up

* polish caching docs.

* Update docs/source/en/optimization/cache.md

Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

* Update docs/source/en/optimization/cache.md

Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

* up

---------

Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

docs/source/en/api/cache.md

@@ -29,7 +29,7 @@ Cache methods speedup diffusion transformers by storing and reusing intermediate
 
 [[autodoc]] apply_faster_cache
 
-### FirstBlockCacheConfig
+## FirstBlockCacheConfig
 
 [[autodoc]] FirstBlockCacheConfig
 

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

@@ -21,7 +21,7 @@ The abstract from the paper is:
 
 *Image generation has recently seen tremendous advances, with diffusion models allowing to synthesize convincing images for a large variety of text prompts. In this article, we propose DiffEdit, a method to take advantage of text-conditioned diffusion models for the task of semantic image editing, where the goal is to edit an image based on a text query. Semantic image editing is an extension of image generation, with the additional constraint that the generated image should be as similar as possible to a given input image. Current editing methods based on diffusion models usually require to provide a mask, making the task much easier by treating it as a conditional inpainting task. In contrast, our main contribution is able to automatically generate a mask highlighting regions of the input image that need to be edited, by contrasting predictions of a diffusion model conditioned on different text prompts. Moreover, we rely on latent inference to preserve content in those regions of interest and show excellent synergies with mask-based diffusion. DiffEdit achieves state-of-the-art editing performance on ImageNet. In addition, we evaluate semantic image editing in more challenging settings, using images from the COCO dataset as well as text-based generated images.*
 
-The original codebase can be found at [Xiang-cd/DiffEdit-stable-diffusion](https://github.com/Xiang-cd/DiffEdit-stable-diffusion), and you can try it out in this [demo](https://blog.problemsolversguild.com/technical/research/2022/11/02/DiffEdit-Implementation.html).
+The original codebase can be found at [Xiang-cd/DiffEdit-stable-diffusion](https://github.com/Xiang-cd/DiffEdit-stable-diffusion), and you can try it out in this [demo](https://blog.problemsolversguild.com/posts/2022-11-02-diffedit-implementation.html).
 
 This pipeline was contributed by [clarencechen](https://github.com/clarencechen). ❤️
 

docs/source/en/modular_diffusers/custom_blocks.md

@@ -140,7 +140,7 @@ class Florence2ImageAnnotatorBlock(ModularPipelineBlocks):
                 type_hint=str,
                 required=True,
                 default="mask_image",
-                description="""Output type from annotation predictions. Availabe options are
+                description="""Output type from annotation predictions. Available options are
                 mask_image:
                     -black and white mask image for the given image based on the task type
                 mask_overlay:
@@ -256,7 +256,7 @@ class Florence2ImageAnnotatorBlock(ModularPipelineBlocks):
                 type_hint=str,
                 required=True,
                 default="mask_image",
-                description="""Output type from annotation predictions. Availabe options are
+                description="""Output type from annotation predictions. Available options are
                 mask_image:
                     -black and white mask image for the given image based on the task type
                 mask_overlay:

docs/source/en/modular_diffusers/loop_sequential_pipeline_blocks.md

@@ -53,7 +53,7 @@ The loop wrapper can pass additional arguments, like current iteration index, to
 
 A loop block is a [`~modular_pipelines.ModularPipelineBlocks`], but the `__call__` method behaves differently.
 
-- It recieves the iteration variable from the loop wrapper.
+- It receives the iteration variable from the loop wrapper.
 - It works directly with the [`~modular_pipelines.BlockState`] instead of the [`~modular_pipelines.PipelineState`].
 - It doesn't require retrieving or updating the [`~modular_pipelines.BlockState`].
 

