update

* Guard ftfy import with is_ftfy_available

* Remove xfail for PRX pipeline tests as they appear to work on transformers>4.57.1

* make style and make quality

.github/workflows/pr_modular_tests.yml

@@ -117,7 +117,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality]"
+        uv pip install -e ".[quality,test]"
         #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]"
+        uv pip install -e ".[quality,test]"
         #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]"
+        uv pip install -e ".[quality,test]"
 
     - name: Environment
       run: |
@@ -242,7 +242,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality]"
+        uv pip install -e ".[quality,test]"
         # 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

@@ -199,11 +199,6 @@ 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

@@ -126,11 +126,6 @@ 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]
-> [Caching](../../optimization/cache) may also speed up inference by storing and reusing intermediate outputs.
+> See the [Caching](../../optimization/cache) guide to speed up inference by storing and reusing intermediate outputs.
 
 ## LoRA for faster inference
 
@@ -190,6 +190,12 @@ 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

@@ -101,6 +101,7 @@ _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",
@@ -221,12 +222,14 @@ 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",
@@ -235,6 +238,7 @@ extras["test"] = deps_list(
     "sentencepiece",
     "scipy",
     "tiktoken",
+    "torchsde",
     "torchvision",
     "transformers",
     "phonemizer",

src/diffusers/dependency_versions_table.py

@@ -8,6 +8,7 @@ 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

@@ -1117,6 +1117,26 @@ 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,
@@ -1134,11 +1154,14 @@ 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,
@@ -2668,16 +2691,17 @@ 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,
@@ -2692,7 +2716,7 @@ def _native_npu_attention(
             query,
             key,
             value,
-            None,
+            attn_mask,
             dropout_p,
             None,
             scale,

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_params: tuple[torch.Tensor, torch.Tensor, torch.Tensor],
+        temb_mod: 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 = temb_mod_params
+        mod_shift, mod_scale, mod_gate = Flux2Modulation.split(temb_mod, 1)[0]
 
         norm_hidden_states = self.norm(hidden_states)
         norm_hidden_states = (1 + mod_scale) * norm_hidden_states + mod_shift
@@ -498,16 +498,18 @@ class Flux2TransformerBlock(nn.Module):
         self,
         hidden_states: torch.Tensor,
         encoder_hidden_states: 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], ...],
+        temb_mod_img: torch.Tensor,
+        temb_mod_txt: 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) = 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
+        (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
+        )
 
         # Img stream
         norm_hidden_states = self.norm1(hidden_states)
@@ -627,15 +629,19 @@ 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) -> tuple[tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...]:
+    def forward(self, temb: 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 * self.mod_param_sets, dim=-1)
+        mod_params = torch.chunk(mod, 3 * 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(self.mod_param_sets))
+        return tuple(mod_params[3 * i : 3 * (i + 1)] for i in range(mod_param_sets))
 
 
 class Flux2Transformer2DModel(
@@ -824,7 +830,7 @@ 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)[0]
+        single_stream_mod = self.single_stream_modulation(temb)
 
         # 2. Input projection for image (hidden_states) and conditioning text (encoder_hidden_states)
         hidden_states = self.x_embedder(hidden_states)
@@ -861,8 +867,8 @@ class Flux2Transformer2DModel(
                 encoder_hidden_states, hidden_states = block(
                     hidden_states=hidden_states,
                     encoder_hidden_states=encoder_hidden_states,
-                    temb_mod_params_img=double_stream_mod_img,
-                    temb_mod_params_txt=double_stream_mod_txt,
+                    temb_mod_img=double_stream_mod_img,
+                    temb_mod_txt=double_stream_mod_txt,
                     image_rotary_emb=concat_rotary_emb,
                     joint_attention_kwargs=joint_attention_kwargs,
                 )
@@ -884,7 +890,7 @@ class Flux2Transformer2DModel(
                 hidden_states = block(
                     hidden_states=hidden_states,
                     encoder_hidden_states=None,
-                    temb_mod_params=single_stream_mod,
+                    temb_mod=single_stream_mod,
                     image_rotary_emb=concat_rotary_emb,
                     joint_attention_kwargs=joint_attention_kwargs,
                 )

src/diffusers/models/transformers/transformer_qwenimage.py

@@ -164,7 +164,11 @@ 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))
+    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),
+    )
     return text_seq_len, per_sample_len, encoder_hidden_states_mask
 
