-
Notifications
You must be signed in to change notification settings - Fork 253
/
Copy pathtest_weights_compression.py
1642 lines (1398 loc) · 62.1 KB
/
test_weights_compression.py
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
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
# Copyright (c) 2025 Intel Corporation
# 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
import os
from contextlib import nullcontext
from typing import Callable, Dict, List, Optional
from unittest.mock import patch
import numpy as np
import openvino.runtime as ov
import pandas as pd
import pytest
from attr import dataclass
from openvino.runtime import opset13 as opset
import nncf
import nncf.openvino.optimized_functions as opt_fns
from nncf import CompressWeightsMode
from nncf import SensitivityMetric
from nncf.common.factory import NNCFGraphFactory
from nncf.common.utils.debug import nncf_debug
from nncf.common.utils.helpers import set_env_variable
from nncf.data.dataset import Dataset
from nncf.experimental.common.tensor_statistics.collectors import AggregatorBase
from nncf.openvino.graph.model_transformer import OVModelTransformer
from nncf.openvino.graph.node_utils import get_const_value_as_numpy_tensor
from nncf.parameters import BackupMode
from nncf.parameters import CompressionFormat
from nncf.quantization import compress_weights
from nncf.quantization.advanced_parameters import AdvancedCompressionParameters
from nncf.quantization.advanced_parameters import AdvancedCompressionParameters as CompressionParams
from nncf.quantization.advanced_parameters import AdvancedGPTQParameters as GPTQParams
from nncf.quantization.advanced_parameters import AdvancedLoraCorrectionParameters as LoraParams
from nncf.quantization.algorithms.weight_compression.config import WeightCompressionConfig
from nncf.quantization.algorithms.weight_compression.config import WeightCompressionParameters
from nncf.quantization.algorithms.weight_compression.mixed_precision import MIXED_PRECISION_CRITERIA
from nncf.quantization.algorithms.weight_compression.openvino_backend import OVWeightCompressionAlgoBackend
from nncf.quantization.algorithms.weight_compression.weight_lowering import do_int_quantization
from nncf.quantization.algorithms.weight_compression.weight_lowering import get_integer_quantization_error
from nncf.quantization.algorithms.weight_compression.weight_lowering import reshape_weight_for_grouped_quantization
from nncf.scopes import IgnoredScope
from nncf.tensor import Tensor
from nncf.tensor import TensorDataType
from tests.cross_fw.shared.comparator import compare_stats
from tests.cross_fw.shared.json import dump_to_json
from tests.cross_fw.shared.json import load_json
from tests.cross_fw.test_templates.template_test_weights_compression import ACTIVATION
from tests.cross_fw.test_templates.template_test_weights_compression import TemplateWeightCompression
from tests.openvino.native.common import get_actual_reference_for_current_openvino
from tests.openvino.native.models import AWQActMatmulModel
from tests.openvino.native.models import AWQMatmulModel
from tests.openvino.native.models import GatherAndMatmulShareData
from tests.openvino.native.models import GatherWithTwoReductionAxes
from tests.openvino.native.models import IdentityMatmul
from tests.openvino.native.models import IntegerModel
from tests.openvino.native.models import MatMul
from tests.openvino.native.models import ModelNamedConsts
from tests.openvino.native.models import OVReferenceModel
from tests.openvino.native.models import SequentialMatmulModel
from tests.openvino.native.models import WeightsModel
from tests.openvino.native.quantization.test_fq_params_calculation import REFERENCE_SCALES_DIR
TEST_MODELS = {
IntegerModel: ["matmul_2_data", "gather_2_data", "matmul_1_data"],
WeightsModel: ["weights_0", "weights_1"],
}
DATA_BASED_SENSITIVITY_METRICS = (
SensitivityMetric.HESSIAN_INPUT_ACTIVATION,
SensitivityMetric.MEAN_ACTIVATION_VARIANCE,
SensitivityMetric.MAX_ACTIVATION_VARIANCE,
SensitivityMetric.MEAN_ACTIVATION_MAGNITUDE,
)
ALL_SENSITIVITY_METRICS = DATA_BASED_SENSITIVITY_METRICS + (SensitivityMetric.WEIGHT_QUANTIZATION_ERROR,)
INT8_MODES = (CompressWeightsMode.INT8, CompressWeightsMode.INT8_SYM, CompressWeightsMode.INT8_ASYM)
INT4_NF4_MODES = (CompressWeightsMode.INT4_SYM, CompressWeightsMode.INT4_ASYM, CompressWeightsMode.NF4)
INT4_MODES = (CompressWeightsMode.INT4_SYM, CompressWeightsMode.INT4_ASYM)
class LMLinearModel(OVReferenceModel):
