forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathcache.py
More file actions
556 lines (417 loc) 路 19.7 KB
/
Copy pathcache.py
File metadata and controls
556 lines (417 loc) 路 19.7 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
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
# coding=utf-8
# Copyright 2025 HuggingFace Inc.
#
# 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 gc
import pytest
import torch
from diffusers.hooks import FasterCacheConfig, FirstBlockCacheConfig, PyramidAttentionBroadcastConfig
from diffusers.hooks.faster_cache import _FASTER_CACHE_BLOCK_HOOK, _FASTER_CACHE_DENOISER_HOOK
from diffusers.hooks.first_block_cache import _FBC_BLOCK_HOOK, _FBC_LEADER_BLOCK_HOOK
from diffusers.hooks.pyramid_attention_broadcast import _PYRAMID_ATTENTION_BROADCAST_HOOK
from diffusers.models.cache_utils import CacheMixin
from ...testing_utils import assert_tensors_close, backend_empty_cache, is_cache, torch_device
def require_cache_mixin(func):
"""Decorator to skip tests if model doesn't use CacheMixin."""
def wrapper(self, *args, **kwargs):
if not issubclass(self.model_class, CacheMixin):
pytest.skip(f"{self.model_class.__name__} does not use CacheMixin.")
return func(self, *args, **kwargs)
return wrapper
class CacheTesterMixin:
"""
Base mixin class providing common test implementations for cache testing.
Cache-specific mixins should:
1. Inherit from their respective config mixin (e.g., PyramidAttentionBroadcastConfigMixin)
2. Inherit from this mixin
3. Define the cache config to use for tests
Expected class attributes:
- model_class: The model class to test (must use CacheMixin)
Expected methods in test classes:
- get_init_dict(): Returns dict of arguments to initialize the model
- get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
Optional overrides:
- cache_input_key: Property returning the input tensor key to vary between passes (default: "hidden_states")
"""
@property
def cache_input_key(self):
return "hidden_states"
def setup_method(self):
gc.collect()
backend_empty_cache(torch_device)
def teardown_method(self):
gc.collect()
backend_empty_cache(torch_device)
def _get_cache_config(self):
"""
Get the cache config for testing.
Should be implemented by subclasses.
"""
raise NotImplementedError("Subclass must implement _get_cache_config")
def _get_hook_names(self):
"""
Get the hook names to check for this cache type.
Should be implemented by subclasses.
Returns a list of hook name strings.
"""
raise NotImplementedError("Subclass must implement _get_hook_names")
def _test_cache_enable_disable_state(self):
"""Test that cache enable/disable updates the is_cache_enabled state correctly."""
init_dict = self.get_init_dict()
model = self.model_class(**init_dict).to(torch_device)
# Initially cache should not be enabled
assert not model.is_cache_enabled, "Cache should not be enabled initially."
config = self._get_cache_config()
# Enable cache
model.enable_cache(config)
assert model.is_cache_enabled, "Cache should be enabled after enable_cache()."
# Disable cache
model.disable_cache()
assert not model.is_cache_enabled, "Cache should not be enabled after disable_cache()."
def _test_cache_double_enable_raises_error(self):
"""Test that enabling cache twice raises an error."""
init_dict = self.get_init_dict()
model = self.model_class(**init_dict).to(torch_device)
config = self._get_cache_config()
model.enable_cache(config)
# Trying to enable again should raise ValueError
with pytest.raises(ValueError, match="Caching has already been enabled"):
model.enable_cache(config)
# Cleanup
model.disable_cache()
def _test_cache_hooks_registered(self):
"""Test that cache hooks are properly registered and removed."""
init_dict = self.get_init_dict()
model = self.model_class(**init_dict).to(torch_device)
config = self._get_cache_config()
hook_names = self._get_hook_names()
model.enable_cache(config)
# Check that at least one hook was registered
hook_count = 0
for module in model.modules():
if hasattr(module, "_diffusers_hook"):
for hook_name in hook_names:
hook = module._diffusers_hook.get_hook(hook_name)
if hook is not None:
hook_count += 1
assert hook_count > 0, f"At least one cache hook should be registered. Hook names: {hook_names}"
# Disable and verify hooks are removed
model.disable_cache()
hook_count_after = 0
for module in model.modules():
if hasattr(module, "_diffusers_hook"):
for hook_name in hook_names:
hook = module._diffusers_hook.get_hook(hook_name)
if hook is not None:
hook_count_after += 1
assert hook_count_after == 0, "Cache hooks should be removed after disable_cache()."
