-
Notifications
You must be signed in to change notification settings - Fork 2.2k
/
Copy pathqkvToContextPlugin.cpp
1541 lines (1382 loc) · 56.2 KB
/
qkvToContextPlugin.cpp
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
/*
* SPDX-FileCopyrightText: Copyright (c) 1993-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* SPDX-License-Identifier: Apache-2.0
*
* 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.
*/
// Need 10.1 for cublasGemmStridedBatchedEx
#include <cuda.h>
#if CUDA_VERSION >= 10010
#include "NvInfer.h"
#include "bertQKVToContextPlugin/fused_multihead_attention/fused_multihead_attention.h"
#include "bertQKVToContextPlugin/fused_multihead_attention_v2/fused_multihead_attention_v2.h"
#include "common/bertCommon.h"
#include "common/serialize.hpp"
#include "mhaRunner.h"
#include "qkvToContextPlugin.h"
#include <cstdint>
#include <cstring>
#include <iostream>
#include <tuple>
#include <vector>
using namespace nvinfer1;
using namespace nvinfer1::plugin;
using namespace nvinfer1::plugin::bert;
using namespace nvinfer1::pluginInternal;
namespace
{
char const* const kQKV_TO_CONTEXT_PLUGIN_VERSION{"4"};
char const* const kQKV_TO_CONTEXT_VAR_SEQLEN_PLUGIN_VERSION{"5"};
char const* const kQKV_TO_CONTEXT_PLUGIN_NAME{"CustomQKVToContextPluginDynamic"};
} // namespace
// Static class fields initialization
PluginFieldCollection QKVToContextPluginDynamicCreator::mFC{};
std::vector<PluginField> QKVToContextPluginDynamicCreator::mPluginAttributes;
REGISTER_TENSORRT_PLUGIN(QKVToContextPluginDynamicCreator);
constexpr uint32_t kIIDX = 0; // index of the input tensor
constexpr uint32_t kMIDX = 1; // index of the mask
// Static class fields initialization
PluginFieldCollection QKVToContextVarSeqlenPluginCreator::mFC{};
std::vector<PluginField> QKVToContextVarSeqlenPluginCreator::mPluginAttributes;
REGISTER_TENSORRT_PLUGIN(QKVToContextVarSeqlenPluginCreator);
QKVToContextPluginDynamic::~QKVToContextPluginDynamic() {}
QKVToContextPluginDynamic::QKVToContextPluginDynamic(const std::string name, const DataType type,
const int32_t hiddenSize, const int32_t numHeads, float const dqProbs, bool hasImask)
: mLayerName(name)
, mS(0)
, mB(0)
, mHeadSize(hiddenSize / numHeads)
, mHiddenSize(hiddenSize)
, mNumHeads(numHeads)
, mType(type)
, mDqProbs(dqProbs)
{
mHasImask = static_cast<int32_t>(hasImask);
mSM = getSMVersion();
}
QKVToContextPluginDynamic::QKVToContextPluginDynamic(const std::string name, const DataType type, const int32_t S,
const int32_t B, const int32_t SM, const int32_t hiddenSize, const int32_t numHeads, float const dqProbs,
bool hasImask, bool hasUnfusedDispatcher, void const* runnerStateBuffer)
: mLayerName(name)
, mS(S)
, mB(B)
, mSM(SM)
, mHeadSize(hiddenSize / numHeads)
, mHiddenSize(hiddenSize)
, mNumHeads(numHeads)
, mType(type)
, mDqProbs(dqProbs)
{
BERT_DEBUG_MSG("MHA Runner Deser");
mHasImask = static_cast<int32_t>(hasImask);
mHasUnfusedDispatcher = static_cast<int32_t>(hasUnfusedDispatcher);
createMHARunner();
if (hasUnfusedDispatcher)
{
PLUGIN_ASSERT(unfusedDispatcher.get());
