Merge branch 'main' into bnb-compile-docs

Author: sayakpaul

src/diffusers/models/embeddings.py

@@ -1149,9 +1149,7 @@ def get_1d_rotary_pos_embed(
 
     theta = theta * ntk_factor
     freqs = (
-        1.0
-        / (theta ** (torch.arange(0, dim, 2, dtype=freqs_dtype, device=pos.device)[: (dim // 2)] / dim))
-        / linear_factor
+        1.0 / (theta ** (torch.arange(0, dim, 2, dtype=freqs_dtype, device=pos.device) / dim)) / linear_factor
     )  # [D/2]
     freqs = torch.outer(pos, freqs)  # type: ignore   # [S, D/2]
     is_npu = freqs.device.type == "npu"

src/diffusers/models/modeling_utils.py

@@ -814,14 +814,43 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
                 Mirror source to resolve accessibility issues if you're downloading a model in China. We do not
                 guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
                 information.
-            device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
+            device_map (`Union[int, str, torch.device]` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
                 A map that specifies where each submodule should go. It doesn't need to be defined for each
                 parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the
                 same device. Defaults to `None`, meaning that the model will be loaded on CPU.
 
+                Examples:
+
+                ```py
+                >>> from diffusers import AutoModel
+                >>> import torch
+
+                >>> # This works.
+                >>> model = AutoModel.from_pretrained(
+                ...     "stabilityai/stable-diffusion-xl-base-1.0", subfolder="unet", device_map="cuda"
+                ... )
+                >>> # This also works (integer accelerator device ID).
+                >>> model = AutoModel.from_pretrained(
+                ...     "stabilityai/stable-diffusion-xl-base-1.0", subfolder="unet", device_map=0
+                ... )
+                >>> # Specifying a supported offloading strategy like "auto" also works.
+                >>> model = AutoModel.from_pretrained(
+                ...     "stabilityai/stable-diffusion-xl-base-1.0", subfolder="unet", device_map="auto"
+                ... )
+                >>> # Specifying a dictionary as `device_map` also works.
+                >>> model = AutoModel.from_pretrained(
+                ...     "stabilityai/stable-diffusion-xl-base-1.0",
+                ...     subfolder="unet",
+                ...     device_map={"": torch.device("cuda")},
+                ... )
+                ```
+
                 Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
                 more information about each option see [designing a device
-                map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
+                map](https://huggingface.co/docs/accelerate/en/concept_guides/big_model_inference#the-devicemap). You
+                can also refer to the [Diffusers-specific
+                documentation](https://huggingface.co/docs/diffusers/main/en/training/distributed_inference#model-sharding)
+                for more concrete examples.
             max_memory (`Dict`, *optional*):
                 A dictionary device identifier for the maximum memory. Will default to the maximum memory available for
                 each GPU and the available CPU RAM if unset.
@@ -1387,7 +1416,7 @@ def _load_pretrained_model(
         low_cpu_mem_usage: bool = True,
         dtype: Optional[Union[str, torch.dtype]] = None,
         keep_in_fp32_modules: Optional[List[str]] = None,
-        device_map: Dict[str, Union[int, str, torch.device]] = None,
+        device_map: Union[str, int, torch.device, Dict[str, Union[int, str, torch.device]]] = None,
         offload_state_dict: Optional[bool] = None,
         offload_folder: Optional[Union[str, os.PathLike]] = None,
         dduf_entries: Optional[Dict[str, DDUFEntry]] = None,

src/diffusers/pipelines/flux/pipeline_flux.py

@@ -898,6 +898,7 @@ def __call__(
             )
 
         # 6. Denoising loop
+        self.scheduler.set_begin_index(0)
         with self.progress_bar(total=num_inference_steps) as progress_bar:
             for i, t in enumerate(timesteps):
                 if self.interrupt:

src/diffusers/pipelines/pipeline_utils.py

@@ -669,14 +669,11 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
                 Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not
                 guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
                 information.
-            device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
-                A map that specifies where each submodule should go. It doesn’t need to be defined for each
-                parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the
-                same device.
-
-                Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
-                more information about each option see [designing a device
-                map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
+            device_map (`str`, *optional*):
+                Strategy that dictates how the different components of a pipeline should be placed on available
+                devices. Currently, only "balanced" `device_map` is supported. Check out
+                [this](https://huggingface.co/docs/diffusers/main/en/tutorials/inference_with_big_models#device-placement)
+                to know more.
             max_memory (`Dict`, *optional*):
                 A dictionary device identifier for the maximum memory. Will default to the maximum memory available for
                 each GPU and the available CPU RAM if unset.