@torch.no_grad()
def _test_cache_inference(self):
"""Test that model can run inference with cache enabled."""
init_dict = self.get_init_dict()
inputs_dict = self.get_dummy_inputs()
model = self.model_class(**init_dict).to(torch_device)
model.eval()
config = self._get_cache_config()
model.enable_cache(config)
# First pass populates the cache
_ = model(**inputs_dict, return_dict=False)[0]
# Create modified inputs for second pass (vary input tensor to simulate denoising)
inputs_dict_step2 = inputs_dict.copy()
if self.cache_input_key in inputs_dict_step2:
inputs_dict_step2[self.cache_input_key] = inputs_dict_step2[self.cache_input_key] + torch.randn_like(
inputs_dict_step2[self.cache_input_key]
)
# Second pass uses cached attention with different inputs (produces approximated output)
output_with_cache = model(**inputs_dict_step2, return_dict=False)[0]
assert output_with_cache is not None, "Model output should not be None with cache enabled."
assert not torch.isnan(output_with_cache).any(), "Model output contains NaN with cache enabled."
# Run same inputs without cache to compare
model.disable_cache()
output_without_cache = model(**inputs_dict_step2, return_dict=False)[0]
# Cached output should be different from non-cached output (due to approximation)
assert not torch.allclose(output_without_cache, output_with_cache, atol=1e-5), (
"Cached output should be different from non-cached output due to cache approximation."
)
@torch.no_grad()
def _test_cache_context_manager(self, atol=1e-5, rtol=0):
"""Test the cache_context context manager properly isolates cache state."""
init_dict = self.get_init_dict()
inputs_dict = self.get_dummy_inputs()
model = self.model_class(**init_dict).to(torch_device)
model.eval()
config = self._get_cache_config()
model.enable_cache(config)
# Run inference in first context
with model.cache_context("context_1"):
output_ctx1 = model(**inputs_dict, return_dict=False)[0]
# Run same inference in second context (cache should be reset)
with model.cache_context("context_2"):
output_ctx2 = model(**inputs_dict, return_dict=False)[0]
# Both contexts should produce the same output (first pass in each)
assert_tensors_close(
output_ctx1,
output_ctx2,
atol=atol,
rtol=rtol,
msg="First pass in different cache contexts should produce the same output.",
)
model.disable_cache()
@torch.no_grad()
def _test_reset_stateful_cache(self):
"""Test that _reset_stateful_cache resets the cache state."""
init_dict = self.get_init_dict()
inputs_dict = self.get_dummy_inputs()
model = self.model_class(**init_dict).to(torch_device)
model.eval()
config = self._get_cache_config()
model.enable_cache(config)
_ = model(**inputs_dict, return_dict=False)[0]
model._reset_stateful_cache()
model.disable_cache()
@is_cache
class PyramidAttentionBroadcastConfigMixin:
"""
Base mixin providing PyramidAttentionBroadcast cache config.
Expected class attributes:
- model_class: The model class to test (must use CacheMixin)
"""
# Default PAB config - can be overridden by subclasses
PAB_CONFIG = {
"spatial_attention_block_skip_range": 2,
}
# Store timestep for callback (must be within default range (100, 800) for skipping to trigger)
_current_timestep = 500
def _get_cache_config(self):
config_kwargs = self.PAB_CONFIG.copy()
config_kwargs["current_timestep_callback"] = lambda: self._current_timestep
return PyramidAttentionBroadcastConfig(**config_kwargs)
def _get_hook_names(self):
return [_PYRAMID_ATTENTION_BROADCAST_HOOK]
@is_cache
class PyramidAttentionBroadcastTesterMixin(PyramidAttentionBroadcastConfigMixin, CacheTesterMixin):
"""
Mixin class for testing PyramidAttentionBroadcast caching on models.
