_validation_prompts == len(
-                negative_validation_prompts
-            ), "The length of negative prompts for validation is greater than the number of validation prompts."
+            assert num_of_validation_prompts == len(negative_validation_prompts), (
+                "The length of negative prompts for validation is greater than the number of validation prompts."
+            )
             args.validation_inference_steps = [args.validation_inference_steps] * num_of_validation_prompts
             args.validation_guidance_scale = [args.validation_guidance_scale] * num_of_validation_prompts
 

examples/research_projects/multi_token_textual_inversion/textual_inversion.py

@@ -830,9 +830,9 @@ def main():
                 # Let's make sure we don't update any embedding weights besides the newly added token
                 index_no_updates = get_mask(tokenizer, accelerator)
                 with torch.no_grad():
-                    accelerator.unwrap_model(text_encoder).get_input_embeddings().weight[
-                        index_no_updates
-                    ] = orig_embeds_params[index_no_updates]
+                    accelerator.unwrap_model(text_encoder).get_input_embeddings().weight[index_no_updates] = (
+                        orig_embeds_params[index_no_updates]
+                    )
 
             # Checks if the accelerator has performed an optimization step behind the scenes
             if accelerator.sync_gradients:

examples/research_projects/onnxruntime/textual_inversion/textual_inversion.py

@@ -886,9 +886,9 @@ def main():
                 index_no_updates[min(placeholder_token_ids) : max(placeholder_token_ids) + 1] = False
 
                 with torch.no_grad():
-                    accelerator.unwrap_model(text_encoder).get_input_embeddings().weight[
-                        index_no_updates
-                    ] = orig_embeds_params[index_no_updates]
+                    accelerator.unwrap_model(text_encoder).get_input_embeddings().weight[index_no_updates] = (
+                        orig_embeds_params[index_no_updates]
+                    )
 
             # Checks if the accelerator has performed an optimization step behind the scenes
             if accelerator.sync_gradients:

examples/research_projects/promptdiffusion/pipeline_prompt_diffusion.py

@@ -663,8 +663,7 @@ class PromptDiffusionPipeline(
                     self.check_image(image, prompt, prompt_embeds)
         else:
             raise ValueError(
-                f"You have passed a list of images of length {len(image_pair)}."
-                f"Make sure the list size equals to two."
+                f"You have passed a list of images of length {len(image_pair)}.Make sure the list size equals to two."
             )
 
         # Check `controlnet_conditioning_scale`

examples/research_projects/pytorch_xla/training/text_to_image/train_text_to_image_xla.py

@@ -173,7 +173,7 @@ class TrainSD:
         if not dataloader_exception:
             xm.wait_device_ops()
             total_time = time.time() - last_time
-            print(f"Average step time: {total_time/(self.args.max_train_steps-measure_start_step)}")
+            print(f"Average step time: {total_time / (self.args.max_train_steps - measure_start_step)}")
         else:
             print("dataloader exception happen, skip result")
             return
@@ -622,7 +622,7 @@ def main(args):
     num_devices_per_host = num_devices // num_hosts
     if xm.is_master_ordinal():
         print("***** Running training *****")
-        print(f"Instantaneous batch size per device = {args.train_batch_size // num_devices_per_host }")
+        print(f"Instantaneous batch size per device = {args.train_batch_size // num_devices_per_host}")
         print(
             f"Total train batch size (w. parallel, distributed & accumulation) = {args.train_batch_size * num_hosts}"
         )

examples/research_projects/scheduled_huber_loss_training/dreambooth/train_dreambooth.py

@@ -1057,7 +1057,7 @@ def main(args):
 
     if args.train_text_encoder and unwrap_model(text_encoder).dtype != torch.float32:
         raise ValueError(
-            f"Text encoder loaded as datatype {unwrap_model(text_encoder).dtype}." f" {low_precision_error_string}"
+            f"Text encoder loaded as datatype {unwrap_model(text_encoder).dtype}. {low_precision_error_string}"
         )
 
     # Enable TF32 for faster training on Ampere GPUs,

examples/research_projects/scheduled_huber_loss_training/dreambooth/train_dreambooth_lora.py

@@ -1021,7 +1021,7 @@ def main(args):
 
         lora_state_dict, network_alphas = 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 = {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")
 

examples/research_projects/scheduled_huber_loss_training/dreambooth/train_dreambooth_lora_sdxl.py

@@ -118,7 +118,7 @@ def save_model_card(
             )
 
     model_description = f"""
-# {'SDXL' if 'playground' not in base_model else 'Playground'} LoRA DreamBooth - {repo_id}
+# {"SDXL" if "playground" not in base_model else "Playground"} LoRA DreamBooth - {repo_id}
 
 <Gallery />
 
@@ -1336,7 +1336,7 @@ def main(args):
 
         lora_state_dict, network_alphas = 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 = {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:

examples/research_projects/scheduled_huber_loss_training/text_to_image/train_text_to_image_lora_sdxl.py

@@ -750,7 +750,7 @@ def main(args):
                 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 = {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:

examples/research_projects/sd3_lora_colab/train_dreambooth_lora_sd3_miniature.py

@@ -765,7 +765,7 @@ def main(args):
         lora_state_dict = StableDiffusion3Pipeline.lora_state_dict(input_dir)
 
         transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
         }
         transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
         incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")

examples/text_to_image/train_text_to_image_lora_sdxl.py

@@ -767,7 +767,7 @@ def main(args):
                 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 = {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:

examples/textual_inversion/textual_inversion.py

@@ -910,9 +910,9 @@ def main():
                 index_no_updates[min(placeholder_token_ids) : max(placeholder_token_ids) + 1] = False
 
                 with torch.no_grad():
-                    accelerator.unwrap_model(text_encoder).get_input_embeddings().weight[
-                        index_no_updates
-                    ] = orig_embeds_params[index_no_updates]
+                    accelerator.unwrap_model(text_encoder).get_input_embeddings().weight[index_no_updates] = (
+                        orig_embeds_params[index_no_updates]
+                    )
 
             # Checks if the accelerator has performed an optimization step behind the scenes
             if accelerator.sync_gradients:

examples/textual_inversion/textual_inversion_sdxl.py

@@ -965,12 +965,12 @@ def main():
                 index_no_updates_2[min(placeholder_token_ids_2) : max(placeholder_token_ids_2) + 1] = False
 
                 with torch.no_grad():
-                    accelerator.unwrap_model(text_encoder_1).get_input_embeddings().weight[
-                        index_no_updates
-                    ] = orig_embeds_params[index_no_updates]
-                    accelerator.unwrap_model(text_encoder_2).get_input_embeddings().weight[
-                        index_no_updates_2
-                    ] = orig_embeds_params_2[index_no_updates_2]
+                    accelerator.unwrap_model(text_encoder_1).get_input_embeddings().weight[index_no_updates] = (
+                        orig_embeds_params[index_no_updates]
+                    )
+                    accelerator.unwrap_model(text_encoder_2).get_input_embeddings().weight[index_no_updates_2] = (
+                        orig_embeds_params_2[index_no_updates_2]
+                    )
 
             # Checks if the accelerator has performed an optimization step behind the scenes
             if accelerator.sync_gradients:

examples/vqgan/test_vqgan.py

@@ -177,7 +177,7 @@ class TextToImage(ExamplesTestsAccelerate):
                 --model_config_name_or_path {vqmodel_config_path}
                 --discriminator_config_name_or_path {discriminator_config_path}
                 --checkpointing_steps=1
-                --resume_from_checkpoint={os.path.join(tmpdir, 'checkpoint-4')}
+                --resume_from_checkpoint={os.path.join(tmpdir, "checkpoint-4")}
                 --output_dir {tmpdir}
                 --seed=0
                 """.split()
@@ -262,7 +262,7 @@ class TextToImage(ExamplesTestsAccelerate):
                 --model_config_name_or_path {vqmodel_config_path}
                 --discriminator_config_name_or_path {discriminator_config_path}
                 --checkpointing_steps=1
-                --resume_from_checkpoint={os.path.join(tmpdir, 'checkpoint-4')}
+                --resume_from_checkpoint={os.path.join(tmpdir, "checkpoint-4")}
                 --output_dir {tmpdir}
                 --use_ema
                 --seed=0
@@ -377,7 +377,7 @@ class TextToImage(ExamplesTestsAccelerate):
                 --discriminator_config_name_or_path {discriminator_config_path}
                 --output_dir {tmpdir}
                 --checkpointing_steps=2
-                --resume_from_checkpoint={os.path.join(tmpdir, 'checkpoint-4')}
+                --resume_from_checkpoint={os.path.join(tmpdir, "checkpoint-4")}
                 --checkpoints_total_limit=2
                 --seed=0
                 """.split()

examples/vqgan/train_vqgan.py

@@ -653,15 +653,15 @@ def main():
     try:
         # Gets the resolution of the timm transformation after centercrop
         timm_centercrop_transform = timm_transform.transforms[1]
-        assert isinstance(
-            timm_centercrop_transform, transforms.CenterCrop
-        ), f"Timm model {timm_model} is currently incompatible with this script. Try vgg19."
+        assert isinstance(timm_centercrop_transform, transforms.CenterCrop), (
+            f"Timm model {timm_model} is currently incompatible with this script. Try vgg19."
+        )
         timm_model_resolution = timm_centercrop_transform.size[0]
         # Gets final normalization
         timm_model_normalization = timm_transform.transforms[-1]
-        assert isinstance(
-            timm_model_normalization, transforms.Normalize
-        ), f"Timm model {timm_model} is currently incompatible with this script. Try vgg19."
+        assert isinstance(timm_model_normalization, transforms.Normalize), (
+            f"Timm model {timm_model} is currently incompatible with this script. Try vgg19."
+        )
     except AssertionError as e:
         raise NotImplementedError(e)
     # Enable flash attention if asked

pyproject.toml

@@ -3,7 +3,7 @@ line-length = 119
 
 [tool.ruff.lint]
 # Never enforce `E501` (line length violations).
-ignore = ["C901", "E501", "E741", "F402", "F823"]
+ignore = ["C901", "E501", "E721", "E741", "F402", "F823"]
 select = ["C", "E", "F", "I", "W"]
 
 # Ignore import violations in all `__init__.py` files.

scripts/convert_amused.py

@@ -468,7 +468,7 @@ def make_vqvae(old_vae):
 
