update

.github/workflows/pr_modular_tests.yml

@@ -117,7 +117,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality,test]"
+        uv pip install -e ".[quality]"
         #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
         uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps

.github/workflows/pr_tests.yml

@@ -114,7 +114,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality,test]"
+        uv pip install -e ".[quality]"
         #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
         uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps
@@ -191,7 +191,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality,test]"
+        uv pip install -e ".[quality]"
 
     - name: Environment
       run: |
@@ -242,7 +242,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality,test]"
+        uv pip install -e ".[quality]"
         # TODO (sayakpaul, DN6): revisit `--no-deps`
         uv pip install -U peft@git+https://github.com/huggingface/peft.git --no-deps
         uv pip install -U tokenizers

.github/workflows/pr_tests_gpu.yml

@@ -199,6 +199,11 @@ jobs:
 
     - name: Install dependencies
       run: |
+        # Install pkgs which depend on setuptools<81 for pkg_resources first with no build isolation
+        uv pip install pip==25.2 setuptools==80.10.2
+        uv pip install --no-build-isolation k-diffusion==0.0.12
+        uv pip install --upgrade pip setuptools
+        # Install the rest as normal
         uv pip install -e ".[quality]"
         uv pip install peft@git+https://github.com/huggingface/peft.git
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git

.github/workflows/push_tests.yml

@@ -126,6 +126,11 @@ jobs:
 
     - name: Install dependencies
       run: |
+        # Install pkgs which depend on setuptools<81 for pkg_resources first with no build isolation
+        uv pip install pip==25.2 setuptools==80.10.2
+        uv pip install --no-build-isolation k-diffusion==0.0.12
+        uv pip install --upgrade pip setuptools
+        # Install the rest as normal
         uv pip install -e ".[quality]"
         uv pip install peft@git+https://github.com/huggingface/peft.git
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git

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

@@ -29,7 +29,7 @@ Qwen-Image comes in the following variants:
 | Qwen-Image-Edit Plus | [Qwen/Qwen-Image-Edit-2509](https://huggingface.co/Qwen/Qwen-Image-Edit-2509) |
 
 > [!TIP]
-> See the [Caching](../../optimization/cache) guide to speed up inference by storing and reusing intermediate outputs.
+> [Caching](../../optimization/cache) may also speed up inference by storing and reusing intermediate outputs.
 
 ## LoRA for faster inference
 
@@ -190,12 +190,6 @@ For detailed benchmark scripts and results, see [this gist](https://gist.github.
   - all
   - __call__
 
-## QwenImageLayeredPipeline
-
-[[autodoc]] QwenImageLayeredPipeline
-  - all
-  - __call__
-
 ## QwenImagePipelineOutput
 
 [[autodoc]] pipelines.qwenimage.pipeline_output.QwenImagePipelineOutput

setup.py

@@ -101,7 +101,6 @@ _deps = [
     "datasets",
     "filelock",
     "flax>=0.4.1",
-    "ftfy",
     "hf-doc-builder>=0.3.0",
     "httpx<1.0.0",
     "huggingface-hub>=0.34.0,<2.0",
@@ -222,14 +221,12 @@ extras["docs"] = deps_list("hf-doc-builder")
 extras["training"] = deps_list("accelerate", "datasets", "protobuf", "tensorboard", "Jinja2", "peft", "timm")
 extras["test"] = deps_list(
     "compel",
-    "ftfy",
     "GitPython",
     "datasets",
     "Jinja2",
     "invisible-watermark",
     "librosa",
     "parameterized",
-    "protobuf",
     "pytest",
     "pytest-timeout",
     "pytest-xdist",
@@ -238,7 +235,6 @@ extras["test"] = deps_list(
     "sentencepiece",
     "scipy",
     "tiktoken",
-    "torchsde",
     "torchvision",
     "transformers",
     "phonemizer",

src/diffusers/dependency_versions_table.py

@@ -8,7 +8,6 @@ deps = {
     "datasets": "datasets",
     "filelock": "filelock",
     "flax": "flax>=0.4.1",
-    "ftfy": "ftfy",
     "hf-doc-builder": "hf-doc-builder>=0.3.0",
     "httpx": "httpx<1.0.0",
     "huggingface-hub": "huggingface-hub>=0.34.0,<2.0",

src/diffusers/models/attention_dispatch.py

@@ -266,10 +266,6 @@ class _HubKernelConfig:
     function_attr: str
     revision: str | None = None
     kernel_fn: Callable | None = None
-    wrapped_forward_attr: str | None = None
-    wrapped_backward_attr: str | None = None
-    wrapped_forward_fn: Callable | None = None
-    wrapped_backward_fn: Callable | None = None
 