Expected class attributes:
- model_class: The model class to test (must use CacheMixin)
Expected methods to be implemented by subclasses:
- get_init_dict(): Returns dict of arguments to initialize the model
- get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
Pytest mark: cache
Use `pytest -m "not cache"` to skip these tests
"""
@require_cache_mixin
def test_pab_cache_enable_disable_state(self):
self._test_cache_enable_disable_state()
@require_cache_mixin
def test_pab_cache_double_enable_raises_error(self):
self._test_cache_double_enable_raises_error()
@require_cache_mixin
def test_pab_cache_hooks_registered(self):
self._test_cache_hooks_registered()
@require_cache_mixin
def test_pab_cache_inference(self):
self._test_cache_inference()
@require_cache_mixin
def test_pab_cache_context_manager(self):
self._test_cache_context_manager()
@require_cache_mixin
def test_pab_reset_stateful_cache(self):
self._test_reset_stateful_cache()
@is_cache
class FirstBlockCacheConfigMixin:
"""
Base mixin providing FirstBlockCache config.
Expected class attributes:
- model_class: The model class to test (must use CacheMixin)
"""
# Default FBC config - can be overridden by subclasses
# Higher threshold makes FBC more aggressive about caching (skips more often)
FBC_CONFIG = {
"threshold": 1.0,
}
def _get_cache_config(self):
return FirstBlockCacheConfig(**self.FBC_CONFIG)
def _get_hook_names(self):
return [_FBC_LEADER_BLOCK_HOOK, _FBC_BLOCK_HOOK]
@is_cache
class FirstBlockCacheTesterMixin(FirstBlockCacheConfigMixin, CacheTesterMixin):
"""
Mixin class for testing FirstBlockCache on models.
Expected class attributes:
- model_class: The model class to test (must use CacheMixin)
Expected methods to be implemented by subclasses:
- get_init_dict(): Returns dict of arguments to initialize the model
- get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
Pytest mark: cache
Use `pytest -m "not cache"` to skip these tests
"""
@torch.no_grad()
def _test_cache_inference(self):
"""Test that model can run inference with FBC cache enabled (requires cache_context)."""
init_dict = self.get_init_dict()
inputs_dict = self.get_dummy_inputs()
model = self.model_class(**init_dict).to(torch_device)
model.eval()
config = self._get_cache_config()
model.enable_cache(config)
# FBC requires cache_context to be set for inference
with model.cache_context("fbc_test"):
# First pass populates the cache
_ = model(**inputs_dict, return_dict=False)[0]
# Create modified inputs for second pass
inputs_dict_step2 = inputs_dict.copy()
if self.cache_input_key in inputs_dict_step2:
inputs_dict_step2[self.cache_input_key] = inputs_dict_step2[self.cache_input_key] + torch.randn_like(
inputs_dict_step2[self.cache_input_key]
)
# Second pass - FBC should skip remaining blocks and use cached residuals
output_with_cache = model(**inputs_dict_step2, return_dict=False)[0]
assert output_with_cache is not None, "Model output should not be None with cache enabled."
assert not torch.isnan(output_with_cache).any(), "Model output contains NaN with cache enabled."
# Run same inputs without cache to compare
model.disable_cache()
output_without_cache = model(**inputs_dict_step2, return_dict=False)[0]
# Cached output should be different from non-cached output (due to approximation)
assert not torch.allclose(output_without_cache, output_with_cache, atol=1e-5), (
"Cached output should be different from non-cached output due to cache approximation."
)
@torch.no_grad()
def _test_reset_stateful_cache(self):
"""Test that _reset_stateful_cache resets the FBC cache state (requires cache_context)."""
init_dict = self.get_init_dict()
inputs_dict = self.get_dummy_inputs()
model = self.model_class(**init_dict).to(torch_device)
model.eval()
config = self._get_cache_config()
model.enable_cache(config)
with model.cache_context("fbc_test"):
_ = model(**inputs_dict, return_dict=False)[0]
model._reset_stateful_cache()
model.disable_cache()
@require_cache_mixin
def test_fbc_cache_enable_disable_state(self):
self._test_cache_enable_disable_state()
@require_cache_mixin
def test_fbc_cache_double_enable_raises_error(self):
self._test_cache_double_enable_raises_error()
@require_cache_mixin
def test_fbc_cache_hooks_registered(self):
self._test_cache_hooks_registered()
@require_cache_mixin
def test_fbc_cache_inference(self):
self._test_cache_inference()
@require_cache_mixin
def test_fbc_cache_context_manager(self):
self._test_cache_context_manager()
@require_cache_mixin
def test_fbc_reset_stateful_cache(self):
self._test_reset_stateful_cache()
@is_cache
class FasterCacheConfigMixin:
"""
Base mixin providing FasterCache config.
