by default, skip loading the componeneets does not have the repo id

docs: improve docstring scheduling_ddpm_flax.py

docker/diffusers-pytorch-cuda/Dockerfile

@@ -2,7 +2,7 @@ FROM nvidia/cuda:12.1.0-runtime-ubuntu20.04
 LABEL maintainer="Hugging Face"
 LABEL repository="diffusers"
 
-ARG PYTHON_VERSION=3.12
+ARG PYTHON_VERSION=3.11
 ENV DEBIAN_FRONTEND=noninteractive
 
 RUN apt-get -y update \
@@ -32,10 +32,12 @@ RUN uv venv --python ${PYTHON_VERSION} --seed ${VIRTUAL_ENV}
 ENV PATH="$VIRTUAL_ENV/bin:$PATH"
 
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
+# Install torch, torchvision, and torchaudio together to ensure compatibility
 RUN uv pip install --no-cache-dir \
     torch \
     torchvision \
-    torchaudio
+    torchaudio \
+    --index-url https://download.pytorch.org/whl/cu121
 
 RUN uv pip install --no-cache-dir "git+https://github.com/huggingface/diffusers.git@main#egg=diffusers[test]"
 

docker/diffusers-pytorch-xformers-cuda/Dockerfile

@@ -2,7 +2,7 @@ FROM nvidia/cuda:12.1.0-runtime-ubuntu20.04
 LABEL maintainer="Hugging Face"
 LABEL repository="diffusers"
 
-ARG PYTHON_VERSION=3.12
+ARG PYTHON_VERSION=3.11
 ENV DEBIAN_FRONTEND=noninteractive
 
 RUN apt-get -y update \
@@ -32,10 +32,12 @@ RUN uv venv --python ${PYTHON_VERSION} --seed ${VIRTUAL_ENV}
 ENV PATH="$VIRTUAL_ENV/bin:$PATH"
 
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
+# Install torch, torchvision, and torchaudio together to ensure compatibility
 RUN uv pip install --no-cache-dir \
     torch \
     torchvision \
-    torchaudio
+    torchaudio \
+    --index-url https://download.pytorch.org/whl/cu121
 
 RUN uv pip install --no-cache-dir "git+https://github.com/huggingface/diffusers.git@main#egg=diffusers[test]"
 

src/diffusers/models/auto_model.py

@@ -18,7 +18,7 @@ from typing import Optional, Union
 from huggingface_hub.utils import validate_hf_hub_args
 
 from ..configuration_utils import ConfigMixin
-from ..utils import logging
+from ..utils import DIFFUSERS_LOAD_ID_FIELDS, logging
 from ..utils.dynamic_modules_utils import get_class_from_dynamic_module, resolve_trust_remote_code
 
 
@@ -220,4 +220,11 @@ class AutoModel(ConfigMixin):
             raise ValueError(f"AutoModel can't find a model linked to {orig_class_name}.")
 
         kwargs = {**load_config_kwargs, **kwargs}
-        return model_cls.from_pretrained(pretrained_model_or_path, **kwargs)
+        model = model_cls.from_pretrained(pretrained_model_or_path, **kwargs)
+
+        load_id_kwargs = {"pretrained_model_name_or_path": pretrained_model_or_path, **kwargs}
+        parts = [load_id_kwargs.get(field, "null") for field in DIFFUSERS_LOAD_ID_FIELDS]
+        load_id = "|".join("null" if p is None else p for p in parts)
+        model._diffusers_load_id = load_id
+
+        return model

src/diffusers/models/resnet.py

@@ -366,7 +366,12 @@ class ResnetBlock2D(nn.Module):
         hidden_states = self.conv2(hidden_states)
 
         if self.conv_shortcut is not None:
-            input_tensor = self.conv_shortcut(input_tensor.contiguous())
+            # Only use contiguous() during training to avoid DDP gradient stride mismatch warning.
+            # In inference mode (eval or no_grad), skip contiguous() for better performance, especially on CPU.
+            # Issue: https://github.com/huggingface/diffusers/issues/12975
+            if self.training:
+                input_tensor = input_tensor.contiguous()
+            input_tensor = self.conv_shortcut(input_tensor)
 