OUTPUT_DIM = 32
HIDDEN_DIM = 16
INPUT_SHAPE = [1, 24, HIDDEN_DIM] # [B, SeqLen, HiddenDim]
def _create_ov_model(
self, transpose_b: bool = True, transpose_a: bool = False, input_shape: Optional[List[int]] = None
):
self._input_shape = self.INPUT_SHAPE if input_shape is None else input_shape
hdim_axis = -2 if transpose_a else -1
self._hidden_dim = self._input_shape[hdim_axis]
input_1 = opset.parameter(self._input_shape, name="Input")
weight_shape = self.get_weight_shape(transpose_b)
data = self._rng.random(weight_shape).astype(np.float32)
matmul = opset.matmul(input_1, data, transpose_a=transpose_a, transpose_b=transpose_b, name="MatMul")
result = opset.result(matmul, name="Result")
result.get_output_tensor(0).set_names(set(["Result"]))
model = ov.Model([result], [input_1])
return model
@property
def hidden_dim(self):
return self._hidden_dim
def get_weight_shape(self, transpose_b: bool = True):
return [self.OUTPUT_DIM, self.hidden_dim] if transpose_b else [self.hidden_dim, self.OUTPUT_DIM]
def get_next_node(node):
target_inputs = node.output(0).get_target_inputs()
assert len(target_inputs) == 1
next_node = next(iter(target_inputs)).get_node()
return next_node
def get_shape_for_second_input(op_with_weights: ov.Node) -> List[int]:
return list(op_with_weights.inputs()[1].get_shape())
def check_int8_node(op: ov.Node, mode: CompressWeightsMode = CompressWeightsMode.INT8_ASYM):
dtype = ov.Type.u8 if mode == CompressWeightsMode.INT8_ASYM else ov.Type.i8
assert op.get_element_type() == dtype
compressed_weight = get_const_value_as_numpy_tensor(op)
stats = {"compressed_weight": compressed_weight}
convert_node = get_next_node(op)
assert convert_node.get_type_name() == "Convert"
if mode == CompressWeightsMode.INT8_ASYM:
sub_node = get_next_node(convert_node)
assert sub_node.get_type_name() == "Subtract"
convert_node = sub_node.input_value(1).get_node()
assert convert_node.get_type_name() == "Convert"
zero_point_node = convert_node.input_value(0).get_node()
zero_point = get_const_value_as_numpy_tensor(zero_point_node)
stats["zero_point"] = zero_point
reduced_weight_shape = list(op.shape)
reduced_weight_shape[-1] = 1
assert list(zero_point_node.shape) == reduced_weight_shape
mul_node = get_next_node(sub_node)
else:
mul_node = get_next_node(convert_node)
assert mul_node.get_type_name() == "Multiply"
scale_node = mul_node.input_value(1).get_node()
scale = get_const_value_as_numpy_tensor(scale_node)
stats["scale"] = scale
return stats
def check_int4_grouped(op: ov.Node, mode: CompressWeightsMode, group_size: int = 7):
dtype = ov.Type.u4 if mode == CompressWeightsMode.INT4_ASYM else ov.Type.i4
assert op.get_element_type() == dtype
weight_shape = op.shape
# NOTE: get_const_value_as_numpy_tensor doesn't work for 4-bit types
assert list(weight_shape)[-1] == group_size
reduced_weight_shape = list(weight_shape)
reduced_weight_shape[-1] = 1
convert_node = get_next_node(op)
assert convert_node.get_type_name() == "Convert"
if mode == CompressWeightsMode.INT4_ASYM:
sub_node = get_next_node(convert_node)
assert sub_node.get_type_name() == "Subtract"
convert_node = sub_node.input_value(1).get_node()
assert convert_node.get_type_name() == "Convert"
zero_point_node = convert_node.input_value(0).get_node()
assert zero_point_node.get_element_type() == dtype
assert list(zero_point_node.shape) == reduced_weight_shape
mul_node = get_next_node(sub_node)
else:
mul_node = get_next_node(convert_node)
assert mul_node.get_type_name() == "Multiply"
scale_node = mul_node.input_value(1).get_node()
assert list(scale_node.shape) == reduced_weight_shape
reshape_node = get_next_node(mul_node)
assert reshape_node.get_type_name() == "Reshape"
convert_node = get_next_node(reshape_node)
assert convert_node.get_type_name() == "Convert"
return {
"scale": get_const_value_as_numpy_tensor(scale_node),
}
def check_nf4_grouped(op: ov.Node, group_size: int = 7):
assert op.get_element_type() == ov.Type.nf4
weight_shape = op.shape
# NOTE: get_const_value_as_numpy_tensor doesn't work for 4-bit types
assert list(weight_shape)[-1] == group_size
reduced_weight_shape = list(weight_shape)
reduced_weight_shape[-1] = 1
convert_node = get_next_node(op)
assert convert_node.get_type_name() == "Convert"
mul_node = get_next_node(convert_node)
assert mul_node.get_type_name() == "Multiply"
scale_node = mul_node.input_value(1).get_node()
assert list(scale_node.shape) == reduced_weight_shape
reshape_node = get_next_node(mul_node)
assert reshape_node.get_type_name() == "Reshape"
