forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathcontext_parallel.py
More file actions
383 lines (308 loc) 路 16.1 KB
/
Copy pathcontext_parallel.py
File metadata and controls
383 lines (308 loc) 路 16.1 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
# 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 copy
import functools
import inspect
from dataclasses import dataclass
from typing import Dict, List, Tuple, Type, Union
import torch
import torch.distributed as dist
if torch.distributed.is_available():
import torch.distributed._functional_collectives as funcol
from ..models._modeling_parallel import (
ContextParallelConfig,
ContextParallelInput,
ContextParallelModelPlan,
ContextParallelOutput,
gather_size_by_comm,
)
from ..utils import get_logger
from ..utils.torch_utils import maybe_allow_in_graph, unwrap_module
from .hooks import HookRegistry, ModelHook
logger = get_logger(__name__) # pylint: disable=invalid-name
_CONTEXT_PARALLEL_INPUT_HOOK_TEMPLATE = "cp_input---{}"
_CONTEXT_PARALLEL_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: ContextParallelConfig,
plan: Dict[str, ContextParallelModelPlan],
) -> None:
"""Apply context parallel on a model."""
logger.debug(f"Applying context parallel with CP mesh: {parallel_config._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_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_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)
def remove_context_parallel(module: torch.nn.Module, plan: Dict[str, ContextParallelModelPlan]) -> None:
for module_id, cp_model_plan in plan.items():
submodule = _get_submodule_by_name(module, module_id)
if not isinstance(submodule, list):
submodule = [submodule]
for m in submodule:
registry = HookRegistry.check_if_exists_or_initialize(m)
if isinstance(cp_model_plan, dict):
hook_name = _CONTEXT_PARALLEL_INPUT_HOOK_TEMPLATE.format(module_id)
elif isinstance(cp_model_plan, (ContextParallelOutput, list, tuple)):
hook_name = _CONTEXT_PARALLEL_OUTPUT_HOOK_TEMPLATE.format(module_id)
else:
raise ValueError(f"Unsupported context parallel model plan type: {type(cp_model_plan)}")
registry.remove_hook(hook_name)
class ContextParallelSplitHook(ModelHook):
def __init__(self, metadata: ContextParallelModelPlan, parallel_config: ContextParallelConfig) -> 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:
logger.warning_once(
f"Expected input tensor to have {cp_input.expected_dims} dimensions, but got {x.dim()} dimensions, split will not be applied."
)
return x
else:
if self.parallel_config.ulysses_anything:
return PartitionAnythingSharder.shard_anything(
x, cp_input.split_dim, self.parallel_config._flattened_mesh
)
return EquipartitionSharder.shard(x, cp_input.split_dim, self.parallel_config._flattened_mesh)
class ContextParallelGatherHook(ModelHook):
def __init__(self, metadata: ContextParallelModelPlan, parallel_config: ContextParallelConfig) -> 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
if self.parallel_config.ulysses_anything:
output[i] = PartitionAnythingSharder.unshard_anything(
output[i], cpm.gather_dim, self.parallel_config._flattened_mesh
)
else:
output[i] = EquipartitionSharder.unshard(
output[i], cpm.gather_dim, self.parallel_config._flattened_mesh
)
return output[0] if is_tensor else tuple(output)
class AllGatherFunction(torch.autograd.Function):
@staticmethod
def forward(ctx, tensor, dim, group):
ctx.dim = dim
ctx.group = group
ctx.world_size = torch.distributed.get_world_size(group)
ctx.rank = torch.distributed.get_rank(group)
return funcol.all_gather_tensor(tensor, dim, group=group)
@staticmethod
def backward(ctx, grad_output):
grad_chunks = torch.chunk(grad_output, ctx.world_size, dim=ctx.dim)
return grad_chunks[ctx.rank], None, None
class EquipartitionSharder:
@classmethod
def shard(cls, tensor: torch.Tensor, dim: int, mesh: torch.distributed.device_mesh.DeviceMesh) -> torch.Tensor:
# NOTE: the following assertion does not have to be true in general. We simply enforce it for now
# because the alternate case has not yet been tested/required for any model.
assert tensor.size()[dim] % mesh.size() == 0, (
"Tensor size along dimension to be sharded must be divisible by mesh size"
)
# The following is not fullgraph compatible with Dynamo (fails in DeviceMesh.get_rank)
# return tensor.chunk(mesh.size(), dim=dim)[mesh.get_rank()]
return tensor.chunk(mesh.size(), dim=dim)[torch.distributed.get_rank(mesh.get_group())]
@classmethod
def unshard(cls, tensor: torch.Tensor, dim: int, mesh: torch.distributed.device_mesh.DeviceMesh) -> torch.Tensor:
tensor = tensor.contiguous()
tensor = AllGatherFunction.apply(tensor, dim, mesh.get_group())
return tensor
class AllGatherAnythingFunction(torch.autograd.Function):
@staticmethod
def forward(ctx, tensor: torch.Tensor, dim: int, group: dist.device_mesh.DeviceMesh):
ctx.dim = dim
ctx.group = group
ctx.world_size = dist.get_world_size(group)
ctx.rank = dist.get_rank(group)
gathered_tensor = _all_gather_anything(tensor, dim, group)
return gathered_tensor
@staticmethod
def backward(ctx, grad_output):
# NOTE: We use `tensor_split` instead of chunk, because the `chunk`
# function may return fewer than the specified number of chunks!
grad_splits = torch.tensor_split(grad_output, ctx.world_size, dim=ctx.dim)
return grad_splits[ctx.rank], None, None
class PartitionAnythingSharder:
@classmethod
def shard_anything(
cls, tensor: torch.Tensor, dim: int, mesh: torch.distributed.device_mesh.DeviceMesh
) -> torch.Tensor:
assert tensor.size()[dim] >= mesh.size(), (
f"Cannot shard tensor of size {tensor.size()} along dim {dim} across mesh of size {mesh.size()}."
)
# NOTE: We use `tensor_split` instead of chunk, because the `chunk`
# function may return fewer than the specified number of chunks!
return tensor.tensor_split(mesh.size(), dim=dim)[dist.get_rank(mesh.get_group())]
@classmethod
def unshard_anything(
cls, tensor: torch.Tensor, dim: int, mesh: torch.distributed.device_mesh.DeviceMesh
) -> torch.Tensor:
tensor = tensor.contiguous()
tensor = AllGatherAnythingFunction.apply(tensor, dim, mesh.get_group())
return tensor
@functools.lru_cache(maxsize=64)
def _fill_gather_shapes(shape: Tuple[int], gather_dims: Tuple[int], dim: int, world_size: int) -> List[List[int]]:
gather_shapes = []
for i in range(world_size):
rank_shape = list(copy.deepcopy(shape))
rank_shape[dim] = gather_dims[i]
gather_shapes.append(rank_shape)
return gather_shapes
@maybe_allow_in_graph
def _all_gather_anything(tensor: torch.Tensor, dim: int, group: dist.device_mesh.DeviceMesh) -> torch.Tensor:
world_size = dist.get_world_size(group=group)
tensor = tensor.contiguous()
shape = tensor.shape
rank_dim = shape[dim]
gather_dims = gather_size_by_comm(rank_dim, group)
gather_shapes = _fill_gather_shapes(tuple(shape), tuple(gather_dims), dim, world_size)
gathered_tensors = [torch.empty(shape, device=tensor.device, dtype=tensor.dtype) for shape in gather_shapes]
dist.all_gather(gathered_tensors, tensor, group=group)
gathered_tensor = torch.cat(gathered_tensors, dim=dim)
return gathered_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__}'")