 
 # Registry for hub-based attention kernels
@@ -284,11 +280,7 @@ _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,
-        wrapped_forward_attr="flash_attn_interface._wrapped_flash_attn_forward",
-        wrapped_backward_attr="flash_attn_interface._wrapped_flash_attn_backward",
+        repo_id="kernels-community/flash-attn2", function_attr="flash_attn_func", revision=None
     ),
     AttentionBackendName.FLASH_VARLEN_HUB: _HubKernelConfig(
         repo_id="kernels-community/flash-attn2", function_attr="flash_attn_varlen_func", revision=None
@@ -613,39 +605,22 @@ 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]
 
-    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):
+    if config.kernel_fn is not None:
         return
 
     try:
         from kernels import get_kernel
 
         kernel_module = get_kernel(config.repo_id, revision=config.revision)
-        if needs_kernel:
-            config.kernel_fn = _resolve_kernel_attr(kernel_module, config.function_attr)
+        kernel_func = getattr(kernel_module, config.function_attr)
 
-        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)
+        # Cache the downloaded kernel function in the config object
+        config.kernel_fn = kernel_func
 
     except Exception as e:
         logger.error(f"An error occurred while fetching kernel '{config.repo_id}' from the Hub: {e}")
@@ -1096,237 +1071,6 @@ 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: torch.Tensor | None = None,
-    dropout_p: float = 0.0,
-    is_causal: bool = False,
-    scale: float | None = None,
-    enable_gqa: bool = False,
-    return_lse: bool = False,
-    _save_ctx: bool = True,
-    _parallel_config: "ParallelConfig" | None = 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: torch.Tensor | None = None,
-    dropout_p: float = 0.0,
-    is_causal: bool = False,
-    scale: float | None = None,
-    enable_gqa: bool = False,
-    return_lse: bool = False,
-    _save_ctx: bool = True,
-    _parallel_config: "ParallelConfig" | None = None,
-    *,
-    window_size: tuple[int, int] = (-1, -1),
-    softcap: float = 0.0,
-    num_splits: int = 1,
-    pack_gqa: bool | None = 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: bool | None = None,
-    deterministic: bool = False,
-    sm_margin: int = 0,
-):
-    query, key, value = ctx.saved_tensors
-    kernel_fn = ctx._hub_kernel
-    # NOTE: Unlike the FA2 hub kernel, the FA3 hub kernel does not expose separate wrapped forward/backward
-    # primitives (no `wrapped_forward_attr`/`wrapped_backward_attr` in its `_HubKernelConfig`). We
-    # therefore rerun the forward pass under `torch.enable_grad()` and differentiate through it with
-    # `torch.autograd.grad()`. This is a second forward pass during backward; it can be avoided once
-    # the FA3 hub exposes a dedicated fused backward kernel (analogous to `_wrapped_flash_attn_backward`
-    # in the FA2 hub), at which point this can be refactored to match `_flash_attention_hub_backward_op`.
-    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,
@@ -1365,46 +1109,6 @@ 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: torch.Tensor | None = None,
-    dropout_p: float = 0.0,
-    is_causal: bool = False,
-    scale: float | None = None,
-    enable_gqa: bool = False,
-    return_lse: bool = False,
-    _save_ctx: bool = True,
-    _parallel_config: "ParallelConfig" | None = 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,
@@ -1413,26 +1117,6 @@ def _sage_attention_backward_op(
     raise NotImplementedError("Backward pass is not implemented for Sage attention.")
 
 
-def _maybe_modify_attn_mask_npu(query: torch.Tensor, key: torch.Tensor, attn_mask: torch.Tensor | None = None):
-    # Skip Attention Mask if all values are 1, `None` mask can speedup the computation
-    if attn_mask is not None and torch.all(attn_mask != 0):
-        attn_mask = None
-
-    # Reshape Attention Mask: [batch_size, seq_len_k] -> [batch_size, 1, sqe_len_q, seq_len_k]
-    # https://www.hiascend.com/document/detail/zh/Pytorch/730/apiref/torchnpuCustomsapi/docs/context/torch_npu-npu_fusion_attention.md
-    if (
-        attn_mask is not None
-        and attn_mask.ndim == 2
-        and attn_mask.shape[0] == query.shape[0]
-        and attn_mask.shape[1] == key.shape[1]
-    ):
-        B, Sq, Skv = attn_mask.shape[0], query.shape[1], key.shape[1]
-        attn_mask = ~attn_mask.to(torch.bool)
-        attn_mask = attn_mask.unsqueeze(1).expand(B, Sq, Skv).unsqueeze(1).contiguous()
-
-    return attn_mask
-
-
 def _npu_attention_forward_op(
     ctx: torch.autograd.function.FunctionCtx,
     query: torch.Tensor,
@@ -1450,14 +1134,11 @@ def _npu_attention_forward_op(
     if return_lse:
         raise ValueError("NPU attention backend does not support setting `return_lse=True`.")
 