convert_node = get_next_node(reshape_node)
assert convert_node.get_type_name() == "Convert"
return {
"scale": get_const_value_as_numpy_tensor(scale_node),
}
def check_int4_sym_grouped(op: ov.Node):
return check_int4_grouped(op, mode=CompressWeightsMode.INT4_SYM)
def check_int4_asym_grouped(op: ov.Node):
return check_int4_grouped(op, mode=CompressWeightsMode.INT4_ASYM)
def check_int8_sym(op: ov.Node):
return check_int8_node(op, mode=CompressWeightsMode.INT8_SYM)
def get_mixed_mapping(primary_fn: Callable, list_layers: List[str]):
mapping = {node_name: check_int8_node for node_name in list_layers}
primary_node_name = TEST_MODELS[IntegerModel][0]
mapping[primary_node_name] = primary_fn
return mapping
@pytest.mark.parametrize(
("mode", "group_size", "check_fn_per_node_map"),
(
(CompressWeightsMode.INT8_ASYM, -1, {node_name: check_int8_node for node_name in TEST_MODELS[IntegerModel]}),
(CompressWeightsMode.INT8_SYM, -1, {node_name: check_int8_sym for node_name in TEST_MODELS[IntegerModel]}),
(CompressWeightsMode.INT4_SYM, 7, get_mixed_mapping(check_int4_sym_grouped, TEST_MODELS[IntegerModel])),
(CompressWeightsMode.INT4_ASYM, 7, get_mixed_mapping(check_int4_asym_grouped, TEST_MODELS[IntegerModel])),
(CompressWeightsMode.NF4, 7, get_mixed_mapping(check_nf4_grouped, TEST_MODELS[IntegerModel])),
),
)
def test_compare_compressed_weights(mode, group_size, check_fn_per_node_map):
model = IntegerModel().ov_model
compressed_model = compress_weights(model, mode=mode, group_size=group_size)
actual_stats = {}
for op in compressed_model.get_ops():
op_name = op.get_friendly_name()
if op.get_type_name() == "Constant" and op_name in check_fn_per_node_map:
check_fn = check_fn_per_node_map[op_name]
actual_stats[op_name] = check_fn(op)
ref_stats_path = get_actual_reference_for_current_openvino(
REFERENCE_SCALES_DIR / f"IntegerModel_compressed_weights_{mode.value}.json"
)
if os.getenv("NNCF_TEST_REGEN_DOT") is not None:
dump_to_json(ref_stats_path, actual_stats)
ref_stats = load_json(ref_stats_path)
compare_stats(ref_stats, actual_stats)
@pytest.mark.parametrize("metric", DATA_BASED_SENSITIVITY_METRICS)
def test_gather_in_4_bit_if_all_layers_with_data(metric):
dim1 = 2 # sequence length dimension
dim2 = 7
max_input_value = 6
model = IntegerModel(dim1=dim1, dim2=dim2, max_input_value=max_input_value, add_batch_dimension=True).ov_model
input_shape = (dim1, dim2, dim1)
n_inputs = input_shape[0] * input_shape[1] * input_shape[2]
n_copies = int(np.ceil(n_inputs / (max_input_value + 1)))
# Rolling is needed to ensure non-zero variance
input_ = [np.roll(np.arange(max_input_value + 1), i + 1) for i in range(n_copies)]
input_ = np.hstack(input_)[:n_inputs]
input_ = input_.reshape(input_shape)
dataset = Dataset([input_])
compressed_model = compress_weights(
model,
mode=CompressWeightsMode.INT4_SYM,
ratio=0.5,
group_size=1,
all_layers=True,
sensitivity_metric=metric,
dataset=dataset,
)
int4_reference_node_names = ["gather_2_data"]
nodes_map = {op.get_friendly_name(): op for op in compressed_model.get_ordered_ops()}
for node_name in int4_reference_node_names:
node = nodes_map[node_name]
assert node.get_type_name() == "Constant"
assert node.get_element_type() == ov.Type.i4
def test_gather_can_be_8_bit_if_all_layers_without_data():
model = IntegerModel().ov_model
compressed_model = compress_weights(
model,
mode=CompressWeightsMode.INT4_SYM,
ratio=0.5,
group_size=1,
all_layers=True,
)
int8_reference_node_names = ["gather_2_data", "gather_2_data/zero_point"]
nodes_map = {op.get_friendly_name(): op for op in compressed_model.get_ordered_ops()}
for node_name in int8_reference_node_names:
node = nodes_map[node_name]
assert node.get_type_name() == "Constant"
assert node.get_element_type() == ov.Type.u8
@pytest.mark.parametrize("mode", (CompressWeightsMode.INT8_SYM, CompressWeightsMode.INT8_ASYM))
def test_conv_in_8_bit_if_mode_8bit(mode):
model = WeightsModel().ov_model
compressed_model = compress_weights(model, mode=mode)
int8_reference_node_names = ["conv_weights_0", "conv_weights_1"]
nodes_map = {op.get_friendly_name(): op for op in compressed_model.get_ordered_ops()}
dtype = ov.Type.u8 if mode == CompressWeightsMode.INT8_ASYM else ov.Type.i8
for node_name in int8_reference_node_names:
node = nodes_map[node_name]
assert node.get_type_name() == "Constant"
assert node.get_element_type() == dtype
@pytest.mark.parametrize("all_layers", (True, False))
def test_conv_in_8_bit_if_mode_4bit(all_layers):