PLUGIN_ASSERT(runnerStateBuffer != nullptr);
auto length = unfusedDispatcher->getSerializationSize();
unfusedDispatcher->deserialize(runnerStateBuffer, length);
}
BERT_DEBUG_MSG("MHA Runner Deser Done");
}
IPluginCapability* QKVToContextPluginDynamic::getCapabilityInterface(PluginCapabilityType type) noexcept
{
try
{
if (type == PluginCapabilityType::kBUILD)
{
return static_cast<IPluginV3OneBuild*>(this);
}
if (type == PluginCapabilityType::kRUNTIME)
{
return static_cast<IPluginV3OneRuntime*>(this);
}
PLUGIN_ASSERT(type == PluginCapabilityType::kCORE);
return static_cast<IPluginV3OneCore*>(this);
}
catch (std::exception const& e)
{
caughtError(e);
}
return nullptr;
}
void QKVToContextPluginDynamic::createMHARunner()
{
if (!fusedDispatcher.get())
{
if (mType == DataType::kHALF)
{
fusedDispatcher.reset(new FusedMHARunnerFP16(mNumHeads, mSM));
}
else if (mType == DataType::kINT8)
{
fusedDispatcher.reset(new FusedMHARunnerInt8(mNumHeads, mSM, mDqProbs));
}
}
if (!unfusedDispatcher.get())
{
unfusedDispatcher.reset(new UnfusedMHARunner(mType, mNumHeads, mSM));
}
}
IPluginV3* QKVToContextPluginDynamic::clone() noexcept
{
BERT_DEBUG_MSG("QKV Clone");
QKVToContextPluginDynamic* ret = nullptr;
mHasUnfusedDispatcher = 0;
char* bufferData = nullptr;
// the workspacesize is 0 if we have not call setup the dispatcher yet.
if (unfusedDispatcher.get() && unfusedDispatcher->getWorkspaceSize())
{
mHasUnfusedDispatcher = 1;
mRunnerStateBuffer.resize(unfusedDispatcher->getSerializationSize());
unfusedDispatcher->serialize(mRunnerStateBuffer.data());
bufferData = mRunnerStateBuffer.data();
}
ret = new QKVToContextPluginDynamic(mLayerName, mType, mS, mB, mSM, mHiddenSize, mNumHeads, mDqProbs,
static_cast<bool>(mHasImask), mHasUnfusedDispatcher, static_cast<void const*>(bufferData));
ret->setPluginNamespace(mNamespace.c_str());
BERT_DEBUG_MSG("QKV Clone done");
return ret;
}
int32_t QKVToContextPluginDynamic::getOutputShapes(DimsExprs const* inputs, int32_t nbInputs,
DimsExprs const* shapeInputs, int32_t nbShapeInputs, DimsExprs* outputs, int32_t nbOutputs,
IExprBuilder& exprBuilder) noexcept
{
try
{
PLUGIN_ASSERT(inputs != nullptr);
PLUGIN_ASSERT(nbInputs == 1 + mHasImask);
PLUGIN_ASSERT(nbShapeInputs == 0);
PLUGIN_ASSERT(outputs != nullptr);
PLUGIN_ASSERT(nbOutputs == 1);
// Input is BxSx3*N*H, output should be BxSxN*H
// Copy over everything
outputs[kIIDX] = inputs[kIIDX];
// Divide last dim by three
auto const* three = exprBuilder.constant(3);
outputs[kIIDX].d[HDIM] = exprBuilder.operation(DimensionOperation::kFLOOR_DIV, *inputs[kIIDX].d[HDIM], *three);
return pluginStatus_t::STATUS_SUCCESS;
}
catch (std::exception const& e)
{
caughtError(e);
}
return pluginStatus_t::STATUS_FAILURE;
}
bool QKVToContextPluginDynamic::supportsFormatCombination(
int32_t pos, DynamicPluginTensorDesc const* inOut, int32_t nbInputs, int32_t /*nbOutputs*/) noexcept
{
PLUGIN_ASSERT(pos >= 0);
PLUGIN_ASSERT(pos < 2 + mHasImask);
PLUGIN_ASSERT(nbInputs == 1 + mHasImask);
auto const* in = inOut;
auto const* out = inOut + nbInputs;