docs/source/en/optimization/cache.md

@@ -68,6 +68,20 @@ config = FasterCacheConfig(
 pipeline.transformer.enable_cache(config)
 ```
 
+## FirstBlockCache
+
+[FirstBlock Cache](https://huggingface.co/docs/diffusers/main/en/api/cache#diffusers.FirstBlockCacheConfig) checks how much the early layers of the denoiser changes from one timestep to the next. If the change is small, the model skips the expensive later layers and reuses the previous output.
+
+```py
+import torch
+from diffusers import DiffusionPipeline
+from diffusers.hooks import apply_first_block_cache, FirstBlockCacheConfig
+
+pipeline = DiffusionPipeline.from_pretrained(
+    "Qwen/Qwen-Image", torch_dtype=torch.bfloat16
+)
+apply_first_block_cache(pipeline.transformer, FirstBlockCacheConfig(threshold=0.2))
+```
 ## TaylorSeer Cache
 
 [TaylorSeer Cache](https://huggingface.co/papers/2403.06923) accelerates diffusion inference by using Taylor series expansions to approximate and cache intermediate activations across denoising steps. The method predicts future outputs based on past computations, reusing them at specified intervals to reduce redundant calculations.
@@ -87,8 +101,7 @@ from diffusers import FluxPipeline, TaylorSeerCacheConfig
 pipe = FluxPipeline.from_pretrained(
     "black-forest-labs/FLUX.1-dev",
     torch_dtype=torch.bfloat16,
-)
-pipe.to("cuda")
+).to("cuda")
 
 config = TaylorSeerCacheConfig(
     cache_interval=5,
@@ -97,4 +110,4 @@ config = TaylorSeerCacheConfig(
     taylor_factors_dtype=torch.bfloat16,
 )
 pipe.transformer.enable_cache(config)
-```
+```

examples/cogvideo/train_cogvideox_image_to_video_lora.py

@@ -149,13 +149,13 @@ def get_args():
         "--validation_prompt",
         type=str,
         default=None,
-        help="One or more prompt(s) that is used during validation to verify that the model is learning. Multiple validation prompts should be separated by the '--validation_prompt_seperator' string.",
+        help="One or more prompt(s) that is used during validation to verify that the model is learning. Multiple validation prompts should be separated by the '--validation_prompt_separator' string.",
     )
     parser.add_argument(
         "--validation_images",
         type=str,
         default=None,
-        help="One or more image path(s) that is used during validation to verify that the model is learning. Multiple validation paths should be separated by the '--validation_prompt_seperator' string. These should correspond to the order of the validation prompts.",
+        help="One or more image path(s) that is used during validation to verify that the model is learning. Multiple validation paths should be separated by the '--validation_prompt_separator' string. These should correspond to the order of the validation prompts.",
     )
     parser.add_argument(
         "--validation_prompt_separator",

examples/cogvideo/train_cogvideox_lora.py

@@ -140,7 +140,7 @@ def get_args():
         "--validation_prompt",
         type=str,
         default=None,
-        help="One or more prompt(s) that is used during validation to verify that the model is learning. Multiple validation prompts should be separated by the '--validation_prompt_seperator' string.",
+        help="One or more prompt(s) that is used during validation to verify that the model is learning. Multiple validation prompts should be separated by the '--validation_prompt_separator' string.",
     )
     parser.add_argument(
         "--validation_prompt_separator",

examples/research_projects/onnxruntime/text_to_image/README.md

@@ -4,7 +4,7 @@ The `train_text_to_image.py` script shows how to fine-tune stable diffusion mode
 
 ___Note___:
 
-___This script is experimental. The script fine-tunes the whole model and often times the model overfits and runs into issues like catastrophic forgetting. It's recommended to try different hyperparamters to get the best result on your dataset.___
