Buckets:

mishig's picture
|
download
raw
164 kB
# LoRA LoRA 是一种快速且轻量级的训练方法,它通过插入并训练数量远少于模型全部参数的参数来实现训练。这会产生更小的文件(约 100 MB),并使快速训练模型以学习新概念变得更加容易。LoRA 权重通常加载到去噪器、文本编码器或两者中。去噪器通常对应于 UNet(例如 [UNet2DConditionModel](/docs/diffusers/v0.43.0/zh/api/models/unet2d-cond#diffusers.UNet2DConditionModel))或 Transformer(例如 [SD3Transformer2DModel](/docs/diffusers/v0.43.0/zh/api/models/sd3_transformer2d#diffusers.SD3Transformer2DModel))。有几种用于加载 LoRA 权重的类: - `StableDiffusionLoraLoaderMixin` 提供了用于加载和卸载、融合和取消融合、启用和禁用等管理 LoRA 权重的函数。该类可与任何模型配合使用。 - `StableDiffusionXLLoraLoaderMixin` 是 `StableDiffusionLoraLoaderMixin` 类的 [Stable Diffusion (SDXL)](../../api/pipelines/stable_diffusion/stable_diffusion_xl) 版本,用于加载和保存 LoRA 权重。它仅能与 SDXL 模型配合使用。 - `SD3LoraLoaderMixin` 为 [Stable Diffusion 3](https://huggingface.co/blog/sd3) 提供了类似的功能。 - `FluxLoraLoaderMixin` 为 [Flux](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux) 提供了类似的功能。 - `CogVideoXLoraLoaderMixin` 为 [CogVideoX](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogvideox) 提供了类似的功能。 - `Mochi1LoraLoaderMixin` 为 [Mochi](https://huggingface.co/docs/diffusers/main/en/api/pipelines/mochi) 提供了类似的功能。 - `AuraFlowLoraLoaderMixin` 为 [AuraFlow](https://huggingface.co/fal/AuraFlow) 提供了类似的功能。 - `LTXVideoLoraLoaderMixin` 为 [LTX-Video](https://huggingface.co/docs/diffusers/main/en/api/pipelines/ltx_video) 提供了类似的功能。 - `SanaLoraLoaderMixin` 为 [Sana](https://huggingface.co/docs/diffusers/main/en/api/pipelines/sana) 提供了类似的功能。 - `HeliosLoraLoaderMixin` 为 [HunyuanVideo](https://huggingface.co/docs/diffusers/main/en/api/pipelines/helios) 提供了类似的功能。 - `HunyuanVideoLoraLoaderMixin` 为 [HunyuanVideo](https://huggingface.co/docs/diffusers/main/en/api/pipelines/hunyuan_video) 提供了类似的功能。 - `Lumina2LoraLoaderMixin` 为 [Lumina2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/lumina2) 提供了类似的功能。 - `WanLoraLoaderMixin` 为 [Wan](https://huggingface.co/docs/diffusers/main/en/api/pipelines/wan) 提供了类似的功能。 - `SkyReelsV2LoraLoaderMixin` 为 [SkyReels-V2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/skyreels_v2) 提供了类似的功能。 - `CogView4LoraLoaderMixin` 为 [CogView4](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogview4) 提供了类似的功能。 - `AmusedLoraLoaderMixin` 用于 `AmusedPipeline`。 - `AnimaLoraLoaderMixin` 为 [Anima](https://huggingface.co/docs/diffusers/main/en/api/pipelines/anima) 提供了类似的功能。 - `AceStepLoraLoaderMixin` 为 [ACE-Step](https://huggingface.co/docs/diffusers/main/en/api/pipelines/ace_step) 提供类似的功能。 - `HiDreamImageLoraLoaderMixin` 为 [HiDream Image](https://huggingface.co/docs/diffusers/main/en/api/pipelines/hidream) 提供了类似的功能 - `QwenImageLoraLoaderMixin` 为 [Qwen Image](https://huggingface.co/docs/diffusers/main/en/api/pipelines/qwen) 提供了类似的功能。 - `ZImageLoraLoaderMixin` 为 [Z-Image](https://huggingface.co/docs/diffusers/main/en/api/pipelines/zimage) 提供了类似的功能。 - `Flux2LoraLoaderMixin` 为 [Flux2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux2) 提供了类似的功能。 - `ErnieImageLoraLoaderMixin` 为 [Ernie-Image](https://huggingface.co/docs/diffusers/main/en/api/pipelines/ernie_image) 提供了类似的功能。 - `LTX2LoraLoaderMixin` 为 [Flux2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/ltx2) 提供了类似的功能。 - `MiniMaxH3LoraLoaderMixin` 为 [MiniMax-H3](https://huggingface.co/docs/diffusers/main/en/api/pipelines/minimax_h3) 提供了类似的功能。 - `PikaH3FlexLoraLoaderMixin` 为 [Pika H3 Flex](https://huggingface.co/docs/diffusers/main/en/api/pipelines/pika_h3_flex) 提供了类似的功能。 - `RFD4ProteinaLoraLoaderMixin` 为 [RFD4-Proteína](https://huggingface.co/docs/diffusers/main/en/api/pipelines/rfd4_proteina) 提供了类似的功能。 - `BriaFibo2LoraLoaderMixin` 为 [fibo-2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/bria_fibo2) 提供了类似的功能。 - `LoraBaseMixin` 提供了一个基类,其中包含用于融合、解融合、卸载 LoRA 等的若干实用方法。 > [!TIP] > 要了解如何加载 LoRA 权重,请参阅 [LoRA](../../tutorials/using_peft_for_inference) 加载指南。 ## RFD4ProteinaLoraLoaderMixin[[diffusers.loaders.RFD4ProteinaLoraLoaderMixin]] #### diffusers.loaders.RFD4ProteinaLoraLoaderMixin[[diffusers.loaders.RFD4ProteinaLoraLoaderMixin]] ```python diffusers.loaders.RFD4ProteinaLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7532) 将 LoRA 层加载到 [RFD4ProteinaTransformer](/docs/diffusers/v0.43.0/zh/api/models/rfd4_proteina_transformer#diffusers.RFD4ProteinaTransformer) 中。 自编码器与去噪器检查点配对使用,且未进行适配,因此 Transformer 是唯一可加载的模块。 #### fuse_lora[[diffusers.loaders.RFD4ProteinaLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7694) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.RFD4ProteinaLoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7631) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.RFD4ProteinaLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7595) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.RFD4ProteinaLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7543) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.RFD4ProteinaLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7658) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.RFD4ProteinaLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7714) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## LoraBaseMixin[[diffusers.loaders.lora_base.LoraBaseMixin]] #### diffusers.loaders.lora_base.LoraBaseMixin[[diffusers.loaders.lora_base.LoraBaseMixin]] ```python diffusers.loaders.lora_base.LoraBaseMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L491) 用于处理 LoRA 的实用类。 #### delete_adapters[[diffusers.loaders.lora_base.LoraBaseMixin.delete_adapters]] ```python delete_adapters(adapter_names: list[str] | str) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L873) **参数:** adapter_names (`list[str, str]`) : 要删除的适配器名称。 从流水线中删除适配器的 LoRA 层。 示例: ```py from diffusers import AutoPipelineForText2Image import torch pipeline = AutoPipelineForText2Image.from_pretrained( "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16 ).to("cuda") pipeline.load_lora_weights( "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_names="cinematic" ) pipeline.delete_adapters("cinematic") ``` #### disable_lora[[diffusers.loaders.lora_base.LoraBaseMixin.disable_lora]] ```python disable_lora() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L813) 禁用管道中处于激活状态的 LoRA 层。 示例: ```py from diffusers import AutoPipelineForText2Image import torch pipeline = AutoPipelineForText2Image.from_pretrained( "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16 ).to("cuda") pipeline.load_lora_weights( "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic" ) pipeline.disable_lora() ``` #### enable_lora[[diffusers.loaders.lora_base.LoraBaseMixin.enable_lora]] ```python enable_lora() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L843) 启用管道中处于激活状态的 LoRA 层。 示例: ```py from diffusers import AutoPipelineForText2Image import torch pipeline = AutoPipelineForText2Image.from_pretrained( "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16 ).to("cuda") pipeline.load_lora_weights( "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic" ) pipeline.enable_lora() ``` #### enable_lora_hotswap[[diffusers.loaders.lora_base.LoraBaseMixin.enable_lora_hotswap]] ```python enable_lora_hotswap(**kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L1020) **参数:** target_rank (`int`) : 所有将加载的适配器中的最高秩。 check_compiled (`str`,*可选*,默认为 `"error"`):如何处理已编译的模型。该检查可能返回以下消息:- "error"(默认):引发错误 - "warn":发出警告 - "ignore":不执行任何操作 在不触发模型重新编译的情况下热插拔适配器,或者在加载的适配器秩不同时进行热插拔。 #### fuse_lora[[diffusers.loaders.lora_base.LoraBaseMixin.fuse_lora]] ```python fuse_lora(components: list[str] | None = None, lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L549) **参数:** components : (`list[str]`): 要融合 LoRA 的 LoRA 可注入组件列表。 lora_scale (`float`,默认值为 1.0):控制 LoRA 参数对输出的影响程度。 safe_fusing (`bool`,默认为 `False`):是否在融合前检查融合权重中的 NaN 值,如果值为 NaN 则不进行融合。 adapter_names (`list[str]`, *可选*) : 用于融合的适配器名称。如果未传入任何值,则所有处于活动状态的适配器都将被融合。 将 LoRA 参数融合到对应块的原始参数中。 示例: ```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) ``` #### get_active_adapters[[diffusers.loaders.lora_base.LoraBaseMixin.get_active_adapters]] ```python get_active_adapters() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L911) 获取当前激活的适配器列表。 示例: ```python from diffusers import DiffusionPipeline pipeline = DiffusionPipeline.from_pretrained( "stabilityai/stable-diffusion-xl-base-1.0", ).to("cuda") pipeline.load_lora_weights("CiroN2022/toy-face", weight_name="toy_face_sdxl.safetensors", adapter_name="toy") pipeline.get_active_adapters() ``` #### get_list_adapters[[diffusers.loaders.lora_base.LoraBaseMixin.get_list_adapters]] ```python get_list_adapters() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L944) 获取管道中当前所有可用适配器的列表。 #### set_adapters[[diffusers.loaders.lora_base.LoraBaseMixin.set_adapters]] ```python set_adapters(adapter_names: list[str] | str, adapter_weights: float | dict[str, float | dict | None] | list[float | dict[str, float | dict | None] | None] | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L698) **参数:** adapter_names (`list[str]` 或 `str`) : 要使用的适配器名称。 adapter_weights (`float`、`dict` 或 `list`,*可选*):适配器的权重。标量会缩放所有支持 LoRA 的组件。字典将组件名称(例如 `"unet"`、`"transformer"` 或 `"text_encoder"`)映射到对应的权重;UNet 组件的权重也可以是一个包含块缩放的嵌套字典。单个标量或字典会应用于每个适配器。列表必须按 `adapter_names` 的顺序,为每个适配器包含一个标量、字典或 `None`。`None` 默认为 `1.0`,包括列表项和组件权重。未指定的组件和 UNet 块也默认为 `1.0`。 设置当前用于流水线的活动适配器。 示例: ```py from diffusers import AutoPipelineForText2Image import torch pipeline = AutoPipelineForText2Image.from_pretrained( "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16 ).to("cuda") pipeline.load_lora_weights( "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic" ) pipeline.load_lora_weights( "troveski/sdxl-lora", weight_name="pytorch_lora_weights.safetensors", adapter_name="illustration" ) pipeline.set_adapters(["cinematic", "illustration"], adapter_weights=[0.5, 0.5]) pipeline.set_adapters( ["cinematic", "illustration"], adapter_weights=[{"unet": 0.5}, {"unet": 0.8, "text_encoder": 0.6}], ) ``` #### set_lora_device[[diffusers.loaders.lora_base.LoraBaseMixin.set_lora_device]] ```python set_lora_device(adapter_names: list[str], device: torch.device | str | int) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L966) **参数:** adapter_names (`list[str]`) : 要发送到该设备的适配器列表。 device (`torch.device | str | int`) : 用于将适配器发送到的设备。可以是 torch 设备、字符串或整数。 将 `adapter_names` 中列出的 LoRA 移动到目标设备。如果您希望加载多个适配器并释放部分 GPU 内存,可以将 LoRA 卸载到 CPU,此操作非常有用。 将 LoRA 适配器卸载到 CPU 后,只要模型的其余部分仍在 GPU 上,这些 LoRA 适配器就无法再用于推理,因为这会导致设备不匹配。在将这些 LoRA 适配器用于推理之前,请务必将设备重新设置回 GPU。 ```python >>> pipe.load_lora_weights(path_1, adapter_name="adapter-1") >>> pipe.load_lora_weights(path_2, adapter_name="adapter-2") >>> pipe.set_adapters("adapter-1") >>> image_1 = pipe(**kwargs) >>> # switch to adapter-2, offload adapter-1 >>> pipeline.set_lora_device(adapter_names=["adapter-1"], device="cpu") >>> pipeline.set_lora_device(adapter_names=["adapter-2"], device="cuda:0") >>> pipe.set_adapters("adapter-2") >>> image_2 = pipe(**kwargs) >>> # switch back to adapter-1, offload adapter-2 >>> pipeline.set_lora_device(adapter_names=["adapter-2"], device="cpu") >>> pipeline.set_lora_device(adapter_names=["adapter-1"], device="cuda:0") >>> pipe.set_adapters("adapter-1") >>> ... ``` #### unfuse_lora[[diffusers.loaders.lora_base.LoraBaseMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L636) **参数:** components (`list[str]`) : 要从中解除 LoRA 融合的、可注入 LoRA 的组件列表。 unfuse_unet (`bool`,默认为 `True`):是否取消融合 UNet LoRA 参数。 unfuse_text_encoder (`bool`,默认为 `True`):是否解绑文本编码器的 LoRA 参数。如果文本编码器未通过 monkey-patch 方式注入 LoRA 参数,则此设置不会生效。 撤销 [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora) 的效果。 #### unload_lora_weights[[diffusers.loaders.lora_base.LoraBaseMixin.unload_lora_weights]] ```python unload_lora_weights() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L526) 卸载 LoRA 参数。 示例: ```python >>> # Assuming `pipeline` is already loaded with the LoRA parameters. >>> pipeline.unload_lora_weights() >>> ... ``` #### write_lora_layers[[diffusers.loaders.lora_base.LoraBaseMixin.write_lora_layers]] ```python write_lora_layers(state_dict: dict[str, torch.Tensor], save_directory: str, is_main_process: bool, weight_name: str, save_function: Callable, safe_serialization: bool, lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L1043) 将 LoRA 层的 state dict(可选包含元数据)写入磁盘。 ## StableDiffusionLoraLoaderMixin[[diffusers.loaders.StableDiffusionLoraLoaderMixin]] #### diffusers.loaders.StableDiffusionLoraLoaderMixin[[diffusers.loaders.StableDiffusionLoraLoaderMixin]] ```python diffusers.loaders.StableDiffusionLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L138) 将 LoRA 层加载到 Stable Diffusion 的 [UNet2DConditionModel](/docs/diffusers/v0.43.0/zh/api/models/unet2d-cond#diffusers.UNet2DConditionModel) 和 [`CLIPTextModel`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel) 中。 #### load_lora_into_text_encoder[[diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_text_encoder]] ```python load_lora_into_text_encoder(state_dict, network_alphas, text_encoder, prefix = None, lora_scale = 1.0, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L408) **参数:** state_dict (`dict`) : 包含 lora 层参数的标准 state dict。键名应额外添加 `text_encoder` 前缀,以区分 unet 的 lora 层。 network_alphas (`dict[str, float]`) : 用于稳定学习和防止下溢的网络 alpha 值。该值与 kohya-ss 训练脚本中的 `--network_alpha` 选项含义相同。请参阅 [此链接](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning)。 text_encoder (`CLIPTextModel`):用于加载 LoRA 层的文本编码器模型。 prefix (`str`) : `state_dict` 中 `text_encoder` 的预期前缀。 lora_scale (`float`) : 在将 LoRA 线性层的输出与常规 LoRA 层的输出相加之前,对 LoRA 线性层输出进行缩放的系数。 adapter_name (`str`,*可选*):用于引用已加载的适配器模型的适配器名称。如果未指定,则使用 `default_{i}`,其中 i 为正在加载的适配器总数。 low_cpu_mem_usage (`bool`,*可选*):通过仅加载预训练的 LoRA 权重而不初始化随机权重,来加快模型加载速度。 hotswap (`bool`,*可选*):参见 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 metadata (`dict`) : 可选的 LoRA 适配器元数据。如果提供,`peft` 的 `LoraConfig` 参数将不会从状态字典中推导得出。 这将把`state_dict`中指定的 LoRA 层加载到`text_encoder`中 #### load_lora_into_unet[[diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet]] ```python load_lora_into_unet(state_dict, network_alphas, unet, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L354) **参数:** state_dict (`dict`) : 包含 lora 层参数的标准 state dict。键可以直接索引到 unet,也可以带有额外的 `unet` 前缀,用于区分文本编码器的 lora 层。 network_alphas (`dict[str, float]`) : 用于稳定学习和防止下溢的网络 alpha 值。该值与 kohya-ss 训练脚本中的 `--network_alpha` 选项含义相同。请参阅 [此链接](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning)。 unet (`UNet2DConditionModel`):要加载 LoRA 层的 UNet 模型。 adapter_name (`str`,*可选*):用于引用已加载的适配器模型的适配器名称。如果未指定,则使用 `default_{i}`,其中 i 为正在加载的适配器总数。 low_cpu_mem_usage (`bool`,*可选*):仅加载预训练的 LoRA 权重而不初始化随机权重,以加快模型加载速度。 hotswap (`bool`,*可选*):参见 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 metadata (`dict`) : 可选的 LoRA 适配器元数据。如果提供,`peft` 的 `LoraConfig` 参数将不会从状态字典中推导得出。 这将把 `state_dict` 中指定的 LoRA 层加载到 `unet` 中。 #### load_lora_weights[[diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L148) **参数:** pretrained_model_name_or_path_or_dict (`str` 或 `os.PathLike` 或 `dict`):参见 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 adapter_name (`str`,*可选*):用于引用已加载的适配器模型的适配器名称。如果未指定,则使用 `default_{i}`,其中 i 为正在加载的适配器总数。 low_cpu_mem_usage (`bool`,*可选*):通过仅加载预训练的 LoRA 权重而不初始化随机权重,来加快模型加载速度。 hotswap (`bool`, *可选*) : 默认为 `False`。是否就地用新加载的适配器替换现有的(LoRA)适配器。这意味着,不会加载额外的适配器,而是将现有适配器的权重替换为新适配器的权重。这种方式可能更快且更节省内存。然而,热替换的主要优势在于,当模型使用 torch.compile 编译时,加载新适配器无需重新编译模型。使用热替换时,传入的 `adapter_name` 应为已加载适配器的名称。 如果新适配器和旧适配器的秩(rank)和/或 LoRA alpha(即缩放)不同,在加载适配器之前需要调用一个额外方法: ```py pipeline = ... # load diffusers pipeline max_rank = ... # the highest rank among all LoRAs that you want to load # call *before* compiling and loading the LoRA adapter pipeline.enable_lora_hotswap(target_rank=max_rank) pipeline.load_lora_weights(file_name) # optionally compile the model now ``` 请注意,目前尚不支持对文本编码器适配器进行热替换。该技术还有一些其他限制,详见此处:https://huggingface.co/docs/peft/main/en/package_reference/hotswap kwargs (`dict`,*可选*):参见 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 将 `pretrained_model_name_or_path_or_dict` 中指定的 LoRA 权重加载到 `self.unet` 和 `self.text_encoder` 中。 所有 kwargs 都会转发到 `self.lora_state_dict`。 有关状态字典加载方式的更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 有关状态字典如何加载到 `self.unet` 的更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 有关状态字典如何加载到 `self.text_encoder` 的更多详细信息,请参阅 [load_lora_into_text_encoder()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_text_encoder)。 #### lora_state_dict[[diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L243) **参数:** pretrained_model_name_or_path_or_dict (`str` 或 `os.PathLike` 或 `dict`) : 可以是以下之一: - 一个字符串,即托管在 Hub 上的预训练模型的 *model id*(例如 `google/ddpm-celebahq-256`)。 - 一个指向 *目录* 的路径(例如 `./my_model_directory`),其中包含使用 [ModelMixin.save_pretrained()](/docs/diffusers/v0.43.0/zh/api/models/overview#diffusers.ModelMixin.save_pretrained) 保存的模型权重。 - 一个 [torch state dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict)。 cache_dir (`str | os.PathLike`,*可选*):如果未使用标准缓存,则已下载的预训练模型配置将缓存到的目录路径。 force_download (`bool`,*可选*,默认为 `False`):是否强制(重新)下载模型权重和配置文件,如果存在缓存版本则覆盖它们。 proxies (`dict[str, str]`,*可选*):一个字典,用于按协议或端点指定要使用的代理服务器,例如 `{'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}`。每次请求都会使用这些代理。 local_files_only (`bool`,*可选*,默认为 `False`):是否仅加载本地模型权重和配置文件。如果设置为 `True`,则不会从 Hub 下载模型。 token (`str` 或 *bool*,*可选*) : 用于远程文件的 HTTP 承载者授权令牌。如果为 `True`,则使用由 `diffusers-cli login` 生成的令牌(存储在 `~/.huggingface` 中)。 revision (`str`,*可选*,默认为 `"main"`):要使用的特定模型版本。它可以是分支名称、标签名称、提交 ID,或 Git 允许的任何标识符。 subfolder (`str`,*可选*,默认为 `""`):模型文件在 Hub 或本地的大型模型仓库中的子文件夹位置。 weight_name (`str`,*可选*,默认为 None):序列化状态字典文件的名称。 return_lora_metadata (`bool`,*可选*,默认为 False):启用后,将额外返回 LoRA 适配器元数据,通常位于 state dict 中。 返回 LoRA 权重和网络 alpha 的状态字典。 > [!WARNING] > 我们有限地支持加载 A1111 格式的 LoRA 检查点。 > > 此函数 实验性功能,未来可能会发生变化。 #### save_lora_weights[[diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, unet_lora_layers: dict[str, typing.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: typing.Callable = None, safe_serialization: bool = True, unet_lora_adapter_metadata = None, text_encoder_lora_adapter_metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L466) **参数:** save_directory (`str` 或 `os.PathLike`):用于保存 LoRA 参数的目录。如果该目录不存在,将自动创建。 unet_lora_layers (`dict[str, torch.nn.Module]` 或 `dict[str, torch.Tensor]`) : 与 `unet` 对应的 LoRA 层的 state dict。 text_encoder_lora_layers (`dict[str, torch.nn.Module]` 或 `dict[str, torch.Tensor]`) : 与 `text_encoder` 对应的 LoRA 层的 state dict。由于它来自 🤗 Transformers,因此必须显式传入文本编码器的 LoRA state dict。 is_main_process (`bool`,*可选*,默认为 `True`):调用此函数的进程是否为主进程。在分布式训练期间,如果需要在所有进程上调用此函数,此参数非常有用。在这种情况下,仅在将 `is_main_process=True` 设置为主进程时,以避免竞态条件。 save_function (`Callable`) : 用于保存状态字典的函数。在分布式训练期间,当您需要用其他方法替换 `torch.save` 时非常有用。可通过环境变量 `DIFFUSERS_SAVE_MODE` 进行配置。 safe_serialization (`bool`,*可选*,默认为 `True`):是否使用 `safetensors` 保存模型,还是使用传统的 PyTorch 方式(通过 `pickle`)。 unet_lora_adapter_metadata : 与 unet 关联的 LoRA 适配器元数据,将随 state dict 一起序列化。 text_encoder_lora_adapter_metadata : 与文本编码器关联的 LoRA 适配器元数据,将随状态字典一起序列化。 保存与 UNet 和文本编码器对应的 LoRA 参数。 ## StableDiffusionXLLoraLoaderMixin[[diffusers.loaders.StableDiffusionXLLoraLoaderMixin]] #### diffusers.loaders.StableDiffusionXLLoraLoaderMixin[[diffusers.loaders.StableDiffusionXLLoraLoaderMixin]] ```python diffusers.loaders.StableDiffusionXLLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L585) 将 LoRA 层加载到 Stable Diffusion XL 的 [UNet2DConditionModel](/docs/diffusers/v0.43.0/zh/api/models/unet2d-cond#diffusers.UNet2DConditionModel)、[`CLIPTextModel`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel) 和 [`CLIPTextModelWithProjection`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection) 中。 #### fuse_lora[[diffusers.loaders.StableDiffusionXLLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['unet', 'text_encoder', 'text_encoder_2'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L938) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_text_encoder[[diffusers.loaders.StableDiffusionXLLoraLoaderMixin.load_lora_into_text_encoder]] ```python load_lora_into_text_encoder(state_dict, network_alphas, text_encoder, prefix = None, lora_scale = 1.0, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L831) **参数:** state_dict (`dict`) : 包含 lora 层参数的标准 state dict。键名应额外添加 `text_encoder` 前缀,以区分 unet 的 lora 层。 network_alphas (`dict[str, float]`) : 用于稳定学习和防止下溢的网络 alpha 值。该值与 kohya-ss 训练脚本中的 `--network_alpha` 选项含义相同。请参阅 [此链接](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning)。 text_encoder (`CLIPTextModel`):用于加载 LoRA 层的文本编码器模型。 prefix (`str`) : `state_dict` 中 `text_encoder` 的预期前缀。 lora_scale (`float`) : 在将 LoRA 线性层的输出与常规 LoRA 层的输出相加之前,对 LoRA 线性层输出进行缩放的系数。 adapter_name (`str`,*可选*):用于引用已加载的适配器模型的适配器名称。如果未指定,则使用 `default_{i}`,其中 i 为正在加载的适配器总数。 low_cpu_mem_usage (`bool`,*可选*):通过仅加载预训练的 LoRA 权重而不初始化随机权重,来加快模型加载速度。 hotswap (`bool`,*可选*):参见 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 metadata (`dict`) : 可选的 LoRA 适配器元数据。