     # assert (old_output == new_output).all()
     print("skipping full vae equivalence check")
-    print(f"vae full diff { (old_output - new_output).float().abs().sum()}")
+    print(f"vae full diff {(old_output - new_output).float().abs().sum()}")
 
     return new_vae
 

scripts/convert_consistency_to_diffusers.py

@@ -239,7 +239,7 @@ def con_pt_to_diffuser(checkpoint_path: str, unet_config):
 
             if i != len(up_block_types) - 1:
                 new_prefix = f"up_blocks.{i}.upsamplers.0"
-                old_prefix = f"output_blocks.{current_layer-1}.1"
+                old_prefix = f"output_blocks.{current_layer - 1}.1"
                 new_checkpoint = convert_resnet(checkpoint, new_checkpoint, old_prefix, new_prefix)
         elif layer_type == "AttnUpBlock2D":
             for j in range(layers_per_block + 1):
@@ -255,7 +255,7 @@ def con_pt_to_diffuser(checkpoint_path: str, unet_config):
 
             if i != len(up_block_types) - 1:
                 new_prefix = f"up_blocks.{i}.upsamplers.0"
-                old_prefix = f"output_blocks.{current_layer-1}.2"
+                old_prefix = f"output_blocks.{current_layer - 1}.2"
                 new_checkpoint = convert_resnet(checkpoint, new_checkpoint, old_prefix, new_prefix)
 
     new_checkpoint["conv_norm_out.weight"] = checkpoint["out.0.weight"]

scripts/convert_dance_diffusion_to_diffusers.py

@@ -261,9 +261,9 @@ def main(args):
 
     model_name = args.model_path.split("/")[-1].split(".")[0]
     if not os.path.isfile(args.model_path):
-        assert (
-            model_name == args.model_path
-        ), f"Make sure to provide one of the official model names {MODELS_MAP.keys()}"
+        assert model_name == args.model_path, (
+            f"Make sure to provide one of the official model names {MODELS_MAP.keys()}"
+        )
         args.model_path = download(model_name)
 
     sample_rate = MODELS_MAP[model_name]["sample_rate"]
@@ -290,9 +290,9 @@ def main(args):
     assert all(k.endswith("kernel") for k in list(diffusers_minus_renamed)), f"Problem with {diffusers_minus_renamed}"
 
     for key, value in renamed_state_dict.items():
-        assert (
-            diffusers_state_dict[key].squeeze().shape == value.squeeze().shape
-        ), f"Shape for {key} doesn't match. Diffusers: {diffusers_state_dict[key].shape} vs. {value.shape}"
+        assert diffusers_state_dict[key].squeeze().shape == value.squeeze().shape, (
+            f"Shape for {key} doesn't match. Diffusers: {diffusers_state_dict[key].shape} vs. {value.shape}"
+        )
         if key == "time_proj.weight":
             value = value.squeeze()
 

scripts/convert_diffusers_to_original_sdxl.py

@@ -52,18 +52,18 @@ for i in range(3):
     for j in range(2):
         # loop over resnets/attentions for downblocks
         hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}."
-        sd_down_res_prefix = f"input_blocks.{3*i + j + 1}.0."
+        sd_down_res_prefix = f"input_blocks.{3 * i + j + 1}.0."
         unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix))
 
         if i > 0:
             hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}."
-            sd_down_atn_prefix = f"input_blocks.{3*i + j + 1}.1."
+            sd_down_atn_prefix = f"input_blocks.{3 * i + j + 1}.1."
             unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix))
 
     for j in range(4):
         # loop over resnets/attentions for upblocks
         hf_up_res_prefix = f"up_blocks.{i}.resnets.{j}."
-        sd_up_res_prefix = f"output_blocks.{3*i + j}.0."
+        sd_up_res_prefix = f"output_blocks.{3 * i + j}.0."
         unet_conversion_map_layer.append((sd_up_res_prefix, hf_up_res_prefix))
 
         if i < 2:
@@ -75,12 +75,12 @@ for i in range(3):
     if i < 3:
         # no downsample in down_blocks.3
         hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv."
-        sd_downsample_prefix = f"input_blocks.{3*(i+1)}.0.op."
+        sd_downsample_prefix = f"input_blocks.{3 * (i + 1)}.0.op."
         unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix))
 
         # no upsample in up_blocks.3
         hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
-        sd_upsample_prefix = f"output_blocks.{3*i + 2}.{1 if i == 0 else 2}."
+        sd_upsample_prefix = f"output_blocks.{3 * i + 2}.{1 if i == 0 else 2}."
         unet_conversion_map_layer.append((sd_upsample_prefix, hf_upsample_prefix))
 unet_conversion_map_layer.append(("output_blocks.2.2.conv.", "output_blocks.2.1.conv."))
 
@@ -89,7 +89,7 @@ sd_mid_atn_prefix = "middle_block.1."
 unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix))
 for j in range(2):
     hf_mid_res_prefix = f"mid_block.resnets.{j}."
-    sd_mid_res_prefix = f"middle_block.{2*j}."
+    sd_mid_res_prefix = f"middle_block.{2 * j}."
     unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix))
 
 
@@ -137,20 +137,20 @@ for i in range(4):
         vae_conversion_map.append((sd_downsample_prefix, hf_downsample_prefix))
 
         hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
-        sd_upsample_prefix = f"up.{3-i}.upsample."
+        sd_upsample_prefix = f"up.{3 - i}.upsample."
         vae_conversion_map.append((sd_upsample_prefix, hf_upsample_prefix))
 
     # up_blocks have three resnets
     # also, up blocks in hf are numbered in reverse from sd
     for j in range(3):
         hf_up_prefix = f"decoder.up_blocks.{i}.resnets.{j}."
-        sd_up_prefix = f"decoder.up.{3-i}.block.{j}."
+        sd_up_prefix = f"decoder.up.{3 - i}.block.{j}."
         vae_conversion_map.append((sd_up_prefix, hf_up_prefix))
 
 # this part accounts for mid blocks in both the encoder and the decoder
 for i in range(2):
     hf_mid_res_prefix = f"mid_block.resnets.{i}."
-    sd_mid_res_prefix = f"mid.block_{i+1}."
+    sd_mid_res_prefix = f"mid.block_{i + 1}."
     vae_conversion_map.append((sd_mid_res_prefix, hf_mid_res_prefix))
 
 

scripts/convert_diffusers_to_original_stable_diffusion.py

@@ -47,36 +47,36 @@ for i in range(4):
     for j in range(2):
         # loop over resnets/attentions for downblocks
         hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}."
-        sd_down_res_prefix = f"input_blocks.{3*i + j + 1}.0."
+        sd_down_res_prefix = f"input_blocks.{3 * i + j + 1}.0."
         unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix))
 
         if i < 3:
             # no attention layers in down_blocks.3
             hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}."
-            sd_down_atn_prefix = f"input_blocks.{3*i + j + 1}.1."
+            sd_down_atn_prefix = f"input_blocks.{3 * i + j + 1}.1."
             unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix))
 
     for j in range(3):
         # loop over resnets/attentions for upblocks
         hf_up_res_prefix = f"up_blocks.{i}.resnets.{j}."
-        sd_up_res_prefix = f"output_blocks.{3*i + j}.0."
+        sd_up_res_prefix = f"output_blocks.{3 * i + j}.0."
         unet_conversion_map_layer.append((sd_up_res_prefix, hf_up_res_prefix))
 
         if i > 0:
             # no attention layers in up_blocks.0
             hf_up_atn_prefix = f"up_blocks.{i}.attentions.{j}."
-            sd_up_atn_prefix = f"output_blocks.{3*i + j}.1."
+            sd_up_atn_prefix = f"output_blocks.{3 * i + j}.1."
             unet_conversion_map_layer.append((sd_up_atn_prefix, hf_up_atn_prefix))
 
     if i < 3:
         # no downsample in down_blocks.3
         hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv."
-        sd_downsample_prefix = f"input_blocks.{3*(i+1)}.0.op."
+        sd_downsample_prefix = f"input_blocks.{3 * (i + 1)}.0.op."
         unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix))
 
         # no upsample in up_blocks.3
         hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
-        sd_upsample_prefix = f"output_blocks.{3*i + 2}.{1 if i == 0 else 2}."
+        sd_upsample_prefix = f"output_blocks.{3 * i + 2}.{1 if i == 0 else 2}."
         unet_conversion_map_layer.append((sd_upsample_prefix, hf_upsample_prefix))
 
 hf_mid_atn_prefix = "mid_block.attentions.0."
@@ -85,7 +85,7 @@ unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix))
 
 for j in range(2):
     hf_mid_res_prefix = f"mid_block.resnets.{j}."
-    sd_mid_res_prefix = f"middle_block.{2*j}."
+    sd_mid_res_prefix = f"middle_block.{2 * j}."
     unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix))
 
 
@@ -133,20 +133,20 @@ for i in range(4):
         vae_conversion_map.append((sd_downsample_prefix, hf_downsample_prefix))
 
         hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
-        sd_upsample_prefix = f"up.{3-i}.upsample."
+        sd_upsample_prefix = f"up.{3 - i}.upsample."
         vae_conversion_map.append((sd_upsample_prefix, hf_upsample_prefix))
 
     # up_blocks have three resnets
     # also, up blocks in hf are numbered in reverse from sd
     for j in range(3):
         hf_up_prefix = f"decoder.up_blocks.{i}.resnets.{j}."
-        sd_up_prefix = f"decoder.up.{3-i}.block.{j}."
+        sd_up_prefix = f"decoder.up.{3 - i}.block.{j}."
         vae_conversion_map.append((sd_up_prefix, hf_up_prefix))
 
 # this part accounts for mid blocks in both the encoder and the decoder
 for i in range(2):
     hf_mid_res_prefix = f"mid_block.resnets.{i}."
-    sd_mid_res_prefix = f"mid.block_{i+1}."
+    sd_mid_res_prefix = f"mid.block_{i + 1}."
     vae_conversion_map.append((sd_mid_res_prefix, hf_mid_res_prefix))
 
