[LoRA] Support HunyuanVideo (#10254)

* 1217

* 1217

* 1217

* update

* reverse

* add test

* update test

* make style

* update

* make style

---------
Co-authored-by: Aryan <aryan@huggingface.co>

src/diffusers/loaders/__init__.py

@@ -70,6 +70,7 @@ if is_torch_available():
             "FluxLoraLoaderMixin",
             "CogVideoXLoraLoaderMixin",
             "Mochi1LoraLoaderMixin",
+            "HunyuanVideoLoraLoaderMixin",
             "SanaLoraLoaderMixin",
         ]
         _import_structure["textual_inversion"] = ["TextualInversionLoaderMixin"]
@@ -90,6 +91,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
                 AmusedLoraLoaderMixin,
                 CogVideoXLoraLoaderMixin,
                 FluxLoraLoaderMixin,
+                HunyuanVideoLoraLoaderMixin,
                 LoraLoaderMixin,
                 LTXVideoLoraLoaderMixin,
                 Mochi1LoraLoaderMixin,

src/diffusers/loaders/lora_pipeline.py

@@ -3870,6 +3870,314 @@ class SanaLoraLoaderMixin(LoraBaseMixin):
         super().unfuse_lora(components=components)
 
 
+class HunyuanVideoLoraLoaderMixin(LoraBaseMixin):
+    r"""
+    Load LoRA layers into [`HunyuanVideoTransformer3DModel`]. Specific to [`HunyuanVideoPipeline`].
+    """
+
+    _lora_loadable_modules = ["transformer"]
+    transformer_name = TRANSFORMER_NAME
+
+    @classmethod
+    @validate_hf_hub_args
+    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.lora_state_dict
+    def lora_state_dict(
+        cls,
+        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
+        **kwargs,
+    ):
+        r"""
+        Return state dict for lora weights and the network alphas.
+
+        <Tip warning={true}>
+
+        We support loading A1111 formatted LoRA checkpoints in a limited capacity.
+
+        This function is experimental and might change in the future.
+
+        </Tip>
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                Can be either:
+
+                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+                      with [`ModelMixin.save_pretrained`].
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+
+        """
+        # Load the main state dict first which has the LoRA layers for either of
+        # transformer and text encoder or both.
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
+        revision = kwargs.pop("revision", None)
+        subfolder = kwargs.pop("subfolder", None)
+        weight_name = kwargs.pop("weight_name", None)
+        use_safetensors = kwargs.pop("use_safetensors", None)
+
+        allow_pickle = False
+        if use_safetensors is None:
+            use_safetensors = True
+            allow_pickle = True
+
+        user_agent = {
+            "file_type": "attn_procs_weights",
+            "framework": "pytorch",
+        }
+
+        state_dict = _fetch_state_dict(
+            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
+            weight_name=weight_name,
+            use_safetensors=use_safetensors,
+            local_files_only=local_files_only,
+            cache_dir=cache_dir,
+            force_download=force_download,
+            proxies=proxies,
+            token=token,
+            revision=revision,
+            subfolder=subfolder,
+            user_agent=user_agent,
+            allow_pickle=allow_pickle,
+        )
+
+        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
+        if is_dora_scale_present:
+            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
+            logger.warning(warn_msg)
+            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}
+
+        return state_dict
+
+    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.load_lora_weights
+    def load_lora_weights(
+        self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], adapter_name=None, **kwargs
+    ):
+        """
+        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
+        `self.text_encoder`. All kwargs are forwarded to `self.lora_state_dict`. See
+        [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
+        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
+        dict is loaded into `self.transformer`.
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+            low_cpu_mem_usage (`bool`, *optional*):
+                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
+                weights.
+            kwargs (`dict`, *optional*):
+                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
+        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
+            raise ValueError(
+                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
+            )
+
+        # if a dict is passed, copy it instead of modifying it inplace
+        if isinstance(pretrained_model_name_or_path_or_dict, dict):
+            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()
+
+        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
+        state_dict = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
+
+        is_correct_format = all("lora" in key for key in state_dict.keys())
+        if not is_correct_format:
+            raise ValueError("Invalid LoRA checkpoint.")
+
+        self.load_lora_into_transformer(
+            state_dict,
+            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
+            adapter_name=adapter_name,
+            _pipeline=self,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+        )
+
+    @classmethod
+    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->HunyuanVideoTransformer3DModel
+    def load_lora_into_transformer(
+        cls, state_dict, transformer, adapter_name=None, _pipeline=None, low_cpu_mem_usage=False
+    ):
+        """