 

src/diffusers/pipelines/pipeline_utils.py

@@ -112,7 +112,7 @@ LIBRARIES = []
 for library in LOADABLE_CLASSES:
     LIBRARIES.append(library)
 
-SUPPORTED_DEVICE_MAP = ["balanced"] + [get_device()]
+SUPPORTED_DEVICE_MAP = ["balanced"] + [get_device(), "cpu"]
 
 logger = logging.get_logger(__name__)
 
@@ -468,8 +468,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         pipeline_is_sequentially_offloaded = any(
             module_is_sequentially_offloaded(module) for _, module in self.components.items()
         )
-
-        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
+        is_pipeline_device_mapped = self._is_pipeline_device_mapped()
         if is_pipeline_device_mapped:
             raise ValueError(
                 "It seems like you have activated a device mapping strategy on the pipeline which doesn't allow explicit device placement using `to()`. You can call `reset_device_map()` to remove the existing device map from the pipeline."
@@ -1188,7 +1187,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         """
         self._maybe_raise_error_if_group_offload_active(raise_error=True)
 
-        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
+        is_pipeline_device_mapped = self._is_pipeline_device_mapped()
         if is_pipeline_device_mapped:
             raise ValueError(
                 "It seems like you have activated a device mapping strategy on the pipeline so calling `enable_model_cpu_offload() isn't allowed. You can call `reset_device_map()` first and then call `enable_model_cpu_offload()`."
@@ -1312,7 +1311,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
         self.remove_all_hooks()
 
-        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
+        is_pipeline_device_mapped = self._is_pipeline_device_mapped()
         if is_pipeline_device_mapped:
             raise ValueError(
                 "It seems like you have activated a device mapping strategy on the pipeline so calling `enable_sequential_cpu_offload() isn't allowed. You can call `reset_device_map()` first and then call `enable_sequential_cpu_offload()`."
@@ -2228,6 +2227,21 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 return True
         return False
 
+    def _is_pipeline_device_mapped(self):
+        # We support passing `device_map="cuda"`, for example. This is helpful, in case
+        # users want to pass `device_map="cpu"` when initializing a pipeline. This explicit declaration is desirable
+        # in limited VRAM environments because quantized models often initialize directly on the accelerator.
+        device_map = self.hf_device_map
+        is_device_type_map = False
+        if isinstance(device_map, str):
+            try:
+                torch.device(device_map)
+                is_device_type_map = True
+            except RuntimeError:
+                pass
+
+        return not is_device_type_map and isinstance(device_map, dict) and len(device_map) > 1
+
 
 class StableDiffusionMixin:
     r"""

src/diffusers/pipelines/prx/pipeline_prx.py

@@ -18,7 +18,6 @@ import re
 import urllib.parse as ul
 from typing import Callable
 
-import ftfy
 import torch
 from transformers import (
     AutoTokenizer,
@@ -34,13 +33,13 @@ 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 (
-    logging,
-    replace_example_docstring,
-)
+from diffusers.utils import is_ftfy_available, 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

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

tests/models/testing_utils/compile.py

@@ -81,7 +81,7 @@ class TorchCompileTesterMixin:
             _ = model(**inputs_dict)
 
     @torch.no_grad()
-    def test_torch_compile_repeated_blocks(self):
+    def test_torch_compile_repeated_blocks(self, recompile_limit=1):
         if self.model_class._repeated_blocks is None:
             pytest.skip("Skipping test as the model class doesn't have `_repeated_blocks` set.")
 