如果提供,`peft` 的 `LoraConfig` 参数将不会从状态字典中推导得出。 这将把`state_dict`中指定的 LoRA 层加载到`text_encoder`中 #### load_lora_into_unet[[diffusers.loaders.StableDiffusionXLLoraLoaderMixin.load_lora_into_unet]] ```python load_lora_into_unet(state_dict, network_alphas, unet, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L776) **参数:** state_dict (`dict`) : 包含 lora 层参数的标准 state dict。键可以直接索引到 unet,也可以带有额外的 `unet` 前缀,用于区分文本编码器的 lora 层。 network_alphas (`dict[str, float]`) : 用于稳定学习和防止下溢的网络 alpha 值。该值与 kohya-ss 训练脚本中的 `--network_alpha` 选项含义相同。请参阅 [此链接](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning)。 unet (`UNet2DConditionModel`):要加载 LoRA 层的 UNet 模型。 adapter_name (`str`,*可选*):用于引用已加载的适配器模型的适配器名称。如果未指定,则使用 `default_{i}`,其中 i 为正在加载的适配器总数。 low_cpu_mem_usage (`bool`,*可选*):仅加载预训练的 LoRA 权重而不初始化随机权重,以加快模型加载速度。 hotswap (`bool`,*可选*):参见 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 metadata (`dict`) : 可选的 LoRA 适配器元数据。如果提供,`peft` 的 `LoraConfig` 参数将不会从状态字典中推导得出。 这将把 `state_dict` 中指定的 LoRA 层加载到 `unet` 中。 #### load_lora_weights[[diffusers.loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L596) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.StableDiffusionXLLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L664) **参数:** pretrained_model_name_or_path_or_dict (`str` 或 `os.PathLike` 或 `dict`) : 可以是以下之一: - 一个字符串,即托管在 Hub 上的预训练模型的 *model id*(例如 `google/ddpm-celebahq-256`)。 - 一个指向 *目录* 的路径(例如 `./my_model_directory`),其中包含使用 [ModelMixin.save_pretrained()](/docs/diffusers/v0.43.0/zh/api/models/overview#diffusers.ModelMixin.save_pretrained) 保存的模型权重。 - 一个 [torch state dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict)。 cache_dir (`str | os.PathLike`,*可选*):如果未使用标准缓存,则已下载的预训练模型配置将缓存到的目录路径。 force_download (`bool`,*可选*,默认为 `False`):是否强制(重新)下载模型权重和配置文件,如果存在缓存版本则覆盖它们。 proxies (`dict[str, str]`,*可选*):一个字典,用于按协议或端点指定要使用的代理服务器,例如 `{'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}`。每次请求都会使用这些代理。 local_files_only (`bool`,*可选*,默认为 `False`):是否仅加载本地模型权重和配置文件。如果设置为 `True`,则不会从 Hub 下载模型。 token (`str` 或 *bool*,*可选*) : 用于远程文件的 HTTP 承载者授权令牌。如果为 `True`,则使用由 `diffusers-cli login` 生成的令牌(存储在 `~/.huggingface` 中)。 revision (`str`,*可选*,默认为 `"main"`):要使用的特定模型版本。它可以是分支名称、标签名称、提交 ID,或 Git 允许的任何标识符。 subfolder (`str`,*可选*,默认为 `""`):模型文件在 Hub 或本地的大型模型仓库中的子文件夹位置。 weight_name (`str`,*可选*,默认为 None):序列化状态字典文件的名称。 return_lora_metadata (`bool`,*可选*,默认为 False):启用后,将额外返回 LoRA 适配器元数据,通常位于 state dict 中。 返回 LoRA 权重和网络 alpha 的状态字典。 > [!WARNING] > 我们有限地支持加载 A1111 格式的 LoRA 检查点。 > > 此函数 实验性功能,未来可能会发生变化。 #### save_lora_weights[[diffusers.loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, unet_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, text_encoder_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, text_encoder_2_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, unet_lora_adapter_metadata = None, text_encoder_lora_adapter_metadata = None, text_encoder_2_lora_adapter_metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L890) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.StableDiffusionXLLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['unet', 'text_encoder', 'text_encoder_2'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L957) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## SD3LoraLoaderMixin[[diffusers.loaders.SD3LoraLoaderMixin]] #### diffusers.loaders.SD3LoraLoaderMixin[[diffusers.loaders.SD3LoraLoaderMixin]] ```python diffusers.loaders.SD3LoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L964) 将 LoRA 层加载到 [SD3Transformer2DModel](/docs/diffusers/v0.43.0/zh/api/models/sd3_transformer2d#diffusers.SD3Transformer2DModel)、[`CLIPTextModel`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel) 和 [`CLIPTextModelWithProjection`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection) 中。 特定于 [StableDiffusion3Pipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/stable_diffusion/stable_diffusion_3#diffusers.StableDiffusion3Pipeline)。 #### fuse_lora[[diffusers.loaders.SD3LoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer', 'text_encoder', 'text_encoder_2'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1225) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_text_encoder[[diffusers.loaders.SD3LoraLoaderMixin.load_lora_into_text_encoder]] ```python load_lora_into_text_encoder(state_dict, network_alphas, text_encoder, prefix = None, lora_scale = 1.0, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1116) **参数:** state_dict (`dict`) : 包含 lora 层参数的标准 state dict。键名应额外添加 `text_encoder` 前缀,以区分 unet 的 lora 层。 network_alphas (`dict[str, float]`) : 用于稳定学习和防止下溢的网络 alpha 值。该值与 kohya-ss 训练脚本中的 `--network_alpha` 选项含义相同。请参阅 [此链接](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning)。 text_encoder (`CLIPTextModel`):用于加载 LoRA 层的文本编码器模型。 prefix (`str`) : `state_dict` 中 `text_encoder` 的预期前缀。 lora_scale (`float`) : 在将 LoRA 线性层的输出与常规 LoRA 层的输出相加之前,对 LoRA 线性层输出进行缩放的系数。 adapter_name (`str`,*可选*):用于引用已加载的适配器模型的适配器名称。如果未指定,则使用 `default_{i}`,其中 i 为正在加载的适配器总数。 low_cpu_mem_usage (`bool`,*可选*):通过仅加载预训练的 LoRA 权重而不初始化随机权重,来加快模型加载速度。 hotswap (`bool`,*可选*):参见 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 metadata (`dict`) : 可选的 LoRA 适配器元数据。如果提供,`peft` 的 `LoraConfig` 参数将不会从状态字典中推导得出。 这将把`state_dict`中指定的 LoRA 层加载到`text_encoder`中 #### load_lora_into_transformer[[diffusers.loaders.SD3LoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1090) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.SD3LoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1031) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.SD3LoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L977) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.SD3LoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, text_encoder_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, text_encoder_2_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata = None, text_encoder_lora_adapter_metadata = None, text_encoder_2_lora_adapter_metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1175) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.SD3LoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer', 'text_encoder', 'text_encoder_2'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1245) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## FluxLoraLoaderMixin[[diffusers.loaders.FluxLoraLoaderMixin]] #### diffusers.loaders.FluxLoraLoaderMixin[[diffusers.loaders.FluxLoraLoaderMixin]] ```python diffusers.loaders.FluxLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1441) 将 LoRA 层加载到 [FluxTransformer2DModel](/docs/diffusers/v0.43.0/zh/api/models/flux_transformer#diffusers.FluxTransformer2DModel)、[`CLIPTextModel`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel) 中。 特定于 [FluxPipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/flux#diffusers.FluxPipeline)。 #### fuse_lora[[diffusers.loaders.FluxLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1879) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### load_lora_into_text_encoder[[diffusers.loaders.FluxLoraLoaderMixin.load_lora_into_text_encoder]] ```python load_lora_into_text_encoder(state_dict, network_alphas, text_encoder, prefix = None, lora_scale = 1.0, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1756) **参数:** state_dict (`dict`) : 包含 lora 层参数的标准 state dict。键名应额外添加 `text_encoder` 前缀,以区分 unet 的 lora 层。 network_alphas (`dict[str, float]`) : 用于稳定学习和防止下溢的网络 alpha 值。该值与 kohya-ss 训练脚本中的 `--network_alpha` 选项含义相同。请参阅 [此链接](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning)。 text_encoder (`CLIPTextModel`):用于加载 LoRA 层的文本编码器模型。 prefix (`str`) : `state_dict` 中 `text_encoder` 的预期前缀。 lora_scale (`float`) : 在将 LoRA 线性层的输出与常规 LoRA 层的输出相加之前,对 LoRA 线性层输出进行缩放的系数。 adapter_name (`str`,*可选*):用于引用已加载的适配器模型的适配器名称。如果未指定,则使用 `default_{i}`,其中 i 为正在加载的适配器总数。 low_cpu_mem_usage (`bool`,*可选*):通过仅加载预训练的 LoRA 权重而不初始化随机权重,来加快模型加载速度。 hotswap (`bool`,*可选*):参见 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 metadata (`dict`) : 可选的 LoRA 适配器元数据。如果提供,`peft` 的 `LoraConfig` 参数将不会从状态字典中推导得出。 这将把`state_dict`中指定的 LoRA 层加载到`text_encoder`中 #### load_lora_into_transformer[[diffusers.loaders.FluxLoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, network_alphas, transformer, adapter_name = None, metadata = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1675) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.FluxLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1579) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.FluxLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], return_alphas: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1454) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.FluxLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.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: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata = None, text_encoder_lora_adapter_metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1815) **参数:** save_directory (`str` 或 `os.PathLike`):用于保存 LoRA 参数的目录。如果该目录不存在,将自动创建。 transformer_lora_layers (`dict[str, torch.nn.Module]` 或 `dict[str, torch.Tensor]`) : 与 `transformer` 对应的 LoRA 层的 state dict。 text_encoder_lora_layers (`dict[str, torch.nn.Module]` 或 `dict[str, torch.Tensor]`) : 与 `text_encoder` 对应的 LoRA 层的 state dict。由于它来自 🤗 Transformers,因此必须显式传入文本编码器的 LoRA state dict。 is_main_process (`bool`,*可选*,默认为 `True`):调用此函数的进程是否为主进程。在分布式训练期间,如果需要在所有进程上调用此函数,此参数非常有用。在这种情况下,仅在将 `is_main_process=True` 设置为主进程时,以避免竞态条件。 save_function (`Callable`) : 用于保存状态字典的函数。在分布式训练期间,当您需要用其他方法替换 `torch.save` 时非常有用。可通过环境变量 `DIFFUSERS_SAVE_MODE` 进行配置。 safe_serialization (`bool`,*可选*,默认为 `True`):是否使用 `safetensors` 保存模型,还是使用传统的 PyTorch 方式(通过 `pickle`)。 transformer_lora_adapter_metadata : 与 transformer 关联的 LoRA 适配器元数据,将随 state dict 一起序列化。 text_encoder_lora_adapter_metadata : 与文本编码器关联的 LoRA 适配器元数据,将随状态字典一起序列化。 保存与 UNet 和文本编码器对应的 LoRA 参数。 #### unfuse_lora[[diffusers.loaders.FluxLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer', 'text_encoder'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1911) **参数:** components (`list[str]`) : 要从中解除 LoRA 融合的、可注入 LoRA 的组件列表。 撤销 [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora) 的效果。 #### unload_lora_weights[[diffusers.loaders.FluxLoraLoaderMixin.unload_lora_weights]] ```python unload_lora_weights(reset_to_overwritten_params = False) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1926) **参数:** reset_to_overwritten_params (`bool`,默认为 `False`):是否将加载了 LoRA 的模块重置为其原始参数。