cogvideox lora training draft

src/diffusers/loaders/__init__.py

@@ -67,6 +67,7 @@ if is_torch_available():
             "StableDiffusionXLLoraLoaderMixin",
             "LoraLoaderMixin",
             "FluxLoraLoaderMixin",
+            "CogVideoXLoraLoaderMixin",
         ]
         _import_structure["textual_inversion"] = ["TextualInversionLoaderMixin"]
         _import_structure["ip_adapter"] = ["IPAdapterMixin"]
@@ -84,6 +85,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             from .ip_adapter import IPAdapterMixin
             from .lora_pipeline import (
                 AmusedLoraLoaderMixin,
+                CogVideoXLoraLoaderMixin,
                 FluxLoraLoaderMixin,
                 LoraLoaderMixin,
                 SD3LoraLoaderMixin,

src/diffusers/loaders/lora_pipeline.py

@@ -2257,6 +2257,468 @@ class AmusedLoraLoaderMixin(StableDiffusionLoraLoaderMixin):
         )
 
 
+class CogVideoXLoraLoaderMixin(LoraBaseMixin):
+    r"""
+    Load LoRA layers into [`CogVideoXTransformer3DModel`],
+    [`T5EncoderModel`](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel). Specific to
+    [`CogVideoX`].
+    """
+
+    _lora_loadable_modules = ["transformer", "text_encoder"]
+    transformer_name = TRANSFORMER_NAME
+    text_encoder_name = TEXT_ENCODER_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 = cls._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,
+        )
+
+        return state_dict
+
+    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`].
+            kwargs (`dict`, *optional*):
+                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.
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        # 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 or "dora_scale" 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,
+        )
+
+        text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k}
+        if len(text_encoder_state_dict) > 0:
+            self.load_lora_into_text_encoder(
+                text_encoder_state_dict,
+                network_alphas=None,
+                text_encoder=self.text_encoder,
+                prefix="text_encoder",
+                lora_scale=self.lora_scale,
+                adapter_name=adapter_name,
+                _pipeline=self,
+            )
+
+    @classmethod
+    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer
+    def load_lora_into_transformer(cls, state_dict, transformer, adapter_name=None, _pipeline=None):
+        """
+        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 (`SD3Transformer2DModel`):
+                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.
+        """
+        from peft import LoraConfig, inject_adapter_in_model, set_peft_model_state_dict
+
+        keys = list(state_dict.keys())
+
+        transformer_keys = [k for k in keys if k.startswith(cls.transformer_name)]
+        state_dict = {
+            k.replace(f"{cls.transformer_name}.", ""): v for k, v in state_dict.items() if k in transformer_keys
+        }
+
+        if len(state_dict.keys()) > 0:
+            # check with first key if is not in peft format
+            first_key = next(iter(state_dict.keys()))
+            if "lora_A" not in first_key:
+                state_dict = convert_unet_state_dict_to_peft(state_dict)
+
+            if adapter_name in getattr(transformer, "peft_config", {}):
+                raise ValueError(
+                    f"Adapter name {adapter_name} already in use in the transformer - please select a new adapter name."
+                )
+
+            rank = {}
+            for key, val in state_dict.items():
+                if "lora_B" in key:
+                    rank[key] = val.shape[1]
+
+            lora_config_kwargs = get_peft_kwargs(rank, network_alpha_dict=None, peft_state_dict=state_dict)
+            if "use_dora" in lora_config_kwargs:
+                if lora_config_kwargs["use_dora"] and is_peft_version("<", "0.9.0"):
+                    raise ValueError(
+                        "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
+                    )
+                else:
+                    lora_config_kwargs.pop("use_dora")
+            lora_config = LoraConfig(**lora_config_kwargs)
+
+            # adapter_name
+            if adapter_name is None:
+                adapter_name = get_adapter_name(transformer)
+
+            # In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
+            # otherwise loading LoRA weights will lead to an error
+            is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
+
+            inject_adapter_in_model(lora_config, transformer, adapter_name=adapter_name)
+            incompatible_keys = set_peft_model_state_dict(transformer, state_dict, adapter_name)
+
+            if incompatible_keys is not None:
+                # check only for unexpected keys
+                unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
+                if unexpected_keys:
+                    logger.warning(
+                        f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
+                        f" {unexpected_keys}. "
+                    )
+
+            # Offload back.
+            if is_model_cpu_offload:
+                _pipeline.enable_model_cpu_offload()
+            elif is_sequential_cpu_offload:
+                _pipeline.enable_sequential_cpu_offload()
+            # Unsafe code />
+
+    @classmethod
+    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.load_lora_into_text_encoder
+    def load_lora_into_text_encoder(
+        cls,
+        state_dict,
+        network_alphas,
+        text_encoder,
+        prefix=None,
+        lora_scale=1.0,
+        adapter_name=None,
+        _pipeline=None,
+    ):
+        """
+        This will load the LoRA layers specified in `state_dict` into `text_encoder`
+
+        Parameters:
+            state_dict (`dict`):
+                A standard state dict containing the lora layer parameters. The key should be prefixed with an
+                additional `text_encoder` to distinguish between unet lora layers.