 

scripts/convert_hunyuandit_controlnet_to_diffusers.py

@@ -21,9 +21,9 @@ def main(args):
     model_config = HunyuanDiT2DControlNetModel.load_config(
         "Tencent-Hunyuan/HunyuanDiT-v1.2-Diffusers", subfolder="transformer"
     )
-    model_config[
-        "use_style_cond_and_image_meta_size"
-    ] = args.use_style_cond_and_image_meta_size  ### version <= v1.1: True; version >= v1.2: False
+    model_config["use_style_cond_and_image_meta_size"] = (
+        args.use_style_cond_and_image_meta_size
+    )  ### version <= v1.1: True; version >= v1.2: False
     print(model_config)
 
     for key in state_dict:

scripts/convert_hunyuandit_to_diffusers.py

@@ -13,15 +13,14 @@ def main(args):
             state_dict = state_dict[args.load_key]
         except KeyError:
             raise KeyError(
-                f"{args.load_key} not found in the checkpoint."
-                f"Please load from the following keys:{state_dict.keys()}"
+                f"{args.load_key} not found in the checkpoint.Please load from the following keys:{state_dict.keys()}"
             )
 
     device = "cuda"
     model_config = HunyuanDiT2DModel.load_config("Tencent-Hunyuan/HunyuanDiT-Diffusers", subfolder="transformer")
-    model_config[
-        "use_style_cond_and_image_meta_size"
-    ] = args.use_style_cond_and_image_meta_size  ### version <= v1.1: True; version >= v1.2: False
+    model_config["use_style_cond_and_image_meta_size"] = (
+        args.use_style_cond_and_image_meta_size
+    )  ### version <= v1.1: True; version >= v1.2: False
 
     # input_size -> sample_size, text_dim -> cross_attention_dim
     for key in state_dict:

scripts/convert_k_upscaler_to_diffusers.py

@@ -142,14 +142,14 @@ def block_to_diffusers_checkpoint(block, checkpoint, block_idx, block_type):
             diffusers_attention_prefix = f"{block_type}_blocks.{block_idx}.attentions.{attention_idx}"
             idx = n * attention_idx + 1 if block_type == "up" else n * attention_idx + 2
             self_attention_prefix = f"{block_prefix}.{idx}"
-            cross_attention_prefix = f"{block_prefix}.{idx }"
+            cross_attention_prefix = f"{block_prefix}.{idx}"
             cross_attention_index = 1 if not attention.add_self_attention else 2
             idx = (
                 n * attention_idx + cross_attention_index
                 if block_type == "up"
                 else n * attention_idx + cross_attention_index + 1
             )
-            cross_attention_prefix = f"{block_prefix}.{idx }"
+            cross_attention_prefix = f"{block_prefix}.{idx}"
 
             diffusers_checkpoint.update(
                 cross_attn_to_diffusers_checkpoint(
@@ -220,9 +220,9 @@ def unet_model_from_original_config(original_config):
 
     block_out_channels = original_config["channels"]
 
-    assert (
-        len(set(original_config["depths"])) == 1
-    ), "UNet2DConditionModel currently do not support blocks with different number of layers"
+    assert len(set(original_config["depths"])) == 1, (
+        "UNet2DConditionModel currently do not support blocks with different number of layers"
+    )
     layers_per_block = original_config["depths"][0]
 
     class_labels_dim = original_config["mapping_cond_dim"]

scripts/convert_mochi_to_diffusers.py

@@ -168,28 +168,28 @@ def convert_mochi_vae_state_dict_to_diffusers(encoder_ckpt_path, decoder_ckpt_pa
     # Convert block_in (MochiMidBlock3D)
     for i in range(3):  # layers_per_block[-1] = 3
         new_state_dict[f"{prefix}block_in.resnets.{i}.norm1.norm_layer.weight"] = decoder_state_dict.pop(
-            f"blocks.0.{i+1}.stack.0.weight"
+            f"blocks.0.{i + 1}.stack.0.weight"
         )
         new_state_dict[f"{prefix}block_in.resnets.{i}.norm1.norm_layer.bias"] = decoder_state_dict.pop(
-            f"blocks.0.{i+1}.stack.0.bias"
+            f"blocks.0.{i + 1}.stack.0.bias"
         )
         new_state_dict[f"{prefix}block_in.resnets.{i}.conv1.conv.weight"] = decoder_state_dict.pop(
-            f"blocks.0.{i+1}.stack.2.weight"
+            f"blocks.0.{i + 1}.stack.2.weight"
         )
         new_state_dict[f"{prefix}block_in.resnets.{i}.conv1.conv.bias"] = decoder_state_dict.pop(
-            f"blocks.0.{i+1}.stack.2.bias"
+            f"blocks.0.{i + 1}.stack.2.bias"
         )
         new_state_dict[f"{prefix}block_in.resnets.{i}.norm2.norm_layer.weight"] = decoder_state_dict.pop(
-            f"blocks.0.{i+1}.stack.3.weight"
+            f"blocks.0.{i + 1}.stack.3.weight"
         )
         new_state_dict[f"{prefix}block_in.resnets.{i}.norm2.norm_layer.bias"] = decoder_state_dict.pop(
-            f"blocks.0.{i+1}.stack.3.bias"
+            f"blocks.0.{i + 1}.stack.3.bias"
         )
         new_state_dict[f"{prefix}block_in.resnets.{i}.conv2.conv.weight"] = decoder_state_dict.pop(
-            f"blocks.0.{i+1}.stack.5.weight"
+            f"blocks.0.{i + 1}.stack.5.weight"
         )
         new_state_dict[f"{prefix}block_in.resnets.{i}.conv2.conv.bias"] = decoder_state_dict.pop(
-            f"blocks.0.{i+1}.stack.5.bias"
+            f"blocks.0.{i + 1}.stack.5.bias"
         )
 
     # Convert up_blocks (MochiUpBlock3D)
@@ -197,33 +197,35 @@ def convert_mochi_vae_state_dict_to_diffusers(encoder_ckpt_path, decoder_ckpt_pa
     for block in range(3):
         for i in range(down_block_layers[block]):
             new_state_dict[f"{prefix}up_blocks.{block}.resnets.{i}.norm1.norm_layer.weight"] = decoder_state_dict.pop(
-                f"blocks.{block+1}.blocks.{i}.stack.0.weight"
+                f"blocks.{block + 1}.blocks.{i}.stack.0.weight"
             )
             new_state_dict[f"{prefix}up_blocks.{block}.resnets.{i}.norm1.norm_layer.bias"] = decoder_state_dict.pop(
-                f"blocks.{block+1}.blocks.{i}.stack.0.bias"
+                f"blocks.{block + 1}.blocks.{i}.stack.0.bias"
             )
             new_state_dict[f"{prefix}up_blocks.{block}.resnets.{i}.conv1.conv.weight"] = decoder_state_dict.pop(
-                f"blocks.{block+1}.blocks.{i}.stack.2.weight"
+                f"blocks.{block + 1}.blocks.{i}.stack.2.weight"
             )
             new_state_dict[f"{prefix}up_blocks.{block}.resnets.{i}.conv1.conv.bias"] = decoder_state_dict.pop(
-                f"blocks.{block+1}.blocks.{i}.stack.2.bias"
+                f"blocks.{block + 1}.blocks.{i}.stack.2.bias"
             )
             new_state_dict[f"{prefix}up_blocks.{block}.resnets.{i}.norm2.norm_layer.weight"] = decoder_state_dict.pop(
-                f"blocks.{block+1}.blocks.{i}.stack.3.weight"
+                f"blocks.{block + 1}.blocks.{i}.stack.3.weight"
             )
             new_state_dict[f"{prefix}up_blocks.{block}.resnets.{i}.norm2.norm_layer.bias"] = decoder_state_dict.pop(
-                f"blocks.{block+1}.blocks.{i}.stack.3.bias"
+                f"blocks.{block + 1}.blocks.{i}.stack.3.bias"
             )
             new_state_dict[f"{prefix}up_blocks.{block}.resnets.{i}.conv2.conv.weight"] = decoder_state_dict.pop(
-                f"blocks.{block+1}.blocks.{i}.stack.5.weight"
+                f"blocks.{block + 1}.blocks.{i}.stack.5.weight"
             )
             new_state_dict[f"{prefix}up_blocks.{block}.resnets.{i}.conv2.conv.bias"] = decoder_state_dict.pop(
-                f"blocks.{block+1}.blocks.{i}.stack.5.bias"
+                f"blocks.{block + 1}.blocks.{i}.stack.5.bias"
             )
         new_state_dict[f"{prefix}up_blocks.{block}.proj.weight"] = decoder_state_dict.pop(
-            f"blocks.{block+1}.proj.weight"
+            f"blocks.{block + 1}.proj.weight"
+        )
+        new_state_dict[f"{prefix}up_blocks.{block}.proj.bias"] = decoder_state_dict.pop(
+            f"blocks.{block + 1}.proj.bias"
         )
-        new_state_dict[f"{prefix}up_blocks.{block}.proj.bias"] = decoder_state_dict.pop(f"blocks.{block+1}.proj.bias")
 
     # Convert block_out (MochiMidBlock3D)
     for i in range(3):  # layers_per_block[0] = 3
@@ -267,133 +269,133 @@ def convert_mochi_vae_state_dict_to_diffusers(encoder_ckpt_path, decoder_ckpt_pa
     # Convert block_in (MochiMidBlock3D)
     for i in range(3):  # layers_per_block[0] = 3
         new_state_dict[f"{prefix}block_in.resnets.{i}.norm1.norm_layer.weight"] = encoder_state_dict.pop(
-            f"layers.{i+1}.stack.0.weight"
+            f"layers.{i + 1}.stack.0.weight"
         )
         new_state_dict[f"{prefix}block_in.resnets.{i}.norm1.norm_layer.bias"] = encoder_state_dict.pop(
-            f"layers.{i+1}.stack.0.bias"
+            f"layers.{i + 1}.stack.0.bias"
         )
         new_state_dict[f"{prefix}block_in.resnets.{i}.conv1.conv.weight"] = encoder_state_dict.pop(
-            f"layers.{i+1}.stack.2.weight"
+            f"layers.{i + 1}.stack.2.weight"
         )
         new_state_dict[f"{prefix}block_in.resnets.{i}.conv1.conv.bias"] = encoder_state_dict.pop(
-            f"layers.{i+1}.stack.2.bias"
+            f"layers.{i + 1}.stack.2.bias"
         )
         new_state_dict[f"{prefix}block_in.resnets.{i}.norm2.norm_layer.weight"] = encoder_state_dict.pop(
-            f"layers.{i+1}.stack.3.weight"
+            f"layers.{i + 1}.stack.3.weight"
         )
         new_state_dict[f"{prefix}block_in.resnets.{i}.norm2.norm_layer.bias"] = encoder_state_dict.pop(
-            f"layers.{i+1}.stack.3.bias"
+            f"layers.{i + 1}.stack.3.bias"
         )
         new_state_dict[f"{prefix}block_in.resnets.{i}.conv2.conv.weight"] = encoder_state_dict.pop(
-            f"layers.{i+1}.stack.5.weight"
+            f"layers.{i + 1}.stack.5.weight"
         )
         new_state_dict[f"{prefix}block_in.resnets.{i}.conv2.conv.bias"] = encoder_state_dict.pop(
-            f"layers.{i+1}.stack.5.bias"
+            f"layers.{i + 1}.stack.5.bias"
         )
 