+        This will load the LoRA layers specified in `state_dict` into `transformer`.
+
+        Parameters:
+            state_dict (`dict`):
+                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
+                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
+                encoder lora layers.
+            transformer (`HunyuanVideoTransformer3DModel`):
+                The Transformer model to load the LoRA layers into.
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+            low_cpu_mem_usage (`bool`, *optional*):
+                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
+                weights.
+        """
+        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
+            raise ValueError(
+                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
+            )
+
+        # Load the layers corresponding to transformer.
+        logger.info(f"Loading {cls.transformer_name}.")
+        transformer.load_lora_adapter(
+            state_dict,
+            network_alphas=None,
+            adapter_name=adapter_name,
+            _pipeline=_pipeline,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+        )
+
+    @classmethod
+    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.save_lora_weights
+    def save_lora_weights(
+        cls,
+        save_directory: Union[str, os.PathLike],
+        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+        is_main_process: bool = True,
+        weight_name: str = None,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+    ):
+        r"""
+        Save the LoRA parameters corresponding to the UNet and text encoder.
+
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to save LoRA parameters to. Will be created if it doesn't exist.
+            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
+                State dict of the LoRA layers corresponding to the `transformer`.
+            is_main_process (`bool`, *optional*, defaults to `True`):
+                Whether the process calling this is the main process or not. Useful during distributed training and you
+                need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+                process to avoid race conditions.
+            save_function (`Callable`):
+                The function to use to save the state dictionary. Useful during distributed training when you need to
+                replace `torch.save` with another method. Can be configured with the environment variable
+                `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
+        """
+        state_dict = {}
+
+        if not transformer_lora_layers:
+            raise ValueError("You must pass `transformer_lora_layers`.")
+
+        if transformer_lora_layers:
+            state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))
+
+        # Save the model
+        cls.write_lora_layers(
+            state_dict=state_dict,
+            save_directory=save_directory,
+            is_main_process=is_main_process,
+            weight_name=weight_name,
+            save_function=save_function,
+            safe_serialization=safe_serialization,
+        )
+
+    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.fuse_lora with unet->transformer
+    def fuse_lora(
+        self,
+        components: List[str] = ["transformer", "text_encoder"],
+        lora_scale: float = 1.0,
+        safe_fusing: bool = False,
+        adapter_names: Optional[List[str]] = None,
+        **kwargs,
+    ):
+        r"""
+        Fuses the LoRA parameters into the original parameters of the corresponding blocks.
+
+        <Tip warning={true}>
+
+        This is an experimental API.
+
+        </Tip>
+
+        Args:
+            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
+            lora_scale (`float`, defaults to 1.0):
+                Controls how much to influence the outputs with the LoRA parameters.
+            safe_fusing (`bool`, defaults to `False`):
+                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
+            adapter_names (`List[str]`, *optional*):
+                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.
+
+        Example:
+
+        ```py
+        from diffusers import DiffusionPipeline
+        import torch
+
+        pipeline = DiffusionPipeline.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ).to("cuda")
+        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
+        pipeline.fuse_lora(lora_scale=0.7)
+        ```
+        """
+        super().fuse_lora(
+            components=components, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names
+        )
+
+    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.unfuse_lora with unet->transformer
+    def unfuse_lora(self, components: List[str] = ["transformer", "text_encoder"], **kwargs):
+        r"""
+        Reverses the effect of
+        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).
+
+        <Tip warning={true}>
+
+        This is an experimental API.
+
+        </Tip>
+
+        Args:
+            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
+            unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
+            unfuse_text_encoder (`bool`, defaults to `True`):
+                Whether to unfuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
+                LoRA parameters then it won't have any effect.
+        """
+        super().unfuse_lora(components=components)
+
+
 class LoraLoaderMixin(StableDiffusionLoraLoaderMixin):
     def __init__(self, *args, **kwargs):
         deprecation_message = "LoraLoaderMixin is deprecated and this will be removed in a future version. Please use `StableDiffusionLoraLoaderMixin`, instead."