+___This script is experimental. The script fine-tunes the whole model and often times the model overfits and runs into issues like catastrophic forgetting. It's recommended to try different hyperparameters to get the best result on your dataset.___
 
 
 ## Running locally with PyTorch

examples/research_projects/sdxl_flax/sdxl_single.py

@@ -18,7 +18,7 @@ cc.initialize_cache("/tmp/sdxl_cache")
 NUM_DEVICES = jax.device_count()
 
 # 1. Let's start by downloading the model and loading it into our pipeline class
-# Adhering to JAX's functional approach, the model's parameters are returned seperatetely and
+# Adhering to JAX's functional approach, the model's parameters are returned separately and
 # will have to be passed to the pipeline during inference
 pipeline, params = FlaxStableDiffusionXLPipeline.from_pretrained(
     "stabilityai/stable-diffusion-xl-base-1.0", revision="refs/pr/95", split_head_dim=True

src/diffusers/models/attention_dispatch.py

@@ -256,6 +256,10 @@ class _HubKernelConfig:
     function_attr: str
     revision: Optional[str] = None
     kernel_fn: Optional[Callable] = None
+    wrapped_forward_attr: Optional[str] = None
+    wrapped_backward_attr: Optional[str] = None
+    wrapped_forward_fn: Optional[Callable] = None
+    wrapped_backward_fn: Optional[Callable] = None
 
 
 # Registry for hub-based attention kernels
@@ -270,7 +274,11 @@ _HUB_KERNELS_REGISTRY: Dict["AttentionBackendName", _HubKernelConfig] = {
         # revision="fake-ops-return-probs",
     ),
     AttentionBackendName.FLASH_HUB: _HubKernelConfig(
-        repo_id="kernels-community/flash-attn2", function_attr="flash_attn_func", revision=None
+        repo_id="kernels-community/flash-attn2",
+        function_attr="flash_attn_func",
+        revision=None,
+        wrapped_forward_attr="flash_attn_interface._wrapped_flash_attn_forward",
+        wrapped_backward_attr="flash_attn_interface._wrapped_flash_attn_backward",
     ),
     AttentionBackendName.FLASH_VARLEN_HUB: _HubKernelConfig(
         repo_id="kernels-community/flash-attn2", function_attr="flash_attn_varlen_func", revision=None
@@ -594,22 +602,39 @@ def _flex_attention_causal_mask_mod(batch_idx, head_idx, q_idx, kv_idx):
 
 
 # ===== Helpers for downloading kernels =====
+def _resolve_kernel_attr(module, attr_path: str):
+    target = module
+    for attr in attr_path.split("."):
+        if not hasattr(target, attr):
+            raise AttributeError(f"Kernel module '{module.__name__}' does not define attribute path '{attr_path}'.")
+        target = getattr(target, attr)
+    return target
+
+
 def _maybe_download_kernel_for_backend(backend: AttentionBackendName) -> None:
     if backend not in _HUB_KERNELS_REGISTRY:
         return
     config = _HUB_KERNELS_REGISTRY[backend]
 
-    if config.kernel_fn is not None:
+    needs_kernel = config.kernel_fn is None
+    needs_wrapped_forward = config.wrapped_forward_attr is not None and config.wrapped_forward_fn is None
+    needs_wrapped_backward = config.wrapped_backward_attr is not None and config.wrapped_backward_fn is None
+
+    if not (needs_kernel or needs_wrapped_forward or needs_wrapped_backward):
         return
 
     try:
         from kernels import get_kernel
 
         kernel_module = get_kernel(config.repo_id, revision=config.revision)
-        kernel_func = getattr(kernel_module, config.function_attr)
+        if needs_kernel:
+            config.kernel_fn = _resolve_kernel_attr(kernel_module, config.function_attr)
 
-        # Cache the downloaded kernel function in the config object
-        config.kernel_fn = kernel_func
+        if needs_wrapped_forward:
+            config.wrapped_forward_fn = _resolve_kernel_attr(kernel_module, config.wrapped_forward_attr)
+
+        if needs_wrapped_backward:
+            config.wrapped_backward_fn = _resolve_kernel_attr(kernel_module, config.wrapped_backward_attr)
 
     except Exception as e:
         logger.error(f"An error occurred while fetching kernel '{config.repo_id}' from the Hub: {e}")