     # Convert down_blocks (MochiDownBlock3D)
     down_block_layers = [3, 4, 6]  # layers_per_block[1], layers_per_block[2], layers_per_block[3]
     for block in range(3):
         new_state_dict[f"{prefix}down_blocks.{block}.conv_in.conv.weight"] = encoder_state_dict.pop(
-            f"layers.{block+4}.layers.0.weight"
+            f"layers.{block + 4}.layers.0.weight"
         )
         new_state_dict[f"{prefix}down_blocks.{block}.conv_in.conv.bias"] = encoder_state_dict.pop(
-            f"layers.{block+4}.layers.0.bias"
+            f"layers.{block + 4}.layers.0.bias"
         )
 
         for i in range(down_block_layers[block]):
             # Convert resnets
-            new_state_dict[
-                f"{prefix}down_blocks.{block}.resnets.{i}.norm1.norm_layer.weight"
-            ] = encoder_state_dict.pop(f"layers.{block+4}.layers.{i+1}.stack.0.weight")
+            new_state_dict[f"{prefix}down_blocks.{block}.resnets.{i}.norm1.norm_layer.weight"] = (
+                encoder_state_dict.pop(f"layers.{block + 4}.layers.{i + 1}.stack.0.weight")
+            )
             new_state_dict[f"{prefix}down_blocks.{block}.resnets.{i}.norm1.norm_layer.bias"] = encoder_state_dict.pop(
-                f"layers.{block+4}.layers.{i+1}.stack.0.bias"
+                f"layers.{block + 4}.layers.{i + 1}.stack.0.bias"
             )
             new_state_dict[f"{prefix}down_blocks.{block}.resnets.{i}.conv1.conv.weight"] = encoder_state_dict.pop(
-                f"layers.{block+4}.layers.{i+1}.stack.2.weight"
+                f"layers.{block + 4}.layers.{i + 1}.stack.2.weight"
             )
             new_state_dict[f"{prefix}down_blocks.{block}.resnets.{i}.conv1.conv.bias"] = encoder_state_dict.pop(
-                f"layers.{block+4}.layers.{i+1}.stack.2.bias"
+                f"layers.{block + 4}.layers.{i + 1}.stack.2.bias"
+            )
+            new_state_dict[f"{prefix}down_blocks.{block}.resnets.{i}.norm2.norm_layer.weight"] = (
+                encoder_state_dict.pop(f"layers.{block + 4}.layers.{i + 1}.stack.3.weight")
             )
-            new_state_dict[
-                f"{prefix}down_blocks.{block}.resnets.{i}.norm2.norm_layer.weight"
-            ] = encoder_state_dict.pop(f"layers.{block+4}.layers.{i+1}.stack.3.weight")
             new_state_dict[f"{prefix}down_blocks.{block}.resnets.{i}.norm2.norm_layer.bias"] = encoder_state_dict.pop(
-                f"layers.{block+4}.layers.{i+1}.stack.3.bias"
+                f"layers.{block + 4}.layers.{i + 1}.stack.3.bias"
             )
             new_state_dict[f"{prefix}down_blocks.{block}.resnets.{i}.conv2.conv.weight"] = encoder_state_dict.pop(
-                f"layers.{block+4}.layers.{i+1}.stack.5.weight"
+                f"layers.{block + 4}.layers.{i + 1}.stack.5.weight"
             )
             new_state_dict[f"{prefix}down_blocks.{block}.resnets.{i}.conv2.conv.bias"] = encoder_state_dict.pop(
-                f"layers.{block+4}.layers.{i+1}.stack.5.bias"
+                f"layers.{block + 4}.layers.{i + 1}.stack.5.bias"
             )
 
             # Convert attentions
-            qkv_weight = encoder_state_dict.pop(f"layers.{block+4}.layers.{i+1}.attn_block.attn.qkv.weight")
+            qkv_weight = encoder_state_dict.pop(f"layers.{block + 4}.layers.{i + 1}.attn_block.attn.qkv.weight")
             q, k, v = qkv_weight.chunk(3, dim=0)
 
             new_state_dict[f"{prefix}down_blocks.{block}.attentions.{i}.to_q.weight"] = q
             new_state_dict[f"{prefix}down_blocks.{block}.attentions.{i}.to_k.weight"] = k
             new_state_dict[f"{prefix}down_blocks.{block}.attentions.{i}.to_v.weight"] = v
             new_state_dict[f"{prefix}down_blocks.{block}.attentions.{i}.to_out.0.weight"] = encoder_state_dict.pop(
-                f"layers.{block+4}.layers.{i+1}.attn_block.attn.out.weight"
+                f"layers.{block + 4}.layers.{i + 1}.attn_block.attn.out.weight"
             )
             new_state_dict[f"{prefix}down_blocks.{block}.attentions.{i}.to_out.0.bias"] = encoder_state_dict.pop(
-                f"layers.{block+4}.layers.{i+1}.attn_block.attn.out.bias"
+                f"layers.{block + 4}.layers.{i + 1}.attn_block.attn.out.bias"
             )
             new_state_dict[f"{prefix}down_blocks.{block}.norms.{i}.norm_layer.weight"] = encoder_state_dict.pop(
-                f"layers.{block+4}.layers.{i+1}.attn_block.norm.weight"
+                f"layers.{block + 4}.layers.{i + 1}.attn_block.norm.weight"
             )
             new_state_dict[f"{prefix}down_blocks.{block}.norms.{i}.norm_layer.bias"] = encoder_state_dict.pop(
-                f"layers.{block+4}.layers.{i+1}.attn_block.norm.bias"
+                f"layers.{block + 4}.layers.{i + 1}.attn_block.norm.bias"
             )
 
     # Convert block_out (MochiMidBlock3D)
     for i in range(3):  # layers_per_block[-1] = 3
         # Convert resnets
         new_state_dict[f"{prefix}block_out.resnets.{i}.norm1.norm_layer.weight"] = encoder_state_dict.pop(
-            f"layers.{i+7}.stack.0.weight"
+            f"layers.{i + 7}.stack.0.weight"
         )
         new_state_dict[f"{prefix}block_out.resnets.{i}.norm1.norm_layer.bias"] = encoder_state_dict.pop(
-            f"layers.{i+7}.stack.0.bias"
+            f"layers.{i + 7}.stack.0.bias"
         )
         new_state_dict[f"{prefix}block_out.resnets.{i}.conv1.conv.weight"] = encoder_state_dict.pop(
-            f"layers.{i+7}.stack.2.weight"
+            f"layers.{i + 7}.stack.2.weight"
         )
         new_state_dict[f"{prefix}block_out.resnets.{i}.conv1.conv.bias"] = encoder_state_dict.pop(
-            f"layers.{i+7}.stack.2.bias"
+            f"layers.{i + 7}.stack.2.bias"
         )
         new_state_dict[f"{prefix}block_out.resnets.{i}.norm2.norm_layer.weight"] = encoder_state_dict.pop(
-            f"layers.{i+7}.stack.3.weight"
+            f"layers.{i + 7}.stack.3.weight"
         )
         new_state_dict[f"{prefix}block_out.resnets.{i}.norm2.norm_layer.bias"] = encoder_state_dict.pop(
-            f"layers.{i+7}.stack.3.bias"
+            f"layers.{i + 7}.stack.3.bias"
         )
         new_state_dict[f"{prefix}block_out.resnets.{i}.conv2.conv.weight"] = encoder_state_dict.pop(
-            f"layers.{i+7}.stack.5.weight"
+            f"layers.{i + 7}.stack.5.weight"
         )
         new_state_dict[f"{prefix}block_out.resnets.{i}.conv2.conv.bias"] = encoder_state_dict.pop(
-            f"layers.{i+7}.stack.5.bias"
+            f"layers.{i + 7}.stack.5.bias"
         )
 
         # Convert attentions
-        qkv_weight = encoder_state_dict.pop(f"layers.{i+7}.attn_block.attn.qkv.weight")
+        qkv_weight = encoder_state_dict.pop(f"layers.{i + 7}.attn_block.attn.qkv.weight")
         q, k, v = qkv_weight.chunk(3, dim=0)
 
         new_state_dict[f"{prefix}block_out.attentions.{i}.to_q.weight"] = q
         new_state_dict[f"{prefix}block_out.attentions.{i}.to_k.weight"] = k
         new_state_dict[f"{prefix}block_out.attentions.{i}.to_v.weight"] = v
         new_state_dict[f"{prefix}block_out.attentions.{i}.to_out.0.weight"] = encoder_state_dict.pop(
-            f"layers.{i+7}.attn_block.attn.out.weight"
+            f"layers.{i + 7}.attn_block.attn.out.weight"
         )
         new_state_dict[f"{prefix}block_out.attentions.{i}.to_out.0.bias"] = encoder_state_dict.pop(
-            f"layers.{i+7}.attn_block.attn.out.bias"
+            f"layers.{i + 7}.attn_block.attn.out.bias"
         )
         new_state_dict[f"{prefix}block_out.norms.{i}.norm_layer.weight"] = encoder_state_dict.pop(
-            f"layers.{i+7}.attn_block.norm.weight"
+            f"layers.{i + 7}.attn_block.norm.weight"
         )
         new_state_dict[f"{prefix}block_out.norms.{i}.norm_layer.bias"] = encoder_state_dict.pop(
-            f"layers.{i+7}.attn_block.norm.bias"
+            f"layers.{i + 7}.attn_block.norm.bias"
         )
 
     # Convert output layers

scripts/convert_original_audioldm2_to_diffusers.py

@@ -662,7 +662,7 @@ def convert_open_clap_checkpoint(checkpoint):
             # replace sequential layers with list
             sequential_layer = re.match(sequential_layers_pattern, key).group(1)
 