model = WeightsModel().ov_model
compressed_model = compress_weights(
model,
mode=CompressWeightsMode.INT4_SYM,
ratio=1,
group_size=1,
all_layers=all_layers,
)
for op in compressed_model.get_ordered_ops():
if op.get_type_name() == "Constant":
if op.get_friendly_name() in [
"conv_weights_0",
"conv_weights_0/zero_point",
"conv_weights_1",
"conv_weights_1/zero_point",
]:
assert ov.Type.u8 == op.get_element_type()
elif op.get_friendly_name() in ["weights_1", "weights_1/zero_point"]:
assert ov.Type.i4 == op.get_element_type()
elif op.get_friendly_name() in ["weights_0", "weights_0/zero_point"]:
dtype = ov.Type.i4 if all_layers else ov.Type.u8
assert dtype == op.get_element_type()
def test_gather_can_be_4_bit_if_all_layers_without_data():
model = IntegerModel().ov_model
compressed_model = compress_weights(
model,
mode=CompressWeightsMode.INT4_SYM,
ratio=1,
group_size=1,
all_layers=True,
)
int4_reference_node_names = ["gather_2_data"]
nodes_map = {op.get_friendly_name(): op for op in compressed_model.get_ordered_ops()}
for node_name in int4_reference_node_names:
node = nodes_map[node_name]
assert node.get_type_name() == "Constant"
assert node.get_element_type() == ov.Type.i4
@pytest.mark.parametrize("metric", ALL_SENSITIVITY_METRICS)
def test_gather_in_8_bit_if_not_all_layers(metric):
model = IntegerModel(add_batch_dimension=True).ov_model
dataset = Dataset([np.ones([1, 7, 1])])
compressed_model = compress_weights(
model,
mode=CompressWeightsMode.INT4_SYM,
ratio=0.5,
group_size=1,
all_layers=False,
sensitivity_metric=metric,
dataset=dataset,
)
int8_reference_node_names = ["gather_2_data", "gather_2_data/zero_point"]
nodes_map = {op.get_friendly_name(): op for op in compressed_model.get_ordered_ops()}
for node_name in int8_reference_node_names:
node = nodes_map[node_name]
assert node.get_type_name() == "Constant"
assert node.get_element_type() == ov.Type.u8
@pytest.mark.parametrize("mode", (CompressWeightsMode.INT8_SYM, CompressWeightsMode.INT8_ASYM))
def test_quantize_Gather_with_multiple_reduction_axes_in_8bit(mode):
model = GatherWithTwoReductionAxes().ov_model
compressed_model = compress_weights(model, mode=mode)
dtype = ov.Type.u8 if mode == CompressWeightsMode.INT8_ASYM else ov.Type.i8
for op in compressed_model.get_ordered_ops():
if op.get_type_name() == "Constant" and op.get_friendly_name() == "gather_1_data":
assert op.get_element_type() == dtype
@pytest.mark.parametrize("mode", (CompressWeightsMode.INT4_SYM, CompressWeightsMode.INT4_ASYM))
@pytest.mark.parametrize("all_layers", (True, False))
def test_quantize_Gather_with_multiple_reduction_axes_if_mode_4bit(mode, all_layers):
model = GatherWithTwoReductionAxes().ov_model
compressed_model = compress_weights(model, mode=mode, all_layers=all_layers)
for op in compressed_model.get_ordered_ops():
if op.get_type_name() == "Constant" and op.get_friendly_name() == "gather_1_data":
assert op.get_element_type() == ov.Type.u8
@pytest.mark.parametrize("mode", (CompressWeightsMode.INT4_SYM, CompressWeightsMode.INT4_ASYM))
def test_shared_gather(mode):
weight_name_vs_type = {
"gather_2_data": ov.Type.u8,
"shared_data": ov.Type.u8,
"matmul_1_data": ov.Type.i4 if mode == CompressWeightsMode.INT4_SYM else ov.Type.u4,
}
model = GatherAndMatmulShareData().ov_model
compressed_model = compress_weights(model, mode, group_size=3)
for op in compressed_model.get_ordered_ops():
op_name = op.get_friendly_name()
if op.get_type_name() == "Constant" and op_name in weight_name_vs_type:
assert op.get_element_type() == weight_name_vs_type[op_name]
@pytest.mark.parametrize("all_layers", (True, False))
def test_shared_gather_all_layers(all_layers):
weight_name_vs_type = {
"gather_2_data": ov.Type.u4 if all_layers else ov.Type.u8,
"shared_data": ov.Type.u4 if all_layers else ov.Type.u8,
"matmul_1_data": ov.Type.u4,
}
model = GatherAndMatmulShareData().ov_model
compressed_model = compress_weights(model, CompressWeightsMode.INT4_ASYM, group_size=-1, all_layers=all_layers)
for op in compressed_model.get_ordered_ops():
op_name = op.get_friendly_name()
if op.get_type_name() == "Constant" and op_name in weight_name_vs_type:
assert op.get_element_type() == weight_name_vs_type[op_name]
@dataclass
class QuantErrorDesc:
weight: List[float]
ref_error: int = 0
axis = (1,)
name: str = ""
atol: float = None
config: WeightCompressionConfig = WeightCompressionConfig()
def __str__(self):
prefix = "exact_match_" if self.ref_error == 0 else ""
name = self.name.replace(" ", "_") if self.name else self.__class__.__name__