-    attn_mask = _maybe_modify_attn_mask_npu(query, key, attn_mask)
-
     out = npu_fusion_attention(
         query,
         key,
         value,
         query.size(2),  # num_heads
-        atten_mask=attn_mask,
         input_layout="BSND",
         pse=None,
         scale=1.0 / math.sqrt(query.shape[-1]) if scale is None else scale,
@@ -2261,7 +1942,7 @@ def _flash_attention(
 @_AttentionBackendRegistry.register(
     AttentionBackendName.FLASH_HUB,
     constraints=[_check_device, _check_qkv_dtype_bf16_or_fp16, _check_shape],
-    supports_context_parallel=True,
+    supports_context_parallel=False,
 )
 def _flash_attention_hub(
     query: torch.Tensor,
@@ -2279,35 +1960,17 @@ def _flash_attention_hub(
         raise ValueError("`attn_mask` is not supported for flash-attn 2.")
 
     func = _HUB_KERNELS_REGISTRY[AttentionBackendName.FLASH_HUB].kernel_fn
-    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
+    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
 
     return (out, lse) if return_lse else out
 
@@ -2454,7 +2117,7 @@ def _flash_attention_3(
 @_AttentionBackendRegistry.register(
     AttentionBackendName._FLASH_3_HUB,
     constraints=[_check_device, _check_qkv_dtype_bf16_or_fp16, _check_shape],
-    supports_context_parallel=True,
+    supports_context_parallel=False,
 )
 def _flash_attention_3_hub(
     query: torch.Tensor,
@@ -2469,68 +2132,33 @@ def _flash_attention_3_hub(
     return_attn_probs: bool = False,
     _parallel_config: "ParallelConfig" | None = 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
-    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,
+    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,
     )
-    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
+    # 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
 
 
 @_AttentionBackendRegistry.register(
@@ -3040,17 +2668,16 @@ def _native_npu_attention(
     return_lse: bool = False,
     _parallel_config: "ParallelConfig" | None = None,
 ) -> torch.Tensor:
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not supported for NPU attention")
     if return_lse:
         raise ValueError("NPU attention backend does not support setting `return_lse=True`.")
     if _parallel_config is None:
-        attn_mask = _maybe_modify_attn_mask_npu(query, key, attn_mask)
-
         out = npu_fusion_attention(
             query,
             key,
             value,
             query.size(2),  # num_heads
-            atten_mask=attn_mask,
             input_layout="BSND",
             pse=None,
             scale=1.0 / math.sqrt(query.shape[-1]) if scale is None else scale,
@@ -3065,7 +2692,7 @@ def _native_npu_attention(
             query,
             key,
             value,
-            attn_mask,
+            None,
             dropout_p,
             None,
             scale,
@@ -3162,7 +2789,7 @@ def _sage_attention(
 @_AttentionBackendRegistry.register(
     AttentionBackendName.SAGE_HUB,
     constraints=[_check_device_cuda, _check_qkv_dtype_bf16_or_fp16, _check_shape],
-    supports_context_parallel=True,
+    supports_context_parallel=False,
 )
 def _sage_attention_hub(
     query: torch.Tensor,
@@ -3190,23 +2817,6 @@ 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/transformers/transformer_flux2.py

@@ -424,7 +424,7 @@ class Flux2SingleTransformerBlock(nn.Module):
         self,
         hidden_states: torch.Tensor,
         encoder_hidden_states: torch.Tensor | None,
-        temb_mod: torch.Tensor,
+        temb_mod_params: tuple[torch.Tensor, torch.Tensor, torch.Tensor],
         image_rotary_emb: tuple[torch.Tensor, torch.Tensor] | None = None,
         joint_attention_kwargs: dict[str, Any] | None = None,
         split_hidden_states: bool = False,
@@ -436,7 +436,7 @@ class Flux2SingleTransformerBlock(nn.Module):
             text_seq_len = encoder_hidden_states.shape[1]
             hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
 
-        mod_shift, mod_scale, mod_gate = Flux2Modulation.split(temb_mod, 1)[0]
+        mod_shift, mod_scale, mod_gate = temb_mod_params
 
         norm_hidden_states = self.norm(hidden_states)
         norm_hidden_states = (1 + mod_scale) * norm_hidden_states + mod_shift
@@ -498,18 +498,16 @@ class Flux2TransformerBlock(nn.Module):
         self,
         hidden_states: torch.Tensor,
         encoder_hidden_states: torch.Tensor,
-        temb_mod_img: torch.Tensor,
-        temb_mod_txt: torch.Tensor,
+        temb_mod_params_img: tuple[tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...],
+        temb_mod_params_txt: tuple[tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...],
         image_rotary_emb: tuple[torch.Tensor, torch.Tensor] | None = None,
         joint_attention_kwargs: dict[str, Any] | None = None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         joint_attention_kwargs = joint_attention_kwargs or {}
 
         # Modulation parameters shape: [1, 1, self.dim]
-        (shift_msa, scale_msa, gate_msa), (shift_mlp, scale_mlp, gate_mlp) = Flux2Modulation.split(temb_mod_img, 2)
-        (c_shift_msa, c_scale_msa, c_gate_msa), (c_shift_mlp, c_scale_mlp, c_gate_mlp) = Flux2Modulation.split(
-            temb_mod_txt, 2
-        )
+        (shift_msa, scale_msa, gate_msa), (shift_mlp, scale_mlp, gate_mlp) = temb_mod_params_img
+        (c_shift_msa, c_scale_msa, c_gate_msa), (c_shift_mlp, c_scale_mlp, c_gate_mlp) = temb_mod_params_txt
 