src/diffusers/loaders/peft.py

@@ -53,6 +53,7 @@ _SET_ADAPTER_SCALE_FN_MAPPING = {
     "FluxTransformer2DModel": lambda model_cls, weights: weights,
     "CogVideoXTransformer3DModel": lambda model_cls, weights: weights,
     "MochiTransformer3DModel": lambda model_cls, weights: weights,
+    "HunyuanVideoTransformer3DModel": lambda model_cls, weights: weights,
     "LTXVideoTransformer3DModel": lambda model_cls, weights: weights,
     "SanaTransformer2DModel": lambda model_cls, weights: weights,
 }

src/diffusers/models/transformers/transformer_hunyuan_video.py

@@ -19,7 +19,8 @@ import torch.nn as nn
 import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
-from ...utils import is_torch_version
+from ...loaders import PeftAdapterMixin
+from ...utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
 from ..attention import FeedForward
 from ..attention_processor import Attention, AttentionProcessor
 from ..embeddings import (
@@ -32,6 +33,9 @@ from ..modeling_utils import ModelMixin
 from ..normalization import AdaLayerNormContinuous, AdaLayerNormZero, AdaLayerNormZeroSingle
 
 
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
 class HunyuanVideoAttnProcessor2_0:
     def __init__(self):
         if not hasattr(F, "scaled_dot_product_attention"):
@@ -496,7 +500,7 @@ class HunyuanVideoTransformerBlock(nn.Module):
         return hidden_states, encoder_hidden_states
 
 
-class HunyuanVideoTransformer3DModel(ModelMixin, ConfigMixin):
+class HunyuanVideoTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
     r"""
     A Transformer model for video-like data used in [HunyuanVideo](https://huggingface.co/tencent/HunyuanVideo).
 
@@ -670,8 +674,24 @@ class HunyuanVideoTransformer3DModel(ModelMixin, ConfigMixin):
         encoder_attention_mask: torch.Tensor,
         pooled_projections: torch.Tensor,
         guidance: torch.Tensor = None,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
         return_dict: bool = True,
     ) -> Union[torch.Tensor, Dict[str, torch.Tensor]]:
+        if attention_kwargs is not None:
+            attention_kwargs = attention_kwargs.copy()
+            lora_scale = attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
         batch_size, num_channels, num_frames, height, width = hidden_states.shape
         p, p_t = self.config.patch_size, self.config.patch_size_t
         post_patch_num_frames = num_frames // p_t
@@ -757,6 +777,10 @@ class HunyuanVideoTransformer3DModel(ModelMixin, ConfigMixin):
         hidden_states = hidden_states.permute(0, 4, 1, 5, 2, 6, 3, 7)
         hidden_states = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)
 
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
         if not return_dict:
             return (hidden_states,)
 

src/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video.py

@@ -20,6 +20,7 @@ import torch
 from transformers import CLIPTextModel, CLIPTokenizer, LlamaModel, LlamaTokenizerFast
 
 from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...loaders import HunyuanVideoLoraLoaderMixin
 from ...models import AutoencoderKLHunyuanVideo, HunyuanVideoTransformer3DModel
 from ...schedulers import FlowMatchEulerDiscreteScheduler
 from ...utils import logging, replace_example_docstring
@@ -132,7 +133,7 @@ def retrieve_timesteps(
     return timesteps, num_inference_steps
 
 
-class HunyuanVideoPipeline(DiffusionPipeline):
+class HunyuanVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoaderMixin):
     r"""
     Pipeline for text-to-video generation using HunyuanVideo.
 
