forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathtorch_utils.py
More file actions
409 lines (325 loc) 路 14.6 KB
/
Copy pathtorch_utils.py
File metadata and controls
409 lines (325 loc) 路 14.6 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
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
# Copyright 2026 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.
"""
PyTorch utilities: Utilities related to PyTorch
"""
from __future__ import annotations
import functools
import os
from typing import Callable, ParamSpec, TypeVar
from . import logging
from .import_utils import (
is_torch_available,
is_torch_mlu_available,
is_torch_neuronx_available,
is_torch_npu_available,
is_torch_version,
)
T = TypeVar("T")
P = ParamSpec("P")
if is_torch_available():
import torch
from torch.fft import fftn, fftshift, ifftn, ifftshift
BACKEND_SUPPORTS_TRAINING = {
"cuda": True,
"xpu": True,
"cpu": True,
"mps": False,
"neuron": False,
"default": True,
}
BACKEND_EMPTY_CACHE = {
"cuda": torch.cuda.empty_cache,
"xpu": torch.xpu.empty_cache,
"cpu": None,
"mps": torch.mps.empty_cache,
"neuron": None,
"default": None,
}
BACKEND_DEVICE_COUNT = {
"cuda": torch.cuda.device_count,
"xpu": torch.xpu.device_count,
"cpu": lambda: 0,
"mps": lambda: 0,
"neuron": lambda: getattr(getattr(torch, "neuron", None), "device_count", lambda: 0)(),
"default": 0,
}
BACKEND_MANUAL_SEED = {
"cuda": torch.cuda.manual_seed,
"xpu": torch.xpu.manual_seed,
"cpu": torch.manual_seed,
"mps": torch.mps.manual_seed,
"neuron": torch.manual_seed,
"default": torch.manual_seed,
}
BACKEND_RESET_PEAK_MEMORY_STATS = {
"cuda": torch.cuda.reset_peak_memory_stats,
"xpu": getattr(torch.xpu, "reset_peak_memory_stats", None),
"cpu": None,
"mps": None,
"neuron": None,
"default": None,
}
BACKEND_RESET_MAX_MEMORY_ALLOCATED = {
"cuda": torch.cuda.reset_max_memory_allocated,
"xpu": getattr(torch.xpu, "reset_peak_memory_stats", None),
"cpu": None,
"mps": None,
"neuron": None,
"default": None,
}
BACKEND_MAX_MEMORY_ALLOCATED = {
"cuda": torch.cuda.max_memory_allocated,
"xpu": getattr(torch.xpu, "max_memory_allocated", None),
"cpu": 0,
"mps": 0,
"neuron": 0,
"default": 0,
}
BACKEND_SYNCHRONIZE = {
"cuda": torch.cuda.synchronize,
"xpu": getattr(torch.xpu, "synchronize", None),
"cpu": None,
"mps": None,
"neuron": getattr(getattr(torch, "neuron", None), "synchronize", None),
"default": None,
}
_FP64_UNSUPPORTED_DEVICES = frozenset({"mps", "npu", "neuron"})
_INT64_UNSUPPORTED_DEVICES = frozenset({"mps", "npu", "neuron"})
_DTYPE_DOWNCAST = {torch.float64: torch.float32, torch.int64: torch.int32}
_DTYPE_UNSUPPORTED_DEVICES = {torch.float64: _FP64_UNSUPPORTED_DEVICES, torch.int64: _INT64_UNSUPPORTED_DEVICES}
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
try:
from torch._dynamo import allow_in_graph as maybe_allow_in_graph
except (ImportError, ModuleNotFoundError):
def maybe_allow_in_graph(cls):
return cls
# This dispatches a defined function according to the accelerator from the function definitions.
def _device_agnostic_dispatch(device: str, dispatch_table: dict[str, callable], *args, **kwargs):
if device not in dispatch_table:
return dispatch_table["default"](*args, **kwargs)
fn = dispatch_table[device]
# Some device agnostic functions return values. Need to guard against 'None' instead at
# user level
if not callable(fn):
return fn
return fn(*args, **kwargs)
# These are callables which automatically dispatch the function specific to the accelerator
def backend_manual_seed(device: str, seed: int):
return _device_agnostic_dispatch(device, BACKEND_MANUAL_SEED, seed)
def backend_synchronize(device: str):
return _device_agnostic_dispatch(device, BACKEND_SYNCHRONIZE)
def backend_empty_cache(device: str):
return _device_agnostic_dispatch(device, BACKEND_EMPTY_CACHE)
def backend_device_count(device: str):
return _device_agnostic_dispatch(device, BACKEND_DEVICE_COUNT)
def backend_reset_peak_memory_stats(device: str):
return _device_agnostic_dispatch(device, BACKEND_RESET_PEAK_MEMORY_STATS)
def backend_reset_max_memory_allocated(device: str):
return _device_agnostic_dispatch(device, BACKEND_RESET_MAX_MEMORY_ALLOCATED)
def backend_max_memory_allocated(device: str):
return _device_agnostic_dispatch(device, BACKEND_MAX_MEMORY_ALLOCATED)