         # Img stream
         norm_hidden_states = self.norm1(hidden_states)
@@ -629,19 +627,15 @@ class Flux2Modulation(nn.Module):
         self.linear = nn.Linear(dim, dim * 3 * self.mod_param_sets, bias=bias)
         self.act_fn = nn.SiLU()
 
-    def forward(self, temb: torch.Tensor) -> torch.Tensor:
+    def forward(self, temb: torch.Tensor) -> tuple[tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...]:
         mod = self.act_fn(temb)
         mod = self.linear(mod)
-        return mod
 
-    @staticmethod
-    # split inside the transformer blocks, to avoid passing tuples into checkpoints https://github.com/huggingface/diffusers/issues/12776
-    def split(mod: torch.Tensor, mod_param_sets: int) -> tuple[tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...]:
         if mod.ndim == 2:
             mod = mod.unsqueeze(1)
-        mod_params = torch.chunk(mod, 3 * mod_param_sets, dim=-1)
+        mod_params = torch.chunk(mod, 3 * self.mod_param_sets, dim=-1)
         # Return tuple of 3-tuples of modulation params shift/scale/gate
-        return tuple(mod_params[3 * i : 3 * (i + 1)] for i in range(mod_param_sets))
+        return tuple(mod_params[3 * i : 3 * (i + 1)] for i in range(self.mod_param_sets))
 
 
 class Flux2Transformer2DModel(
@@ -830,7 +824,7 @@ class Flux2Transformer2DModel(
 
         double_stream_mod_img = self.double_stream_modulation_img(temb)
         double_stream_mod_txt = self.double_stream_modulation_txt(temb)
-        single_stream_mod = self.single_stream_modulation(temb)
+        single_stream_mod = self.single_stream_modulation(temb)[0]
 
         # 2. Input projection for image (hidden_states) and conditioning text (encoder_hidden_states)
         hidden_states = self.x_embedder(hidden_states)
@@ -867,8 +861,8 @@ class Flux2Transformer2DModel(
                 encoder_hidden_states, hidden_states = block(
                     hidden_states=hidden_states,
                     encoder_hidden_states=encoder_hidden_states,
-                    temb_mod_img=double_stream_mod_img,
-                    temb_mod_txt=double_stream_mod_txt,
+                    temb_mod_params_img=double_stream_mod_img,
+                    temb_mod_params_txt=double_stream_mod_txt,
                     image_rotary_emb=concat_rotary_emb,
                     joint_attention_kwargs=joint_attention_kwargs,
                 )
@@ -890,7 +884,7 @@ class Flux2Transformer2DModel(
                 hidden_states = block(
                     hidden_states=hidden_states,
                     encoder_hidden_states=None,
-                    temb_mod=single_stream_mod,
+                    temb_mod_params=single_stream_mod,
                     image_rotary_emb=concat_rotary_emb,
                     joint_attention_kwargs=joint_attention_kwargs,
                 )

src/diffusers/models/transformers/transformer_qwenimage.py

@@ -164,11 +164,7 @@ def compute_text_seq_len_from_mask(
     position_ids = torch.arange(text_seq_len, device=encoder_hidden_states.device, dtype=torch.long)
     active_positions = torch.where(encoder_hidden_states_mask, position_ids, position_ids.new_zeros(()))
     has_active = encoder_hidden_states_mask.any(dim=1)
-    per_sample_len = torch.where(
-        has_active,
-        active_positions.max(dim=1).values + 1,
-        torch.as_tensor(text_seq_len, device=encoder_hidden_states.device),
-    )
+    per_sample_len = torch.where(has_active, active_positions.max(dim=1).values + 1, torch.as_tensor(text_seq_len))
     return text_seq_len, per_sample_len, encoder_hidden_states_mask
 
 

src/diffusers/pipelines/prx/pipeline_prx.py

@@ -18,6 +18,7 @@ import re
 import urllib.parse as ul
 from typing import Callable
 
+import ftfy
 import torch
 from transformers import (
     AutoTokenizer,
@@ -33,13 +34,13 @@ from diffusers.models.transformers.transformer_prx import PRXTransformer2DModel
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline
 from diffusers.pipelines.prx.pipeline_output import PRXPipelineOutput
 from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
-from diffusers.utils import is_ftfy_available, logging, replace_example_docstring
+from diffusers.utils import (
+    logging,
+    replace_example_docstring,
+)
 from diffusers.utils.torch_utils import randn_tensor
 
 
-if is_ftfy_available():
-    import ftfy
-
 DEFAULT_RESOLUTION = 512
 
 ASPECT_RATIO_256_BIN = {

src/diffusers/schedulers/scheduling_flow_match_lcm.py

@@ -14,7 +14,6 @@
 
 import math
 from dataclasses import dataclass
-from typing import Literal
 
 import numpy as np
 import torch
@@ -42,7 +41,7 @@ class FlowMatchLCMSchedulerOutput(BaseOutput):
             denoising loop.
     """
 