int32_t packedSize = getMHAMaskPackedSize(mSM, mType, in->desc.dims.d[SDIM]);
// we only support int8 IO in fused mha runner, and we only support fused mha runner on Xavier, Turing and Ampere
if (mType == DataType::kINT8)
{
if (!elem(mSM, {kSM_75, kSM_80, kSM_86, kSM_87, kSM_89, kSM_90, kSM_100, kSM_120}))
{
gLogError << "INT8 IO is only supported on Turing, Ampere, Hopper and Blackwell for plugin "
<< kQKV_TO_CONTEXT_PLUGIN_NAME << std::endl;
return false;
}
if (in->desc.dims.d[SDIM] == -1)
{
gLogError << "INT8 IO not support dynamic shape in sequence dimension for plugin "
<< kQKV_TO_CONTEXT_PLUGIN_NAME << std::endl;
return false;
}
if (packedSize == unfusedMaskSize)
{
gLogError << "INT8 IO only support sequence length 128,384 for plugin " << kQKV_TO_CONTEXT_PLUGIN_NAME
<< std::endl;
return false;
}
}
if (pos == 0)
{
bool isFormatSupported = in->desc.format == TensorFormat::kLINEAR;
if (mType == DataType::kINT8)
{
if (in->desc.dims.d[HDIM] % 32U == 0)
{
isFormatSupported = in->desc.format == TensorFormat::kCHW32;
}
else
{
isFormatSupported = in->desc.format == TensorFormat::kCHW4;
}
}
// must not check descriptions > pos
return (in->desc.type == mType) && // precision
isFormatSupported && // format
(in->desc.dims.nbDims == 5) && // num dims
((in->desc.dims.d[HDIM] % 3U) == 0) && // see getOutputDimensions
((in->desc.dims.d[3]) == 1) && // for fc
((in->desc.dims.d[4]) == 1) // for fc
;
}
// pos==1
if ((mHasImask && pos == 1)) // pos 1 is the mask
{
auto const* inMask = &inOut[1].desc;
if (inMask->dims.d[1] != -1 && inMask->dims.d[1] != packedSize)
{
gLogError << "CustomEmbLayerNormPluginDynamic returned mask with pack size " << inMask->dims.d[1]
<< ", but " << kQKV_TO_CONTEXT_PLUGIN_NAME << " expects mask pack size " << packedSize
<< std::endl;
return false;
}
// detect full mask and check that it was produced
return (inMask->type == DataType::kINT32) && // precision
(inMask->format == TensorFormat::kLINEAR) && // format
(inMask->dims.nbDims == 2) && // Bx2*maskSize
(inMask->dims.d[0] == in->desc.dims.d[BDIM]);
}
if (!mHasImask || pos == 2) // output pos
{
bool isFormatSupported = out->desc.format == TensorFormat::kLINEAR;
if (mType == DataType::kINT8)
{
if (out->desc.dims.d[HDIM] % 32U == 0)
{
isFormatSupported = out->desc.format == TensorFormat::kCHW32;
}
else
{
isFormatSupported = out->desc.format == TensorFormat::kCHW4;
}
}
return (in->desc.type == out->desc.type) && // precision
isFormatSupported && // format
(out->desc.dims.nbDims == 5) && // num dims
((in->desc.dims.d[HDIM] / 3) == (out->desc.dims.d[HDIM])) && // div 3
((out->desc.dims.d[3]) == 1) && // for fc
((out->desc.dims.d[4]) == 1) && // for fc
((out->desc.dims.d[BDIM]) == in->desc.dims.d[BDIM]) && // check B
((out->desc.dims.d[SDIM]) == in->desc.dims.d[SDIM]) // check S
;
}
return false;
}
int32_t QKVToContextPluginDynamic::onShapeChange(
PluginTensorDesc const* in, int32_t nbInputs, PluginTensorDesc const* out, int32_t nbOutputs) noexcept
{
try
{
PLUGIN_ASSERT(in != nullptr);
PLUGIN_ASSERT(nbInputs == 1 + mHasImask);
PLUGIN_ASSERT(nbOutputs == 1);