@@ -1060,6 +1085,231 @@ def _flash_attention_backward_op(
     return grad_query, grad_key, grad_value
 
 
+def _flash_attention_hub_forward_op(
+    ctx: torch.autograd.function.FunctionCtx,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
+    dropout_p: float = 0.0,
+    is_causal: bool = False,
+    scale: Optional[float] = None,
+    enable_gqa: bool = False,
+    return_lse: bool = False,
+    _save_ctx: bool = True,
+    _parallel_config: Optional["ParallelConfig"] = None,
+):
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not yet supported for flash-attn hub kernels.")
+    if enable_gqa:
+        raise ValueError("`enable_gqa` is not yet supported for flash-attn hub kernels.")
+
+    config = _HUB_KERNELS_REGISTRY[AttentionBackendName.FLASH_HUB]
+    wrapped_forward_fn = config.wrapped_forward_fn
+    wrapped_backward_fn = config.wrapped_backward_fn
+    if wrapped_forward_fn is None or wrapped_backward_fn is None:
+        raise RuntimeError(
+            "Flash attention hub kernels must expose `_wrapped_flash_attn_forward` and `_wrapped_flash_attn_backward` "
+            "for context parallel execution."
+        )
+
+    if scale is None:
+        scale = query.shape[-1] ** (-0.5)
+
+    window_size = (-1, -1)
+    softcap = 0.0
+    alibi_slopes = None
+    deterministic = False
+    grad_enabled = any(x.requires_grad for x in (query, key, value))
+
+    if grad_enabled or (_parallel_config is not None and _parallel_config.context_parallel_config._world_size > 1):
+        dropout_p = dropout_p if dropout_p > 0 else 1e-30
+
+    with torch.set_grad_enabled(grad_enabled):
+        out, lse, S_dmask, rng_state = wrapped_forward_fn(
+            query,
+            key,
+            value,
+            dropout_p,
+            scale,
+            is_causal,
+            window_size[0],
+            window_size[1],
+            softcap,
+            alibi_slopes,
+            return_lse,
+        )
+        lse = lse.permute(0, 2, 1).contiguous()
+
+    if _save_ctx:
+        ctx.save_for_backward(query, key, value, out, lse, rng_state)
+        ctx.dropout_p = dropout_p
+        ctx.scale = scale
+        ctx.is_causal = is_causal
+        ctx.window_size = window_size
+        ctx.softcap = softcap
+        ctx.alibi_slopes = alibi_slopes
+        ctx.deterministic = deterministic
+
+    return (out, lse) if return_lse else out
+
+
+def _flash_attention_hub_backward_op(
+    ctx: torch.autograd.function.FunctionCtx,
+    grad_out: torch.Tensor,
+    *args,
+    **kwargs,
+):
+    config = _HUB_KERNELS_REGISTRY[AttentionBackendName.FLASH_HUB]
+    wrapped_backward_fn = config.wrapped_backward_fn
+    if wrapped_backward_fn is None:
+        raise RuntimeError(
+            "Flash attention hub kernels must expose `_wrapped_flash_attn_backward` for context parallel execution."
+        )
+
+    query, key, value, out, lse, rng_state = ctx.saved_tensors
+    grad_query, grad_key, grad_value = torch.empty_like(query), torch.empty_like(key), torch.empty_like(value)
+
+    _ = wrapped_backward_fn(
+        grad_out,
+        query,
+        key,
+        value,
+        out,
+        lse,
+        grad_query,
+        grad_key,
+        grad_value,
+        ctx.dropout_p,
+        ctx.scale,
+        ctx.is_causal,
+        ctx.window_size[0],
+        ctx.window_size[1],
+        ctx.softcap,
+        ctx.alibi_slopes,
+        ctx.deterministic,
+        rng_state,
+    )
+
+    grad_query = grad_query[..., : grad_out.shape[-1]]
+    grad_key = grad_key[..., : grad_out.shape[-1]]
+    grad_value = grad_value[..., : grad_out.shape[-1]]
+
+    return grad_query, grad_key, grad_value
+
+
+def _flash_attention_3_hub_forward_op(
+    ctx: torch.autograd.function.FunctionCtx,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
+    dropout_p: float = 0.0,
+    is_causal: bool = False,
+    scale: Optional[float] = None,
+    enable_gqa: bool = False,
+    return_lse: bool = False,