-    prev_sample: torch.Tensor
+    prev_sample: torch.FloatTensor
 
 
 class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
@@ -80,11 +79,11 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         use_beta_sigmas (`bool`, defaults to False):
             Whether to use beta sigmas for step sizes in the noise schedule during sampling.
         time_shift_type (`str`, defaults to "exponential"):
-            The type of dynamic resolution-dependent timestep shifting to apply.
-        scale_factors (`list[float]`, *optional*, defaults to `None`):
+            The type of dynamic resolution-dependent timestep shifting to apply. Either "exponential" or "linear".
+        scale_factors ('list', defaults to None)
             It defines how to scale the latents at which predictions are made.
-        upscale_mode (`str`, *optional*, defaults to "bicubic"):
-            Upscaling method, applied if scale-wise generation is considered.
+        upscale_mode ('str', defaults to 'bicubic')
+            Upscaling method, applied if scale-wise generation is considered
     """
 
     _compatibles = []
@@ -102,33 +101,16 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         max_image_seq_len: int = 4096,
         invert_sigmas: bool = False,
         shift_terminal: float | None = None,
-        use_karras_sigmas: bool | None = False,
-        use_exponential_sigmas: bool | None = False,
-        use_beta_sigmas: bool | None = False,
-        time_shift_type: Literal["exponential", "linear"] = "exponential",
+        use_karras_sigmas: bool = False,
+        use_exponential_sigmas: bool = False,
+        use_beta_sigmas: bool = False,
+        time_shift_type: str = "exponential",
         scale_factors: list[float] | None = None,
-        upscale_mode: Literal[
-            "nearest",
-            "linear",
-            "bilinear",
-            "bicubic",
-            "trilinear",
-            "area",
-            "nearest-exact",
-        ] = "bicubic",
+        upscale_mode: str = "bicubic",
     ):
         if self.config.use_beta_sigmas and not is_scipy_available():
             raise ImportError("Make sure to install scipy if you want to use beta sigmas.")
-        if (
-            sum(
-                [
-                    self.config.use_beta_sigmas,
-                    self.config.use_exponential_sigmas,
-                    self.config.use_karras_sigmas,
-                ]
-            )
-            > 1
-        ):
+        if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1:
             raise ValueError(
                 "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used."
             )
@@ -180,7 +162,7 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         return self._begin_index
 
     # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
-    def set_begin_index(self, begin_index: int = 0) -> None:
+    def set_begin_index(self, begin_index: int = 0):
         """
         Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
 
@@ -190,18 +172,18 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         """
         self._begin_index = begin_index
 
-    def set_shift(self, shift: float) -> None:
+    def set_shift(self, shift: float):
         self._shift = shift
 
-    def set_scale_factors(self, scale_factors: list[float], upscale_mode: str) -> None:
+    def set_scale_factors(self, scale_factors: list, upscale_mode):
         """
         Sets scale factors for a scale-wise generation regime.
 
         Args:
-            scale_factors (`list[float]`):
-                The scale factors for each step.
+            scale_factors (`list`):
+                The scale factors for each step
             upscale_mode (`str`):
-                Upscaling method.
+                Upscaling method
         """
         self._scale_factors = scale_factors
         self._upscale_mode = upscale_mode
@@ -256,18 +238,16 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
 
         return sample
 
-    def _sigma_to_t(self, sigma: float | torch.FloatTensor) -> float | torch.FloatTensor:
+    def _sigma_to_t(self, sigma):
         return sigma * self.config.num_train_timesteps
 
-    def time_shift(
-        self, mu: float, sigma: float, t: float | np.ndarray | torch.Tensor
-    ) -> float | np.ndarray | torch.Tensor:
+    def time_shift(self, mu: float, sigma: float, t: torch.Tensor):
         if self.config.time_shift_type == "exponential":
             return self._time_shift_exponential(mu, sigma, t)
         elif self.config.time_shift_type == "linear":
             return self._time_shift_linear(mu, sigma, t)
 
-    def stretch_shift_to_terminal(self, t: np.ndarray | torch.Tensor) -> np.ndarray | torch.Tensor:
+    def stretch_shift_to_terminal(self, t: torch.Tensor) -> torch.Tensor:
         r"""
         Stretches and shifts the timestep schedule to ensure it terminates at the configured `shift_terminal` config
         value.
@@ -276,13 +256,12 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         https://github.com/Lightricks/LTX-Video/blob/a01a171f8fe3d99dce2728d60a73fecf4d4238ae/ltx_video/schedulers/rf.py#L51
 
         Args:
-            t (`torch.Tensor` or `np.ndarray`):
-                A tensor or numpy array of timesteps to be stretched and shifted.
+            t (`torch.Tensor`):
+                A tensor of timesteps to be stretched and shifted.
 