PluginTensorDesc const& inDesc = in[kIIDX];
TRT_UNUSED inDesc;
PLUGIN_ASSERT(out != nullptr);
PluginTensorDesc const& outDesc = out[0];
TRT_UNUSED outDesc;
PLUGIN_ASSERT(mType == inDesc.type);
PLUGIN_ASSERT(mType == outDesc.type);
PLUGIN_ASSERT(inDesc.dims.d[BDIM] == outDesc.dims.d[BDIM]);
PLUGIN_ASSERT(inDesc.dims.d[SDIM] == outDesc.dims.d[SDIM]);
PLUGIN_ASSERT(inDesc.dims.d[HDIM] == 3 * outDesc.dims.d[HDIM]);
if (mHasImask)
{
PluginTensorDesc const& maskDesc = in[kMIDX];
TRT_UNUSED maskDesc;
PLUGIN_ASSERT(maskDesc.dims.d[0] == inDesc.dims.d[BDIM]);
}
createMHARunner();
// mS and mB that are set by configurePlugin() may be stale
mS = inDesc.dims.d[SDIM];
mB = inDesc.dims.d[BDIM];
PLUGIN_ASSERT(mS);
PLUGIN_ASSERT(mB);
if (fusedDispatcher.get() && fusedDispatcher->isValid(mHeadSize, mS))
{
fusedDispatcher->setup(mS, mB, mHeadSize);
}
else
{
unfusedDispatcher->setup(mS, mB, mHeadSize);
}
return pluginStatus_t::STATUS_SUCCESS;
}
catch (std::exception const& e)
{
caughtError(e);
}
return pluginStatus_t::STATUS_FAILURE;
}
int32_t QKVToContextPluginDynamic::configurePlugin(
DynamicPluginTensorDesc const* in, int32_t nbInputs, DynamicPluginTensorDesc const* out, int32_t nbOutputs) noexcept
{
try
{
PLUGIN_ASSERT(in != nullptr);
PLUGIN_ASSERT(nbInputs == 1 + mHasImask);
PLUGIN_ASSERT(nbOutputs == 1);
PluginTensorDesc const& inDesc = in[kIIDX].desc;
TRT_UNUSED inDesc;
PLUGIN_ASSERT(out != nullptr);
PluginTensorDesc const& outDesc = out->desc;
TRT_UNUSED outDesc;
PLUGIN_ASSERT(mType == inDesc.type);
PLUGIN_ASSERT(mType == outDesc.type);
PLUGIN_ASSERT(inDesc.dims.d[BDIM] == outDesc.dims.d[BDIM]);
PLUGIN_ASSERT(inDesc.dims.d[SDIM] == outDesc.dims.d[SDIM]);
PLUGIN_ASSERT(inDesc.dims.d[HDIM] == 3 * outDesc.dims.d[HDIM]);
if (mHasImask)
{
PluginTensorDesc const& maskDesc = in[kMIDX].desc;
TRT_UNUSED maskDesc;
PLUGIN_ASSERT(maskDesc.dims.d[0] == inDesc.dims.d[BDIM]);
}
createMHARunner();
const int32_t S = inDesc.dims.d[SDIM];
const int32_t B = inDesc.dims.d[BDIM] <= 0 ? in->max.d[BDIM] : inDesc.dims.d[BDIM];
if (S <= 0)
{
// in dynamic shape build stage, we setup with max sequence that cannot fused
const int32_t Smin = in->min.d[SDIM];
const int32_t Smax = in->max.d[SDIM];
if (fusedDispatcher.get())
{
for (int32_t i = Smax; i >= Smin; --i)
{
if (!fusedDispatcher->isValid(mHeadSize, i))
{
unfusedDispatcher->setup(i, B, mHeadSize);
mS = i;
mB = B;
break;
}
}
}
else
{
unfusedDispatcher->setup(Smax, B, mHeadSize);
mS = Smax;
mB = B;
}
}
else
{
// in inference stage or in static shape build stage
if (fusedDispatcher.get() && fusedDispatcher->isValid(mHeadSize, S))
{
fusedDispatcher->setup(S, B, mHeadSize);
}
else
{
unfusedDispatcher->setup(S, B, mHeadSize);
}
mS = S;
mB = B;
}
return pluginStatus_t::STATUS_SUCCESS;
}
catch (std::exception const& e)
{
caughtError(e);
}
return pluginStatus_t::STATUS_FAILURE;
}
size_t QKVToContextPluginDynamic::getWorkspaceSize(DynamicPluginTensorDesc const* /*inputs*/, int32_t /*nbInputs*/,
DynamicPluginTensorDesc const* /*outputs*/, int32_t /*nbOutputs*/) const noexcept
{