+    _save_ctx: bool = True,
+    _parallel_config: Optional["ParallelConfig"] = None,
+    *,
+    window_size: Tuple[int, int] = (-1, -1),
+    softcap: float = 0.0,
+    num_splits: int = 1,
+    pack_gqa: Optional[bool] = None,
+    deterministic: bool = False,
+    sm_margin: int = 0,
+):
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not yet supported for flash-attn 3 hub kernels.")
+    if dropout_p != 0.0:
+        raise ValueError("`dropout_p` is not yet supported for flash-attn 3 hub kernels.")
+    if enable_gqa:
+        raise ValueError("`enable_gqa` is not yet supported for flash-attn 3 hub kernels.")
+
+    func = _HUB_KERNELS_REGISTRY[AttentionBackendName._FLASH_3_HUB].kernel_fn
+    out = func(
+        q=query,
+        k=key,
+        v=value,
+        softmax_scale=scale,
+        causal=is_causal,
+        qv=None,
+        q_descale=None,
+        k_descale=None,
+        v_descale=None,
+        window_size=window_size,
+        softcap=softcap,
+        num_splits=num_splits,
+        pack_gqa=pack_gqa,
+        deterministic=deterministic,
+        sm_margin=sm_margin,
+        return_attn_probs=return_lse,
+    )
+
+    lse = None
+    if return_lse:
+        out, lse = out
+        lse = lse.permute(0, 2, 1).contiguous()
+
+    if _save_ctx:
+        ctx.save_for_backward(query, key, value)
+        ctx.scale = scale
+        ctx.is_causal = is_causal
+        ctx._hub_kernel = func
+
+    return (out, lse) if return_lse else out
+
+
+def _flash_attention_3_hub_backward_op(
+    ctx: torch.autograd.function.FunctionCtx,
+    grad_out: torch.Tensor,
+    *args,
+    window_size: Tuple[int, int] = (-1, -1),
+    softcap: float = 0.0,
+    num_splits: int = 1,
+    pack_gqa: Optional[bool] = None,
+    deterministic: bool = False,
+    sm_margin: int = 0,
+):
+    query, key, value = ctx.saved_tensors
+    kernel_fn = ctx._hub_kernel
+    with torch.enable_grad():
+        query_r = query.detach().requires_grad_(True)
+        key_r = key.detach().requires_grad_(True)
+        value_r = value.detach().requires_grad_(True)
+
+        out = kernel_fn(
+            q=query_r,
+            k=key_r,
+            v=value_r,
+            softmax_scale=ctx.scale,
+            causal=ctx.is_causal,
+            qv=None,
+            q_descale=None,
+            k_descale=None,
+            v_descale=None,
+            window_size=window_size,
+            softcap=softcap,
+            num_splits=num_splits,
+            pack_gqa=pack_gqa,
+            deterministic=deterministic,
+            sm_margin=sm_margin,
+            return_attn_probs=False,
+        )
+        if isinstance(out, tuple):
+            out = out[0]
+
+        grad_query, grad_key, grad_value = torch.autograd.grad(
+            out,
+            (query_r, key_r, value_r),
+            grad_out,
+            retain_graph=False,
+            allow_unused=False,
+        )
+
+    return grad_query, grad_key, grad_value
+
+
 def _sage_attention_forward_op(
     ctx: torch.autograd.function.FunctionCtx,
     query: torch.Tensor,
@@ -1098,6 +1348,46 @@ def _sage_attention_forward_op(
     return (out, lse) if return_lse else out
 
 
+def _sage_attention_hub_forward_op(
+    ctx: torch.autograd.function.FunctionCtx,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
+    dropout_p: float = 0.0,
+    is_causal: bool = False,
+    scale: Optional[float] = None,
+    enable_gqa: bool = False,
+    return_lse: bool = False,
+    _save_ctx: bool = True,
+    _parallel_config: Optional["ParallelConfig"] = None,
+):
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not yet supported for Sage attention.")
+    if dropout_p > 0.0:
+        raise ValueError("`dropout_p` is not yet supported for Sage attention.")
+    if enable_gqa:
+        raise ValueError("`enable_gqa` is not yet supported for Sage attention.")
+
+    func = _HUB_KERNELS_REGISTRY[AttentionBackendName.SAGE_HUB].kernel_fn
+    out = func(
+        q=query,
+        k=key,