-            key = key.replace(f"sequential.{sequential_layer}.", f"layers.{int(sequential_layer)//3}.linear.")
+            key = key.replace(f"sequential.{sequential_layer}.", f"layers.{int(sequential_layer) // 3}.linear.")
         elif re.match(text_projection_pattern, key):
             projecton_layer = int(re.match(text_projection_pattern, key).group(1))
 

scripts/convert_original_audioldm_to_diffusers.py

@@ -636,7 +636,7 @@ def convert_open_clap_checkpoint(checkpoint):
             # replace sequential layers with list
             sequential_layer = re.match(sequential_layers_pattern, key).group(1)
 
-            key = key.replace(f"sequential.{sequential_layer}.", f"layers.{int(sequential_layer)//3}.linear.")
+            key = key.replace(f"sequential.{sequential_layer}.", f"layers.{int(sequential_layer) // 3}.linear.")
         elif re.match(text_projection_pattern, key):
             projecton_layer = int(re.match(text_projection_pattern, key).group(1))
 

scripts/convert_original_musicldm_to_diffusers.py

@@ -642,7 +642,7 @@ def convert_open_clap_checkpoint(checkpoint):
             # replace sequential layers with list
             sequential_layer = re.match(sequential_layers_pattern, key).group(1)
 
-            key = key.replace(f"sequential.{sequential_layer}.", f"layers.{int(sequential_layer)//3}.linear.")
+            key = key.replace(f"sequential.{sequential_layer}.", f"layers.{int(sequential_layer) // 3}.linear.")
         elif re.match(text_projection_pattern, key):
             projecton_layer = int(re.match(text_projection_pattern, key).group(1))
 

scripts/convert_stable_audio.py

@@ -95,18 +95,18 @@ def convert_stable_audio_state_dict_to_diffusers(state_dict, num_autoencoder_lay
             # get idx of the layer
             idx = int(new_key.split("coder.layers.")[1].split(".")[0])
 
-            new_key = new_key.replace(f"coder.layers.{idx}", f"coder.block.{idx-1}")
+            new_key = new_key.replace(f"coder.layers.{idx}", f"coder.block.{idx - 1}")
 
             if "encoder" in new_key:
                 for i in range(3):
-                    new_key = new_key.replace(f"block.{idx-1}.layers.{i}", f"block.{idx-1}.res_unit{i+1}")
-                new_key = new_key.replace(f"block.{idx-1}.layers.3", f"block.{idx-1}.snake1")
-                new_key = new_key.replace(f"block.{idx-1}.layers.4", f"block.{idx-1}.conv1")
+                    new_key = new_key.replace(f"block.{idx - 1}.layers.{i}", f"block.{idx - 1}.res_unit{i + 1}")
+                new_key = new_key.replace(f"block.{idx - 1}.layers.3", f"block.{idx - 1}.snake1")
+                new_key = new_key.replace(f"block.{idx - 1}.layers.4", f"block.{idx - 1}.conv1")
             else:
                 for i in range(2, 5):
-                    new_key = new_key.replace(f"block.{idx-1}.layers.{i}", f"block.{idx-1}.res_unit{i-1}")
-                new_key = new_key.replace(f"block.{idx-1}.layers.0", f"block.{idx-1}.snake1")
-                new_key = new_key.replace(f"block.{idx-1}.layers.1", f"block.{idx-1}.conv_t1")
+                    new_key = new_key.replace(f"block.{idx - 1}.layers.{i}", f"block.{idx - 1}.res_unit{i - 1}")
+                new_key = new_key.replace(f"block.{idx - 1}.layers.0", f"block.{idx - 1}.snake1")
+                new_key = new_key.replace(f"block.{idx - 1}.layers.1", f"block.{idx - 1}.conv_t1")
 
             new_key = new_key.replace("layers.0.beta", "snake1.beta")
             new_key = new_key.replace("layers.0.alpha", "snake1.alpha")
@@ -118,9 +118,9 @@ def convert_stable_audio_state_dict_to_diffusers(state_dict, num_autoencoder_lay
             new_key = new_key.replace("layers.3.weight_", "conv2.weight_")
 
             if idx == num_autoencoder_layers + 1:
-                new_key = new_key.replace(f"block.{idx-1}", "snake1")
+                new_key = new_key.replace(f"block.{idx - 1}", "snake1")
             elif idx == num_autoencoder_layers + 2:
-                new_key = new_key.replace(f"block.{idx-1}", "conv2")
+                new_key = new_key.replace(f"block.{idx - 1}", "conv2")
 
         else:
             new_key = new_key

scripts/convert_svd_to_diffusers.py

@@ -381,9 +381,9 @@ def convert_ldm_unet_checkpoint(
 
         # TODO resnet time_mixer.mix_factor
         if f"input_blocks.{i}.0.time_mixer.mix_factor" in unet_state_dict:
-            new_checkpoint[
-                f"down_blocks.{block_id}.resnets.{layer_in_block_id}.time_mixer.mix_factor"
-            ] = unet_state_dict[f"input_blocks.{i}.0.time_mixer.mix_factor"]
+            new_checkpoint[f"down_blocks.{block_id}.resnets.{layer_in_block_id}.time_mixer.mix_factor"] = (
+                unet_state_dict[f"input_blocks.{i}.0.time_mixer.mix_factor"]
+            )
 
         if len(attentions):
             paths = renew_attention_paths(attentions)
@@ -478,9 +478,9 @@ def convert_ldm_unet_checkpoint(
             )
 
             if f"output_blocks.{i}.0.time_mixer.mix_factor" in unet_state_dict:
-                new_checkpoint[
-                    f"up_blocks.{block_id}.resnets.{layer_in_block_id}.time_mixer.mix_factor"
-                ] = unet_state_dict[f"output_blocks.{i}.0.time_mixer.mix_factor"]
+                new_checkpoint[f"up_blocks.{block_id}.resnets.{layer_in_block_id}.time_mixer.mix_factor"] = (
+                    unet_state_dict[f"output_blocks.{i}.0.time_mixer.mix_factor"]
+                )
 
             output_block_list = {k: sorted(v) for k, v in output_block_list.items()}
             if ["conv.bias", "conv.weight"] in output_block_list.values():

scripts/convert_vq_diffusion_to_diffusers.py

@@ -51,9 +51,9 @@ PORTED_VQVAES = ["image_synthesis.modeling.codecs.image_codec.patch_vqgan.PatchV
 
 
 def vqvae_model_from_original_config(original_config):
-    assert (
-        original_config["target"] in PORTED_VQVAES
-    ), f"{original_config['target']} has not yet been ported to diffusers."
+    assert original_config["target"] in PORTED_VQVAES, (
+        f"{original_config['target']} has not yet been ported to diffusers."
+    )
 
     original_config = original_config["params"]
 
@@ -464,15 +464,15 @@ PORTED_CONTENT_EMBEDDINGS = ["image_synthesis.modeling.embeddings.dalle_mask_ima
 def transformer_model_from_original_config(
     original_diffusion_config, original_transformer_config, original_content_embedding_config
 ):
-    assert (
-        original_diffusion_config["target"] in PORTED_DIFFUSIONS
-    ), f"{original_diffusion_config['target']} has not yet been ported to diffusers."
-    assert (
-        original_transformer_config["target"] in PORTED_TRANSFORMERS
-    ), f"{original_transformer_config['target']} has not yet been ported to diffusers."
-    assert (
-        original_content_embedding_config["target"] in PORTED_CONTENT_EMBEDDINGS
-    ), f"{original_content_embedding_config['target']} has not yet been ported to diffusers."
+    assert original_diffusion_config["target"] in PORTED_DIFFUSIONS, (
+        f"{original_diffusion_config['target']} has not yet been ported to diffusers."
+    )
+    assert original_transformer_config["target"] in PORTED_TRANSFORMERS, (
+        f"{original_transformer_config['target']} has not yet been ported to diffusers."
+    )
+    assert original_content_embedding_config["target"] in PORTED_CONTENT_EMBEDDINGS, (
+        f"{original_content_embedding_config['target']} has not yet been ported to diffusers."
+    )
 
     original_diffusion_config = original_diffusion_config["params"]
     original_transformer_config = original_transformer_config["params"]

setup.py

@@ -122,7 +122,7 @@ _deps = [
     "pytest-timeout",
     "pytest-xdist",
     "python>=3.8.0",
-    "ruff==0.1.5",
+    "ruff==0.9.10",
     "safetensors>=0.3.1",
     "sentencepiece>=0.1.91,!=0.1.92",
     "GitPython<3.1.19",

src/diffusers/__init__.py

@@ -155,6 +155,7 @@ else:
             "AutoencoderKLWan",
             "AutoencoderOobleck",
             "AutoencoderTiny",
+            "AutoModel",
             "CacheMixin",
             "CogVideoXTransformer3DModel",
             "CogView3PlusTransformer2DModel",
@@ -197,6 +198,7 @@ else:
             "T2IAdapter",
             "T5FilmDecoder",
             "Transformer2DModel",
+            "TransformerTemporalModel",
             "UNet1DModel",
             "UNet2DConditionModel",
             "UNet2DModel",
@@ -731,6 +733,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AutoencoderKLWan,
             AutoencoderOobleck,
             AutoencoderTiny,
+            AutoModel,
             CacheMixin,
             CogVideoXTransformer3DModel,
             CogView3PlusTransformer2DModel,
@@ -772,6 +775,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             T2IAdapter,
             T5FilmDecoder,
             Transformer2DModel,
+            TransformerTemporalModel,
             UNet1DModel,
             UNet2DConditionModel,
             UNet2DModel,

src/diffusers/dependency_versions_table.py

@@ -29,7 +29,7 @@ deps = {
     "pytest-timeout": "pytest-timeout",
     "pytest-xdist": "pytest-xdist",
     "python": "python>=3.8.0",
-    "ruff": "ruff==0.1.5",
+    "ruff": "ruff==0.9.10",
     "safetensors": "safetensors>=0.3.1",
     "sentencepiece": "sentencepiece>=0.1.91,!=0.1.92",
     "GitPython": "GitPython<3.1.19",

src/diffusers/hooks/group_offloading.py

@@ -56,6 +56,7 @@ class ModuleGroup:
         buffers: Optional[List[torch.Tensor]] = None,
         non_blocking: bool = False,
         stream: Optional[torch.cuda.Stream] = None,
+        record_stream: Optional[bool] = False,
         low_cpu_mem_usage=False,
         onload_self: bool = True,
     ) -> None:
@@ -68,11 +69,14 @@ class ModuleGroup:
         self.buffers = buffers or []
         self.non_blocking = non_blocking or stream is not None
         self.stream = stream
+        self.record_stream = record_stream
         self.onload_self = onload_self
         self.low_cpu_mem_usage = low_cpu_mem_usage
-
         self.cpu_param_dict = self._init_cpu_param_dict()
 