return prefix + name
SCALE_1 = 1.2
SCALE_2 = 3.4
SCALE_3 = 5.6
SCALE_4 = 7.8
LINSPACE = np.arange(0, 256, 17, dtype=np.float32)
TWO_ROWS_LINSPACE = np.vstack((LINSPACE * SCALE_1, LINSPACE * SCALE_2))
LINSPACE_INT4_ASYM = np.arange(0, 16, dtype=np.float32)
TWO_ROWS_LINSPACE_INT4_ASYM = np.vstack((LINSPACE_INT4_ASYM * SCALE_1, LINSPACE_INT4_ASYM * SCALE_2))
LINSPACE_INT4_SYM = np.arange(-8, 7, dtype=np.float32)
TWO_ROWS_LINSPACE_INT4_SYM = np.vstack((LINSPACE_INT4_SYM * SCALE_1, LINSPACE_INT4_SYM * SCALE_2))
TWO_OTHER_ROWS_LINSPACE_INT4_SYM = np.vstack((LINSPACE_INT4_SYM * SCALE_3, LINSPACE_INT4_SYM * SCALE_4))
TWO_GROUPS_IN_TWO_ROWS_SYM = np.hstack((TWO_ROWS_LINSPACE_INT4_SYM, TWO_OTHER_ROWS_LINSPACE_INT4_SYM))
TWO_OTHER_ROWS_LINSPACE_INT4_ASYM = np.vstack((LINSPACE_INT4_ASYM * SCALE_3, LINSPACE_INT4_ASYM * SCALE_4))
TWO_GROUPS_IN_TWO_ROWS_ASYM = np.hstack((TWO_ROWS_LINSPACE_INT4_ASYM, TWO_OTHER_ROWS_LINSPACE_INT4_ASYM))
int4_sym_config = WeightCompressionConfig(mode=CompressWeightsMode.INT4_SYM, group_size=-1)
int4_asym_config = WeightCompressionConfig(mode=CompressWeightsMode.INT4_ASYM, group_size=-1)
int4_sym_grouped_config = WeightCompressionConfig(mode=CompressWeightsMode.INT4_SYM, group_size=15)
int4_asym_grouped_config = WeightCompressionConfig(mode=CompressWeightsMode.INT4_ASYM, group_size=16)
LIST_DESCS = [
# zero error
QuantErrorDesc(name="2 rows of scaled [0, 255] linspace", weight=TWO_ROWS_LINSPACE),
QuantErrorDesc(name="2 rows of scaled [-7, 7] linspace", weight=TWO_ROWS_LINSPACE_INT4_SYM, config=int4_sym_config),
QuantErrorDesc(
name="2 rows of scaled [0, 15] linspace", weight=TWO_ROWS_LINSPACE_INT4_ASYM, config=int4_asym_config
),
QuantErrorDesc(
name="two groups in two rows sym",
weight=TWO_GROUPS_IN_TWO_ROWS_SYM,
config=int4_sym_grouped_config,
),
QuantErrorDesc(
name="two groups in two rows asym",
weight=TWO_GROUPS_IN_TWO_ROWS_ASYM,
config=int4_asym_grouped_config,
),
# non-zero error
QuantErrorDesc(name="2 rows scaled [1, 254] linspace", weight=TWO_ROWS_LINSPACE[:, 1:-1], ref_error=239, atol=1),
QuantErrorDesc(
name="2 columns of scaled [0, 255] linspace", weight=np.transpose(TWO_ROWS_LINSPACE), ref_error=46818, atol=1
),
QuantErrorDesc(
name="2 rows of scaled [0, 15] linspace for sym",
weight=TWO_ROWS_LINSPACE_INT4_ASYM,
config=int4_sym_config,
ref_error=4.12,
atol=1,
),
QuantErrorDesc(
name="2 columns of of scaled [0, 15] linspace for sym",
weight=np.transpose(TWO_ROWS_LINSPACE_INT4_ASYM),
config=int4_sym_config,
ref_error=0.63,
atol=1,
),
QuantErrorDesc(
name="2 rows [1,14] linspace for asym",
weight=TWO_ROWS_LINSPACE_INT4_ASYM[:, 1:-1],
config=int4_asym_config,
ref_error=1.49,
atol=1,
),
QuantErrorDesc(
name="2 columns of [0-15] linspace for asym",
weight=np.transpose(TWO_ROWS_LINSPACE_INT4_ASYM),
config=int4_asym_config,
ref_error=162,
atol=1,
),
]
@pytest.mark.parametrize("desc", LIST_DESCS, ids=map(str, LIST_DESCS))
def test_quantization_error_calculation(desc: QuantErrorDesc):
weight = Tensor(desc.weight)
axis = 1
actual_error = get_integer_quantization_error(weight, axis, desc.config)
ref_error = desc.ref_error
atol = desc.atol if desc.atol is not None else 1e-8
assert np.allclose(actual_error, ref_error, atol=atol)
WEIGHTS_2x4 = np.array([[-4, -3, -2, -1], [0, 11, 2, 3]]) # [2, 4]
WEIGHTS_abs_max = np.array([4, 2, 11, 3]) # [4]
@dataclass
class CalculateScaleDesc:
weight: np.array
ref_scale: np.array
axis: int
group_size: int
CALCULATE_SCALE_DESCS = [
CalculateScaleDesc(weight=WEIGHTS_2x4, ref_scale=WEIGHTS_abs_max.reshape([2, 2, 1]), axis=1, group_size=2),
CalculateScaleDesc(weight=WEIGHTS_2x4, ref_scale=np.abs(WEIGHTS_2x4).reshape([2, 1, 4]), axis=0, group_size=1),
CalculateScaleDesc(
weight=WEIGHTS_2x4.reshape([1, 2, 4, 1]),
ref_scale=WEIGHTS_abs_max.reshape([1, 2, 2, 1, 1]),
axis=2,
group_size=2,
),
CalculateScaleDesc(
weight=WEIGHTS_2x4.reshape([1, 2, 4, 1]),
ref_scale=np.abs(WEIGHTS_2x4.reshape([1, 2, 4, 1])),
axis=0,
group_size=1,
),
CalculateScaleDesc(
weight=WEIGHTS_2x4.reshape([2, 2, 2]), ref_scale=WEIGHTS_abs_max.reshape([2, 2, 1, 1]), axis=2, group_size=2
),
CalculateScaleDesc(
weight=WEIGHTS_2x4.reshape([2, 2, 2]),
ref_scale=np.array([4, 3, 2, 11]).reshape([2, 1, 1, 2]),
axis=1,
group_size=2,
),
CalculateScaleDesc(
weight=WEIGHTS_2x4.reshape([2, 2, 2]),
ref_scale=np.array([4, 11, 2, 3]).reshape([1, 1, 2, 2]),
axis=0,
group_size=2,
),
]
@pytest.mark.parametrize(
("ignored_scope", "num_compressed"),
(
(IgnoredScope(types=["MatMul"]), 1),