+        v=value,
+        tensor_layout="NHD",
+        is_causal=is_causal,
+        sm_scale=scale,
+        return_lse=return_lse,
+    )
+
+    lse = None
+    if return_lse:
+        out, lse, *_ = out
+        lse = lse.permute(0, 2, 1).contiguous()
+
+    return (out, lse) if return_lse else out
+
+
 def _sage_attention_backward_op(
     ctx: torch.autograd.function.FunctionCtx,
     grad_out: torch.Tensor,
@@ -1512,7 +1802,7 @@ def _flash_attention(
 @_AttentionBackendRegistry.register(
     AttentionBackendName.FLASH_HUB,
     constraints=[_check_device, _check_qkv_dtype_bf16_or_fp16, _check_shape],
-    supports_context_parallel=False,
+    supports_context_parallel=True,
 )
 def _flash_attention_hub(
     query: torch.Tensor,
@@ -1530,17 +1820,35 @@ def _flash_attention_hub(
         raise ValueError("`attn_mask` is not supported for flash-attn 2.")
 
     func = _HUB_KERNELS_REGISTRY[AttentionBackendName.FLASH_HUB].kernel_fn
-    out = func(
-        q=query,
-        k=key,
-        v=value,
-        dropout_p=dropout_p,
-        softmax_scale=scale,
-        causal=is_causal,
-        return_attn_probs=return_lse,
-    )
-    if return_lse:
-        out, lse, *_ = out
+    if _parallel_config is None:
+        out = func(
+            q=query,
+            k=key,
+            v=value,
+            dropout_p=dropout_p,
+            softmax_scale=scale,
+            causal=is_causal,
+            return_attn_probs=return_lse,
+        )
+        if return_lse:
+            out, lse, *_ = out
+    else:
+        out = _templated_context_parallel_attention(
+            query,
+            key,
+            value,
+            None,
+            dropout_p,
+            is_causal,
+            scale,
+            False,
+            return_lse,
+            forward_op=_flash_attention_hub_forward_op,
+            backward_op=_flash_attention_hub_backward_op,
+            _parallel_config=_parallel_config,
+        )
+        if return_lse:
+            out, lse = out
 
     return (out, lse) if return_lse else out
 
@@ -1687,7 +1995,7 @@ def _flash_attention_3(
 @_AttentionBackendRegistry.register(
     AttentionBackendName._FLASH_3_HUB,
     constraints=[_check_device, _check_qkv_dtype_bf16_or_fp16, _check_shape],
-    supports_context_parallel=False,
+    supports_context_parallel=True,
 )
 def _flash_attention_3_hub(
     query: torch.Tensor,
@@ -1702,33 +2010,68 @@ def _flash_attention_3_hub(
     return_attn_probs: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
-    if _parallel_config:
-        raise NotImplementedError(f"{AttentionBackendName._FLASH_3_HUB.value} is not implemented for parallelism yet.")
     if attn_mask is not None:
         raise ValueError("`attn_mask` is not supported for flash-attn 3.")
 
     func = _HUB_KERNELS_REGISTRY[AttentionBackendName._FLASH_3_HUB].kernel_fn
-    out = func(
-        q=query,
-        k=key,
-        v=value,
-        softmax_scale=scale,
-        causal=is_causal,
-        qv=None,
-        q_descale=None,
-        k_descale=None,
-        v_descale=None,
+    if _parallel_config is None:
+        out = func(
+            q=query,
+            k=key,
+            v=value,
+            softmax_scale=scale,
+            causal=is_causal,
+            qv=None,
+            q_descale=None,
+            k_descale=None,
+            v_descale=None,
+            window_size=window_size,
+            softcap=softcap,
+            num_splits=1,
+            pack_gqa=None,
+            deterministic=deterministic,
+            sm_margin=0,
+            return_attn_probs=return_attn_probs,
+        )
+        return (out[0], out[1]) if return_attn_probs else out
+
+    forward_op = functools.partial(
+        _flash_attention_3_hub_forward_op,
         window_size=window_size,
         softcap=softcap,
         num_splits=1,
         pack_gqa=None,
         deterministic=deterministic,
         sm_margin=0,
-        return_attn_probs=return_attn_probs,