+        if self.stream is None and self.record_stream:
+            raise ValueError("`record_stream` cannot be True when `stream` is None.")
+
     def _init_cpu_param_dict(self):
         cpu_param_dict = {}
         if self.stream is None:
@@ -112,6 +116,8 @@ class ModuleGroup:
     def onload_(self):
         r"""Onloads the group of modules to the onload_device."""
         context = nullcontext() if self.stream is None else torch.cuda.stream(self.stream)
+        current_stream = torch.cuda.current_stream() if self.record_stream else None
+
         if self.stream is not None:
             # Wait for previous Host->Device transfer to complete
             self.stream.synchronize()
@@ -122,14 +128,22 @@ class ModuleGroup:
                     for group_module in self.modules:
                         for param in group_module.parameters():
                             param.data = pinned_memory[param].to(self.onload_device, non_blocking=self.non_blocking)
+                            if self.record_stream:
+                                param.data.record_stream(current_stream)
                         for buffer in group_module.buffers():
                             buffer.data = pinned_memory[buffer].to(self.onload_device, non_blocking=self.non_blocking)
+                            if self.record_stream:
+                                buffer.data.record_stream(current_stream)
 
                     for param in self.parameters:
                         param.data = pinned_memory[param].to(self.onload_device, non_blocking=self.non_blocking)
+                        if self.record_stream:
+                            param.data.record_stream(current_stream)
 
                     for buffer in self.buffers:
                         buffer.data = pinned_memory[buffer].to(self.onload_device, non_blocking=self.non_blocking)
+                        if self.record_stream:
+                            buffer.data.record_stream(current_stream)
 
             else:
                 for group_module in self.modules:
@@ -143,11 +157,14 @@ class ModuleGroup:
 
                 for buffer in self.buffers:
                     buffer.data = buffer.data.to(self.onload_device, non_blocking=self.non_blocking)
+                    if self.record_stream:
+                        buffer.data.record_stream(current_stream)
 
     def offload_(self):
         r"""Offloads the group of modules to the offload_device."""
         if self.stream is not None:
-            torch.cuda.current_stream().synchronize()
+            if not self.record_stream:
+                torch.cuda.current_stream().synchronize()
             for group_module in self.modules:
                 for param in group_module.parameters():
                     param.data = self.cpu_param_dict[param]
@@ -331,6 +348,7 @@ def apply_group_offloading(
     num_blocks_per_group: Optional[int] = None,
     non_blocking: bool = False,
     use_stream: bool = False,
+    record_stream: bool = False,
     low_cpu_mem_usage: bool = False,
 ) -> None:
     r"""
@@ -378,6 +396,10 @@ def apply_group_offloading(
         use_stream (`bool`, defaults to `False`):
             If True, offloading and onloading is done asynchronously using a CUDA stream. This can be useful for
             overlapping computation and data transfer.
+        record_stream (`bool`, defaults to `False`): When enabled with `use_stream`, it marks the current tensor
+            as having been used by this stream. It is faster at the expense of slightly more memory usage. Refer to the
+            [PyTorch official docs](https://pytorch.org/docs/stable/generated/torch.Tensor.record_stream.html) more
+            details.
         low_cpu_mem_usage (`bool`, defaults to `False`):
             If True, the CPU memory usage is minimized by pinning tensors on-the-fly instead of pre-pinning them. This
             option only matters when using streamed CPU offloading (i.e. `use_stream=True`). This can be useful when
@@ -417,11 +439,24 @@ def apply_group_offloading(
             raise ValueError("num_blocks_per_group must be provided when using offload_type='block_level'.")
 
         _apply_group_offloading_block_level(
-            module, num_blocks_per_group, offload_device, onload_device, non_blocking, stream, low_cpu_mem_usage
+            module=module,
+            num_blocks_per_group=num_blocks_per_group,
+            offload_device=offload_device,
+            onload_device=onload_device,
+            non_blocking=non_blocking,
+            stream=stream,
+            record_stream=record_stream,
+            low_cpu_mem_usage=low_cpu_mem_usage,
         )
     elif offload_type == "leaf_level":
         _apply_group_offloading_leaf_level(
-            module, offload_device, onload_device, non_blocking, stream, low_cpu_mem_usage
+            module=module,
+            offload_device=offload_device,
+            onload_device=onload_device,
+            non_blocking=non_blocking,
+            stream=stream,
+            record_stream=record_stream,
+            low_cpu_mem_usage=low_cpu_mem_usage,
         )
     else:
         raise ValueError(f"Unsupported offload_type: {offload_type}")
@@ -434,6 +469,7 @@ def _apply_group_offloading_block_level(
     onload_device: torch.device,
     non_blocking: bool,
     stream: Optional[torch.cuda.Stream] = None,
+    record_stream: Optional[bool] = False,
     low_cpu_mem_usage: bool = False,
 ) -> None:
     r"""
@@ -453,6 +489,14 @@ def _apply_group_offloading_block_level(
         stream (`torch.cuda.Stream`, *optional*):
             If provided, offloading and onloading is done asynchronously using the provided stream. This can be useful
             for overlapping computation and data transfer.
+        record_stream (`bool`, defaults to `False`): When enabled with `use_stream`, it marks the current tensor
+            as having been used by this stream. It is faster at the expense of slightly more memory usage. Refer to the
+            [PyTorch official docs](https://pytorch.org/docs/stable/generated/torch.Tensor.record_stream.html) more
+            details.
+        low_cpu_mem_usage (`bool`, defaults to `False`):
+            If True, the CPU memory usage is minimized by pinning tensors on-the-fly instead of pre-pinning them. This
+            option only matters when using streamed CPU offloading (i.e. `use_stream=True`). This can be useful when
+            the CPU memory is a bottleneck but may counteract the benefits of using streams.
     """
 
     # Create module groups for ModuleList and Sequential blocks
@@ -475,6 +519,7 @@ def _apply_group_offloading_block_level(
                 onload_leader=current_modules[0],
                 non_blocking=non_blocking,
                 stream=stream,
+                record_stream=record_stream,
                 low_cpu_mem_usage=low_cpu_mem_usage,
                 onload_self=stream is None,
             )
@@ -512,6 +557,7 @@ def _apply_group_offloading_block_level(
         buffers=buffers,
         non_blocking=False,
         stream=None,
+        record_stream=False,
         onload_self=True,
     )
     next_group = matched_module_groups[0] if len(matched_module_groups) > 0 else None
@@ -524,6 +570,7 @@ def _apply_group_offloading_leaf_level(
     onload_device: torch.device,
     non_blocking: bool,
     stream: Optional[torch.cuda.Stream] = None,
+    record_stream: Optional[bool] = False,
     low_cpu_mem_usage: bool = False,
 ) -> None:
     r"""
@@ -545,6 +592,14 @@ def _apply_group_offloading_leaf_level(
         stream (`torch.cuda.Stream`, *optional*):
             If provided, offloading and onloading is done asynchronously using the provided stream. This can be useful
             for overlapping computation and data transfer.
+        record_stream (`bool`, defaults to `False`): When enabled with `use_stream`, it marks the current tensor
+            as having been used by this stream. It is faster at the expense of slightly more memory usage. Refer to the
+            [PyTorch official docs](https://pytorch.org/docs/stable/generated/torch.Tensor.record_stream.html) more
+            details.
+        low_cpu_mem_usage (`bool`, defaults to `False`):
+            If True, the CPU memory usage is minimized by pinning tensors on-the-fly instead of pre-pinning them. This
+            option only matters when using streamed CPU offloading (i.e. `use_stream=True`). This can be useful when
+            the CPU memory is a bottleneck but may counteract the benefits of using streams.
     """
 
     # Create module groups for leaf modules and apply group offloading hooks
@@ -560,6 +615,7 @@ def _apply_group_offloading_leaf_level(
             onload_leader=submodule,
             non_blocking=non_blocking,
             stream=stream,
+            record_stream=record_stream,
             low_cpu_mem_usage=low_cpu_mem_usage,
             onload_self=True,
         )
@@ -605,6 +661,7 @@ def _apply_group_offloading_leaf_level(
             buffers=buffers,
             non_blocking=non_blocking,
             stream=stream,
+            record_stream=record_stream,
             low_cpu_mem_usage=low_cpu_mem_usage,
             onload_self=True,
         )
@@ -624,6 +681,7 @@ def _apply_group_offloading_leaf_level(
             buffers=None,
             non_blocking=False,
             stream=None,
+            record_stream=False,
             low_cpu_mem_usage=low_cpu_mem_usage,
             onload_self=True,
         )

src/diffusers/loaders/ip_adapter.py

@@ -295,8 +295,7 @@ class IPAdapterMixin:
             ):
                 if len(scale_configs) != len(attn_processor.scale):
                     raise ValueError(
-                        f"Cannot assign {len(scale_configs)} scale_configs to "
-                        f"{len(attn_processor.scale)} IP-Adapter."
+                        f"Cannot assign {len(scale_configs)} scale_configs to {len(attn_processor.scale)} IP-Adapter."
                     )
                 elif len(scale_configs) == 1:
                     scale_configs = scale_configs * len(attn_processor.scale)

src/diffusers/loaders/lora_base.py

@@ -316,6 +316,7 @@ def _load_lora_into_text_encoder(
     adapter_name=None,
     _pipeline=None,
     low_cpu_mem_usage=False,
+    hotswap: bool = False,
 ):
     if not USE_PEFT_BACKEND:
         raise ValueError("PEFT backend is required for this method.")
@@ -341,6 +342,10 @@ def _load_lora_into_text_encoder(
     # their prefixes.
     prefix = text_encoder_name if prefix is None else prefix
 