@@ -92,7 +92,6 @@ class TorchCompileTesterMixin:
         model.eval()
         model.compile_repeated_blocks(fullgraph=True)
 
-        recompile_limit = 1
         if self.model_class.__name__ == "UNet2DConditionModel":
             recompile_limit = 2
 

tests/models/testing_utils/lora.py

@@ -375,7 +375,7 @@ class LoraHotSwappingForModelTesterMixin:
             # additionally check if dynamic compilation works.
             if different_shapes is not None:
                 for height, width in different_shapes:
-                    new_inputs_dict = self.prepare_dummy_input(height=height, width=width)
+                    new_inputs_dict = self.get_dummy_inputs(height=height, width=width)
                     _ = model(**new_inputs_dict)
             else:
                 output0_after = model(**inputs_dict)["sample"]
@@ -390,7 +390,7 @@ class LoraHotSwappingForModelTesterMixin:
         with torch.inference_mode():
             if different_shapes is not None:
                 for height, width in different_shapes:
-                    new_inputs_dict = self.prepare_dummy_input(height=height, width=width)
+                    new_inputs_dict = self.get_dummy_inputs(height=height, width=width)
                     _ = model(**new_inputs_dict)
             else:
                 output1_after = model(**inputs_dict)["sample"]

tests/models/testing_utils/quantization.py

@@ -628,6 +628,21 @@ class BitsAndBytesTesterMixin(BitsAndBytesConfigMixin, QuantizationTesterMixin):
         """Test that quantized models can be used for training with adapters."""
         self._test_quantization_training(BitsAndBytesConfigMixin.BNB_CONFIGS["4bit_nf4"])
 
+    @pytest.mark.parametrize(
+        "config_name",
+        list(BitsAndBytesConfigMixin.BNB_CONFIGS.keys()),
+        ids=list(BitsAndBytesConfigMixin.BNB_CONFIGS.keys()),
+    )
+    def test_cpu_device_map(self, config_name):
+        config_kwargs = BitsAndBytesConfigMixin.BNB_CONFIGS[config_name]
+        model_quantized = self._create_quantized_model(config_kwargs, device_map="cpu")
+
+        assert hasattr(model_quantized, "hf_device_map"), "Model should have hf_device_map attribute"
+        assert model_quantized.hf_device_map is not None, "hf_device_map should not be None"
+        assert model_quantized.device == torch.device("cpu"), (
+            f"Model should be on CPU, but is on {model_quantized.device}"
+        )
+
 
 @is_quantization
 @is_quanto

tests/models/transformers/test_models_transformer_wan_vace.py

@@ -147,22 +147,7 @@ class TestWanVACETransformer3DCompile(WanVACETransformer3DTesterConfig, TorchCom
     def test_torch_compile_repeated_blocks(self):
         # WanVACE has two block types (WanTransformerBlock and WanVACETransformerBlock),
         # so we need recompile_limit=2 instead of the default 1.
-        import torch._dynamo
-        import torch._inductor.utils
-
-        init_dict = self.get_init_dict()
-        inputs_dict = self.get_dummy_inputs()
-
-        model = self.model_class(**init_dict).to(torch_device)
-        model.eval()
-        model.compile_repeated_blocks(fullgraph=True)
-
-        with (
-            torch._inductor.utils.fresh_inductor_cache(),
-            torch._dynamo.config.patch(recompile_limit=2),
-        ):
-            _ = model(**inputs_dict)
-            _ = model(**inputs_dict)
+        super().test_torch_compile_repeated_blocks(recompile_limit=2)
 
 
 class TestWanVACETransformer3DBitsAndBytes(WanVACETransformer3DTesterConfig, BitsAndBytesTesterMixin):

tests/pipelines/allegro/test_allegro.py

@@ -158,6 +158,10 @@ class AllegroPipelineFastTests(PipelineTesterMixin, PyramidAttentionBroadcastTes
     def test_save_load_optional_components(self):
         pass
 
