update

Author: a-r-r-o-w

src/diffusers/hooks/__init__.py

@@ -16,6 +16,7 @@
 
 
 if is_torch_available():
+    from .context_parallel import apply_context_parallel
     from .faster_cache import FasterCacheConfig, apply_faster_cache
     from .first_block_cache import FirstBlockCacheConfig, apply_first_block_cache
     from .group_offloading import apply_group_offloading

src/diffusers/hooks/context_parallel.py

@@ -0,0 +1,275 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from dataclasses import dataclass
+from typing import Dict, List, Type, Union
+
+import torch
+import torch.distributed._functional_collectives as funcol
+
+from ..models._modeling_parallel import (
+    ContextParallelInput,
+    ContextParallelModelPlan,
+    ContextParallelOutput,
+    ParallelConfig,
+)
+from ..models.attention_dispatch import _parallel_context
+from ..utils import get_logger
+from ..utils.torch_utils import unwrap_module
+from .hooks import HookRegistry, ModelHook
+
+
+logger = get_logger(__name__)  # pylint: disable=invalid-name
+
+_CONTEXT_PARALLEL_MODEL_HOOK = "context_parallel_model_hook"
+_CONTEXT_PARALLEL_SUBMODULE_INPUT_HOOK_TEMPLATE = "cp_input---{}"
+_CONTEXT_PARALLEL_SUBMODULE_OUTPUT_HOOK_TEMPLATE = "cp_output---{}"
+
+
+# TODO(aryan): consolidate with ._helpers.TransformerBlockMetadata
+@dataclass
+class ModuleForwardMetadata:
+    cached_parameter_indices: Dict[str, int] = None
+    _cls: Type = None
+
+    def _get_parameter_from_args_kwargs(self, identifier: str, args=(), kwargs=None):
+        kwargs = kwargs or {}
+
+        if identifier in kwargs:
+            return kwargs[identifier], True, None
+
+        if self.cached_parameter_indices is not None:
+            index = self.cached_parameter_indices.get(identifier, None)
+            if index is None:
+                raise ValueError(f"Parameter '{identifier}' not found in cached indices.")
+            return args[index], False, index
+
+        if self._cls is None:
+            raise ValueError("Model class is not set for metadata.")
+
+        parameters = list(inspect.signature(self._cls.forward).parameters.keys())
+        parameters = parameters[1:]  # skip `self`
+        self.cached_parameter_indices = {param: i for i, param in enumerate(parameters)}
+
+        if identifier not in self.cached_parameter_indices:
+            raise ValueError(f"Parameter '{identifier}' not found in function signature but was requested.")
+
+        index = self.cached_parameter_indices[identifier]
+
+        if index >= len(args):
+            raise ValueError(f"Expected {index} arguments but got {len(args)}.")
+
+        return args[index], False, index
+
+
+def apply_context_parallel(
+    module: torch.nn.Module,
+    parallel_config: ParallelConfig,
+    plan: Dict[str, ContextParallelModelPlan],
+) -> None:
+    """Apply context parallel on a model."""
+    logger.debug(f"Applying context parallel with CP mesh: {parallel_config.cp_mesh} and plan: {plan}")
+
+    for module_id, cp_model_plan in plan.items():
+        submodule = _get_submodule_by_name(module, module_id)
+        if not isinstance(submodule, list):
+            submodule = [submodule]
+
+        logger.debug(f"Applying ContextParallelHook to {module_id=} identifying a total of {len(submodule)} modules")
+
+        for m in submodule:
+            if isinstance(cp_model_plan, dict):
+                hook = ContextParallelSplitHook(cp_model_plan, parallel_config)
+                hook_name = _CONTEXT_PARALLEL_SUBMODULE_INPUT_HOOK_TEMPLATE.format(module_id)
+            elif isinstance(cp_model_plan, (ContextParallelOutput, list, tuple)):
+                if isinstance(cp_model_plan, ContextParallelOutput):
+                    cp_model_plan = [cp_model_plan]
+                if not all(isinstance(x, ContextParallelOutput) for x in cp_model_plan):
+                    raise ValueError(f"Expected all elements of cp_model_plan to be CPOutput, but got {cp_model_plan}")
+                hook = ContextParallelGatherHook(cp_model_plan, parallel_config)