     )
-    # When `return_attn_probs` is True, the above returns a tuple of
-    # actual outputs and lse.
-    return (out[0], out[1]) if return_attn_probs else out
+    backward_op = functools.partial(
+        _flash_attention_3_hub_backward_op,
+        window_size=window_size,
+        softcap=softcap,
+        num_splits=1,
+        pack_gqa=None,
+        deterministic=deterministic,
+        sm_margin=0,
+    )
+    out = _templated_context_parallel_attention(
+        query,
+        key,
+        value,
+        None,
+        0.0,
+        is_causal,
+        scale,
+        False,
+        return_attn_probs,
+        forward_op=forward_op,
+        backward_op=backward_op,
+        _parallel_config=_parallel_config,
+    )
+    if return_attn_probs:
+        out, lse = out
+        return out, lse
+
+    return out
 
 
 @_AttentionBackendRegistry.register(
@@ -2309,7 +2652,7 @@ def _sage_attention(
 @_AttentionBackendRegistry.register(
     AttentionBackendName.SAGE_HUB,
     constraints=[_check_device_cuda, _check_qkv_dtype_bf16_or_fp16, _check_shape],
-    supports_context_parallel=False,
+    supports_context_parallel=True,
 )
 def _sage_attention_hub(
     query: torch.Tensor,
@@ -2337,6 +2680,23 @@ def _sage_attention_hub(
         )
         if return_lse:
             out, lse, *_ = out
+    else:
+        out = _templated_context_parallel_attention(
+            query,
+            key,
+            value,
+            None,
+            0.0,
+            is_causal,
+            scale,
+            False,
+            return_lse,
+            forward_op=_sage_attention_hub_forward_op,
+            backward_op=_sage_attention_backward_op,
+            _parallel_config=_parallel_config,
+        )
+        if return_lse:
+            out, lse = out
 
     return (out, lse) if return_lse else out
 

src/diffusers/models/cache_utils.py

@@ -41,9 +41,11 @@ class CacheMixin:
         Enable caching techniques on the model.
 
         Args:
-            config (`Union[PyramidAttentionBroadcastConfig]`):
+            config (`Union[PyramidAttentionBroadcastConfig, FasterCacheConfig, FirstBlockCacheConfig]`):
                 The configuration for applying the caching technique. Currently supported caching techniques are:
                     - [`~hooks.PyramidAttentionBroadcastConfig`]
+                    - [`~hooks.FasterCacheConfig`]
+                    - [`~hooks.FirstBlockCacheConfig`]
 
         Example:
 

src/diffusers/modular_pipelines/qwenimage/before_denoise.py

@@ -455,7 +455,7 @@ class QwenImageSetTimestepsStep(ModularPipelineBlocks):
 
     @property
     def description(self) -> str:
-        return "Step that sets the the scheduler's timesteps for text-to-image generation. Should be run after prepare latents step."
+        return "Step that sets the scheduler's timesteps for text-to-image generation. Should be run after prepare latents step."
 
     @property
     def expected_components(self) -> List[ComponentSpec]:
@@ -579,7 +579,7 @@ class QwenImageSetTimestepsWithStrengthStep(ModularPipelineBlocks):
 
     @property
     def description(self) -> str:
-        return "Step that sets the the scheduler's timesteps for image-to-image generation, and inpainting. Should be run after prepare latents step."
+        return "Step that sets the scheduler's timesteps for image-to-image generation, and inpainting. Should be run after prepare latents step."
 
     @property
     def expected_components(self) -> List[ComponentSpec]:

src/diffusers/pipelines/lucy/pipeline_lucy_edit.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 #
 # Modifications by Decart AI Team:
-# - Based on pipeline_wan.py, but with supports recieving a condition video appended to the channel dimension.
+# - Based on pipeline_wan.py, but with supports receiving a condition video appended to the channel dimension.
 
 import html
 from typing import Any, Callable, Dict, List, Optional, Tuple, Union