+    @unittest.skip("Decoding without tiling is not yet implemented")
+    def test_pipeline_with_accelerator_device_map(self):
+        pass
+
     def test_inference(self):
         device = "cpu"
 

tests/pipelines/kandinsky/test_kandinsky_combined.py

@@ -34,9 +34,7 @@ enable_full_determinism()
 
 class KandinskyPipelineCombinedFastTests(PipelineTesterMixin, unittest.TestCase):
     pipeline_class = KandinskyCombinedPipeline
-    params = [
-        "prompt",
-    ]
+    params = ["prompt"]
     batch_params = ["prompt", "negative_prompt"]
     required_optional_params = [
         "generator",
@@ -148,6 +146,10 @@ class KandinskyPipelineCombinedFastTests(PipelineTesterMixin, unittest.TestCase)
     def test_dict_tuple_outputs_equivalent(self):
         super().test_dict_tuple_outputs_equivalent(expected_max_difference=5e-4)
 
+    @unittest.skip("Test not supported.")
+    def test_pipeline_with_accelerator_device_map(self):
+        pass
+
 
 class KandinskyPipelineImg2ImgCombinedFastTests(PipelineTesterMixin, unittest.TestCase):
     pipeline_class = KandinskyImg2ImgCombinedPipeline
@@ -264,6 +266,10 @@ class KandinskyPipelineImg2ImgCombinedFastTests(PipelineTesterMixin, unittest.Te
     def test_save_load_optional_components(self):
         super().test_save_load_optional_components(expected_max_difference=5e-4)
 
+    @unittest.skip("Test not supported.")
+    def test_pipeline_with_accelerator_device_map(self):
+        pass
+
 
 class KandinskyPipelineInpaintCombinedFastTests(PipelineTesterMixin, unittest.TestCase):
     pipeline_class = KandinskyInpaintCombinedPipeline
@@ -384,3 +390,7 @@ class KandinskyPipelineInpaintCombinedFastTests(PipelineTesterMixin, unittest.Te
 
     def test_save_load_local(self):
         super().test_save_load_local(expected_max_difference=5e-3)
+
+    @unittest.skip("Test not supported.")
+    def test_pipeline_with_accelerator_device_map(self):
+        pass

tests/pipelines/kandinsky2_2/test_kandinsky_combined.py

@@ -36,9 +36,7 @@ enable_full_determinism()
 
 class KandinskyV22PipelineCombinedFastTests(PipelineTesterMixin, unittest.TestCase):
     pipeline_class = KandinskyV22CombinedPipeline
-    params = [
-        "prompt",
-    ]
+    params = ["prompt"]
     batch_params = ["prompt", "negative_prompt"]
     required_optional_params = [
         "generator",
@@ -70,12 +68,7 @@ class KandinskyV22PipelineCombinedFastTests(PipelineTesterMixin, unittest.TestCa
     def get_dummy_inputs(self, device, seed=0):
         prior_dummy = PriorDummies()
         inputs = prior_dummy.get_dummy_inputs(device=device, seed=seed)
-        inputs.update(
-            {
-                "height": 64,
-                "width": 64,
-            }
-        )
+        inputs.update({"height": 64, "width": 64})
         return inputs
 
     def test_kandinsky(self):
@@ -155,12 +148,18 @@ class KandinskyV22PipelineCombinedFastTests(PipelineTesterMixin, unittest.TestCa
     def test_save_load_optional_components(self):
         super().test_save_load_optional_components(expected_max_difference=5e-3)
 
+    @unittest.skip("Test not supported.")
     def test_callback_inputs(self):
         pass
 
+    @unittest.skip("Test not supported.")
     def test_callback_cfg(self):
         pass
 
+    @unittest.skip("Test not supported.")
+    def test_pipeline_with_accelerator_device_map(self):
+        pass
+
 
 class KandinskyV22PipelineImg2ImgCombinedFastTests(PipelineTesterMixin, unittest.TestCase):
     pipeline_class = KandinskyV22Img2ImgCombinedPipeline
@@ -279,12 +278,18 @@ class KandinskyV22PipelineImg2ImgCombinedFastTests(PipelineTesterMixin, unittest
     def save_load_local(self):
         super().test_save_load_local(expected_max_difference=5e-3)
 