+                hook_name = _CONTEXT_PARALLEL_SUBMODULE_OUTPUT_HOOK_TEMPLATE.format(module_id)
+            else:
+                raise ValueError(f"Unsupported context parallel model plan type: {type(cp_model_plan)}")
+            registry = HookRegistry.check_if_exists_or_initialize(m)
+            registry.register_hook(hook, hook_name)
+
+    registry = HookRegistry.check_if_exists_or_initialize(module)
+    hook = ContextParallelModelHook(parallel_config)
+    registry.register_hook(hook, _CONTEXT_PARALLEL_MODEL_HOOK)
+
+
+class ContextParallelModelHook(ModelHook):
+    def __init__(self, parallel_config: ParallelConfig) -> None:
+        super().__init__()
+        self.parallel_config = parallel_config
+
+    def new_forward(self, module: torch.nn.Module, *args, **kwargs):
+        with _parallel_context(self.parallel_config):
+            return self.fn_ref.original_forward(*args, **kwargs)
+
+
+class ContextParallelSplitHook(ModelHook):
+    def __init__(self, metadata: ContextParallelModelPlan, parallel_config: ParallelConfig) -> None:
+        super().__init__()
+        self.metadata = metadata
+        self.parallel_config = parallel_config
+        self.module_forward_metadata = None
+
+    def initialize_hook(self, module):
+        cls = unwrap_module(module).__class__
+        self.module_forward_metadata = ModuleForwardMetadata(_cls=cls)
+        return module
+
+    def pre_forward(self, module, *args, **kwargs):
+        args_list = list(args)
+
+        for name, cpm in self.metadata.items():
+            if isinstance(cpm, ContextParallelInput) and cpm.split_output:
+                continue
+
+            # Maybe the parameter was passed as a keyword argument
+            input_val, is_kwarg, index = self.module_forward_metadata._get_parameter_from_args_kwargs(
+                name, args_list, kwargs
+            )
+
+            if input_val is None:
+                continue
+
+            # The input_val may be a tensor or list/tuple of tensors. In certain cases, user may specify to shard
+            # the output instead of input for a particular layer by setting split_output=True
+            if isinstance(input_val, torch.Tensor):
+                input_val = self._prepare_cp_input(input_val, cpm)
+            elif isinstance(input_val, (list, tuple)):
+                if len(input_val) != len(cpm):
+                    raise ValueError(
+                        f"Expected input model plan to have {len(input_val)} elements, but got {len(cpm)}."
+                    )
+                sharded_input_val = []
+                for i, x in enumerate(input_val):
+                    if torch.is_tensor(x) and not cpm[i].split_output:
+                        x = self._prepare_cp_input(x, cpm[i])
+                    sharded_input_val.append(x)
+                input_val = sharded_input_val
+            else:
+                raise ValueError(f"Unsupported input type: {type(input_val)}")
+
+            if is_kwarg:
+                kwargs[name] = input_val
+            elif index is not None and index < len(args_list):
+                args_list[index] = input_val
+            else:
+                raise ValueError(
+                    f"An unexpected error occurred while processing the input '{name}'. Please open an "
+                    f"issue at https://github.com/huggingface/diffusers/issues and provide a minimal reproducible "
+                    f"example along with the full stack trace."
+                )
+
+        return tuple(args_list), kwargs
+
+    def post_forward(self, module, output):
+        is_tensor = isinstance(output, torch.Tensor)
+        is_tensor_list = isinstance(output, (list, tuple)) and all(isinstance(x, torch.Tensor) for x in output)
+
+        if not is_tensor and not is_tensor_list:
+            raise ValueError(f"Expected output to be a tensor or a list/tuple of tensors, but got {type(output)}.")
+
+        output = [output] if is_tensor else list(output)
