add custom mesh support

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/_modeling_parallel.py

@@ -59,6 +59,12 @@ class ContextParallelConfig:
         rotate_method (`str`, *optional*, defaults to `"allgather"`):
             Method to use for rotating key/value states across devices in ring attention. Currently, only `"allgather"`
             is supported.
+        mesh (`torch.distributed.device_mesh.DeviceMesh`, *optional*):
+            A custom device mesh to use for context parallelism. If provided, this mesh will be used instead of
+            creating a new one. This is useful when combining context parallelism with other parallelism strategies
+            (e.g., FSDP, tensor parallelism) that share the same device mesh. The mesh must have both "ring" and
+            "ulysses" dimensions. Use size 1 for dimensions not being used (e.g., `mesh_shape=(2, 1, 4)` with
+            `mesh_dim_names=("ring", "ulysses", "fsdp")` for ring attention only with FSDP).
 
     """
 
@@ -67,6 +73,7 @@ class ContextParallelConfig:
     convert_to_fp32: bool = True
     # TODO: support alltoall
     rotate_method: Literal["allgather", "alltoall"] = "allgather"
+    mesh: Optional[torch.distributed.device_mesh.DeviceMesh] = None
 
     _rank: int = None
     _world_size: int = None
@@ -115,7 +122,7 @@ class ContextParallelConfig:
                 f"The product of `ring_degree` ({self.ring_degree}) and `ulysses_degree` ({self.ulysses_degree}) must not exceed the world size ({world_size})."
             )
 
-        self._flattened_mesh = self._mesh._flatten()
+        self._flattened_mesh = self._mesh["ring", "ulysses"]._flatten()
         self._ring_mesh = self._mesh["ring"]
         self._ulysses_mesh = self._mesh["ulysses"]
         self._ring_local_rank = self._ring_mesh.get_local_rank()

src/diffusers/models/modeling_utils.py

@@ -1569,7 +1569,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         mesh = None
         if config.context_parallel_config is not None:
             cp_config = config.context_parallel_config
-            mesh = torch.distributed.device_mesh.init_device_mesh(
+            mesh = cp_config.mesh or torch.distributed.device_mesh.init_device_mesh(
                 device_type=device_type,
                 mesh_shape=cp_config.mesh_shape,
                 mesh_dim_names=cp_config.mesh_dim_names,

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"""