+    @unittest.skip("Test not supported.")
     def test_callback_inputs(self):
         pass
 
+    @unittest.skip("Test not supported.")
     def test_callback_cfg(self):
         pass
 
+    @unittest.skip("Test not supported.")
+    def test_pipeline_with_accelerator_device_map(self):
+        pass
+
 
 class KandinskyV22PipelineInpaintCombinedFastTests(PipelineTesterMixin, unittest.TestCase):
     pipeline_class = KandinskyV22InpaintCombinedPipeline
@@ -411,3 +416,7 @@ class KandinskyV22PipelineInpaintCombinedFastTests(PipelineTesterMixin, unittest
 
     def test_callback_cfg(self):
         pass
+
+    @unittest.skip("`device_map` is not yet supported for connected pipelines.")
+    def test_pipeline_with_accelerator_device_map(self):
+        pass

tests/pipelines/kandinsky2_2/test_kandinsky_inpaint.py

@@ -296,6 +296,9 @@ class KandinskyV22InpaintPipelineFastTests(PipelineTesterMixin, unittest.TestCas
         output = pipe(**inputs)[0]
         assert output.abs().sum() == 0
 
+    def test_pipeline_with_accelerator_device_map(self):
+        super().test_pipeline_with_accelerator_device_map(expected_max_difference=5e-3)
+
 
 @slow
 @require_torch_accelerator

tests/pipelines/kandinsky3/test_kandinsky3_img2img.py

@@ -194,6 +194,9 @@ class Kandinsky3Img2ImgPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
     def test_save_load_dduf(self):
         super().test_save_load_dduf(atol=1e-3, rtol=1e-3)
 
+    def test_pipeline_with_accelerator_device_map(self):
+        super().test_pipeline_with_accelerator_device_map(expected_max_difference=5e-3)
+
 
 @slow
 @require_torch_accelerator

tests/pipelines/prx/test_pipeline_prx.py

@@ -1,7 +1,6 @@
 import unittest
 
 import numpy as np
-import pytest
 import torch
 from transformers import AutoTokenizer
 from transformers.models.t5gemma.configuration_t5gemma import T5GemmaConfig, T5GemmaModuleConfig
@@ -11,17 +10,11 @@ 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/test_pipelines_common.py

@@ -2355,7 +2355,6 @@ class PipelineTesterMixin:
                     f"Component '{name}' has dtype {component.dtype} but expected {expected_dtype}",
                 )
 
-    @require_torch_accelerator
     def test_pipeline_with_accelerator_device_map(self, expected_max_difference=1e-4):
         components = self.get_dummy_components()
         pipe = self.pipeline_class(**components)

tests/pipelines/visualcloze/test_pipeline_visualcloze_combined.py

@@ -342,3 +342,7 @@ class VisualClozePipelineFastTests(unittest.TestCase, PipelineTesterMixin):
         self.assertLess(
             max_diff, expected_max_diff, "The output of the fp16 pipeline changed after saving and loading."
         )
+
+    @unittest.skip("Test not supported.")
+    def test_pipeline_with_accelerator_device_map(self):
+        pass

tests/pipelines/visualcloze/test_pipeline_visualcloze_generation.py

@@ -310,3 +310,7 @@ class VisualClozeGenerationPipelineFastTests(unittest.TestCase, PipelineTesterMi
     @unittest.skip("Skipped due to missing layout_prompt. Needs further investigation.")
     def test_encode_prompt_works_in_isolation(self, extra_required_param_value_dict=None, atol=0.0001, rtol=0.0001):
         pass
+
+    @unittest.skip("Needs to be revisited later.")
+    def test_pipeline_with_accelerator_device_map(self, expected_max_difference=0.0001):
+        pass