// only unfused kernel need workspace, and we need larger workspace for larger sequence length
// we have already setup unfusedDispatcher with max sequence in configurePlugin
// if unfusedDispatcher is not initialized in configurePlugin
PLUGIN_ASSERT(unfusedDispatcher.get());
return unfusedDispatcher->getWorkspaceSize();
}
// IPluginV2Ext Methods
int32_t QKVToContextPluginDynamic::getOutputDataTypes(
DataType* outputTypes, int32_t nbOutputs, DataType const* inputTypes, int32_t nbInputs) const noexcept
{
try
{
PLUGIN_ASSERT(
inputTypes[0] == DataType::kFLOAT || inputTypes[0] == DataType::kHALF || inputTypes[0] == DataType::kINT8);
outputTypes[0] = inputTypes[0];
return pluginStatus_t::STATUS_SUCCESS;
}
catch (std::exception const& e)
{
caughtError(e);
}
return pluginStatus_t::STATUS_FAILURE;
}
void QKVToContextPluginDynamic::setCublasResources(std::shared_ptr<CublasWrapper> cublasWrapper)
{
mCublasWrapper = cublasWrapper;
// The shared cublasWrapper resource owns the handle.
// but `this` instance has a non-owning pointer to the handle.
// Note that the cublasWrapper inits the handle and checks for nullptr
// so we don't have to do that here.
mCublasHandle = mCublasWrapper->getCublasHandle();
}
IPluginV3* QKVToContextPluginDynamic::attachToContext(IPluginResourceContext* context) noexcept
{
try
{
auto p = static_cast<QKVToContextPluginDynamic*>(clone());
// the clone has shared ownership of underling cublasWrapper instance
// that is mapped to current context
p->setCublasResources(createPluginCublasWrapper(context));
return p;
}
catch (const std::exception& e)
{
caughtError(e);
}
return nullptr;
}
char const* QKVToContextPluginDynamic::getPluginVersion() const noexcept
{
return kQKV_TO_CONTEXT_PLUGIN_VERSION;
}
int32_t QKVToContextPluginDynamic::getNbOutputs() const noexcept
{
return 1;
}
char const* QKVToContextPluginDynamic::getPluginName() const noexcept
{
return kQKV_TO_CONTEXT_PLUGIN_NAME;
}
void QKVToContextPluginDynamic::setPluginNamespace(char const* libNamespace) noexcept
{
mNamespace = libNamespace;
}
char const* QKVToContextPluginDynamic::getPluginNamespace() const noexcept
{
return mNamespace.c_str();
}
int32_t QKVToContextPluginDynamic::enqueue(PluginTensorDesc const* inputDesc, PluginTensorDesc const* outputDesc,
void const* const* inputs, void* const* outputs, void* workspace, cudaStream_t stream) noexcept
{
PLUGIN_VALIDATE(inputDesc != nullptr && outputDesc != nullptr && inputs != nullptr && outputs != nullptr);
PLUGIN_ASSERT(mS == inputDesc->dims.d[SDIM]);
PLUGIN_ASSERT(mB == inputDesc->dims.d[BDIM]);
try
{
void const* const maskPtr = mHasImask ? inputs[1] : nullptr;
if (mHasImask && fusedDispatcher.get() && fusedDispatcher->isValid(mHeadSize, inputDesc->dims.d[SDIM]))
{
fusedDispatcher->run(
inputDesc[0], outputDesc[0], inputs[0], maskPtr, outputs[0], workspace, stream, mCublasHandle);
}
else
{
PLUGIN_VALIDATE(unfusedDispatcher.get(), "The Unfused MHARunner is uninitialized, no MHARunner available!");
PLUGIN_VALIDATE(mType != DataType::kINT8, "The Unfused MHARunner does not support INT8!");