+            network_alphas (`Dict[str, float]`):
+                See `LoRALinearLayer` for more details.
+            text_encoder (`T5EncoderModel`):
+                The text encoder model to load the LoRA layers into.
+            prefix (`str`):
+                Expected prefix of the `text_encoder` in the `state_dict`.
+            lora_scale (`float`):
+                How much to scale the output of the lora linear layer before it is added with the output of the regular
+                lora layer.
+            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.
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        from peft import LoraConfig
+
+        # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
+        # then the `state_dict` keys should have `self.unet_name` and/or `self.text_encoder_name` as
+        # their prefixes.
+        keys = list(state_dict.keys())
+        prefix = cls.text_encoder_name if prefix is None else prefix
+
+        # Safe prefix to check with.
+        if any(cls.text_encoder_name in key for key in keys):
+            # Load the layers corresponding to text encoder and make necessary adjustments.
+            text_encoder_keys = [k for k in keys if k.startswith(prefix) and k.split(".")[0] == prefix]
+            text_encoder_lora_state_dict = {
+                k.replace(f"{prefix}.", ""): v for k, v in state_dict.items() if k in text_encoder_keys
+            }
+
+            if len(text_encoder_lora_state_dict) > 0:
+                logger.info(f"Loading {prefix}.")
+                rank = {}
+                text_encoder_lora_state_dict = convert_state_dict_to_diffusers(text_encoder_lora_state_dict)
+
+                # convert state dict
+                text_encoder_lora_state_dict = convert_state_dict_to_peft(text_encoder_lora_state_dict)
+
+                for name, _ in text_encoder_attn_modules(text_encoder):
+                    for module in ("out_proj", "q_proj", "k_proj", "v_proj"):
+                        rank_key = f"{name}.{module}.lora_B.weight"
+                        if rank_key not in text_encoder_lora_state_dict:
+                            continue
+                        rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]
+
+                for name, _ in text_encoder_mlp_modules(text_encoder):
+                    for module in ("fc1", "fc2"):
+                        rank_key = f"{name}.{module}.lora_B.weight"
+                        if rank_key not in text_encoder_lora_state_dict:
+                            continue
+                        rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]
+
+                if network_alphas is not None:
+                    alpha_keys = [
+                        k for k in network_alphas.keys() if k.startswith(prefix) and k.split(".")[0] == prefix
+                    ]
+                    network_alphas = {
+                        k.replace(f"{prefix}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
+                    }
+
+                lora_config_kwargs = get_peft_kwargs(rank, network_alphas, text_encoder_lora_state_dict, is_unet=False)
+                if "use_dora" in lora_config_kwargs:
+                    if lora_config_kwargs["use_dora"]:
+                        if is_peft_version("<", "0.9.0"):
+                            raise ValueError(
+                                "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
+                            )
+                    else:
+                        if is_peft_version("<", "0.9.0"):
+                            lora_config_kwargs.pop("use_dora")
+                lora_config = LoraConfig(**lora_config_kwargs)
+
+                # adapter_name
+                if adapter_name is None:
+                    adapter_name = get_adapter_name(text_encoder)
+
+                is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
+
+                # inject LoRA layers and load the state dict
+                # in transformers we automatically check whether the adapter name is already in use or not
+                text_encoder.load_adapter(
+                    adapter_name=adapter_name,
+                    adapter_state_dict=text_encoder_lora_state_dict,
+                    peft_config=lora_config,
+                )
+
+                # scale LoRA layers with `lora_scale`
+                scale_lora_layers(text_encoder, weight=lora_scale)
+
+                text_encoder.to(device=text_encoder.device, dtype=text_encoder.dtype)
+
+                # Offload back.
+                if is_model_cpu_offload:
+                    _pipeline.enable_model_cpu_offload()
+                elif is_sequential_cpu_offload:
+                    _pipeline.enable_sequential_cpu_offload()
+                # Unsafe code />
+
+    @classmethod
+    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.save_lora_weights with unet->transformer
+    def save_lora_weights(
+        cls,
+        save_directory: Union[str, os.PathLike],
+        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+        text_encoder_lora_layers: Dict[str, torch.nn.Module] = 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`.
+            text_encoder_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
+                State dict of the LoRA layers corresponding to the `text_encoder`. Must explicitly pass the text
+                encoder LoRA state dict because it comes from 🤗 Transformers.
+            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 or text_encoder_lora_layers):
+            raise ValueError("You must pass at least one of `transformer_lora_layers` and `text_encoder_lora_layers`.")
+
+        if transformer_lora_layers:
+            state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))
+
+        if text_encoder_lora_layers:
+            state_dict.update(cls.pack_weights(text_encoder_lora_layers, cls.text_encoder_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
+        )
+
+    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.
+        """
+        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