(IgnoredScope(types=["Gather"]), 2),
(IgnoredScope(names=["MatMul_1"]), 2),
(IgnoredScope(patterns=["MatMul_\\d"]), 1),
),
)
def test_weight_compress_with_ignored_scope(ignored_scope, num_compressed):
model = IntegerModel().ov_model
compressed_model = compress_weights(model, ignored_scope=ignored_scope)
ref_compressed_weights = TEST_MODELS[IntegerModel]
act_num = 0
for op in compressed_model.get_ops():
if (
op.get_type_name() == "Constant"
and op.get_friendly_name() in ref_compressed_weights
and op.get_element_type() == ov.Type.u8
):
act_num += 1
assert act_num == num_compressed
@pytest.mark.parametrize("desc", CALCULATE_SCALE_DESCS)
def test_calculate_scale_per_group(desc: CalculateScaleDesc):
reshaped_weight, reduction_axis = reshape_weight_for_grouped_quantization(
desc.weight, reduction_axes=desc.axis, group_size=desc.group_size
)
act_scale = np.max(np.abs(reshaped_weight), axis=reduction_axis, keepdims=True) # [a1, r//gs, 1, a2]
assert np.allclose(act_scale, desc.ref_scale)
def test_raise_error_for_many_axes():
with pytest.raises(nncf.UnsupportedModelError):
reshape_weight_for_grouped_quantization(WEIGHTS_2x4, reduction_axes=(0, 1), group_size=1)
@pytest.mark.parametrize("mode", INT8_MODES)
@pytest.mark.parametrize(
"params",
(
{"ratio": 0.5},
{"group_size": 64},
{"all_layers": True},
{"all_layers": False},
*({"sensitivity_metric": metric} for metric in ALL_SENSITIVITY_METRICS),
{"dataset": "anything"},
{"scale_estimation": True},
{"lora_correction": True},
{"gptq": True},
{"awq": True},
{"backup_mode": BackupMode.NONE},
{"backup_mode": BackupMode.INT8_ASYM},
{"backup_mode": BackupMode.INT8_SYM},
{"compression_format": CompressionFormat.FQ},
{"compression_format": CompressionFormat.FQ_LORA},
{"advanced_parameters": AdvancedCompressionParameters(statistics_path="anything")},
),
)
def test_raise_error_with_unsupported_params_for_int8(mode, params):
with pytest.raises(nncf.ParameterNotSupportedError):
compress_weights(ov.Model([], []), mode=mode, **params)
@pytest.mark.parametrize("mode", INT4_NF4_MODES)
@pytest.mark.parametrize(
"params",
(
{"dataset": "anything", "lora_correction": True, "gptq": True},
{"compression_format": CompressionFormat.FQ},
{"compression_format": CompressionFormat.FQ_LORA},
),
)
def test_raise_error_with_unsupported_params_for_int4(mode, params):
with pytest.raises(nncf.ParameterNotSupportedError):
compress_weights(ov.Model([], []), mode=mode, **params)
@pytest.mark.parametrize(
"algo",
(
"lora_correction",
"awq",
"scale_estimation",
"gptq",
),
)
def test_raise_error_with_unsupported_params_for_e2m1(algo):
with pytest.raises(nncf.ParameterNotSupportedError):
compress_weights(ov.Model([], []), dataset="anything", mode=CompressWeightsMode.E2M1, **{algo: True})
@pytest.mark.parametrize("mode", INT4_NF4_MODES)
@pytest.mark.parametrize(
"algo",
(
"lora_correction",
"awq",
"scale_estimation",
"gptq",
),
)
def test_raise_error_with_unsupported_params_for_empty_dataset(mode, algo):
with pytest.raises(nncf.ParameterNotSupportedError):
compress_weights(ov.Model([], []), dataset=None, mode=mode, **{algo: True})
@pytest.mark.parametrize("mode", INT4_NF4_MODES)
@pytest.mark.parametrize("metric", DATA_BASED_SENSITIVITY_METRICS)
def test_raise_error_with_data_metric_and_without_dataset(mode, metric):
model = IntegerModel().ov_model
with pytest.raises(nncf.ValidationError):
compress_weights(model, mode=mode, sensitivity_metric=metric, group_size=-1, ratio=0.8)
@pytest.mark.parametrize("mode", INT4_NF4_MODES)
def test_call_max_var_criterion_with_dataset_by_default(mocker, mode):
model = IntegerModel(add_batch_dimension=True).ov_model
dataset = Dataset([np.ones([1, 7, 1])])
criterion_cls = MIXED_PRECISION_CRITERIA.get(SensitivityMetric.MAX_ACTIVATION_VARIANCE)
scores_spy = mocker.spy(criterion_cls, "_calc_sensitivity")
compress_weights(model, mode=mode, ratio=0.8, group_size=-1, dataset=dataset)
scores_spy.assert_called()
@pytest.mark.parametrize("mode", INT4_MODES)
def test_call_max_var_criterion_with_dataset_by_default_awq(mode):
model = AWQMatmulModel().ov_model
dataset = Dataset([np.ones([1, 8, 8])])
compress_weights(model, mode=mode, ratio=1.0, group_size=2, dataset=dataset, awq=True)
@pytest.mark.parametrize("mode", INT4_NF4_MODES)
def test_call_max_var_criterion_with_dataset_awq_for_compressed_model(mode):
model = AWQMatmulModel(is_int8=True).ov_model
dataset = Dataset([np.ones([1, 8, 8])])
compress_weights(model, mode=mode, ratio=1.0, group_size=2, dataset=dataset, awq=True)