         Returns:
-            `torch.Tensor` or `np.ndarray`:
-                A tensor or numpy array of adjusted timesteps such that the final value equals
-                `self.config.shift_terminal`.
+            `torch.Tensor`:
+                A tensor of adjusted timesteps such that the final value equals `self.config.shift_terminal`.
         """
         one_minus_z = 1 - t
         scale_factor = one_minus_z[-1] / (1 - self.config.shift_terminal)
@@ -291,12 +270,12 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
 
     def set_timesteps(
         self,
-        num_inference_steps: int | None = None,
-        device: str | torch.device | None = None,
+        num_inference_steps: int = None,
+        device: str | torch.device = None,
         sigmas: list[float] | None = None,
-        mu: float | None = None,
+        mu: float = None,
         timesteps: list[float] | None = None,
-    ) -> None:
+    ):
         """
         Sets the discrete timesteps used for the diffusion chain (to be run before inference).
 
@@ -338,45 +317,43 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         is_timesteps_provided = timesteps is not None
 
         if is_timesteps_provided:
-            timesteps = np.array(timesteps).astype(np.float32)  # type: ignore
+            timesteps = np.array(timesteps).astype(np.float32)
 
         if sigmas is None:
             if timesteps is None:
-                timesteps = np.linspace(  # type: ignore
-                    self._sigma_to_t(self.sigma_max),
-                    self._sigma_to_t(self.sigma_min),
-                    num_inference_steps,
+                timesteps = np.linspace(
+                    self._sigma_to_t(self.sigma_max), self._sigma_to_t(self.sigma_min), num_inference_steps
                 )
-            sigmas = timesteps / self.config.num_train_timesteps  # type: ignore
+            sigmas = timesteps / self.config.num_train_timesteps
         else:
-            sigmas = np.array(sigmas).astype(np.float32)  # type: ignore
+            sigmas = np.array(sigmas).astype(np.float32)
             num_inference_steps = len(sigmas)
 
         # 2. Perform timestep shifting. Either no shifting is applied, or resolution-dependent shifting of
         #    "exponential" or "linear" type is applied
         if self.config.use_dynamic_shifting:
-            sigmas = self.time_shift(mu, 1.0, sigmas)  # type: ignore
+            sigmas = self.time_shift(mu, 1.0, sigmas)
         else:
-            sigmas = self.shift * sigmas / (1 + (self.shift - 1) * sigmas)  # type: ignore
+            sigmas = self.shift * sigmas / (1 + (self.shift - 1) * sigmas)
 
         # 3. If required, stretch the sigmas schedule to terminate at the configured `shift_terminal` value
         if self.config.shift_terminal:
-            sigmas = self.stretch_shift_to_terminal(sigmas)  # type: ignore
+            sigmas = self.stretch_shift_to_terminal(sigmas)
 
         # 4. If required, convert sigmas to one of karras, exponential, or beta sigma schedules
         if self.config.use_karras_sigmas:
-            sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)  # type: ignore
+            sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
         elif self.config.use_exponential_sigmas:
-            sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps)  # type: ignore
+            sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
         elif self.config.use_beta_sigmas:
-            sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)  # type: ignore
+            sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
 
         # 5. Convert sigmas and timesteps to tensors and move to specified device
-        sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device)  # type: ignore
+        sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device)
         if not is_timesteps_provided:
-            timesteps = sigmas * self.config.num_train_timesteps  # type: ignore
+            timesteps = sigmas * self.config.num_train_timesteps
         else:
-            timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32, device=device)  # type: ignore
+            timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32, device=device)
 
         # 6. Append the terminal sigma value.
         #    If a model requires inverted sigma schedule for denoising but timesteps without inversion, the
@@ -393,11 +370,7 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         self._step_index = None
         self._begin_index = None
 
-    def index_for_timestep(
-        self,
-        timestep: float | torch.Tensor,
-        schedule_timesteps: torch.Tensor | None = None,
-    ) -> int:
+    def index_for_timestep(self, timestep, schedule_timesteps=None):
         if schedule_timesteps is None:
             schedule_timesteps = self.timesteps
 
@@ -409,9 +382,9 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         # case we start in the middle of the denoising schedule (e.g. for image-to-image)
         pos = 1 if len(indices) > 1 else 0
 
-        return int(indices[pos].item())
+        return indices[pos].item()
 
-    def _init_step_index(self, timestep: float | torch.Tensor) -> None:
+    def _init_step_index(self, timestep):
         if self.begin_index is None:
             if isinstance(timestep, torch.Tensor):
                 timestep = timestep.to(self.timesteps.device)
@@ -486,12 +459,7 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
                 size = [round(self._scale_factors[self._step_index] * size) for size in self._init_size]
                 x0_pred = torch.nn.functional.interpolate(x0_pred, size=size, mode=self._upscale_mode)
 