         output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
 

src/diffusers/models/transformers/transformer_longcat_image.py

@@ -23,7 +23,7 @@ from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
 from ...utils import logging
 from ...utils.torch_utils import maybe_allow_in_graph
-from ..attention import AttentionModuleMixin, FeedForward
+from ..attention import AttentionMixin, AttentionModuleMixin, FeedForward
 from ..attention_dispatch import dispatch_attention_fn
 from ..cache_utils import CacheMixin
 from ..embeddings import TimestepEmbedding, Timesteps, apply_rotary_emb, get_1d_rotary_pos_embed
@@ -400,6 +400,7 @@ class LongCatImageTransformer2DModel(
     PeftAdapterMixin,
     FromOriginalModelMixin,
     CacheMixin,
+    AttentionMixin,
 ):
     """
     The Transformer model introduced in Longcat-Image.

src/diffusers/modular_pipelines/modular_pipeline.py

@@ -2142,6 +2142,8 @@ class ModularPipeline(ConfigMixin, PushToHubMixin):
                 name
                 for name in self._component_specs.keys()
                 if self._component_specs[name].default_creation_method == "from_pretrained"
+                and self._component_specs[name].pretrained_model_name_or_path is not None
+                and getattr(self, name, None) is None
             ]
         elif isinstance(names, str):
             names = [names]

src/diffusers/modular_pipelines/modular_pipeline_utils.py

@@ -15,14 +15,14 @@
 import inspect
 import re
 from collections import OrderedDict
-from dataclasses import dataclass, field, fields
+from dataclasses import dataclass, field
 from typing import Any, Dict, List, Literal, Optional, Type, Union
 
 import torch
 
 from ..configuration_utils import ConfigMixin, FrozenDict
 from ..loaders.single_file_utils import _is_single_file_path_or_url
-from ..utils import is_torch_available, logging
+from ..utils import DIFFUSERS_LOAD_ID_FIELDS, is_torch_available, logging
 
 
 if is_torch_available():
@@ -185,7 +185,7 @@ class ComponentSpec:
         """
         Return the names of all loading‐related fields (i.e. those whose field.metadata["loading"] is True).
         """
-        return [f.name for f in fields(cls) if f.metadata.get("loading", False)]
+        return DIFFUSERS_LOAD_ID_FIELDS.copy()
 
     @property
     def load_id(self) -> str:
@@ -197,7 +197,7 @@ class ComponentSpec:
             return "null"
         parts = [getattr(self, k) for k in self.loading_fields()]
         parts = ["null" if p is None else p for p in parts]
-        return "|".join(p for p in parts if p)
+        return "|".join(parts)
 
     @classmethod
     def decode_load_id(cls, load_id: str) -> Dict[str, Optional[str]]:

src/diffusers/pipelines/pipeline_utils.py

@@ -111,7 +111,7 @@ LIBRARIES = []
 for library in LOADABLE_CLASSES:
     LIBRARIES.append(library)
 
-SUPPORTED_DEVICE_MAP = ["balanced"] + [get_device(), "cpu"]
+SUPPORTED_DEVICE_MAP = ["balanced"] + [get_device()]
 
 logger = logging.get_logger(__name__)
 
@@ -467,7 +467,8 @@ 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._is_pipeline_device_mapped()
+
+        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
         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."
@@ -1186,7 +1187,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         """
         self._maybe_raise_error_if_group_offload_active(raise_error=True)
 
-        is_pipeline_device_mapped = self._is_pipeline_device_mapped()
+        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
         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()`."
@@ -1310,7 +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._is_pipeline_device_mapped()
+        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
         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()`."
@@ -2199,21 +2200,6 @@ 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/schedulers/scheduling_ddim_parallel.py

@@ -101,7 +101,7 @@ def betas_for_alpha_bar(
 
 
 # Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
-def rescale_zero_terminal_snr(betas):
+def rescale_zero_terminal_snr(betas: torch.Tensor) -> torch.Tensor:
     """
     Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1)
 
@@ -266,7 +266,7 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
         """
         return sample
 