@pytest.mark.parametrize("mode", INT4_NF4_MODES)
def test_call_max_var_criterion_with_dataset_awq_neg_group_size(mode):
model = AWQMatmulModel().ov_model
dataset = Dataset([np.ones([1, 8, 8])])
compress_weights(model, mode=mode, ratio=1.0, group_size=-1, dataset=dataset, awq=True)
def test_data_type_for_num_weights(mocker):
stub = mocker.stub()
params = WeightCompressionParameters(stub, stub, stub, np.int32(1), stub)
assert isinstance(params.num_weights, np.uint64)
def check_compressed_matmul_subgraph(start_node, activation_dtype, weight_dtype, is_adapter=False):
# Weight scale should be in fp16 nevertheless the weight data type
assert start_node.input_value(1).get_node().get_element_type() == ov.Type.f16
next_node = start_node
if not is_adapter:
# lora adapters are int8 - no reshape for group quantization, it's only for int4 weights
next_node = get_next_node(start_node)
assert next_node.get_type_name() == "Reshape"
next_node = get_next_node(next_node)
if activation_dtype == ov.Type.f16:
# There should be no convert node after multiply if both weights and activations are in f16
assert next_node.get_type_name() != "Convert"
else:
assert next_node.get_type_name() == "Convert"
# In case precision of weight and activation were equal, but not f16, the convert node is manually inserted
# In case of lora adapter the convert is always manually inserted
if (activation_dtype == weight_dtype and weight_dtype != ov.Type.f16) or is_adapter:
ref_name = start_node.get_friendly_name() + "/convert"
assert next_node.get_friendly_name() == ref_name
@pytest.mark.parametrize(
"weight_dtype, activation_dtype",
[
(ov.Type.f32, ov.Type.f32),
(ov.Type.f16, ov.Type.f32),
(ov.Type.bf16, ov.Type.f32),
(ov.Type.f16, ov.Type.f16),
(ov.Type.bf16, ov.Type.bf16),
],
ids=[
"w32a32",
"w16a32",
"wb16a32",
"w16a16",
"wb16a16",
],
)
class TestActivationWeightDtype:
@staticmethod
def test_compression_for_different_dtypes(weight_dtype, activation_dtype):
model = IdentityMatmul(weights_dtype=weight_dtype, activation_dtype=activation_dtype).ov_model
compressed_model = compress_weights(
model, mode=CompressWeightsMode.INT4_SYM, ratio=1, group_size=1, all_layers=True
)
name_to_node_map = {op.get_friendly_name(): op for op in compressed_model.get_ops()}
scale_multiply_node = name_to_node_map["weights/fq_weights_1"]
check_compressed_matmul_subgraph(scale_multiply_node, activation_dtype, weight_dtype)
@staticmethod
def test_set_weight(weight_dtype, activation_dtype, tmp_path):
model = GatherAndMatmulShareData(weights_dtype=weight_dtype, activation_dtype=activation_dtype).ov_model
ov_backend = OVWeightCompressionAlgoBackend(model)
graph = NNCFGraphFactory.create(model)
node_key = "5 MatMul_2"
weight_node = graph.get_node_by_key(node_key)
weight_port_id = 1
weight = ov_backend.get_weight(weight_node, weight_port_id, model, graph)
scaled_weight = weight * 2
ov_backend.set_weight(weight_node, weight_port_id, model, graph, scaled_weight)
ov.save_model(model, tmp_path / "model.xml", compress_to_fp16=True)
DATASET_SIZE = 5
@pytest.mark.parametrize(
("dataset_size", "subset_size", "ref_size"),
(
(DATASET_SIZE, 1, 1),
(DATASET_SIZE, DATASET_SIZE, DATASET_SIZE),
(DATASET_SIZE, DATASET_SIZE + 1, DATASET_SIZE),
),
)
@pytest.mark.parametrize(
("compression_args", "multiplier_of_calls"),
[
({"mode": CompressWeightsMode.INT4_ASYM, "ratio": 1}, 0), # data-free, no reducers
({"mode": CompressWeightsMode.INT4_ASYM, "ratio": 1, "awq": True}, 2), # mean & shape reducer for AWQ
(
{"mode": CompressWeightsMode.INT4_ASYM, "ratio": 0.5, "awq": True},
3,
), # 2 - for AWQ + 1 - for Mixed Precision
(
{
"mode": CompressWeightsMode.INT4_ASYM,
"ratio": 0.5,
"sensitivity_metric": nncf.SensitivityMetric.HESSIAN_INPUT_ACTIVATION,
},
1,
), # 1 reducer for mixed precision
(
{
"mode": CompressWeightsMode.INT4_ASYM,
"ratio": 0.5,
"sensitivity_metric": nncf.SensitivityMetric.MEAN_ACTIVATION_VARIANCE,
},
1,
), # 1 reducer for mixed precision
(
{
"mode": CompressWeightsMode.INT4_ASYM,
"ratio": 0.5,
"sensitivity_metric": nncf.SensitivityMetric.MAX_ACTIVATION_VARIANCE,
},
1,
), # 1 reducer for mixed precision
(
{
"mode": CompressWeightsMode.INT4_ASYM,
"ratio": 0.5,
"sensitivity_metric": nncf.SensitivityMetric.MEAN_ACTIVATION_MAGNITUDE,
},
1,
), # 1 reducer for mixed precision
(
{
"mode": CompressWeightsMode.INT4_ASYM,
"ratio": 0.5,
"sensitivity_metric": nncf.SensitivityMetric.WEIGHT_QUANTIZATION_ERROR,
},
0,
), # 0 - data-free method
],
)