+        for index, cpm in self.metadata.items():
+            if not isinstance(cpm, ContextParallelInput) or not cpm.split_output:
+                continue
+            if index >= len(output):
+                raise ValueError(f"Index {index} out of bounds for output of length {len(output)}.")
+            current_output = output[index]
+            current_output = self._prepare_cp_input(current_output, cpm)
+            output[index] = current_output
+
+        return output[0] if is_tensor else tuple(output)
+
+    def _prepare_cp_input(self, x: torch.Tensor, cp_input: ContextParallelInput) -> torch.Tensor:
+        if cp_input.expected_dims is not None and x.dim() != cp_input.expected_dims:
+            raise ValueError(
+                f"Expected input tensor to have {cp_input.expected_dims} dimensions, but got {x.dim()} dimensions."
+            )
+        return EquipartitionSharder.shard(x, cp_input.split_dim, self.parallel_config._flattened_mesh)
+
+
+class ContextParallelGatherHook(ModelHook):
+    def __init__(self, metadata: ContextParallelModelPlan, parallel_config: ParallelConfig) -> None:
+        super().__init__()
+        self.metadata = metadata
+        self.parallel_config = parallel_config
+
+    def post_forward(self, module, output):
+        is_tensor = isinstance(output, torch.Tensor)
+
+        if is_tensor:
+            output = [output]
+        elif not (isinstance(output, (list, tuple)) and all(isinstance(x, torch.Tensor) for x in output)):
+            raise ValueError(f"Expected output to be a tensor or a list/tuple of tensors, but got {type(output)}.")
+
+        output = list(output)
+
+        if len(output) != len(self.metadata):
+            raise ValueError(f"Expected output to have {len(self.metadata)} elements, but got {len(output)}.")
+
+        for i, cpm in enumerate(self.metadata):
+            if cpm is None:
+                continue
+            output[i] = EquipartitionSharder.unshard(output[i], cpm.gather_dim, self.parallel_config._flattened_mesh)
+
+        return output[0] if is_tensor else tuple(output)
+
+
+class EquipartitionSharder:
+    @classmethod
+    @torch.compiler.disable
+    def shard(cls, tensor: torch.Tensor, dim: int, mesh: torch.distributed.device_mesh.DeviceMesh) -> torch.Tensor:
+        assert tensor.size()[dim] % mesh.size() == 0
+        return tensor.chunk(mesh.size(), dim=dim)[mesh.get_rank()]
+
+    @classmethod
+    @torch.compiler.disable
+    def unshard(cls, tensor: torch.Tensor, dim: int, mesh: torch.distributed.device_mesh.DeviceMesh) -> torch.Tensor:
+        tensor = tensor.contiguous()
+        tensor = funcol.all_gather_tensor(tensor, dim, group=mesh.get_group())
+        return tensor
+
+
+def _get_submodule_by_name(model: torch.nn.Module, name: str) -> Union[torch.nn.Module, List[torch.nn.Module]]:
+    if name.count("*") > 1:
+        raise ValueError("Wildcard '*' can only be used once in the name")
+    return _find_submodule_by_name(model, name)
+
+
+def _find_submodule_by_name(model: torch.nn.Module, name: str) -> Union[torch.nn.Module, List[torch.nn.Module]]:
+    if name == "":
+        return model
+    first_atom, remaining_name = name.split(".", 1) if "." in name else (name, "")
+    if first_atom == "*":
+        if not isinstance(model, torch.nn.ModuleList):
+            raise ValueError("Wildcard '*' can only be used with ModuleList")
+        submodules = []
+        for submodule in model:
+            subsubmodules = _find_submodule_by_name(submodule, remaining_name)
+            if not isinstance(subsubmodules, list):
+                subsubmodules = [subsubmodules]
+            submodules.extend(subsubmodules)
+        return submodules
+    else:
+        if hasattr(model, first_atom):
+            submodule = getattr(model, first_atom)
+            return _find_submodule_by_name(submodule, remaining_name)
+        else:
+            raise ValueError(f"'{first_atom}' is not a submodule of '{model.__class__.__name__}'")