@@ -447,6 +448,10 @@ class HunyuanVideoPipeline(DiffusionPipeline):
     def num_timesteps(self):
         return self._num_timesteps
 
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
     @property
     def interrupt(self):
         return self._interrupt
@@ -471,6 +476,7 @@ class HunyuanVideoPipeline(DiffusionPipeline):
         prompt_attention_mask: Optional[torch.Tensor] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
         callback_on_step_end: Optional[
             Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
         ] = None,
@@ -525,6 +531,10 @@ class HunyuanVideoPipeline(DiffusionPipeline):
                 The output format of the generated image. Choose between `PIL.Image` or `np.array`.
             return_dict (`bool`, *optional*, defaults to `True`):
                 Whether or not to return a [`HunyuanVideoPipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
             clip_skip (`int`, *optional*):
                 Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
                 the output of the pre-final layer will be used for computing the prompt embeddings.
@@ -562,6 +572,7 @@ class HunyuanVideoPipeline(DiffusionPipeline):
         )
 
         self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
         self._interrupt = False
 
         device = self._execution_device
@@ -640,6 +651,7 @@ class HunyuanVideoPipeline(DiffusionPipeline):
                     encoder_attention_mask=prompt_attention_mask,
                     pooled_projections=pooled_prompt_embeds,
                     guidance=guidance,
+                    attention_kwargs=attention_kwargs,
                     return_dict=False,
                 )[0]
 