@@ -33,6 +33,7 @@ _SET_ADAPTER_SCALE_FN_MAPPING = {
     "UNetMotionModel": _maybe_expand_lora_scales,
     "SD3Transformer2DModel": lambda model_cls, weights: weights,
     "FluxTransformer2DModel": lambda model_cls, weights: weights,
+    "CogVideoXTransformer3DModel": lambda model_cls, weights: weights,
 }
 
 

src/diffusers/models/transformers/cogvideox_transformer_3d.py

@@ -19,6 +19,7 @@ import torch
 from torch import nn
 
 from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import PeftAdapterMixin
 from ...utils import is_torch_version, logging
 from ...utils.torch_utils import maybe_allow_in_graph
 from ..attention import Attention, FeedForward
@@ -152,7 +153,7 @@ class CogVideoXBlock(nn.Module):
         return hidden_states, encoder_hidden_states
 
 
-class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
+class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
     """
     A Transformer model for video-like data in [CogVideoX](https://github.com/THUDM/CogVideo).
 

src/diffusers/pipelines/cogvideo/pipeline_cogvideox.py

@@ -22,11 +22,19 @@ import torch
 from transformers import T5EncoderModel, T5Tokenizer
 
 from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...loaders import CogVideoXLoraLoaderMixin
 from ...models import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel
 from ...models.embeddings import get_3d_rotary_pos_embed
 from ...pipelines.pipeline_utils import DiffusionPipeline
 from ...schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
-from ...utils import BaseOutput, logging, replace_example_docstring
+from ...utils import (
+    USE_PEFT_BACKEND,
+    BaseOutput,
+    logging,
+    replace_example_docstring,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
 from ...utils.torch_utils import randn_tensor
 from ...video_processor import VideoProcessor
 
@@ -151,7 +159,7 @@ class CogVideoXPipelineOutput(BaseOutput):
     frames: torch.Tensor
 
 
-class CogVideoXPipeline(DiffusionPipeline):
+class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
     r"""
     Pipeline for text-to-video generation using CogVideoX.
 
@@ -258,6 +266,7 @@ class CogVideoXPipeline(DiffusionPipeline):
         max_sequence_length: int = 226,
         device: Optional[torch.device] = None,
         dtype: Optional[torch.dtype] = None,
+        lora_scale: Optional[float] = None,
     ):
         r"""
         Encodes the prompt into text encoder hidden states.
@@ -284,9 +293,20 @@ class CogVideoXPipeline(DiffusionPipeline):
                 torch device
             dtype: (`torch.dtype`, *optional*):
                 torch dtype
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
         """
         device = device or self._execution_device
 
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, CogVideoXLoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if self.text_encoder is not None and USE_PEFT_BACKEND:
+                scale_lora_layers(self.text_encoder, lora_scale)
+
         prompt = [prompt] if isinstance(prompt, str) else prompt
         if prompt is not None:
             batch_size = len(prompt)
@@ -326,6 +346,11 @@ class CogVideoXPipeline(DiffusionPipeline):
                 dtype=dtype,
             )
 
+        if self.text_encoder is not None:
+            if isinstance(self, CogVideoXLoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder, lora_scale)
+
         return prompt_embeds, negative_prompt_embeds
 
     def prepare_latents(
@@ -507,6 +532,7 @@ class CogVideoXPipeline(DiffusionPipeline):
         ] = None,
         callback_on_step_end_tensor_inputs: List[str] = ["latents"],
         max_sequence_length: int = 226,
+        lora_scale: Optional[float] = None,
     ) -> Union[CogVideoXPipelineOutput, Tuple]:
         """
         Function invoked when calling the pipeline for generation.
@@ -634,6 +660,7 @@ class CogVideoXPipeline(DiffusionPipeline):
             negative_prompt_embeds=negative_prompt_embeds,
             max_sequence_length=max_sequence_length,
             device=device,
+            lora_scale=lora_scale,
         )
         if do_classifier_free_guidance:
             prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)