src/diffusers/models/_modeling_parallel.py

@@ -0,0 +1,105 @@
+# Experimental parallelism support for Diffusers.
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Dict, List, Literal, Optional, Tuple, Union
+
+import torch
+
+from ..utils import get_logger
+
+
+logger = get_logger(__name__)  # pylint: disable=invalid-name
+
+
+# TODO(aryan): add support for the following:
+# - Unified Attention
+# - More dispatcher attention backends
+# - CFG/Data Parallel
+# - Tensor Parallel
+
+
+@dataclass
+class ParallelConfig:
+    rank: int
+    world_size: int
+    ring_degree: int
+    ulysses_degree: int
+    device: torch.device
+    cp_mesh: torch.distributed.device_mesh.DeviceMesh
+
+    # Whether to convert output and LSE to float32 for ring attention numerical stability
+    convert_to_fp32: bool = True
+    # TODO: support alltoall
+    rotate_method: Literal["allgather", "alltoall"] = "allgather"
+
+    _flattened_mesh: torch.distributed.device_mesh.DeviceMesh = None
+    _ring_mesh: torch.distributed.device_mesh.DeviceMesh = None
+    _ulysses_mesh: torch.distributed.device_mesh.DeviceMesh = None
+    _ring_local_rank: int = None
+    _ulysses_local_rank: int = None
+
+    def __post_init__(self):
+        if self.rotate_method != "allgather":
+            raise ValueError(f"Only rotate_method='allgather' is supported for now, but got {self.rotate_method}.")
+        if self._flattened_mesh is None:
+            self._flattened_mesh = self.cp_mesh._flatten()
+        if self._ring_mesh is None:
+            self._ring_mesh = self.cp_mesh["ring"]
+        if self._ulysses_mesh is None:
+            self._ulysses_mesh = self.cp_mesh["ulysses"]
+        if self._ring_local_rank is None:
+            self._ring_local_rank = self._ring_mesh.get_local_rank()
+        if self._ulysses_local_rank is None:
+            self._ulysses_local_rank = self._ulysses_mesh.get_local_rank()
+
+
+@dataclass(frozen=True)
+class ContextParallelInput:
+    split_dim: int
+    expected_dims: Optional[int] = None
+    split_output: bool = False
+
+    def __repr__(self):
+        return f"ContextParallelInput(split_dim={self.split_dim}, expected_dims={self.expected_dims}, split_output={self.split_output})"
+
+
+@dataclass(frozen=True)
+class ContextParallelOutput:
+    gather_dim: int
+    expected_dims: Optional[int] = None
+
+    def __repr__(self):
+        return f"ContextParallelOutput(gather_dim={self.gather_dim}, expected_dims={self.expected_dims})"
+
+
+# A dictionary where keys denote the input to be split across context parallel region, and the
+# value denotes the sharding configuration.
+# If the key is a string, it denotes the name of the parameter in the forward function.
+# If the key is an integer, split_output must be set to True, and it denotes the index of the output
+# to be split across context parallel region.
+ContextParallelInputType = Dict[
+    Union[str, int], Union[ContextParallelInput, List[ContextParallelInput], Tuple[ContextParallelInput, ...]]
+]
+
+# A dictionary where keys denote the output to be gathered across context parallel region, and the
+# value denotes the gathering configuration.
+ContextParallelOutputType = Union[
+    ContextParallelOutput, List[ContextParallelOutput], Tuple[ContextParallelOutput, ...]
+]
+
+# A dictionary where keys denote the module id, and the value denotes how the inputs/outputs of
+# the module should be split/gathered across context parallel region.
+ContextParallelModelPlan = Dict[str, Union[ContextParallelInputType, ContextParallelOutputType]]