src/diffusers/pipelines/wan/pipeline_wan.py

@@ -388,8 +388,10 @@ def __call__(
 
         Args:
             prompt (`str` or `List[str]`, *optional*):
-                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
-                instead.
+                The prompt or prompts to guide the image generation. If not defined, pass `prompt_embeds` instead.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to avoid during image generation. If not defined, pass `negative_prompt_embeds`
+                instead. Ignored when not using guidance (`guidance_scale` < `1`).
             height (`int`, defaults to `480`):
                 The height in pixels of the generated image.
             width (`int`, defaults to `832`):
@@ -434,8 +436,9 @@ def __call__(
                 The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                 will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                 `._callback_tensor_inputs` attribute of your pipeline class.
-            autocast_dtype (`torch.dtype`, *optional*, defaults to `torch.bfloat16`):
-                The dtype to use for the torch.amp.autocast.
+            max_sequence_length (`int`, defaults to `512`):
+                The maximum sequence length of the text encoder. If the prompt is longer than this, it will be
+                truncated. If the prompt is shorter, it will be padded to this length.
 
         Examples:
 

src/diffusers/pipelines/wan/pipeline_wan_i2v.py

@@ -562,12 +562,10 @@ def __call__(
                 The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                 will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                 `._callback_tensor_inputs` attribute of your pipeline class.
-            max_sequence_length (`int`, *optional*, defaults to `512`):
-                The maximum sequence length of the prompt.
-            shift (`float`, *optional*, defaults to `5.0`):
-                The shift of the flow.
-            autocast_dtype (`torch.dtype`, *optional*, defaults to `torch.bfloat16`):
-                The dtype to use for the torch.amp.autocast.
+            max_sequence_length (`int`, defaults to `512`):
+                The maximum sequence length of the text encoder. If the prompt is longer than this, it will be
+                truncated. If the prompt is shorter, it will be padded to this length.
+
         Examples:
 
         Returns:

src/diffusers/pipelines/wan/pipeline_wan_vace.py

@@ -687,8 +687,33 @@ def __call__(
 
         Args:
             prompt (`str` or `List[str]`, *optional*):
-                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`
                 instead.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            video (`List[PIL.Image.Image]`, *optional*):
+                The input video or videos to be used as a starting point for the generation. The video should be a list
+                of PIL images, a numpy array, or a torch tensor. Currently, the pipeline only supports generating one
+                video at a time.
+            mask (`List[PIL.Image.Image]`, *optional*):
+                The input mask defines which video regions to condition on and which to generate. Black areas in the
+                mask indicate conditioning regions, while white areas indicate regions for generation. The mask should
+                be a list of PIL images, a numpy array, or a torch tensor. Currently supports generating a single video
+                at a time.
+            reference_images (`List[PIL.Image.Image]`, *optional*):
+                A list of one or more reference images as extra conditioning for the generation. For example, if you
+                are trying to inpaint a video to change the character, you can pass reference images of the new
+                character here. Refer to the Diffusers [examples](https://github.com/huggingface/diffusers/pull/11582)
+                and original [user
+                guide](https://github.com/ali-vilab/VACE/blob/0897c6d055d7d9ea9e191dce763006664d9780f8/UserGuide.md)
+                for a full list of supported tasks and use cases.
+            conditioning_scale (`float`, `List[float]`, `torch.Tensor`, defaults to `1.0`):
+                The conditioning scale to be applied when adding the control conditioning latent stream to the
+                denoising latent stream in each control layer of the model. If a float is provided, it will be applied
+                uniformly to all layers. If a list or tensor is provided, it should have the same length as the number
+                of control layers in the model (`len(transformer.config.vace_layers)`).
             height (`int`, defaults to `480`):
                 The height in pixels of the generated image.
             width (`int`, defaults to `832`):
@@ -733,8 +758,9 @@ def __call__(
                 The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                 will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                 `._callback_tensor_inputs` attribute of your pipeline class.
-            autocast_dtype (`torch.dtype`, *optional*, defaults to `torch.bfloat16`):
-                The dtype to use for the torch.amp.autocast.
+            max_sequence_length (`int`, defaults to `512`):
+                The maximum sequence length of the text encoder. If the prompt is longer than this, it will be
+                truncated. If the prompt is shorter, it will be padded to this length.
 