def test_number_of_reduced_statistics_for_subset_size(
mocker, dataset_size, subset_size, ref_size, compression_args, multiplier_of_calls
):
model = IdentityMatmul().ov_model
dataset = Dataset([np.array(ACTIVATION)] * dataset_size)
stats_spy = mocker.spy(AggregatorBase, "register_reduced_input")
compress_weights(model, dataset=dataset, subset_size=subset_size, **compression_args)
assert stats_spy.call_count == ref_size * multiplier_of_calls
def test_default_subset_value():
default_value = inspect.signature(compress_weights).parameters["subset_size"].default
assert default_value == 128
@pytest.mark.parametrize("subset_size", (-1, 0))
def test_invalid_subset_size(subset_size):
model = IdentityMatmul().ov_model
dataset = Dataset([np.array(ACTIVATION)])
with pytest.raises(nncf.ValidationError):
compress_weights(model, mode=CompressWeightsMode.INT4_ASYM, ratio=0.5, dataset=dataset, subset_size=subset_size)
def test_duplicate_names_generation():
model = ModelNamedConsts().ov_model
compressed_model = compress_weights(model)
op_names = set()
for op in compressed_model.get_ops():
name = op.get_friendly_name()
assert name not in op_names
op_names.add(name)
@pytest.mark.parametrize(
("mode", "compressed_weight_dtype"),
(
(CompressWeightsMode.INT4_SYM, TensorDataType.int8),
(CompressWeightsMode.INT4_ASYM, TensorDataType.uint8),
(CompressWeightsMode.NF4, TensorDataType.float32),
),
)
def test_call_max_var_criterion_with_dataset_by_default_scale_estimation(mode, compressed_weight_dtype, mocker):
model = AWQMatmulModel().ov_model
dataset = Dataset([np.ones([1, 8, 8])])
from nncf.quantization.algorithms.weight_compression import scale_estimation
from nncf.quantization.algorithms.weight_compression.algorithm import ScaleEstimation
se_spy = mocker.spy(ScaleEstimation, "apply")
tzm_spy = mocker.spy(scale_estimation, "get_target_zero_mask")
compress_weights(model, mode=mode, ratio=1.0, group_size=2, dataset=dataset, scale_estimation=True)
assert se_spy.call_count == 1
assert tzm_spy.call_args_list[0][0][0].dtype == compressed_weight_dtype
@pytest.mark.parametrize("mode", INT4_NF4_MODES)
def test_call_max_var_criterion_with_dataset_scale_estimation_for_compressed_model(mode):
model = AWQMatmulModel(is_int8=True).ov_model
dataset = Dataset([np.ones([1, 8, 8])])
compress_weights(model, mode=mode, ratio=1.0, group_size=2, dataset=dataset, scale_estimation=True)
@pytest.mark.parametrize("mode", INT4_NF4_MODES)
def test_call_max_var_criterion_with_dataset_scale_estimation_neg_group_size(mode):
model = AWQMatmulModel().ov_model
dataset = Dataset([np.ones([1, 8, 8])])
compress_weights(model, mode=mode, ratio=1.0, group_size=-1, dataset=dataset, scale_estimation=True)
@pytest.mark.parametrize("mode", INT4_NF4_MODES)
def test_call_gptq(mode):
model = AWQMatmulModel().ov_model
dataset = Dataset([np.ones([1, 8, 8])])
compress_weights(model, mode=mode, ratio=1.0, group_size=2, dataset=dataset, gptq=True)
@pytest.mark.parametrize("mode", INT4_NF4_MODES)
def test_call_gptq_with_dataset_scale_estimation_neg_group_size(mode):
model = AWQMatmulModel().ov_model
dataset = Dataset([np.ones([1, 8, 8])])
compress_weights(model, mode=mode, ratio=1.0, group_size=-1, dataset=dataset, gptq=True, scale_estimation=True)
# TODO(andreyanufr) Waiting for the e2m1 in OV release
@pytest.mark.xfail
@pytest.mark.parametrize(
("mode", "all_layers", "ratio", "ref_ids"),
(
(SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, True, 1, [0, 1, 2, 3, 4]),
(SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, True, 0.8, [0, 3, 4]),
(SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, True, 0.4, [0]),
(SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, True, 0.2, []),
(SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, False, 1, [0, 1, 2, 3]),
(SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, False, 0.8, [0, 1, 3]),
(SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, False, 0.4, [0]),
(SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, False, 0.2, []),
(SensitivityMetric.HESSIAN_INPUT_ACTIVATION, True, 0.8, [0, 1, 2]),
(SensitivityMetric.HESSIAN_INPUT_ACTIVATION, False, 0.8, [0, 1, 2]),
(SensitivityMetric.MEAN_ACTIVATION_VARIANCE, True, 0.8, [0, 1, 2]),
(SensitivityMetric.MEAN_ACTIVATION_VARIANCE, False, 0.8, [0, 1, 2]),
(SensitivityMetric.MAX_ACTIVATION_VARIANCE, True, 0.8, [0, 1, 2]),
(SensitivityMetric.MAX_ACTIVATION_VARIANCE, False, 0.8, [0, 1, 2]),
(SensitivityMetric.MEAN_ACTIVATION_MAGNITUDE, True, 0.8, [0, 1, 2]),