tests/lora/test_lora_layers_hunyuanvideo.py

+# Copyright 2024 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import sys
+import unittest
+
+import numpy as np
+import pytest
+import torch
+from transformers import CLIPTextModel, CLIPTokenizer, LlamaModel, LlamaTokenizerFast
+
+from diffusers import (
+    AutoencoderKLHunyuanVideo,
+    FlowMatchEulerDiscreteScheduler,
+    HunyuanVideoPipeline,
+    HunyuanVideoTransformer3DModel,
+)
+from diffusers.utils.testing_utils import (
+    floats_tensor,
+    is_torch_version,
+    require_peft_backend,
+    skip_mps,
+    torch_device,
+)
+
+
+sys.path.append(".")
+
+from utils import PeftLoraLoaderMixinTests, check_if_lora_correctly_set  # noqa: E402
+
+
+@require_peft_backend
+@skip_mps
+class HunyuanVideoLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
+    pipeline_class = HunyuanVideoPipeline
+    scheduler_cls = FlowMatchEulerDiscreteScheduler
+    scheduler_classes = [FlowMatchEulerDiscreteScheduler]
+    scheduler_kwargs = {}
+
+    transformer_kwargs = {
+        "in_channels": 4,
+        "out_channels": 4,
+        "num_attention_heads": 2,
+        "attention_head_dim": 10,
+        "num_layers": 1,
+        "num_single_layers": 1,
+        "num_refiner_layers": 1,
+        "patch_size": 1,
+        "patch_size_t": 1,
+        "guidance_embeds": True,
+        "text_embed_dim": 16,
+        "pooled_projection_dim": 8,
+        "rope_axes_dim": (2, 4, 4),
+    }
+    transformer_cls = HunyuanVideoTransformer3DModel
+    vae_kwargs = {
+        "in_channels": 3,
+        "out_channels": 3,
+        "latent_channels": 4,
+        "down_block_types": (
+            "HunyuanVideoDownBlock3D",
+            "HunyuanVideoDownBlock3D",
+            "HunyuanVideoDownBlock3D",
+            "HunyuanVideoDownBlock3D",
+        ),
+        "up_block_types": (
+            "HunyuanVideoUpBlock3D",
+            "HunyuanVideoUpBlock3D",
+            "HunyuanVideoUpBlock3D",
+            "HunyuanVideoUpBlock3D",
+        ),
+        "block_out_channels": (8, 8, 8, 8),
+        "layers_per_block": 1,
+        "act_fn": "silu",
+        "norm_num_groups": 4,
+        "scaling_factor": 0.476986,
+        "spatial_compression_ratio": 8,
+        "temporal_compression_ratio": 4,
+        "mid_block_add_attention": True,
+    }
+    vae_cls = AutoencoderKLHunyuanVideo
+    has_two_text_encoders = True
+    tokenizer_cls, tokenizer_id, tokenizer_subfolder = (
+        LlamaTokenizerFast,
+        "hf-internal-testing/tiny-random-hunyuanvideo",
+        "tokenizer",
+    )
+    tokenizer_2_cls, tokenizer_2_id, tokenizer_2_subfolder = (
+        CLIPTokenizer,
+        "hf-internal-testing/tiny-random-hunyuanvideo",
+        "tokenizer_2",
+    )
+    text_encoder_cls, text_encoder_id, text_encoder_subfolder = (
+        LlamaModel,
+        "hf-internal-testing/tiny-random-hunyuanvideo",
+        "text_encoder",
+    )
+    text_encoder_2_cls, text_encoder_2_id, text_encoder_2_subfolder = (
+        CLIPTextModel,
+        "hf-internal-testing/tiny-random-hunyuanvideo",
+        "text_encoder_2",
+    )
+
+    @property
+    def output_shape(self):
+        return (1, 9, 32, 32, 3)
+
+    def get_dummy_inputs(self, with_generator=True):
+        batch_size = 1
+        sequence_length = 16
+        num_channels = 4
+        num_frames = 9
+        num_latent_frames = 3  # (num_frames - 1) // temporal_compression_ratio + 1
+        sizes = (4, 4)
+
+        generator = torch.manual_seed(0)
+        noise = floats_tensor((batch_size, num_latent_frames, num_channels) + sizes)
+        input_ids = torch.randint(1, sequence_length, size=(batch_size, sequence_length), generator=generator)
+
+        pipeline_inputs = {
+            "prompt": "",
+            "num_frames": num_frames,
+            "num_inference_steps": 1,
+            "guidance_scale": 6.0,
+            "height": 32,
+            "width": 32,
+            "max_sequence_length": sequence_length,
+            "prompt_template": {"template": "{}", "crop_start": 0},
+            "output_type": "np",
+        }
+        if with_generator:
+            pipeline_inputs.update({"generator": generator})
+
+        return noise, input_ids, pipeline_inputs
+
+    @pytest.mark.xfail(
+        condition=torch.device(torch_device).type == "cpu" and is_torch_version(">=", "2.5"),
+        reason="Test currently fails on CPU and PyTorch 2.5.1 but not on PyTorch 2.4.1.",
+        strict=True,
+    )
+    def test_lora_fuse_nan(self):
+        for scheduler_cls in self.scheduler_classes:
+            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
+            pipe = self.pipeline_class(**components)
+            pipe = pipe.to(torch_device)
+            pipe.set_progress_bar_config(disable=None)
+            _, _, inputs = self.get_dummy_inputs(with_generator=False)
+
+            pipe.transformer.add_adapter(denoiser_lora_config, "adapter-1")
+
+            self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in denoiser")
+
+            # corrupt one LoRA weight with `inf` values
+            with torch.no_grad():
+                pipe.transformer.transformer_blocks[0].attn.to_q.lora_A["adapter-1"].weight += float("inf")
+
+            # with `safe_fusing=True` we should see an Error
+            with self.assertRaises(ValueError):
+                pipe.fuse_lora(components=self.pipeline_class._lora_loadable_modules, safe_fusing=True)
+
+            # without we should not see an error, but every image will be black
+            pipe.fuse_lora(components=self.pipeline_class._lora_loadable_modules, safe_fusing=False)
+
+            out = pipe(
+                prompt=inputs["prompt"],
+                height=inputs["height"],
+                width=inputs["width"],
+                num_frames=inputs["num_frames"],
+                num_inference_steps=inputs["num_inference_steps"],
+                max_sequence_length=inputs["max_sequence_length"],
+                output_type="np",
+            )[0]
+
+            self.assertTrue(np.isnan(out).all())
+
+    def test_simple_inference_with_text_lora_denoiser_fused_multi(self):
+        super().test_simple_inference_with_text_lora_denoiser_fused_multi(expected_atol=9e-3)
+
+    def test_simple_inference_with_text_denoiser_lora_unfused(self):
+        super().test_simple_inference_with_text_denoiser_lora_unfused(expected_atol=9e-3)
+
+    # TODO(aryan): Fix the following test
+    @unittest.skip("This test fails with an error I haven't been able to debug yet.")
+    def test_simple_inference_save_pretrained(self):
+        pass
+
+    @unittest.skip("Not supported in HunyuanVideo.")
+    def test_simple_inference_with_text_denoiser_block_scale(self):
+        pass
+
+    @unittest.skip("Not supported in HunyuanVideo.")
+    def test_simple_inference_with_text_denoiser_block_scale_for_all_dict_options(self):
+        pass
+
+    @unittest.skip("Not supported in HunyuanVideo.")
+    def test_modify_padding_mode(self):
+        pass
+
+    @unittest.skip("Text encoder LoRA is not supported in HunyuanVideo.")
+    def test_simple_inference_with_partial_text_lora(self):
+        pass
+
+    @unittest.skip("Text encoder LoRA is not supported in HunyuanVideo.")
+    def test_simple_inference_with_text_lora(self):
+        pass
+
+    @unittest.skip("Text encoder LoRA is not supported in HunyuanVideo.")
+    def test_simple_inference_with_text_lora_and_scale(self):
+        pass
+
+    @unittest.skip("Text encoder LoRA is not supported in HunyuanVideo.")
+    def test_simple_inference_with_text_lora_fused(self):
+        pass
+
+    @unittest.skip("Text encoder LoRA is not supported in HunyuanVideo.")
+    def test_simple_inference_with_text_lora_save_load(self):
+        pass