         Examples:
 

src/diffusers/pipelines/wan/pipeline_wan_video2video.py

@@ -419,12 +419,7 @@ def prepare_latents(
         )
 
         if latents is None:
-            if isinstance(generator, list):
-                init_latents = [
-                    retrieve_latents(self.vae.encode(video[i].unsqueeze(0)), generator[i]) for i in range(batch_size)
-                ]
-            else:
-                init_latents = [retrieve_latents(self.vae.encode(vid.unsqueeze(0)), generator) for vid in video]
+            init_latents = [retrieve_latents(self.vae.encode(vid.unsqueeze(0)), sample_mode="argmax") for vid in video]
 
             init_latents = torch.cat(init_latents, dim=0).to(dtype)
 
@@ -441,7 +436,7 @@ def prepare_latents(
             if hasattr(self.scheduler, "add_noise"):
                 latents = self.scheduler.add_noise(init_latents, noise, timestep)
             else:
-                latents = self.scheduelr.scale_noise(init_latents, timestep, noise)
+                latents = self.scheduler.scale_noise(init_latents, timestep, noise)
         else:
             latents = latents.to(device)
 
@@ -513,7 +508,7 @@ def __call__(
 
         Args:
             prompt (`str` or `List[str]`, *optional*):
-                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`
                 instead.
             height (`int`, defaults to `480`):
                 The height in pixels of the generated image.
@@ -530,6 +525,8 @@ def __call__(
                 of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting
                 `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to
                 the text `prompt`, usually at the expense of lower image quality.
+            strength (`float`, defaults to `0.8`):
+                Higher strength leads to more differences between original image and generated video.
             num_videos_per_prompt (`int`, *optional*, defaults to 1):
                 The number of images to generate per prompt.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
@@ -559,8 +556,9 @@ def __call__(
                 The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                 will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                 `._callback_tensor_inputs` attribute of your pipeline class.
-            autocast_dtype (`torch.dtype`, *optional*, defaults to `torch.bfloat16`):
-                The dtype to use for the torch.amp.autocast.
+            max_sequence_length (`int`, defaults to `512`):
+                The maximum sequence length of the text encoder. If the prompt is longer than this, it will be
+                truncated. If the prompt is shorter, it will be padded to this length.
 
         Examples:
 

src/diffusers/quantizers/quantization_config.py

@@ -493,7 +493,7 @@ def __init__(self, quant_type: str, modules_to_not_convert: Optional[List[str]]
         TORCHAO_QUANT_TYPE_METHODS = self._get_torchao_quant_type_to_method()
         if self.quant_type not in TORCHAO_QUANT_TYPE_METHODS.keys():
             is_floating_quant_type = self.quant_type.startswith("float") or self.quant_type.startswith("fp")
-            if is_floating_quant_type and not self._is_cuda_capability_atleast_8_9():
+            if is_floating_quant_type and not self._is_xpu_or_cuda_capability_atleast_8_9():
                 raise ValueError(
                     f"Requested quantization type: {self.quant_type} is not supported on GPUs with CUDA capability <= 8.9. You "
                     f"can check the CUDA capability of your GPU using `torch.cuda.get_device_capability()`."
@@ -645,7 +645,7 @@ def generate_fpx_quantization_types(bits: int):
             QUANTIZATION_TYPES.update(INT8_QUANTIZATION_TYPES)
             QUANTIZATION_TYPES.update(UINTX_QUANTIZATION_DTYPES)
 
-            if cls._is_cuda_capability_atleast_8_9():
+            if cls._is_xpu_or_cuda_capability_atleast_8_9():
                 QUANTIZATION_TYPES.update(FLOATX_QUANTIZATION_TYPES)
 
             return QUANTIZATION_TYPES
@@ -655,14 +655,16 @@ def generate_fpx_quantization_types(bits: int):
             )
 
     @staticmethod
-    def _is_cuda_capability_atleast_8_9() -> bool:
-        if not torch.cuda.is_available():
-            raise RuntimeError("TorchAO requires a CUDA compatible GPU and installation of PyTorch.")
-
-        major, minor = torch.cuda.get_device_capability()
-        if major == 8:
-            return minor >= 9
-        return major >= 9
+    def _is_xpu_or_cuda_capability_atleast_8_9() -> bool:
+        if torch.cuda.is_available():
+            major, minor = torch.cuda.get_device_capability()
+            if major == 8:
+                return minor >= 9
+            return major >= 9
+        elif torch.xpu.is_available():
+            return True
+        else:
+            raise RuntimeError("TorchAO requires a CUDA compatible GPU or Intel XPU and installation of PyTorch.")
 
     def get_apply_tensor_subclass(self):
         TORCHAO_QUANT_TYPE_METHODS = self._get_torchao_quant_type_to_method()

src/diffusers/utils/import_utils.py

@@ -99,6 +99,7 @@ def _is_package_available(pkg_name: str, get_dist_name: bool = False) -> Tuple[b
 else:
     logger.info("Disabling PyTorch because USE_TORCH is set")
     _torch_available = False
+    _torch_version = "N/A"
 
 _jax_version = "N/A"
 _flax_version = "N/A"