# These are callables which return boolean behaviour flags and can be used to specify some
# device agnostic alternative where the feature is unsupported.
def backend_supports_training(device: str):
if not is_torch_available():
return False
if device not in BACKEND_SUPPORTS_TRAINING:
device = "default"
return BACKEND_SUPPORTS_TRAINING[device]
def maybe_adjust_dtype_for_device(dtype: "torch.dtype", device: "torch.device") -> "torch.dtype":
unsupported = _DTYPE_UNSUPPORTED_DEVICES.get(dtype)
return _DTYPE_DOWNCAST[dtype] if unsupported and device.type in unsupported else dtype
def randn_tensor(
shape: tuple | list,
generator: list["torch.Generator"] | "torch.Generator" | None = None,
device: str | "torch.device" | None = None,
dtype: "torch.dtype" | None = None,
layout: "torch.layout" | None = None,
):
"""A helper function to create random tensors on the desired `device` with the desired `dtype`. When
passing a list of generators, you can seed each batch size individually. If CPU generators are passed, the tensor
is always created on the CPU.
"""
# device on which tensor is created defaults to device
if isinstance(device, str):
device = torch.device(device)
rand_device = device
batch_size = shape[0]
layout = layout or torch.strided
device = device or torch.device("cpu")
# Neuron (XLA) does not support creating random tensors directly on device; always use CPU
if device.type == "neuron":
rand_device = torch.device("cpu")
if generator is not None:
gen_device_type = generator.device.type if not isinstance(generator, list) else generator[0].device.type
if gen_device_type != device.type and gen_device_type == "cpu":
rand_device = "cpu"
if device.type not in ("mps", "neuron"):
logger.info(
f"The passed generator was created on 'cpu' even though a tensor on {device} was expected."
f" Tensors will be created on 'cpu' and then moved to {device}. Note that one can probably"
f" slightly speed up this function by passing a generator that was created on the {device} device."
)
elif gen_device_type != device.type and gen_device_type == "cuda":
raise ValueError(f"Cannot generate a {device} tensor from a generator of type {gen_device_type}.")
# make sure generator list of length 1 is treated like a non-list
if isinstance(generator, list) and len(generator) == 1:
generator = generator[0]
if isinstance(generator, list):
shape = (1,) + shape[1:]
latents = [
torch.randn(shape, generator=generator[i], device=rand_device, dtype=dtype, layout=layout)
for i in range(batch_size)
]
latents = torch.cat(latents, dim=0).to(device)
else:
latents = torch.randn(shape, generator=generator, device=rand_device, dtype=dtype, layout=layout).to(device)
return latents
def is_compiled_module(module) -> bool:
"""Check whether the module was compiled with torch.compile()"""
if is_torch_version("<", "2.0.0") or not hasattr(torch, "_dynamo"):
return False
return isinstance(module, torch._dynamo.eval_frame.OptimizedModule)
def unwrap_module(module):
"""Unwraps a module if it was compiled with torch.compile()"""
return module._orig_mod if is_compiled_module(module) else module
def fourier_filter(x_in: "torch.Tensor", threshold: int, scale: int) -> "torch.Tensor":
"""Fourier filter as introduced in FreeU (https://huggingface.co/papers/2309.11497).
This version of the method comes from here:
https://github.com/huggingface/diffusers/pull/5164#issuecomment-1732638706
"""
x = x_in
B, C, H, W = x.shape
# Non-power of 2 images must be float32
if (W & (W - 1)) != 0 or (H & (H - 1)) != 0:
x = x.to(dtype=torch.float32)
# fftn does not support bfloat16, and produces the experimental ComplexHalf
# dtype (torch.complex32) when given float16, which is numerically unstable
# and triggers a UserWarning. Upcast any non-float32 dtype to float32.
elif x.dtype != torch.float32:
x = x.to(dtype=torch.float32)
# FFT
x_freq = fftn(x, dim=(-2, -1))
x_freq = fftshift(x_freq, dim=(-2, -1))
B, C, H, W = x_freq.shape
mask = torch.ones((B, C, H, W), device=x.device)
crow, ccol = H // 2, W // 2
mask[..., crow - threshold : crow + threshold, ccol - threshold : ccol + threshold] = scale
x_freq = x_freq * mask
# IFFT
x_freq = ifftshift(x_freq, dim=(-2, -1))
x_filtered = ifftn(x_freq, dim=(-2, -1)).real
return x_filtered.to(dtype=x_in.dtype)
def apply_freeu(
resolution_idx: int, hidden_states: "torch.Tensor", res_hidden_states: "torch.Tensor", **freeu_kwargs
) -> tuple["torch.Tensor", "torch.Tensor"]:
"""Applies the FreeU mechanism as introduced in https:
//arxiv.org/abs/2309.11497. Adapted from the official code repository: https://github.com/ChenyangSi/FreeU.