-    def _get_variance(self, timestep, prev_timestep=None):
+    def _get_variance(self, timestep: int, prev_timestep: Optional[int] = None) -> torch.Tensor:
         if prev_timestep is None:
             prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps
 
@@ -279,7 +279,7 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
 
         return variance
 
-    def _batch_get_variance(self, t, prev_t):
+    def _batch_get_variance(self, t: torch.Tensor, prev_t: torch.Tensor) -> torch.Tensor:
         alpha_prod_t = self.alphas_cumprod[t]
         alpha_prod_t_prev = self.alphas_cumprod[torch.clip(prev_t, min=0)]
         alpha_prod_t_prev[prev_t < 0] = torch.tensor(1.0)
@@ -335,7 +335,7 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
         return sample
 
     # Copied from diffusers.schedulers.scheduling_ddim.DDIMScheduler.set_timesteps
-    def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+    def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None) -> None:
         """
         Sets the discrete timesteps used for the diffusion chain (to be run before inference).
 
@@ -392,7 +392,7 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
         sample: torch.Tensor,
         eta: float = 0.0,
         use_clipped_model_output: bool = False,
-        generator=None,
+        generator: Optional[torch.Generator] = None,
         variance_noise: Optional[torch.Tensor] = None,
         return_dict: bool = True,
     ) -> Union[DDIMParallelSchedulerOutput, Tuple]:
@@ -406,11 +406,13 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
             sample (`torch.Tensor`):
                 current instance of sample being created by diffusion process.
             eta (`float`): weight of noise for added noise in diffusion step.
-            use_clipped_model_output (`bool`): if `True`, compute "corrected" `model_output` from the clipped
-                predicted original sample. Necessary because predicted original sample is clipped to [-1, 1] when
-                `self.config.clip_sample` is `True`. If no clipping has happened, "corrected" `model_output` would
-                coincide with the one provided as input and `use_clipped_model_output` will have not effect.
-            generator: random number generator.
+            use_clipped_model_output (`bool`, defaults to `False`):
+                If `True`, compute "corrected" `model_output` from the clipped predicted original sample. This
+                correction is necessary because the predicted original sample is clipped to [-1, 1] when
+                `self.config.clip_sample` is `True`. If no clipping occurred, the "corrected" `model_output` matches
+                the input and `use_clipped_model_output` has no effect.
+            generator (`torch.Generator`, *optional*):
+                Random number generator.
             variance_noise (`torch.Tensor`): instead of generating noise for the variance using `generator`, we
                 can directly provide the noise for the variance itself. This is useful for methods such as
                 CycleDiffusion. (https://huggingface.co/papers/2210.05559)
@@ -496,7 +498,10 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
 
             if variance_noise is None:
                 variance_noise = randn_tensor(
-                    model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype
+                    model_output.shape,
+                    generator=generator,
+                    device=model_output.device,
+                    dtype=model_output.dtype,
                 )
             variance = std_dev_t * variance_noise
 
@@ -513,7 +518,7 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
     def batch_step_no_noise(
         self,
         model_output: torch.Tensor,
-        timesteps: List[int],
+        timesteps: torch.Tensor,
         sample: torch.Tensor,
         eta: float = 0.0,
         use_clipped_model_output: bool = False,
@@ -528,7 +533,7 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
 
         Args:
             model_output (`torch.Tensor`): direct output from learned diffusion model.
-            timesteps (`List[int]`):
+            timesteps (`torch.Tensor`):
                 current discrete timesteps in the diffusion chain. This is now a list of integers.
             sample (`torch.Tensor`):
                 current instance of sample being created by diffusion process.
@@ -696,5 +701,5 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
         velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
         return velocity
 
-    def __len__(self):
+    def __len__(self) -> int:
         return self.config.num_train_timesteps

src/diffusers/schedulers/scheduling_ddpm_flax.py

@@ -22,6 +22,7 @@ import jax
 import jax.numpy as jnp
 
 from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import logging
 from .scheduling_utils_flax import (
     CommonSchedulerState,
     FlaxKarrasDiffusionSchedulers,
@@ -32,6 +33,9 @@ from .scheduling_utils_flax import (
 )
 