Expected class attributes:
- model_class: The model class to test (must use CacheMixin)
"""
# Default FasterCache config - can be overridden by subclasses
FASTER_CACHE_CONFIG = {
"spatial_attention_block_skip_range": 2,
"spatial_attention_timestep_skip_range": (-1, 901),
"tensor_format": "BCHW",
}
def _get_cache_config(self, current_timestep_callback=None):
config_kwargs = self.FASTER_CACHE_CONFIG.copy()
if current_timestep_callback is None:
current_timestep_callback = lambda: 1000 # noqa: E731
config_kwargs["current_timestep_callback"] = current_timestep_callback
return FasterCacheConfig(**config_kwargs)
def _get_hook_names(self):
return [_FASTER_CACHE_DENOISER_HOOK, _FASTER_CACHE_BLOCK_HOOK]
@is_cache
class FasterCacheTesterMixin(FasterCacheConfigMixin, CacheTesterMixin):
"""
Mixin class for testing FasterCache on models.
Note: FasterCache is designed for pipeline-level inference with proper CFG batch handling
and timestep management. Inference tests are skipped at model level - FasterCache should
be tested via pipeline tests (e.g., FluxPipeline, HunyuanVideoPipeline).
Expected class attributes:
- model_class: The model class to test (must use CacheMixin)
Expected methods to be implemented by subclasses:
- get_init_dict(): Returns dict of arguments to initialize the model
- get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
Pytest mark: cache
Use `pytest -m "not cache"` to skip these tests
"""
@torch.no_grad()
def _test_cache_inference(self):
"""Test that model can run inference with FasterCache enabled."""
init_dict = self.get_init_dict()
inputs_dict = self.get_dummy_inputs()
model = self.model_class(**init_dict).to(torch_device)
model.eval()
current_timestep = [1000]
config = self._get_cache_config(current_timestep_callback=lambda: current_timestep[0])
model.enable_cache(config)
# First pass with timestep outside skip range - computes and populates cache
current_timestep[0] = 1000
_ = model(**inputs_dict, return_dict=False)[0]
# Move timestep inside skip range so subsequent passes use cache
current_timestep[0] = 500
# Create modified inputs for second pass
inputs_dict_step2 = inputs_dict.copy()
if self.cache_input_key in inputs_dict_step2:
inputs_dict_step2[self.cache_input_key] = inputs_dict_step2[self.cache_input_key] + torch.randn_like(
inputs_dict_step2[self.cache_input_key]
)
# Second pass uses cached attention with different inputs
output_with_cache = model(**inputs_dict_step2, return_dict=False)[0]
assert output_with_cache is not None, "Model output should not be None with cache enabled."
assert not torch.isnan(output_with_cache).any(), "Model output contains NaN with cache enabled."
# Run same inputs without cache to compare
model.disable_cache()
output_without_cache = model(**inputs_dict_step2, return_dict=False)[0]
# Cached output should be different from non-cached output (due to approximation)
assert not torch.allclose(output_without_cache, output_with_cache, atol=1e-5), (
"Cached output should be different from non-cached output due to cache approximation."
)
@torch.no_grad()
def _test_reset_stateful_cache(self):
"""Test that _reset_stateful_cache resets the FasterCache state."""
init_dict = self.get_init_dict()
inputs_dict = self.get_dummy_inputs()
model = self.model_class(**init_dict).to(torch_device)
model.eval()
config = self._get_cache_config()
model.enable_cache(config)
_ = model(**inputs_dict, return_dict=False)[0]
model._reset_stateful_cache()
model.disable_cache()
@require_cache_mixin
def test_faster_cache_enable_disable_state(self):
self._test_cache_enable_disable_state()
@require_cache_mixin
def test_faster_cache_double_enable_raises_error(self):
self._test_cache_double_enable_raises_error()
@require_cache_mixin
def test_faster_cache_hooks_registered(self):
self._test_cache_hooks_registered()
@require_cache_mixin
def test_faster_cache_inference(self):
self._test_cache_inference()
@require_cache_mixin
def test_faster_cache_context_manager(self):
self._test_cache_context_manager()
@require_cache_mixin
def test_faster_cache_reset_stateful_cache(self):
self._test_reset_stateful_cache()