Args:
resolution_idx (`int`): Integer denoting the UNet block where FreeU is being applied.
hidden_states (`torch.Tensor`): Inputs to the underlying block.
res_hidden_states (`torch.Tensor`): Features from the skip block corresponding to the underlying block.
s1 (`float`): Scaling factor for stage 1 to attenuate the contributions of the skip features.
s2 (`float`): Scaling factor for stage 2 to attenuate the contributions of the skip features.
b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
"""
if resolution_idx == 0:
num_half_channels = hidden_states.shape[1] // 2
hidden_states[:, :num_half_channels] = hidden_states[:, :num_half_channels] * freeu_kwargs["b1"]
res_hidden_states = fourier_filter(res_hidden_states, threshold=1, scale=freeu_kwargs["s1"])
if resolution_idx == 1:
num_half_channels = hidden_states.shape[1] // 2
hidden_states[:, :num_half_channels] = hidden_states[:, :num_half_channels] * freeu_kwargs["b2"]
res_hidden_states = fourier_filter(res_hidden_states, threshold=1, scale=freeu_kwargs["s2"])
return hidden_states, res_hidden_states
def get_torch_cuda_device_capability():
if torch.cuda.is_available():
device = torch.device("cuda")
compute_capability = torch.cuda.get_device_capability(device)
compute_capability = f"{compute_capability[0]}.{compute_capability[1]}"
return float(compute_capability)
else:
return None
@functools.lru_cache
def get_device():
if torch.cuda.is_available():
return "cuda"
elif is_torch_npu_available():
return "npu"
elif hasattr(torch, "xpu") and torch.xpu.is_available():
return "xpu"
elif torch.backends.mps.is_available():
return "mps"
elif is_torch_mlu_available():
return "mlu"
elif is_torch_neuronx_available() and hasattr(torch, "neuron") and torch.neuron.is_available():
return "neuron"
else:
return "cpu"
def empty_device_cache(device_type: str | None = None):
if device_type is None:
device_type = get_device()
if device_type in ["cpu"]:
return
device_mod = getattr(torch, device_type, torch.cuda)
device_mod.empty_cache()
def device_synchronize(device_type: str | None = None):
if device_type is None:
device_type = get_device()
device_mod = getattr(torch, device_type, torch.cuda)
device_mod.synchronize()
def enable_full_determinism():
"""
Helper function for reproducible behavior during distributed training. See
- https://pytorch.org/docs/stable/notes/randomness.html for pytorch
"""
# Enable PyTorch deterministic mode. This potentially requires either the environment
# variable 'CUDA_LAUNCH_BLOCKING' or 'CUBLAS_WORKSPACE_CONFIG' to be set,
# depending on the CUDA version, so we set them both here
os.environ["CUDA_LAUNCH_BLOCKING"] = "1"
os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":16:8"
torch.use_deterministic_algorithms(True)
# Enable CUDNN deterministic mode
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
torch.backends.cuda.matmul.allow_tf32 = False
def disable_full_determinism():
os.environ["CUDA_LAUNCH_BLOCKING"] = "0"
os.environ["CUBLAS_WORKSPACE_CONFIG"] = ""
torch.use_deterministic_algorithms(False)
@functools.wraps(functools.lru_cache)
def lru_cache_unless_export(maxsize=128, typed=False):
def outer_wrapper(fn: Callable[P, T]):
cached = functools.lru_cache(maxsize=maxsize, typed=typed)(fn)
if is_torch_version("<", "2.7.0"):
return cached
@functools.wraps(fn)
def inner_wrapper(*args: P.args, **kwargs: P.kwargs):
compiler = getattr(torch, "compiler", None)
is_exporting = bool(compiler and hasattr(compiler, "is_exporting") and compiler.is_exporting())
is_compiling = bool(compiler and hasattr(compiler, "is_compiling") and compiler.is_compiling())
# Fallback for older builds where compiler.is_compiling is unavailable.
if not is_compiling:
dynamo = getattr(torch, "_dynamo", None)
if dynamo is not None and hasattr(dynamo, "is_compiling"):
is_compiling = dynamo.is_compiling()
if is_exporting or is_compiling:
return fn(*args, **kwargs)
return cached(*args, **kwargs)
return inner_wrapper
return outer_wrapper
if is_torch_available():
torch_device = get_device()