-        noise = randn_tensor(
-            x0_pred.shape,
-            generator=generator,
-            device=x0_pred.device,
-            dtype=x0_pred.dtype,
-        )
+        noise = randn_tensor(x0_pred.shape, generator=generator, device=x0_pred.device, dtype=x0_pred.dtype)
         prev_sample = (1 - sigma_next) * x0_pred + sigma_next * noise
 
         # upon completion increase step index by one
@@ -505,7 +473,7 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         return FlowMatchLCMSchedulerOutput(prev_sample=prev_sample)
 
     # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
-    def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor:
+    def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor:
         """
         Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative
         Models](https://huggingface.co/papers/2206.00364).
@@ -626,15 +594,11 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         )
         return sigmas
 
-    def _time_shift_exponential(
-        self, mu: float, sigma: float, t: float | np.ndarray | torch.Tensor
-    ) -> float | np.ndarray | torch.Tensor:
+    def _time_shift_exponential(self, mu, sigma, t):
         return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
 
-    def _time_shift_linear(
-        self, mu: float, sigma: float, t: float | np.ndarray | torch.Tensor
-    ) -> float | np.ndarray | torch.Tensor:
+    def _time_shift_linear(self, mu, sigma, t):
         return mu / (mu + (1 / t - 1) ** sigma)
 
-    def __len__(self) -> int:
+    def __len__(self):
         return self.config.num_train_timesteps

tests/pipelines/cogvideo/test_cogvideox.py

@@ -18,7 +18,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKLCogVideoX, CogVideoXPipeline, CogVideoXTransformer3DModel, DDIMScheduler
 
@@ -117,7 +117,9 @@ class CogVideoXPipelineFastTests(
 
         torch.manual_seed(0)
         scheduler = DDIMScheduler()
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config.tie_word_embeddings = False
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {

tests/pipelines/controlnet_flux/test_controlnet_flux.py

@@ -19,7 +19,7 @@ import unittest
 import numpy as np
 import torch
 from huggingface_hub import hf_hub_download
-from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast
+from transformers import AutoConfig, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast
 
 from diffusers import (
     AutoencoderKL,
@@ -97,7 +97,9 @@ class FluxControlNetPipelineFastTests(unittest.TestCase, PipelineTesterMixin, Fl
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config.tie_word_embeddings = False
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = T5TokenizerFast.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/controlnet_sd3/test_controlnet_sd3.py

@@ -18,7 +18,14 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel
+from transformers import (
+    AutoConfig,
+    AutoTokenizer,
+    CLIPTextConfig,
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    T5EncoderModel,
+)
 
 from diffusers import (
     AutoencoderKL,
@@ -117,7 +124,9 @@ class StableDiffusion3ControlNetPipelineFastTests(unittest.TestCase, PipelineTes
         text_encoder_2 = CLIPTextModelWithProjection(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_3 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config.tie_word_embeddings = False
+        text_encoder_3 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

tests/pipelines/flux/test_pipeline_flux_control.py

@@ -3,7 +3,7 @@ import unittest
 import numpy as np
 import torch
 from PIL import Image
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, FluxControlPipeline, FluxTransformer2DModel
 
@@ -53,7 +53,9 @@ class FluxControlPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config.tie_word_embeddings = False
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/flux/test_pipeline_flux_control_img2img.py

@@ -3,7 +3,7 @@ import unittest
 import numpy as np
 import torch
 from PIL import Image
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKL,
@@ -57,7 +57,9 @@ class FluxControlImg2ImgPipelineFastTests(unittest.TestCase, PipelineTesterMixin
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config.tie_word_embeddings = False
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/flux/test_pipeline_flux_control_inpaint.py

@@ -3,7 +3,7 @@ import unittest
 import numpy as np
 import torch
 from PIL import Image
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKL,
@@ -58,7 +58,9 @@ class FluxControlInpaintPipelineFastTests(unittest.TestCase, PipelineTesterMixin
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config.tie_word_embeddings = False
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/flux/test_pipeline_flux_fill.py

@@ -3,7 +3,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, FluxFillPipeline, FluxTransformer2DModel
 
@@ -58,7 +58,9 @@ class FluxFillPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config.tie_word_embeddings = False
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/flux/test_pipeline_flux_img2img.py

@@ -3,7 +3,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, FluxImg2ImgPipeline, FluxTransformer2DModel
 
@@ -55,7 +55,9 @@ class FluxImg2ImgPipelineFastTests(unittest.TestCase, PipelineTesterMixin, FluxI
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config.tie_word_embeddings = False
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/flux/test_pipeline_flux_inpaint.py

@@ -3,7 +3,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, FluxInpaintPipeline, FluxTransformer2DModel
 
@@ -55,7 +55,9 @@ class FluxInpaintPipelineFastTests(unittest.TestCase, PipelineTesterMixin, FluxI
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config.tie_word_embeddings = False
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/flux/test_pipeline_flux_kontext.py

@@ -3,7 +3,7 @@ import unittest
 import numpy as np
 import PIL.Image
 import torch
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKL,
@@ -79,7 +79,9 @@ class FluxKontextPipelineFastTests(
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config.tie_word_embeddings = False
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/flux/test_pipeline_flux_kontext_inpaint.py