unfusedDispatcher->run(
inputDesc[0], outputDesc[0], inputs[0], maskPtr, outputs[0], workspace, stream, mCublasHandle);
}
}
catch (std::exception const& e)
{
caughtError(e);
return -1;
}
return 0;
}
PluginFieldCollection const* QKVToContextPluginDynamic::getFieldsToSerialize() noexcept
{
mDataToSerialize.clear();
mDataToSerialize.emplace_back("type_id", &mType, PluginFieldType::kINT32, 1);
mDataToSerialize.emplace_back("hidden_size", &mHiddenSize, PluginFieldType::kINT32, 1);
mDataToSerialize.emplace_back("num_heads", &mNumHeads, PluginFieldType::kINT32, 1);
mDataToSerialize.emplace_back("has_mask", &mHasImask, PluginFieldType::kINT32, 1);
mDataToSerialize.emplace_back("S", &mS, PluginFieldType::kINT32, 1);
mDataToSerialize.emplace_back("B", &mB, PluginFieldType::kINT32, 1);
mDataToSerialize.emplace_back("SM", &mSM, PluginFieldType::kINT32, 1);
if (unfusedDispatcher.get() && unfusedDispatcher->getWorkspaceSize())
{
mHasUnfusedDispatcher = 1;
mRunnerStateBuffer.resize(unfusedDispatcher->getSerializationSize());
unfusedDispatcher->serialize(mRunnerStateBuffer.data());
mDataToSerialize.emplace_back("runnerStateBuffer", (void const*) mRunnerStateBuffer.data(),
PluginFieldType::kUNKNOWN, mRunnerStateBuffer.size());
}
else
{
mHasUnfusedDispatcher = 0;
}
mDataToSerialize.emplace_back("hasUnfusedDispatcher", &mHasUnfusedDispatcher, PluginFieldType::kINT32, 1);
if (mDqProbs >= 0)
{
mDataToSerialize.emplace_back("dq_probs", &mDqProbs, PluginFieldType::kFLOAT32, 1);
}
mFCToSerialize.nbFields = mDataToSerialize.size();
mFCToSerialize.fields = mDataToSerialize.data();
return &mFCToSerialize;
}
QKVToContextPluginDynamicCreator::QKVToContextPluginDynamicCreator()
{
mPluginAttributes.emplace_back(PluginField("type_id", nullptr, PluginFieldType::kINT32, 1));
mPluginAttributes.emplace_back(PluginField("hidden_size", nullptr, PluginFieldType::kINT32, 1));
mPluginAttributes.emplace_back(PluginField("num_heads", nullptr, PluginFieldType::kINT32, 1));
mPluginAttributes.emplace_back(PluginField("has_mask", nullptr, PluginFieldType::kINT32, 1));
mPluginAttributes.emplace_back(PluginField("dq_probs", nullptr, PluginFieldType::kFLOAT32, 1));
mFC.nbFields = mPluginAttributes.size();
mFC.fields = mPluginAttributes.data();
}
char const* QKVToContextPluginDynamicCreator::getPluginName() const noexcept
{
return kQKV_TO_CONTEXT_PLUGIN_NAME;
}
char const* QKVToContextPluginDynamicCreator::getPluginVersion() const noexcept
{
return kQKV_TO_CONTEXT_PLUGIN_VERSION;
}
PluginFieldCollection const* QKVToContextPluginDynamicCreator::getFieldNames() noexcept
{
return &mFC;
}
IPluginV3* QKVToContextPluginDynamicCreator::createPlugin(
char const* name, PluginFieldCollection const* fc, TensorRTPhase phase) noexcept
{
try
{
BERT_DEBUG_MSG("Creating QKV2ContextPlugin...");
PLUGIN_VALIDATE(fc != nullptr);
int32_t hiddenSize = 0;
// Since numHeads must always exist or validateRequiredAttributes will fail,
// we can set numHeads to -1 so that static analysis tools don't warn about
// a division by zero in QKVToContextPluginDynamic constructor.