请参阅 [Flux 文档](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux) 以了解更多信息。 卸载 LoRA 参数。 示例: ```python >>> # Assuming `pipeline` is already loaded with the LoRA parameters. >>> pipeline.unload_lora_weights() >>> ... ``` ## Flux2LoraLoaderMixin[[diffusers.loaders.Flux2LoraLoaderMixin]] #### diffusers.loaders.Flux2LoraLoaderMixin[[diffusers.loaders.Flux2LoraLoaderMixin]] ```python diffusers.loaders.Flux2LoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5773) 将 LoRA 层加载到 [Flux2Transformer2DModel](/docs/diffusers/v0.43.0/zh/api/models/flux2_transformer#diffusers.Flux2Transformer2DModel) 中。专用于 [Flux2Pipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/flux2#diffusers.Flux2Pipeline)。 #### fuse_lora[[diffusers.loaders.Flux2LoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5957) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.Flux2LoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5894) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.Flux2LoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5859) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.Flux2LoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5781) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.Flux2LoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5921) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.Flux2LoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5977) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## Flux3LoraLoaderMixin[[diffusers.loaders.Flux3LoraLoaderMixin]] #### diffusers.loaders.Flux3LoraLoaderMixin[[diffusers.loaders.Flux3LoraLoaderMixin]] ```python diffusers.loaders.Flux3LoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5984) #### fuse_lora[[diffusers.loaders.Flux3LoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6139) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.Flux3LoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6076) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.Flux3LoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6040) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### save_lora_weights[[diffusers.loaders.Flux3LoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6103) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.Flux3LoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6159) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## BriaFibo2LoraLoaderMixin[[diffusers.loaders.BriaFibo2LoraLoaderMixin]] #### diffusers.loaders.BriaFibo2LoraLoaderMixin[[diffusers.loaders.BriaFibo2LoraLoaderMixin]] ```python diffusers.loaders.BriaFibo2LoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6166) 将 LoRA 层加载到 [BriaFibo2Transformer2DModel](/docs/diffusers/v0.43.0/zh/api/models/transformer_bria_fibo2#diffusers.BriaFibo2Transformer2DModel) 中。专用于 [BriaFibo2Pipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/bria_fibo2#diffusers.BriaFibo2Pipeline)。 #### fuse_lora[[diffusers.loaders.BriaFibo2LoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6328) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.BriaFibo2LoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6265) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.BriaFibo2LoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6229) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.BriaFibo2LoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6174) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.BriaFibo2LoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6292) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.BriaFibo2LoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6348) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## ErnieImageLoraLoaderMixin[[diffusers.loaders.ErnieImageLoraLoaderMixin]] #### diffusers.loaders.ErnieImageLoraLoaderMixin[[diffusers.loaders.ErnieImageLoraLoaderMixin]] ```python diffusers.loaders.ErnieImageLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6551) 将 LoRA 层加载到 [ErnieImageTransformer2DModel](/docs/diffusers/v0.43.0/zh/api/models/ernie_image_transformer2d#diffusers.ErnieImageTransformer2DModel) 中。专用于 `ErnieImagePipeline`。 #### fuse_lora[[diffusers.loaders.ErnieImageLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6724) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.ErnieImageLoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6661) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.ErnieImageLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6625) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.ErnieImageLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6559) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.ErnieImageLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6688) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.ErnieImageLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6744) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## LTX2LoraLoaderMixin[[diffusers.loaders.LTX2LoraLoaderMixin]] #### diffusers.loaders.LTX2LoraLoaderMixin[[diffusers.loaders.LTX2LoraLoaderMixin]] ```python diffusers.loaders.LTX2LoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2919) 将 LoRA 层加载到 [LTX2VideoTransformer3DModel](/docs/diffusers/v0.43.0/zh/api/models/ltx2_video_transformer3d#diffusers.LTX2VideoTransformer3DModel) 中。专用于 [LTX2Pipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/ltx2#diffusers.LTX2Pipeline)。 #### fuse_lora[[diffusers.loaders.LTX2LoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3108) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.LTX2LoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None, prefix: str = 'transformer') ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3044) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.LTX2LoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2992) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.LTX2LoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2928) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.LTX2LoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3072) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.LTX2LoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3128) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## CogVideoXLoraLoaderMixin[[diffusers.loaders.CogVideoXLoraLoaderMixin]] #### diffusers.loaders.CogVideoXLoraLoaderMixin[[diffusers.loaders.CogVideoXLoraLoaderMixin]] ```python diffusers.loaders.CogVideoXLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2353) 将 LoRA 层加载到 [CogVideoXTransformer3DModel](/docs/diffusers/v0.43.0/zh/api/models/cogvideox_transformer3d#diffusers.CogVideoXTransformer3DModel) 中。专用于 [CogVideoXPipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/cogvideox#diffusers.CogVideoXPipeline)。 #### fuse_lora[[diffusers.loaders.CogVideoXLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2512) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.CogVideoXLoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2451) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.CogVideoXLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2416) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.CogVideoXLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2361) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.CogVideoXLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2478) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.CogVideoXLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2531) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## Mochi1LoraLoaderMixin[[diffusers.loaders.Mochi1LoraLoaderMixin]] #### diffusers.loaders.Mochi1LoraLoaderMixin[[diffusers.loaders.Mochi1LoraLoaderMixin]] ```python diffusers.loaders.Mochi1LoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2538) 将 LoRA 层加载到 [MochiTransformer3DModel](/docs/diffusers/v0.43.0/zh/api/models/mochi_transformer3d#diffusers.MochiTransformer3DModel) 中。专用于 [MochiPipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/mochi#diffusers.MochiPipeline)。 #### fuse_lora[[diffusers.loaders.Mochi1LoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2700) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.Mochi1LoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2637) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.Mochi1LoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2601) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.Mochi1LoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2546) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.Mochi1LoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2664) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.Mochi1LoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2720) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## AuraFlowLoraLoaderMixin[[diffusers.loaders.AuraFlowLoraLoaderMixin]] #### diffusers.loaders.AuraFlowLoraLoaderMixin[[diffusers.loaders.AuraFlowLoraLoaderMixin]] ```python diffusers.loaders.AuraFlowLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1252) 将 LoRA 层加载到 [AuraFlowTransformer2DModel](/docs/diffusers/v0.43.0/zh/api/models/aura_flow_transformer2d#diffusers.AuraFlowTransformer2DModel) 中。专用于 [AuraFlowPipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/aura_flow#diffusers.AuraFlowPipeline)。 #### fuse_lora[[diffusers.loaders.AuraFlowLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1414) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.AuraFlowLoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1351) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.AuraFlowLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1315) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.AuraFlowLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1260) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.AuraFlowLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1378) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.AuraFlowLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer', 'text_encoder'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L1434) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## LTXVideoLoraLoaderMixin[[diffusers.loaders.LTXVideoLoraLoaderMixin]] #### diffusers.loaders.LTXVideoLoraLoaderMixin[[diffusers.loaders.LTXVideoLoraLoaderMixin]] ```python diffusers.loaders.LTXVideoLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2727) 将 LoRA 层加载到 [LTXVideoTransformer3DModel](/docs/diffusers/v0.43.0/zh/api/models/ltx_video_transformer3d#diffusers.LTXVideoTransformer3DModel) 中。