tests/lora/utils.py

@@ -89,12 +89,12 @@ class PeftLoraLoaderMixinTests:
 
     has_two_text_encoders = False
     has_three_text_encoders = False
-    text_encoder_cls, text_encoder_id = None, None
-    text_encoder_2_cls, text_encoder_2_id = None, None
-    text_encoder_3_cls, text_encoder_3_id = None, None
-    tokenizer_cls, tokenizer_id = None, None
-    tokenizer_2_cls, tokenizer_2_id = None, None
-    tokenizer_3_cls, tokenizer_3_id = None, None
+    text_encoder_cls, text_encoder_id, text_encoder_subfolder = None, None, None
+    text_encoder_2_cls, text_encoder_2_id, text_encoder_2_subfolder = None, None, None
+    text_encoder_3_cls, text_encoder_3_id, text_encoder_3_subfolder = None, None, None
+    tokenizer_cls, tokenizer_id, tokenizer_subfolder = None, None, None
+    tokenizer_2_cls, tokenizer_2_id, tokenizer_2_subfolder = None, None, None
+    tokenizer_3_cls, tokenizer_3_id, tokenizer_3_subfolder = None, None, None
 
     unet_kwargs = None
     transformer_cls = None
@@ -124,16 +124,26 @@ class PeftLoraLoaderMixinTests:
         torch.manual_seed(0)
         vae = self.vae_cls(**self.vae_kwargs)
 
-        text_encoder = self.text_encoder_cls.from_pretrained(self.text_encoder_id)
-        tokenizer = self.tokenizer_cls.from_pretrained(self.tokenizer_id)
+        text_encoder = self.text_encoder_cls.from_pretrained(
+            self.text_encoder_id, subfolder=self.text_encoder_subfolder
+        )
+        tokenizer = self.tokenizer_cls.from_pretrained(self.tokenizer_id, subfolder=self.tokenizer_subfolder)
 
         if self.text_encoder_2_cls is not None:
-            text_encoder_2 = self.text_encoder_2_cls.from_pretrained(self.text_encoder_2_id)
-            tokenizer_2 = self.tokenizer_2_cls.from_pretrained(self.tokenizer_2_id)
+            text_encoder_2 = self.text_encoder_2_cls.from_pretrained(
+                self.text_encoder_2_id, subfolder=self.text_encoder_2_subfolder
+            )
+            tokenizer_2 = self.tokenizer_2_cls.from_pretrained(
+                self.tokenizer_2_id, subfolder=self.tokenizer_2_subfolder
+            )
 
         if self.text_encoder_3_cls is not None:
-            text_encoder_3 = self.text_encoder_3_cls.from_pretrained(self.text_encoder_3_id)
-            tokenizer_3 = self.tokenizer_3_cls.from_pretrained(self.tokenizer_3_id)
+            text_encoder_3 = self.text_encoder_3_cls.from_pretrained(
+                self.text_encoder_3_id, subfolder=self.text_encoder_3_subfolder
+            )
+            tokenizer_3 = self.tokenizer_3_cls.from_pretrained(
+                self.tokenizer_3_id, subfolder=self.tokenizer_3_subfolder
+            )
 
         text_lora_config = LoraConfig(
             r=rank,