int32_t numHeads{-1};
bool hasMask = false;
int32_t typeId = -1;
int32_t s = -1;
int32_t b = -1;
int32_t sm = -1;
bool hasUnfusedDispatcher = false;
void const* runnerStateBuffer = nullptr;
float dqProbs = -1;
PLUGIN_VALIDATE(fc->fields != nullptr);
if (phase == TensorRTPhase::kBUILD)
{
plugin::validateRequiredAttributesExist({"type_id", "hidden_size", "num_heads", "has_mask"}, fc);
}
else
{
PLUGIN_ASSERT(phase == TensorRTPhase::kRUNTIME);
plugin::validateRequiredAttributesExist(
{"type_id", "S", "B", "hidden_size", "num_heads", "has_mask", "SM", "hasUnfusedDispatcher"}, fc);
}
for (int32_t i = 0; i < fc->nbFields; i++)
{
PLUGIN_VALIDATE(fc->fields[i].name != nullptr);
PLUGIN_VALIDATE(fc->fields[i].data != nullptr);
std::string field_name(fc->fields[i].name);
if (field_name.compare("type_id") == 0)
{
typeId = *static_cast<int32_t const*>(fc->fields[i].data);
PLUGIN_VALIDATE(typeId >= 0 && typeId <= 2, ("QKV: Invalid TypeId " + std::to_string(typeId)).c_str());
BERT_DEBUG_VALUE("Building typeId: ", typeId);
}
else if (field_name.compare("hidden_size") == 0)
{
hiddenSize = *static_cast<int32_t const*>(fc->fields[i].data);
PLUGIN_VALIDATE(hiddenSize > 0, ("QKV: Invalid hiddenSize " + std::to_string(hiddenSize)).c_str());
BERT_DEBUG_VALUE("Building hiddenSize: ", hiddenSize);
}
else if (field_name.compare("num_heads") == 0)
{
numHeads = *static_cast<int32_t const*>(fc->fields[i].data);
PLUGIN_VALIDATE(numHeads > 0, ("QKV: Invalid numHeads " + std::to_string(numHeads)).c_str());
BERT_DEBUG_VALUE("Building numHeads: ", numHeads);
}
else if (field_name.compare("has_mask") == 0)
{
auto hasMaskValue = *static_cast<int32_t const*>(fc->fields[i].data);
PLUGIN_VALIDATE(hasMaskValue == 0 || hasMaskValue == 1,
("QKV: Invalid hasMask " + std::to_string(hasMaskValue)).c_str());
hasMask = static_cast<bool>(hasMaskValue);
BERT_DEBUG_VALUE("Building hasMask: ", hasMask);
}
else if (field_name.compare("dq_probs") == 0)
{
dqProbs = *static_cast<float const*>(fc->fields[i].data);
PLUGIN_VALIDATE(dqProbs > 0.0F, ("QKV: Invalid dqProbs " + std::to_string(dqProbs)).c_str());
BERT_DEBUG_VALUE("Building dqProbs: ", dqProbs);
}
else if (field_name.compare("S") == 0)
{
PLUGIN_ASSERT(phase == TensorRTPhase::kRUNTIME);
s = *static_cast<int32_t const*>(fc->fields[i].data);
BERT_DEBUG_VALUE("Building S: ", s);
}
else if (field_name.compare("B") == 0)
{
PLUGIN_ASSERT(phase == TensorRTPhase::kRUNTIME);
b = *static_cast<int32_t const*>(fc->fields[i].data);