专用于 [LTXPipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/ltx_video#diffusers.LTXPipeline)。 #### fuse_lora[[diffusers.loaders.LTXVideoLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2892) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.LTXVideoLoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2829) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.LTXVideoLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2793) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.LTXVideoLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2735) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.LTXVideoLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2856) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.LTXVideoLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2912) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## SanaLoraLoaderMixin[[diffusers.loaders.SanaLoraLoaderMixin]] #### diffusers.loaders.SanaLoraLoaderMixin[[diffusers.loaders.SanaLoraLoaderMixin]] ```python diffusers.loaders.SanaLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3135) 将 LoRA 层加载到 [SanaTransformer2DModel](/docs/diffusers/v0.43.0/zh/api/models/sana_transformer2d#diffusers.SanaTransformer2DModel) 中。专用于 [SanaPipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/sana#diffusers.SanaPipeline)。 #### fuse_lora[[diffusers.loaders.SanaLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3297) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.SanaLoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3234) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.SanaLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3198) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.SanaLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3143) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.SanaLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3261) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.SanaLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3317) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## HeliosLoraLoaderMixin[[diffusers.loaders.HeliosLoraLoaderMixin]] #### diffusers.loaders.HeliosLoraLoaderMixin[[diffusers.loaders.HeliosLoraLoaderMixin]] ```python diffusers.loaders.HeliosLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3324) 将 LoRA 层加载到 [HeliosTransformer3DModel](/docs/diffusers/v0.43.0/zh/api/models/helios_transformer3d#diffusers.HeliosTransformer3DModel) 中。专用于 [HeliosPipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/helios#diffusers.HeliosPipeline) 和 [HeliosPyramidPipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/helios#diffusers.HeliosPyramidPipeline)。 #### fuse_lora[[diffusers.loaders.HeliosLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3487) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.HeliosLoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3424) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.HeliosLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3390) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.HeliosLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3332) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.HeliosLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3451) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.HeliosLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3507) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## HunyuanVideoLoraLoaderMixin[[diffusers.loaders.HunyuanVideoLoraLoaderMixin]] #### diffusers.loaders.HunyuanVideoLoraLoaderMixin[[diffusers.loaders.HunyuanVideoLoraLoaderMixin]] ```python diffusers.loaders.HunyuanVideoLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3514) 将 LoRA 层加载到 [HunyuanVideoTransformer3DModel](/docs/diffusers/v0.43.0/zh/api/models/hunyuan_video_transformer_3d#diffusers.HunyuanVideoTransformer3DModel) 中。专用于 [HunyuanVideoPipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/hunyuan_video#diffusers.HunyuanVideoPipeline)。 #### fuse_lora[[diffusers.loaders.HunyuanVideoLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3679) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.HunyuanVideoLoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3616) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.HunyuanVideoLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3580) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.HunyuanVideoLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3522) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.HunyuanVideoLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3643) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.HunyuanVideoLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3699) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## Lumina2LoraLoaderMixin[[diffusers.loaders.Lumina2LoraLoaderMixin]] #### diffusers.loaders.Lumina2LoraLoaderMixin[[diffusers.loaders.Lumina2LoraLoaderMixin]] ```python diffusers.loaders.Lumina2LoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3706) 将 LoRA 层加载到 [Lumina2Transformer2DModel](/docs/diffusers/v0.43.0/zh/api/models/lumina2_transformer2d#diffusers.Lumina2Transformer2DModel) 中。专用于 [Lumina2Pipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/lumina2#diffusers.Lumina2Pipeline)。 #### fuse_lora[[diffusers.loaders.Lumina2LoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3872) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.Lumina2LoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3809) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.Lumina2LoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3773) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.Lumina2LoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3714) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.Lumina2LoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3836) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.Lumina2LoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3892) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## CogView4LoraLoaderMixin[[diffusers.loaders.CogView4LoraLoaderMixin]] #### diffusers.loaders.CogView4LoraLoaderMixin[[diffusers.loaders.CogView4LoraLoaderMixin]] ```python diffusers.loaders.CogView4LoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4617) 将 LoRA 层加载到 [WanTransformer3DModel](/docs/diffusers/v0.43.0/zh/api/models/wan_transformer_3d#diffusers.WanTransformer3DModel) 中。专用于 [CogView4Pipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/cogview4#diffusers.CogView4Pipeline)。 #### fuse_lora[[diffusers.loaders.CogView4LoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4779) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.CogView4LoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4716) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.CogView4LoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4680) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.CogView4LoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4625) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.CogView4LoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4743) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.CogView4LoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4799) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## WanLoraLoaderMixin[[diffusers.loaders.WanLoraLoaderMixin]] #### diffusers.loaders.WanLoraLoaderMixin[[diffusers.loaders.WanLoraLoaderMixin]] ```python diffusers.loaders.WanLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4088) 将 LoRA 层加载到 [WanTransformer3DModel](/docs/diffusers/v0.43.0/zh/api/models/wan_transformer_3d#diffusers.WanTransformer3DModel) 中。专用于 [WanPipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/wan#diffusers.WanPipeline) 和 `[WanImageToVideoPipeline`。 #### fuse_lora[[diffusers.loaders.WanLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4324) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.WanLoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4261) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.WanLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4202) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.WanLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4096) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.WanLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4288) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.WanLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4344) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## SkyReelsV2LoraLoaderMixin[[diffusers.loaders.SkyReelsV2LoraLoaderMixin]] #### diffusers.loaders.SkyReelsV2LoraLoaderMixin[[diffusers.loaders.SkyReelsV2LoraLoaderMixin]] ```python diffusers.loaders.SkyReelsV2LoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4351) 将 LoRA 层加载到 [SkyReelsV2Transformer3DModel](/docs/diffusers/v0.43.0/zh/api/models/skyreels_v2_transformer_3d#diffusers.SkyReelsV2Transformer3DModel) 中。 #### fuse_lora[[diffusers.loaders.SkyReelsV2LoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4590) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.SkyReelsV2LoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4527) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.SkyReelsV2LoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4467) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.SkyReelsV2LoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4359) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.SkyReelsV2LoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4554) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.SkyReelsV2LoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4610) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## AmusedLoraLoaderMixin[[diffusers.loaders.AmusedLoraLoaderMixin]] #### diffusers.loaders.AmusedLoraLoaderMixin[[diffusers.loaders.AmusedLoraLoaderMixin]] ```python diffusers.loaders.AmusedLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2206) #### load_lora_into_transformer[[diffusers.loaders.AmusedLoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, network_alphas, transformer, adapter_name = None, metadata = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2211) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### save_lora_weights[[diffusers.loaders.AmusedLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, text_encoder_lora_layers: dict[str, torch.nn.Module] = None, transformer_lora_layers: dict[str, torch.nn.Module] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L2298) **参数:** save_directory (`str` 或 `os.PathLike`):用于保存 LoRA 参数的目录。