+    # Safe prefix to check with.
+    if hotswap and any(text_encoder_name in key for key in state_dict.keys()):
+        raise ValueError("At the moment, hotswapping is not supported for text encoders, please pass `hotswap=False`.")
+
     # Load the layers corresponding to text encoder and make necessary adjustments.
     if prefix is not None:
         state_dict = {k[len(f"{prefix}.") :]: v for k, v in state_dict.items() if k.startswith(f"{prefix}.")}
@@ -908,3 +913,23 @@ class LoraBaseMixin:
         # property function that returns the lora scale which can be set at run time by the pipeline.
         # if _lora_scale has not been set, return 1
         return self._lora_scale if hasattr(self, "_lora_scale") else 1.0
+
+    def enable_lora_hotswap(self, **kwargs) -> None:
+        """Enables the possibility to hotswap LoRA adapters.
+
+        Calling this method is only required when hotswapping adapters and if the model is compiled or if the ranks of
+        the loaded adapters differ.
+
+        Args:
+            target_rank (`int`):
+                The highest rank among all the adapters that will be loaded.
+            check_compiled (`str`, *optional*, defaults to `"error"`):
+                How to handle the case when the model is already compiled, which should generally be avoided. The
+                options are:
+                  - "error" (default): raise an error
+                  - "warn": issue a warning
+                  - "ignore": do nothing
+        """
+        for key, component in self.components.items():
+            if hasattr(component, "enable_lora_hotswap") and (key in self._lora_loadable_modules):
+                component.enable_lora_hotswap(**kwargs)

src/diffusers/loaders/lora_conversion_utils.py

@@ -13,15 +13,22 @@
 # limitations under the License.
 
 import re
+from typing import List
 
 import torch
 
-from ..utils import is_peft_version, logging
+from ..utils import is_peft_version, logging, state_dict_all_zero
 
 
 logger = logging.get_logger(__name__)
 
 
+def swap_scale_shift(weight):
+    shift, scale = weight.chunk(2, dim=0)
+    new_weight = torch.cat([scale, shift], dim=0)
+    return new_weight
+
+
 def _maybe_map_sgm_blocks_to_diffusers(state_dict, unet_config, delimiter="_", block_slice_pos=5):
     # 1. get all state_dict_keys
     all_keys = list(state_dict.keys())
@@ -177,9 +184,9 @@ def _convert_non_diffusers_lora_to_diffusers(state_dict, unet_name="unet", text_
             # Store DoRA scale if present.
             if dora_present_in_unet:
                 dora_scale_key_to_replace = "_lora.down." if "_lora.down." in diffusers_name else ".lora.down."
-                unet_state_dict[
-                    diffusers_name.replace(dora_scale_key_to_replace, ".lora_magnitude_vector.")
-                ] = state_dict.pop(key.replace("lora_down.weight", "dora_scale"))
+                unet_state_dict[diffusers_name.replace(dora_scale_key_to_replace, ".lora_magnitude_vector.")] = (
+                    state_dict.pop(key.replace("lora_down.weight", "dora_scale"))
+                )
 
         # Handle text encoder LoRAs.
         elif lora_name.startswith(("lora_te_", "lora_te1_", "lora_te2_")):
@@ -199,13 +206,13 @@ def _convert_non_diffusers_lora_to_diffusers(state_dict, unet_name="unet", text_
                     "_lora.down." if "_lora.down." in diffusers_name else ".lora_linear_layer."
                 )
                 if lora_name.startswith(("lora_te_", "lora_te1_")):
-                    te_state_dict[
-                        diffusers_name.replace(dora_scale_key_to_replace_te, ".lora_magnitude_vector.")
-                    ] = state_dict.pop(key.replace("lora_down.weight", "dora_scale"))
+                    te_state_dict[diffusers_name.replace(dora_scale_key_to_replace_te, ".lora_magnitude_vector.")] = (
+                        state_dict.pop(key.replace("lora_down.weight", "dora_scale"))
+                    )
                 elif lora_name.startswith("lora_te2_"):
-                    te2_state_dict[
-                        diffusers_name.replace(dora_scale_key_to_replace_te, ".lora_magnitude_vector.")
-                    ] = state_dict.pop(key.replace("lora_down.weight", "dora_scale"))
+                    te2_state_dict[diffusers_name.replace(dora_scale_key_to_replace_te, ".lora_magnitude_vector.")] = (
+                        state_dict.pop(key.replace("lora_down.weight", "dora_scale"))
+                    )
 
         # Store alpha if present.
         if lora_name_alpha in state_dict:
@@ -313,6 +320,7 @@ def _convert_text_encoder_lora_key(key, lora_name):
         # Be aware that this is the new diffusers convention and the rest of the code might
         # not utilize it yet.
         diffusers_name = diffusers_name.replace(".lora.", ".lora_linear_layer.")
+
     return diffusers_name
 
 
@@ -331,8 +339,7 @@ def _get_alpha_name(lora_name_alpha, diffusers_name, alpha):
 
 
 # The utilities under `_convert_kohya_flux_lora_to_diffusers()`
-# are taken from https://github.com/kohya-ss/sd-scripts/blob/a61cf73a5cb5209c3f4d1a3688dd276a4dfd1ecb/networks/convert_flux_lora.py
-# All credits go to `kohya-ss`.
+# are adapted from https://github.com/kohya-ss/sd-scripts/blob/a61cf73a5cb5209c3f4d1a3688dd276a4dfd1ecb/networks/convert_flux_lora.py
 def _convert_kohya_flux_lora_to_diffusers(state_dict):
     def _convert_to_ai_toolkit(sds_sd, ait_sd, sds_key, ait_key):
         if sds_key + ".lora_down.weight" not in sds_sd:
@@ -341,7 +348,8 @@ def _convert_kohya_flux_lora_to_diffusers(state_dict):
 
         # scale weight by alpha and dim
         rank = down_weight.shape[0]
-        alpha = sds_sd.pop(sds_key + ".alpha").item()  # alpha is scalar
+        default_alpha = torch.tensor(rank, dtype=down_weight.dtype, device=down_weight.device, requires_grad=False)
+        alpha = sds_sd.pop(sds_key + ".alpha", default_alpha).item()  # alpha is scalar
         scale = alpha / rank  # LoRA is scaled by 'alpha / rank' in forward pass, so we need to scale it back here
 
         # calculate scale_down and scale_up to keep the same value. if scale is 4, scale_down is 2 and scale_up is 2
@@ -362,7 +370,10 @@ def _convert_kohya_flux_lora_to_diffusers(state_dict):
         sd_lora_rank = down_weight.shape[0]
 
         # scale weight by alpha and dim
-        alpha = sds_sd.pop(sds_key + ".alpha")
+        default_alpha = torch.tensor(
+            sd_lora_rank, dtype=down_weight.dtype, device=down_weight.device, requires_grad=False
+        )
+        alpha = sds_sd.pop(sds_key + ".alpha", default_alpha)
         scale = alpha / sd_lora_rank
 
         # calculate scale_down and scale_up
@@ -516,10 +527,103 @@ def _convert_kohya_flux_lora_to_diffusers(state_dict):
                 f"transformer.single_transformer_blocks.{i}.norm.linear",
             )
 