@@ -3,7 +3,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKL,
@@ -79,7 +79,9 @@ class FluxKontextInpaintPipelineFastTests(
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config.tie_word_embeddings = False
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/hidream_image/test_pipeline_hidream.py

@@ -18,6 +18,7 @@ import unittest
 import numpy as np
 import torch
 from transformers import (
+    AutoConfig,
     AutoTokenizer,
     CLIPTextConfig,
     CLIPTextModelWithProjection,
@@ -94,7 +95,9 @@ class HiDreamImagePipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         text_encoder_2 = CLIPTextModelWithProjection(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_3 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config.tie_word_embeddings = False
+        text_encoder_3 = T5EncoderModel(config)
 
         torch.manual_seed(0)
         text_encoder_4 = LlamaForCausalLM.from_pretrained("hf-internal-testing/tiny-random-LlamaForCausalLM")

tests/pipelines/hunyuandit/test_hunyuan_dit.py

@@ -19,7 +19,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, BertModel, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, BertModel, T5EncoderModel
 
 from diffusers import AutoencoderKL, DDPMScheduler, HunyuanDiT2DModel, HunyuanDiTPipeline
 
@@ -74,7 +74,10 @@ class HunyuanDiTPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         scheduler = DDPMScheduler()
         text_encoder = BertModel.from_pretrained("hf-internal-testing/tiny-random-BertModel")
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-BertModel")
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        torch.manual_seed(0)
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config.tie_word_embeddings = False
+        text_encoder_2 = T5EncoderModel(config)
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {

tests/pipelines/ltx/test_ltx.py

@@ -17,7 +17,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKLLTXVideo, FlowMatchEulerDiscreteScheduler, LTXPipeline, LTXVideoTransformer3DModel
 
@@ -88,7 +88,9 @@ class LTXPipelineFastTests(PipelineTesterMixin, FirstBlockCacheTesterMixin, unit
 
         torch.manual_seed(0)
         scheduler = FlowMatchEulerDiscreteScheduler()
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config.tie_word_embeddings = False
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {

tests/pipelines/prx/test_pipeline_prx.py

@@ -1,6 +1,7 @@
 import unittest
 
 import numpy as np
+import pytest
 import torch
 from transformers import AutoTokenizer
 from transformers.models.t5gemma.configuration_t5gemma import T5GemmaConfig, T5GemmaModuleConfig
@@ -10,11 +11,17 @@ from diffusers.models import AutoencoderDC, AutoencoderKL
 from diffusers.models.transformers.transformer_prx import PRXTransformer2DModel
 from diffusers.pipelines.prx.pipeline_prx import PRXPipeline
 from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import is_transformers_version
 
 from ..pipeline_params import TEXT_TO_IMAGE_PARAMS
 from ..test_pipelines_common import PipelineTesterMixin
 
 
+@pytest.mark.xfail(
+    condition=is_transformers_version(">", "4.57.1"),
+    reason="See https://github.com/huggingface/diffusers/pull/12456#issuecomment-3424228544",
+    strict=False,
+)
 class PRXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
     pipeline_class = PRXPipeline
     params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}

tests/pipelines/stable_diffusion_3/test_pipeline_stable_diffusion_3_img2img.py

@@ -4,7 +4,14 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel
+from transformers import (
+    AutoConfig,
+    AutoTokenizer,
+    CLIPTextConfig,
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    T5EncoderModel,
+)
 
 from diffusers import (
     AutoencoderKL,
@@ -73,7 +80,10 @@ class StableDiffusion3Img2ImgPipelineFastTests(PipelineLatentTesterMixin, unitte
         torch.manual_seed(0)
         text_encoder_2 = CLIPTextModelWithProjection(clip_text_encoder_config)
 
-        text_encoder_3 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        torch.manual_seed(0)
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config.tie_word_embeddings = False
+        text_encoder_3 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

tests/pipelines/wan/test_wan_22_image_to_video.py

@@ -18,7 +18,7 @@ import unittest
 import numpy as np
 import torch
 from PIL import Image
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKLWan, UniPCMultistepScheduler, WanImageToVideoPipeline, WanTransformer3DModel
 
@@ -64,7 +64,11 @@ class Wan22ImageToVideoPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
 
         torch.manual_seed(0)
         scheduler = UniPCMultistepScheduler(prediction_type="flow_prediction", use_flow_sigmas=True, flow_shift=3.0)
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        torch.manual_seed(0)
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config.tie_word_embeddings = False
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         torch.manual_seed(0)
@@ -248,7 +252,11 @@ class Wan225BImageToVideoPipelineFastTests(PipelineTesterMixin, unittest.TestCas
 
         torch.manual_seed(0)
         scheduler = UniPCMultistepScheduler(prediction_type="flow_prediction", use_flow_sigmas=True, flow_shift=3.0)
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        torch.manual_seed(0)
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config.tie_word_embeddings = False
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         torch.manual_seed(0)