 
+logger = logging.get_logger(__name__)
+
+
 @flax.struct.dataclass
 class DDPMSchedulerState:
     common: CommonSchedulerState
@@ -42,7 +46,12 @@ class DDPMSchedulerState:
     num_inference_steps: Optional[int] = None
 
     @classmethod
-    def create(cls, common: CommonSchedulerState, init_noise_sigma: jnp.ndarray, timesteps: jnp.ndarray):
+    def create(
+        cls,
+        common: CommonSchedulerState,
+        init_noise_sigma: jnp.ndarray,
+        timesteps: jnp.ndarray,
+    ):
         return cls(common=common, init_noise_sigma=init_noise_sigma, timesteps=timesteps)
 
 
@@ -105,6 +114,10 @@ class FlaxDDPMScheduler(FlaxSchedulerMixin, ConfigMixin):
         prediction_type: str = "epsilon",
         dtype: jnp.dtype = jnp.float32,
     ):
+        logger.warning(
+            "Flax classes are deprecated and will be removed in Diffusers v1.0.0. We "
+            "recommend migrating to PyTorch classes or pinning your version of Diffusers."
+        )
         self.dtype = dtype
 
     def create_state(self, common: Optional[CommonSchedulerState] = None) -> DDPMSchedulerState:
@@ -123,7 +136,10 @@ class FlaxDDPMScheduler(FlaxSchedulerMixin, ConfigMixin):
         )
 
     def scale_model_input(
-        self, state: DDPMSchedulerState, sample: jnp.ndarray, timestep: Optional[int] = None
+        self,
+        state: DDPMSchedulerState,
+        sample: jnp.ndarray,
+        timestep: Optional[int] = None,
     ) -> jnp.ndarray:
         """
         Args:

src/diffusers/utils/__init__.py

@@ -23,6 +23,7 @@ from .constants import (
     DEFAULT_HF_PARALLEL_LOADING_WORKERS,
     DEPRECATED_REVISION_ARGS,
     DIFFUSERS_DYNAMIC_MODULE_NAME,
+    DIFFUSERS_LOAD_ID_FIELDS,
     FLAX_WEIGHTS_NAME,
     GGUF_FILE_EXTENSION,
     HF_ENABLE_PARALLEL_LOADING,

src/diffusers/utils/constants.py

@@ -73,3 +73,11 @@ DECODE_ENDPOINT_HUNYUAN_VIDEO = "https://o7ywnmrahorts457.us-east-1.aws.endpoint
 ENCODE_ENDPOINT_SD_V1 = "https://qc6479g0aac6qwy9.us-east-1.aws.endpoints.huggingface.cloud/"
 ENCODE_ENDPOINT_SD_XL = "https://xjqqhmyn62rog84g.us-east-1.aws.endpoints.huggingface.cloud/"
 ENCODE_ENDPOINT_FLUX = "https://ptccx55jz97f9zgo.us-east-1.aws.endpoints.huggingface.cloud/"
+
+
+DIFFUSERS_LOAD_ID_FIELDS = [
+    "pretrained_model_name_or_path",
+    "subfolder",
+    "variant",
+    "revision",
+]

tests/pipelines/kandinsky2_2/test_kandinsky_inpaint.py

@@ -248,6 +248,9 @@ class KandinskyV22InpaintPipelineFastTests(PipelineTesterMixin, unittest.TestCas
     def test_float16_inference(self):
         super().test_float16_inference(expected_max_diff=5e-1)
 
+    def test_save_load_dduf(self):
+        super().test_save_load_dduf(atol=1e-3, rtol=1e-3)
+
     @is_flaky()
     def test_model_cpu_offload_forward_pass(self):
         super().test_inference_batch_single_identical(expected_max_diff=8e-4)

tests/pipelines/kandinsky3/test_kandinsky3_img2img.py

@@ -191,6 +191,9 @@ class Kandinsky3Img2ImgPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
     def test_inference_batch_single_identical(self):
         super().test_inference_batch_single_identical(expected_max_diff=1e-2)
 
+    def test_save_load_dduf(self):
+        super().test_save_load_dduf(atol=1e-3, rtol=1e-3)
+
 
 @slow
 @require_torch_accelerator