+        # TODO: alphas.
+        def assign_remaining_weights(assignments, source):
+            for lora_key in ["lora_A", "lora_B"]:
+                orig_lora_key = "lora_down" if lora_key == "lora_A" else "lora_up"
+                for target_fmt, source_fmt, transform in assignments:
+                    target_key = target_fmt.format(lora_key=lora_key)
+                    source_key = source_fmt.format(orig_lora_key=orig_lora_key)
+                    value = source.pop(source_key)
+                    if transform:
+                        value = transform(value)
+                    ait_sd[target_key] = value
+
+        if any("guidance_in" in k for k in sds_sd):
+            assign_remaining_weights(
+                [
+                    (
+                        "time_text_embed.guidance_embedder.linear_1.{lora_key}.weight",
+                        "lora_unet_guidance_in_in_layer.{orig_lora_key}.weight",
+                        None,
+                    ),
+                    (
+                        "time_text_embed.guidance_embedder.linear_2.{lora_key}.weight",
+                        "lora_unet_guidance_in_out_layer.{orig_lora_key}.weight",
+                        None,
+                    ),
+                ],
+                sds_sd,
+            )
+
+        if any("img_in" in k for k in sds_sd):
+            assign_remaining_weights(
+                [
+                    ("x_embedder.{lora_key}.weight", "lora_unet_img_in.{orig_lora_key}.weight", None),
+                ],
+                sds_sd,
+            )
+
+        if any("txt_in" in k for k in sds_sd):
+            assign_remaining_weights(
+                [
+                    ("context_embedder.{lora_key}.weight", "lora_unet_txt_in.{orig_lora_key}.weight", None),
+                ],
+                sds_sd,
+            )
+
+        if any("time_in" in k for k in sds_sd):
+            assign_remaining_weights(
+                [
+                    (
+                        "time_text_embed.timestep_embedder.linear_1.{lora_key}.weight",
+                        "lora_unet_time_in_in_layer.{orig_lora_key}.weight",
+                        None,
+                    ),
+                    (
+                        "time_text_embed.timestep_embedder.linear_2.{lora_key}.weight",
+                        "lora_unet_time_in_out_layer.{orig_lora_key}.weight",
+                        None,
+                    ),
+                ],
+                sds_sd,
+            )
+
+        if any("vector_in" in k for k in sds_sd):
+            assign_remaining_weights(
+                [
+                    (
+                        "time_text_embed.text_embedder.linear_1.{lora_key}.weight",
+                        "lora_unet_vector_in_in_layer.{orig_lora_key}.weight",
+                        None,
+                    ),
+                    (
+                        "time_text_embed.text_embedder.linear_2.{lora_key}.weight",
+                        "lora_unet_vector_in_out_layer.{orig_lora_key}.weight",
+                        None,
+                    ),
+                ],
+                sds_sd,
+            )
+
+        if any("final_layer" in k for k in sds_sd):
+            # Notice the swap in processing for "final_layer".
+            assign_remaining_weights(
+                [
+                    (
+                        "norm_out.linear.{lora_key}.weight",
+                        "lora_unet_final_layer_adaLN_modulation_1.{orig_lora_key}.weight",
+                        swap_scale_shift,
+                    ),
+                    ("proj_out.{lora_key}.weight", "lora_unet_final_layer_linear.{orig_lora_key}.weight", None),
+                ],
+                sds_sd,
+            )
+
         remaining_keys = list(sds_sd.keys())
         te_state_dict = {}
         if remaining_keys:
-            if not all(k.startswith("lora_te") for k in remaining_keys):
+            if not all(k.startswith(("lora_te", "lora_te1")) for k in remaining_keys):
                 raise ValueError(f"Incompatible keys detected: \n\n {', '.join(remaining_keys)}")
             for key in remaining_keys:
                 if not key.endswith("lora_down.weight"):
@@ -680,10 +784,98 @@ def _convert_kohya_flux_lora_to_diffusers(state_dict):
     if has_peft_state_dict:
         state_dict = {k: v for k, v in state_dict.items() if k.startswith("transformer.")}
         return state_dict
+
     # Another weird one.
     has_mixture = any(
         k.startswith("lora_transformer_") and ("lora_down" in k or "lora_up" in k or "alpha" in k) for k in state_dict
     )
+
+    # ComfyUI.
+    if not has_mixture:
+        state_dict = {k.replace("diffusion_model.", "lora_unet_"): v for k, v in state_dict.items()}
+        state_dict = {k.replace("text_encoders.clip_l.transformer.", "lora_te_"): v for k, v in state_dict.items()}
+
+        has_position_embedding = any("position_embedding" in k for k in state_dict)
+        if has_position_embedding:
+            zero_status_pe = state_dict_all_zero(state_dict, "position_embedding")
+            if zero_status_pe:
+                logger.info(
+                    "The `position_embedding` LoRA params are all zeros which make them ineffective. "
+                    "So, we will purge them out of the curret state dict to make loading possible."
+                )
+
+            else:
+                logger.info(
+                    "The state_dict has position_embedding LoRA params and we currently do not support them. "
+                    "Open an issue if you need this supported - https://github.com/huggingface/diffusers/issues/new."
+                )
+            state_dict = {k: v for k, v in state_dict.items() if "position_embedding" not in k}
+
+        has_t5xxl = any(k.startswith("text_encoders.t5xxl.transformer.") for k in state_dict)
+        if has_t5xxl:
+            zero_status_t5 = state_dict_all_zero(state_dict, "text_encoders.t5xxl")
+            if zero_status_t5:
+                logger.info(
+                    "The `t5xxl` LoRA params are all zeros which make them ineffective. "
+                    "So, we will purge them out of the curret state dict to make loading possible."
+                )
+            else:
+                logger.info(
+                    "T5-xxl keys found in the state dict, which are currently unsupported. We will filter them out."
+                    "Open an issue if this is a problem - https://github.com/huggingface/diffusers/issues/new."
+                )
+            state_dict = {k: v for k, v in state_dict.items() if not k.startswith("text_encoders.t5xxl.transformer.")}
+
+        has_diffb = any("diff_b" in k and k.startswith(("lora_unet_", "lora_te_")) for k in state_dict)
+        if has_diffb:
+            zero_status_diff_b = state_dict_all_zero(state_dict, ".diff_b")
+            if zero_status_diff_b:
+                logger.info(
+                    "The `diff_b` LoRA params are all zeros which make them ineffective. "
+                    "So, we will purge them out of the curret state dict to make loading possible."
+                )
+            else:
+                logger.info(
+                    "`diff_b` keys found in the state dict which are currently unsupported. "
+                    "So, we will filter out those keys. Open an issue if this is a problem - "
+                    "https://github.com/huggingface/diffusers/issues/new."
+                )
+            state_dict = {k: v for k, v in state_dict.items() if ".diff_b" not in k}
+
+        has_norm_diff = any(".norm" in k and ".diff" in k for k in state_dict)
+        if has_norm_diff:
+            zero_status_diff = state_dict_all_zero(state_dict, ".diff")
+            if zero_status_diff:
+                logger.info(
+                    "The `diff` LoRA params are all zeros which make them ineffective. "
+                    "So, we will purge them out of the curret state dict to make loading possible."
+                )
+            else:
+                logger.info(
+                    "Normalization diff keys found in the state dict which are currently unsupported. "
+                    "So, we will filter out those keys. Open an issue if this is a problem - "
+                    "https://github.com/huggingface/diffusers/issues/new."
+                )
+            state_dict = {k: v for k, v in state_dict.items() if ".norm" not in k and ".diff" not in k}
+
+        limit_substrings = ["lora_down", "lora_up"]
+        if any("alpha" in k for k in state_dict):
+            limit_substrings.append("alpha")
+
+        state_dict = {
+            _custom_replace(k, limit_substrings): v
+            for k, v in state_dict.items()
+            if k.startswith(("lora_unet_", "lora_te_"))
+        }
+
+        if any("text_projection" in k for k in state_dict):
+            logger.info(
+                "`text_projection` keys found in the `state_dict` which are unexpected. "
+                "So, we will filter out those keys. Open an issue if this is a problem - "
+                "https://github.com/huggingface/diffusers/issues/new."
+            )
+            state_dict = {k: v for k, v in state_dict.items() if "text_projection" not in k}
+
     if has_mixture:
         return _convert_mixture_state_dict_to_diffusers(state_dict)
 
@@ -798,6 +990,26 @@ def _convert_xlabs_flux_lora_to_diffusers(old_state_dict):
     return new_state_dict
 
 
+def _custom_replace(key: str, substrings: List[str]) -> str:
+    # Replaces the "."s with "_"s upto the `substrings`.
+    # Example:
+    # lora_unet.foo.bar.lora_A.weight -> lora_unet_foo_bar.lora_A.weight
+    pattern = "(" + "|".join(re.escape(sub) for sub in substrings) + ")"
+
+    match = re.search(pattern, key)
+    if match:
+        start_sub = match.start()
+        if start_sub > 0 and key[start_sub - 1] == ".":
+            boundary = start_sub - 1
+        else:
+            boundary = start_sub
+        left = key[:boundary].replace(".", "_")
+        right = key[boundary:]
+        return left + right
+    else:
+        return key.replace(".", "_")
+
+
 def _convert_bfl_flux_control_lora_to_diffusers(original_state_dict):
     converted_state_dict = {}
     original_state_dict_keys = list(original_state_dict.keys())
@@ -806,28 +1018,23 @@ def _convert_bfl_flux_control_lora_to_diffusers(original_state_dict):
     inner_dim = 3072
     mlp_ratio = 4.0
 
-    def swap_scale_shift(weight):
-        shift, scale = weight.chunk(2, dim=0)
-        new_weight = torch.cat([scale, shift], dim=0)
-        return new_weight
-
     for lora_key in ["lora_A", "lora_B"]:
         ## time_text_embed.timestep_embedder <-  time_in
-        converted_state_dict[
-            f"time_text_embed.timestep_embedder.linear_1.{lora_key}.weight"
-        ] = original_state_dict.pop(f"time_in.in_layer.{lora_key}.weight")
+        converted_state_dict[f"time_text_embed.timestep_embedder.linear_1.{lora_key}.weight"] = (
+            original_state_dict.pop(f"time_in.in_layer.{lora_key}.weight")
+        )
         if f"time_in.in_layer.{lora_key}.bias" in original_state_dict_keys:
-            converted_state_dict[
-                f"time_text_embed.timestep_embedder.linear_1.{lora_key}.bias"
-            ] = original_state_dict.pop(f"time_in.in_layer.{lora_key}.bias")
+            converted_state_dict[f"time_text_embed.timestep_embedder.linear_1.{lora_key}.bias"] = (
+                original_state_dict.pop(f"time_in.in_layer.{lora_key}.bias")
+            )
 
-        converted_state_dict[
-            f"time_text_embed.timestep_embedder.linear_2.{lora_key}.weight"
-        ] = original_state_dict.pop(f"time_in.out_layer.{lora_key}.weight")
+        converted_state_dict[f"time_text_embed.timestep_embedder.linear_2.{lora_key}.weight"] = (
+            original_state_dict.pop(f"time_in.out_layer.{lora_key}.weight")
+        )
         if f"time_in.out_layer.{lora_key}.bias" in original_state_dict_keys:
-            converted_state_dict[
-                f"time_text_embed.timestep_embedder.linear_2.{lora_key}.bias"
-            ] = original_state_dict.pop(f"time_in.out_layer.{lora_key}.bias")
+            converted_state_dict[f"time_text_embed.timestep_embedder.linear_2.{lora_key}.bias"] = (
+                original_state_dict.pop(f"time_in.out_layer.{lora_key}.bias")
+            )
 
         ## time_text_embed.text_embedder <- vector_in
         converted_state_dict[f"time_text_embed.text_embedder.linear_1.{lora_key}.weight"] = original_state_dict.pop(
@@ -849,21 +1056,21 @@ def _convert_bfl_flux_control_lora_to_diffusers(original_state_dict):
         # guidance
         has_guidance = any("guidance" in k for k in original_state_dict)
         if has_guidance:
-            converted_state_dict[
-                f"time_text_embed.guidance_embedder.linear_1.{lora_key}.weight"
-            ] = original_state_dict.pop(f"guidance_in.in_layer.{lora_key}.weight")
+            converted_state_dict[f"time_text_embed.guidance_embedder.linear_1.{lora_key}.weight"] = (
+                original_state_dict.pop(f"guidance_in.in_layer.{lora_key}.weight")
+            )
             if f"guidance_in.in_layer.{lora_key}.bias" in original_state_dict_keys:
-                converted_state_dict[
-                    f"time_text_embed.guidance_embedder.linear_1.{lora_key}.bias"
-                ] = original_state_dict.pop(f"guidance_in.in_layer.{lora_key}.bias")
+                converted_state_dict[f"time_text_embed.guidance_embedder.linear_1.{lora_key}.bias"] = (
+                    original_state_dict.pop(f"guidance_in.in_layer.{lora_key}.bias")
+                )
 
-            converted_state_dict[
-                f"time_text_embed.guidance_embedder.linear_2.{lora_key}.weight"
-            ] = original_state_dict.pop(f"guidance_in.out_layer.{lora_key}.weight")
+            converted_state_dict[f"time_text_embed.guidance_embedder.linear_2.{lora_key}.weight"] = (
+                original_state_dict.pop(f"guidance_in.out_layer.{lora_key}.weight")
+            )
             if f"guidance_in.out_layer.{lora_key}.bias" in original_state_dict_keys:
-                converted_state_dict[
-                    f"time_text_embed.guidance_embedder.linear_2.{lora_key}.bias"
-                ] = original_state_dict.pop(f"guidance_in.out_layer.{lora_key}.bias")
+                converted_state_dict[f"time_text_embed.guidance_embedder.linear_2.{lora_key}.bias"] = (
+                    original_state_dict.pop(f"guidance_in.out_layer.{lora_key}.bias")
+                )
 
         # context_embedder
         converted_state_dict[f"context_embedder.{lora_key}.weight"] = original_state_dict.pop(