如果该目录不存在,将自动创建。 unet_lora_layers (`dict[str, torch.nn.Module]` 或 `dict[str, torch.Tensor]`) : 与 `unet` 对应的 LoRA 层的 state dict。 text_encoder_lora_layers (`dict[str, torch.nn.Module]` 或 `dict[str, torch.Tensor]`) : 与 `text_encoder` 对应的 LoRA 层的 state dict。由于它来自 🤗 Transformers,因此必须显式传入文本编码器的 LoRA state dict。 is_main_process (`bool`,*可选*,默认为 `True`):调用此函数的进程是否为主进程。在分布式训练期间,如果需要在所有进程上调用此函数,此参数非常有用。在这种情况下,仅在将 `is_main_process=True` 设置为主进程时,以避免竞态条件。 save_function (`Callable`) : 用于保存状态字典的函数。在分布式训练期间,当您需要用其他方法替换 `torch.save` 时非常有用。可通过环境变量 `DIFFUSERS_SAVE_MODE` 进行配置。 safe_serialization (`bool`,*可选*,默认为 `True`):是否使用 `safetensors` 保存模型,还是使用传统的 PyTorch 方式(通过 `pickle`)。 保存与 UNet 和文本编码器对应的 LoRA 参数。 ## AnimaLoraLoaderMixin[[diffusers.loaders.AnimaLoraLoaderMixin]] #### diffusers.loaders.AnimaLoraLoaderMixin[[diffusers.loaders.AnimaLoraLoaderMixin]] ```python diffusers.loaders.AnimaLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5587) 将 LoRA 层加载到 [CosmosTransformer3DModel](/docs/diffusers/v0.43.0/zh/api/models/cosmos_transformer3d#diffusers.CosmosTransformer3DModel) 和 [AnimaTextConditioner](/docs/diffusers/v0.43.0/zh/api/pipelines/anima#diffusers.AnimaTextConditioner) 中。 #### fuse_lora[[diffusers.loaders.AnimaLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer', 'text_conditioner'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5747) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_weights[[diffusers.loaders.AnimaLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5652) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.AnimaLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5596) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### unfuse_lora[[diffusers.loaders.AnimaLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer', 'text_conditioner'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5766) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## AceStepLoraLoaderMixin[[diffusers.loaders.AceStepLoraLoaderMixin]] #### diffusers.loaders.AceStepLoraLoaderMixin[[diffusers.loaders.AceStepLoraLoaderMixin]] ```python diffusers.loaders.AceStepLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6941) 将 LoRA 层加载到 [AceStepTransformer1DModel](/docs/diffusers/v0.43.0/zh/api/models/ace_step_transformer#diffusers.AceStepTransformer1DModel) 中。专用于 [AceStepPipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/ace_step#diffusers.AceStepPipeline)。 #### fuse_lora[[diffusers.loaders.AceStepLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7105) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.AceStepLoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7042) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.AceStepLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7006) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.AceStepLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6949) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.AceStepLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7069) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.AceStepLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7125) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## HiDreamImageLoraLoaderMixin[[diffusers.loaders.HiDreamImageLoraLoaderMixin]] #### diffusers.loaders.HiDreamImageLoraLoaderMixin[[diffusers.loaders.HiDreamImageLoraLoaderMixin]] ```python diffusers.loaders.HiDreamImageLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4806) 将 LoRA 层加载到 [HiDreamImageTransformer2DModel](/docs/diffusers/v0.43.0/zh/api/models/hidream_image_transformer#diffusers.HiDreamImageTransformer2DModel) 中。专用于 [HiDreamImagePipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/hidream#diffusers.HiDreamImagePipeline)。 #### fuse_lora[[diffusers.loaders.HiDreamImageLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4971) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.HiDreamImageLoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4908) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.HiDreamImageLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4872) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.HiDreamImageLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4814) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.HiDreamImageLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4935) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.HiDreamImageLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4991) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## QwenImageLoraLoaderMixin[[diffusers.loaders.QwenImageLoraLoaderMixin]] #### diffusers.loaders.QwenImageLoraLoaderMixin[[diffusers.loaders.QwenImageLoraLoaderMixin]] ```python diffusers.loaders.QwenImageLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4998) 将 LoRA 层加载到 [QwenImageTransformer2DModel](/docs/diffusers/v0.43.0/zh/api/models/qwenimage_transformer2d#diffusers.QwenImageTransformer2DModel) 中。专用于 [QwenImagePipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/qwenimage#diffusers.QwenImagePipeline)。 #### fuse_lora[[diffusers.loaders.QwenImageLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5166) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.QwenImageLoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5103) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.QwenImageLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5067) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.QwenImageLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5006) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.QwenImageLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5130) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.QwenImageLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5186) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## ZImageLoraLoaderMixin[[diffusers.loaders.ZImageLoraLoaderMixin]] #### diffusers.loaders.ZImageLoraLoaderMixin[[diffusers.loaders.ZImageLoraLoaderMixin]] ```python diffusers.loaders.ZImageLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5387) 将 LoRA 层加载到 [ZImageTransformer2DModel](/docs/diffusers/v0.43.0/zh/api/models/z_image_transformer2d#diffusers.ZImageTransformer2DModel) 中。专用于 [ZImagePipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/z_image#diffusers.ZImagePipeline)。 #### fuse_lora[[diffusers.loaders.ZImageLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5560) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.ZImageLoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5497) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.ZImageLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5461) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.ZImageLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5395) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.ZImageLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5524) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.ZImageLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5580) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## CosmosLoraLoaderMixin[[diffusers.loaders.CosmosLoraLoaderMixin]] #### diffusers.loaders.CosmosLoraLoaderMixin[[diffusers.loaders.CosmosLoraLoaderMixin]] ```python diffusers.loaders.CosmosLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6751) 将 LoRA 层加载到 [CosmosTransformer3DModel](/docs/diffusers/v0.43.0/zh/api/models/cosmos_transformer3d#diffusers.CosmosTransformer3DModel) 中,专用于 [Cosmos2_5_PredictBasePipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/cosmos#diffusers.Cosmos2_5_PredictBasePipeline)。 #### fuse_lora[[diffusers.loaders.CosmosLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6914) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.CosmosLoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6851) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.CosmosLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6815) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.CosmosLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6760) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.CosmosLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6878) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.CosmosLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6934) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## KandinskyLoraLoaderMixin[[diffusers.loaders.KandinskyLoraLoaderMixin]] #### diffusers.loaders.KandinskyLoraLoaderMixin[[diffusers.loaders.KandinskyLoraLoaderMixin]] ```python diffusers.loaders.KandinskyLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3899) 将 LoRA 层加载到 `Kandinsky5Transformer3DModel` 中, #### fuse_lora[[diffusers.loaders.KandinskyLoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4061) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.KandinskyLoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3998) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.KandinskyLoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3962) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.KandinskyLoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L3907) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.KandinskyLoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4025) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.KandinskyLoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L4081) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## Ideogram4LoraLoaderMixin[[diffusers.loaders.Ideogram4LoraLoaderMixin]] #### diffusers.loaders.Ideogram4LoraLoaderMixin[[diffusers.loaders.Ideogram4LoraLoaderMixin]] ```python diffusers.loaders.Ideogram4LoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6355) 将 LoRA 层加载到 [Ideogram4Transformer2DModel](/docs/diffusers/v0.43.0/zh/api/models/ideogram4_transformer2d#diffusers.Ideogram4Transformer2DModel) 中。专用于 [Ideogram4Pipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/ideogram4#diffusers.Ideogram4Pipeline)。 #### fuse_lora[[diffusers.loaders.Ideogram4LoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6524) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.Ideogram4LoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6461) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.Ideogram4LoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6425) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.Ideogram4LoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6363) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.Ideogram4LoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6488) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.Ideogram4LoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L6544) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## Krea2LoraLoaderMixin[[diffusers.loaders.Krea2LoraLoaderMixin]] #### diffusers.loaders.Krea2LoraLoaderMixin[[diffusers.loaders.Krea2LoraLoaderMixin]] ```python diffusers.loaders.Krea2LoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5193) 将 LoRA 层加载到 [Krea2Transformer2DModel](/docs/diffusers/v0.43.0/zh/api/models/krea2_transformer2d#diffusers.Krea2Transformer2DModel) 中。专用于 [Krea2Pipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/krea2#diffusers.Krea2Pipeline)。 #### fuse_lora[[diffusers.loaders.Krea2LoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5360) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.Krea2LoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5297) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.Krea2LoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5261) 有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.Krea2LoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5201) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.Krea2LoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5324) 有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.Krea2LoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L5380) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## MiniMaxH3LoraLoaderMixin[[diffusers.loaders.MiniMaxH3LoraLoaderMixin]] #### diffusers.loaders.MiniMaxH3LoraLoaderMixin[[diffusers.loaders.MiniMaxH3LoraLoaderMixin]] ```python diffusers.loaders.MiniMaxH3LoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7132) 将 LoRA 层加载到 [MiniMaxH3Transformer3DModel](/docs/diffusers/v0.43.0/zh/api/models/minimax_h3_transformer3d#diffusers.MiniMaxH3Transformer3DModel) 中。专用于 [MiniMaxH3ModularPipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/minimax_h3#diffusers.MiniMaxH3ModularPipeline)。 MiniMax-H3 包含两个具有相同模块名称的独立 DiT 分区(针对其中一个加载的 LoRA 会被加载到另一个中,从而导致输出质量下降),因此路由是显式的:转换后的 state dict 目标为 `transformer.`;要访问 `transformer_ref`,请使用带有 `transformer_ref.` 前缀的文件,或设置 `load_into_transformer_ref=True`。 DiffSynth-Studio LoRA(例如 [DiffSynth-Studio/MiniMax-H3-LoRA-LineartAnime](https://huggingface.co/DiffSynth-Studio/MiniMax-H3-LoRA-LineartAnime))是针对原始检查点中按头交错的融合 QKV 进行训练的,并在转换时进行解交错。其 fp32 因子会使未融合的 LoRA 路径以 fp32 进行计算;在加载模型后对模型执行 `.to(torch.bfloat16)` 可恢复 bf16 内存预算。 针对剪枝检查点([Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3) 中的 `*_pruned_*` 文件;[joyfox/MiniMax-H3-Turbo](https://huggingface.co/joyfox/MiniMax-H3-Turbo) 即为其中之一)训练的 LoRA 会因尺寸不匹配而加载失败。不含 `alpha` 的文件会以 `alpha == rank` 加载(缩放系数为 1.0,这是例如 [larryvrh/MiniMax-H3-Turbo-Lora](https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora) 所采用的约定);如果 `__metadata__` 中存在 `alpha` 条目(例如 [lightx2v/Minimax-h3-Turbo](https://huggingface.co/lightx2v/Minimax-h3-Turbo) 的 8 步文件),则会优先采用该值。 #### fuse_lora[[diffusers.loaders.MiniMaxH3LoraLoaderMixin.fuse_lora]] ```python fuse_lora(components: list[str] = ['transformer', 'transformer_ref'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7489) 有关更多详细信息,请参阅 `fuse_lora()`。 #### load_lora_into_transformer[[diffusers.loaders.MiniMaxH3LoraLoaderMixin.load_lora_into_transformer]] ```python load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7389) 有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_into_transformer_ref[[diffusers.loaders.MiniMaxH3LoraLoaderMixin.load_lora_into_transformer_ref]] ```python load_lora_into_transformer_ref(state_dict, transformer_ref, prefix, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7416) 将 LoRA 层加载到 `transformer_ref` 分区中。`prefix` 是键所携带的组件名称,对于指定了分区名称的文件,该值为 `transformer_ref`;对于通过 `load_into_transformer_ref=True` 路由到此处的文件,该值为 `transformer`。更多详情请参见 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。 #### load_lora_weights[[diffusers.loaders.MiniMaxH3LoraLoaderMixin.load_lora_weights]] ```python load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, load_into_transformer_ref: bool = False, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7274) **参数:** load_into_transformer_ref (`bool`,默认为 `False`):将带有 `transformer.` 前缀的层加载到 `transformer_ref` 分区中——即 `ref2va` 工作流执行去噪所使用的分区——而不是 `transformer`。仅当两个分区都已加载时才需要此设置:如果流水线仅持有 `transformer_ref`,它会自动路由到该分区。 将 LoRA 层加载到 `transformer` 中,或者在设置 `load_into_transformer_ref=True` 时加载到 `transformer_ref` 中。有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。 #### lora_state_dict[[diffusers.loaders.MiniMaxH3LoraLoaderMixin.lora_state_dict]] ```python lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7159) 有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。 #### save_lora_weights[[diffusers.loaders.MiniMaxH3LoraLoaderMixin.save_lora_weights]] ```python save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, transformer_ref_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None, transformer_ref_lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7446) 保存一个或两个 MiniMax-H3 分区的 LoRA 层。LoRA 所属的分区无法从其键中恢复,因此这是发布记录该信息的 H3 LoRA 的唯一方式。更多信息请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。 #### unfuse_lora[[diffusers.loaders.MiniMaxH3LoraLoaderMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] = ['transformer', 'transformer_ref'], **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7508) 有关更多详细信息,请参阅 `unfuse_lora()`。 ## PikaH3FlexLoraLoaderMixin[[diffusers.loaders.PikaH3FlexLoraLoaderMixin]] #### diffusers.loaders.PikaH3FlexLoraLoaderMixin[[diffusers.loaders.PikaH3FlexLoraLoaderMixin]] ```python diffusers.loaders.PikaH3FlexLoraLoaderMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7515) 将 LoRA 层加载到 [MiniMaxH3Transformer3DModel](/docs/diffusers/v0.43.0/zh/api/models/minimax_h3_transformer3d#diffusers.MiniMaxH3Transformer3DModel) 中。专用于 [PikaH3FlexModularPipeline](/docs/diffusers/v0.43.0/zh/api/pipelines/pika_h3_flex#diffusers.PikaH3FlexModularPipeline)。 Pika H3 Flex 版本*是*针对两个 MiniMax-H3 transformer 分区的一对 rank-128 适配器,以 diffusers 模块名称的纯 PEFT 转储形式发布(`transformer` 对应 `fl2va/adapter.safetensors`,`transformer_ref` 对应 `ref2va/adapter.safetensors`)。这些不含 alpha 的文件在 `alpha == rank`(缩放系数 1.0)时加载,这是适配器自身的约定;其加载、路由和保存逻辑均遵循 `MiniMaxH3LoraLoaderMixin`:`fl2va` 适配器加载到 `transformer`,`ref2va` 适配器加载到 `transformer_ref`——当流水线仅持有该分区时(`workflow="ref2va"`)单独加载,当流水线同时持有两个分区时则需设置 `load_into_transformer_ref=True`。 ```py pipe.load_lora_weights("Pika-Labs/Pika-H3-Flex", weight_name="fl2va/adapter.safetensors") ``` ## LoraBaseMixin[[diffusers.loaders.lora_base.LoraBaseMixin]] #### diffusers.loaders.lora_base.LoraBaseMixin[[diffusers.loaders.lora_base.LoraBaseMixin]] ```python diffusers.loaders.lora_base.LoraBaseMixin() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L491) 用于处理 LoRA 的实用类。 #### delete_adapters[[diffusers.loaders.lora_base.LoraBaseMixin.delete_adapters]] ```python delete_adapters(adapter_names: list[str] | str) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L873) **参数:** adapter_names (`list[str, str]`) : 要删除的适配器名称。 从流水线中删除适配器的 LoRA 层。 示例: ```py from diffusers import AutoPipelineForText2Image import torch pipeline = AutoPipelineForText2Image.from_pretrained( "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16 ).to("cuda") pipeline.load_lora_weights( "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_names="cinematic" ) pipeline.delete_adapters("cinematic") ``` #### disable_lora[[diffusers.loaders.lora_base.LoraBaseMixin.disable_lora]] ```python disable_lora() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L813) 禁用管道中处于激活状态的 LoRA 层。 示例: ```py from diffusers import AutoPipelineForText2Image import torch pipeline = AutoPipelineForText2Image.from_pretrained( "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16 ).to("cuda") pipeline.load_lora_weights( "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic" ) pipeline.disable_lora() ``` #### enable_lora[[diffusers.loaders.lora_base.LoraBaseMixin.enable_lora]] ```python enable_lora() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L843) 启用管道中处于激活状态的 LoRA 层。 示例: ```py from diffusers import AutoPipelineForText2Image import torch pipeline = AutoPipelineForText2Image.from_pretrained( "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16 ).to("cuda") pipeline.load_lora_weights( "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic" ) pipeline.enable_lora() ``` #### enable_lora_hotswap[[diffusers.loaders.lora_base.LoraBaseMixin.enable_lora_hotswap]] ```python enable_lora_hotswap(**kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L1020) **参数:** target_rank (`int`) : 所有将加载的适配器中的最高秩。 check_compiled (`str`,*可选*,默认为 `"error"`):如何处理已编译的模型。该检查可能返回以下消息:- "error"(默认):引发错误 - "warn":发出警告 - "ignore":不执行任何操作 在不触发模型重新编译的情况下热插拔适配器,或者在加载的适配器秩不同时进行热插拔。 #### fuse_lora[[diffusers.loaders.lora_base.LoraBaseMixin.fuse_lora]] ```python fuse_lora(components: list[str] | None = None, lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L549) **参数:** components : (`list[str]`): 要融合 LoRA 的 LoRA 可注入组件列表。 lora_scale (`float`,默认值为 1.0):控制 LoRA 参数对输出的影响程度。 safe_fusing (`bool`,默认为 `False`):是否在融合前检查融合权重中的 NaN 值,如果值为 NaN 则不进行融合。 adapter_names (`list[str]`, *可选*) : 用于融合的适配器名称。如果未传入任何值,则所有处于活动状态的适配器都将被融合。 将 LoRA 参数融合到对应块的原始参数中。 示例: ```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) ``` #### get_active_adapters[[diffusers.loaders.lora_base.LoraBaseMixin.get_active_adapters]] ```python get_active_adapters() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L911) 获取当前激活的适配器列表。 示例: ```python from diffusers import DiffusionPipeline pipeline = DiffusionPipeline.from_pretrained( "stabilityai/stable-diffusion-xl-base-1.0", ).to("cuda") pipeline.load_lora_weights("CiroN2022/toy-face", weight_name="toy_face_sdxl.safetensors", adapter_name="toy") pipeline.get_active_adapters() ``` #### get_list_adapters[[diffusers.loaders.lora_base.LoraBaseMixin.get_list_adapters]] ```python get_list_adapters() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L944) 获取管道中当前所有可用适配器的列表。 #### set_adapters[[diffusers.loaders.lora_base.LoraBaseMixin.set_adapters]] ```python set_adapters(adapter_names: list[str] | str, adapter_weights: float | dict[str, float | dict | None] | list[float | dict[str, float | dict | None] | None] | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L698) **参数:** adapter_names (`list[str]` 或 `str`) : 要使用的适配器名称。 adapter_weights (`float`、`dict` 或 `list`,*可选*):适配器的权重。标量会缩放所有支持 LoRA 的组件。字典将组件名称(例如 `"unet"`、`"transformer"` 或 `"text_encoder"`)映射到对应的权重;UNet 组件的权重也可以是一个包含块缩放的嵌套字典。单个标量或字典会应用于每个适配器。列表必须按 `adapter_names` 的顺序,为每个适配器包含一个标量、字典或 `None`。`None` 默认为 `1.0`,包括列表项和组件权重。未指定的组件和 UNet 块也默认为 `1.0`。 设置当前用于流水线的活动适配器。 示例: ```py from diffusers import AutoPipelineForText2Image import torch pipeline = AutoPipelineForText2Image.from_pretrained( "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16 ).to("cuda") pipeline.load_lora_weights( "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic" ) pipeline.load_lora_weights( "troveski/sdxl-lora", weight_name="pytorch_lora_weights.safetensors", adapter_name="illustration" ) pipeline.set_adapters(["cinematic", "illustration"], adapter_weights=[0.5, 0.5]) pipeline.set_adapters( ["cinematic", "illustration"], adapter_weights=[{"unet": 0.5}, {"unet": 0.8, "text_encoder": 0.6}], ) ``` #### set_lora_device[[diffusers.loaders.lora_base.LoraBaseMixin.set_lora_device]] ```python set_lora_device(adapter_names: list[str], device: torch.device | str | int) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L966) **参数:** adapter_names (`list[str]`) : 要发送到该设备的适配器列表。 device (`torch.device | str | int`) : 用于将适配器发送到的设备。可以是 torch 设备、字符串或整数。 将 `adapter_names` 中列出的 LoRA 移动到目标设备。如果您希望加载多个适配器并释放部分 GPU 内存,可以将 LoRA 卸载到 CPU,此操作非常有用。 将 LoRA 适配器卸载到 CPU 后,只要模型的其余部分仍在 GPU 上,这些 LoRA 适配器就无法再用于推理,因为这会导致设备不匹配。在将这些 LoRA 适配器用于推理之前,请务必将设备重新设置回 GPU。 ```python >>> pipe.load_lora_weights(path_1, adapter_name="adapter-1") >>> pipe.load_lora_weights(path_2, adapter_name="adapter-2") >>> pipe.set_adapters("adapter-1") >>> image_1 = pipe(**kwargs) >>> # switch to adapter-2, offload adapter-1 >>> pipeline.set_lora_device(adapter_names=["adapter-1"], device="cpu") >>> pipeline.set_lora_device(adapter_names=["adapter-2"], device="cuda:0") >>> pipe.set_adapters("adapter-2") >>> image_2 = pipe(**kwargs) >>> # switch back to adapter-1, offload adapter-2 >>> pipeline.set_lora_device(adapter_names=["adapter-2"], device="cpu") >>> pipeline.set_lora_device(adapter_names=["adapter-1"], device="cuda:0") >>> pipe.set_adapters("adapter-1") >>> ... ``` #### unfuse_lora[[diffusers.loaders.lora_base.LoraBaseMixin.unfuse_lora]] ```python unfuse_lora(components: list[str] | None = None, **kwargs) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L636) **参数:** components (`list[str]`) : 要从中解除 LoRA 融合的、可注入 LoRA 的组件列表。 unfuse_unet (`bool`,默认为 `True`):是否取消融合 UNet LoRA 参数。 unfuse_text_encoder (`bool`,默认为 `True`):是否解绑文本编码器的 LoRA 参数。如果文本编码器未通过 monkey-patch 方式注入 LoRA 参数,则此设置不会生效。 撤销 [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora) 的效果。 #### unload_lora_weights[[diffusers.loaders.lora_base.LoraBaseMixin.unload_lora_weights]] ```python unload_lora_weights() ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L526) 卸载 LoRA 参数。 示例: ```python >>> # Assuming `pipeline` is already loaded with the LoRA parameters. >>> pipeline.unload_lora_weights() >>> ... ``` #### write_lora_layers[[diffusers.loaders.lora_base.LoraBaseMixin.write_lora_layers]] ```python write_lora_layers(state_dict: dict[str, torch.Tensor], save_directory: str, is_main_process: bool, weight_name: str, save_function: Callable, safe_serialization: bool, lora_adapter_metadata: dict | None = None) ``` [源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_base.py#L1043) 将 LoRA 层的 state dict(可选包含元数据)写入磁盘。
# LoRA
LoRA 是一种快速且轻量级的训练方法,它通过插入并训练数量远少于模型全部参数的参数来实现训练。这会产生更小的文件(约 100 MB),并使快速训练模型以学习新概念变得更加容易。LoRA 权重通常加载到去噪器、文本编码器或两者中。去噪器通常对应于 UNet(例如 [UNet2DConditionModel](/docs/diffusers/v0.43.0/zh/api/models/unet2d-cond#diffusers.UNet2DConditionModel))或 Transformer(例如 [SD3Transformer2DModel](/docs/diffusers/v0.43.0/zh/api/models/sd3_transformer2d#diffusers.SD3Transformer2DModel))。有几种用于加载 LoRA 权重的类:
- `StableDiffusionLoraLoaderMixin` 提供了用于加载和卸载、融合和取消融合、启用和禁用等管理 LoRA 权重的函数。该类可与任何模型配合使用。
- `StableDiffusionXLLoraLoaderMixin` 是 `StableDiffusionLoraLoaderMixin` 类的 [Stable Diffusion (SDXL)](../../api/pipelines/stable_diffusion/stable_diffusion_xl) 版本,用于加载和保存 LoRA 权重。它仅能与 SDXL 模型配合使用。
- `SD3LoraLoaderMixin` 为 [Stable Diffusion 3](https://huggingface.co/blog/sd3) 提供了类似的功能。
- `FluxLoraLoaderMixin` 为 [Flux](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux) 提供了类似的功能。
- `CogVideoXLoraLoaderMixin` 为 [CogVideoX](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogvideox) 提供了类似的功能。
- `Mochi1LoraLoaderMixin` 为 [Mochi](https://huggingface.co/docs/diffusers/main/en/api/pipelines/mochi) 提供了类似的功能。
- `AuraFlowLoraLoaderMixin` 为 [AuraFlow](https://huggingface.co/fal/AuraFlow) 提供了类似的功能。
- `LTXVideoLoraLoaderMixin` 为 [LTX-Video](https://huggingface.co/docs/diffusers/main/en/api/pipelines/ltx_video) 提供了类似的功能。
- `SanaLoraLoaderMixin` 为 [Sana](https://huggingface.co/docs/diffusers/main/en/api/pipelines/sana) 提供了类似的功能。
- `HeliosLoraLoaderMixin` 为 [HunyuanVideo](https://huggingface.co/docs/diffusers/main/en/api/pipelines/helios) 提供了类似的功能。
- `HunyuanVideoLoraLoaderMixin` 为 [HunyuanVideo](https://huggingface.co/docs/diffusers/main/en/api/pipelines/hunyuan_video) 提供了类似的功能。
- `Lumina2LoraLoaderMixin` 为 [Lumina2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/lumina2) 提供了类似的功能。
- `WanLoraLoaderMixin` 为 [Wan](https://huggingface.co/docs/diffusers/main/en/api/pipelines/wan) 提供了类似的功能。
- `SkyReelsV2LoraLoaderMixin` 为 [SkyReels-V2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/skyreels_v2) 提供了类似的功能。
- `CogView4LoraLoaderMixin` 为 [CogView4](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogview4) 提供了类似的功能。
- `AmusedLoraLoaderMixin` 用于 `AmusedPipeline`。
- `AnimaLoraLoaderMixin` 为 [Anima](https://huggingface.co/docs/diffusers/main/en/api/pipelines/anima) 提供了类似的功能。
- `AceStepLoraLoaderMixin` 为 [ACE-Step](https://huggingface.co/docs/diffusers/main/en/api/pipelines/ace_step) 提供类似的功能。
- `HiDreamImageLoraLoaderMixin` 为 [HiDream Image](https://huggingface.co/docs/diffusers/main/en/api/pipelines/hidream) 提供了类似的功能
- `QwenImageLoraLoaderMixin` 为 [Qwen Image](https://huggingface.co/docs/diffusers/main/en/api/pipelines/qwen) 提供了类似的功能。
- `ZImageLoraLoaderMixin` 为 [Z-Image](https://huggingface.co/docs/diffusers/main/en/api/pipelines/zimage) 提供了类似的功能。
- `Flux2LoraLoaderMixin` 为 [Flux2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux2) 提供了类似的功能。
- `ErnieImageLoraLoaderMixin` 为 [Ernie-Image](https://huggingface.co/docs/diffusers/main/en/api/pipelines/ernie_image) 提供了类似的功能。
- `LTX2LoraLoaderMixin` 为 [Flux2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/ltx2) 提供了类似的功能。
- `MiniMaxH3LoraLoaderMixin` 为 [MiniMax-H3](https://huggingface.co/docs/diffusers/main/en/api/pipelines/minimax_h3) 提供了类似的功能。
- `PikaH3FlexLoraLoaderMixin` 为 [Pika H3 Flex](https://huggingface.co/docs/diffusers/main/en/api/pipelines/pika_h3_flex) 提供了类似的功能。
- `RFD4ProteinaLoraLoaderMixin` 为 [RFD4-Proteína](https://huggingface.co/docs/diffusers/main/en/api/pipelines/rfd4_proteina) 提供了类似的功能。
- `BriaFibo2LoraLoaderMixin` 为 [fibo-2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/bria_fibo2) 提供了类似的功能。
- `LoraBaseMixin` 提供了一个基类,其中包含用于融合、解融合、卸载 LoRA 等的若干实用方法。
> [!TIP]
> 要了解如何加载 LoRA 权重,请参阅 [LoRA](../../tutorials/using_peft_for_inference) 加载指南。
## RFD4ProteinaLoraLoaderMixin[[diffusers.loaders.RFD4ProteinaLoraLoaderMixin]]
#### diffusers.loaders.RFD4ProteinaLoraLoaderMixin[[diffusers.loaders.RFD4ProteinaLoraLoaderMixin]]
```python
diffusers.loaders.RFD4ProteinaLoraLoaderMixin()
```
[源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7532)
将 LoRA 层加载到 [RFD4ProteinaTransformer](/docs/diffusers/v0.43.0/zh/api/models/rfd4_proteina_transformer#diffusers.RFD4ProteinaTransformer) 中。
自编码器与去噪器检查点配对使用,且未进行适配,因此 Transformer 是唯一可加载的模块。
#### fuse_lora[[diffusers.loaders.RFD4ProteinaLoraLoaderMixin.fuse_lora]]
```python
fuse_lora(components: list[str] = ['transformer'], lora_scale: float = 1.0, safe_fusing: bool = False, adapter_names: list[str] | None = None, **kwargs)
```
[源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7694)
有关更多详细信息,请参阅 `fuse_lora()`。
#### load_lora_into_transformer[[diffusers.loaders.RFD4ProteinaLoraLoaderMixin.load_lora_into_transformer]]
```python
load_lora_into_transformer(state_dict, transformer, adapter_name = None, _pipeline = None, low_cpu_mem_usage = False, hotswap: bool = False, metadata = None)
```
[源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7631)
有关更多详细信息,请参阅 [load_lora_into_unet()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet)。
#### load_lora_weights[[diffusers.loaders.RFD4ProteinaLoraLoaderMixin.load_lora_weights]]
```python
load_lora_weights(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], adapter_name: str | None = None, hotswap: bool = False, **kwargs)
```
[源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7595)
有关更多详细信息,请参阅 [load_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.load_lora_weights)。
#### lora_state_dict[[diffusers.loaders.RFD4ProteinaLoraLoaderMixin.lora_state_dict]]
```python
lora_state_dict(pretrained_model_name_or_path_or_dict: str | dict[str, torch.Tensor], **kwargs)
```
[源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7543)
有关更多详细信息,请参阅 [lora_state_dict()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.lora_state_dict)。
#### save_lora_weights[[diffusers.loaders.RFD4ProteinaLoraLoaderMixin.save_lora_weights]]
```python
save_lora_weights(save_directory: str | os.PathLike, transformer_lora_layers: dict[str, typing.Union[torch.nn.Module, torch.Tensor]] = None, is_main_process: bool = True, weight_name: str = None, save_function: typing.Callable = None, safe_serialization: bool = True, transformer_lora_adapter_metadata: dict | None = None)
```
[源代码](https://github.com/huggingface/diffusers/blob/v0.43.0/src/diffusers/loaders/lora_pipeline.py#L7658)
有关更多信息,请参阅 [save_lora_weights()](/docs/diffusers/v0.43.0/zh/api/loaders/lora#diffusers.loaders.StableDiffusionLoraLoaderMixin.save_lora_weights)。

Xet Storage Details

Size:
164 kB
·
Xet hash:
324d15f4300c7baecffa9f1e3494982745c472ba7da7b70be1bd20eb87d25370

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.