BERT_DEBUG_VALUE("Building B: ", b);
}
else if (field_name.compare("SM") == 0)
{
PLUGIN_ASSERT(phase == TensorRTPhase::kRUNTIME);
sm = *static_cast<int32_t const*>(fc->fields[i].data);
BERT_DEBUG_VALUE("Building SM: ", sm);
}
else if (field_name.compare("hasUnfusedDispatcher") == 0)
{
PLUGIN_ASSERT(phase == TensorRTPhase::kRUNTIME);
auto hasUnfusedDispatcherValue = *static_cast<int32_t const*>(fc->fields[i].data);
PLUGIN_VALIDATE(hasUnfusedDispatcherValue == 0 || hasUnfusedDispatcherValue == 1,
("QKV: Invalid hasUnfusedDispatcher " + std::to_string(hasUnfusedDispatcherValue)).c_str());
hasUnfusedDispatcher = static_cast<bool>(hasUnfusedDispatcherValue);
BERT_DEBUG_VALUE("Building hasUnfusedDispatcher: ", hasUnfusedDispatcher);
}
else if (field_name.compare("runnerStateBuffer") == 0)
{
PLUGIN_ASSERT(phase == TensorRTPhase::kRUNTIME);
runnerStateBuffer = static_cast<void const*>(fc->fields[i].data);
}
}
BERT_DEBUG_MSG("Building the Plugin...");
auto type = static_cast<DataType>(typeId);
if (type == DataType::kINT8 && dqProbs < 0)
{
BERT_DEBUG_MSG("Using default scale factor");
dqProbs = 1.F / 127.F;
}
if (phase == TensorRTPhase::kBUILD)
{
return new QKVToContextPluginDynamic(name, type, hiddenSize, numHeads, dqProbs, hasMask);
}
PLUGIN_VALIDATE(s != -1, "invalid S during runtime plugin creation");
PLUGIN_VALIDATE(b != -1, "invalid B during runtime plugin creation");
PLUGIN_VALIDATE(sm != -1, "invalid SM during runtime plugin creation");
if (hasUnfusedDispatcher == 1)
{
PLUGIN_VALIDATE(runnerStateBuffer != nullptr, "invalid runnerStateBuffer during runtime plugin creation");
}
else
{
PLUGIN_VALIDATE(runnerStateBuffer == nullptr, "invalid runnerStateBuffer during runtime plugin creation");
}
return new QKVToContextPluginDynamic(
name, type, s, b, sm, hiddenSize, numHeads, dqProbs, hasMask, hasUnfusedDispatcher, runnerStateBuffer);
}
catch (std::exception const& e)
{
caughtError(e);
}
return nullptr;
}
void QKVToContextPluginDynamicCreator::setPluginNamespace(char const* libNamespace) noexcept
{
mNamespace = libNamespace;
}
char const* QKVToContextPluginDynamicCreator::getPluginNamespace() const noexcept
{
return mNamespace.c_str();
}
///// QKVToContextVarSeqlenPlugin (CustomQKVToContextPluginDynamic v5) ////
QKVToContextVarSeqlenPlugin::~QKVToContextVarSeqlenPlugin() {}
QKVToContextVarSeqlenPlugin::QKVToContextVarSeqlenPlugin(std::string const name, DataType const type,
int32_t const hiddenSize, int32_t const numHeads, float const dqProbs, bool hasImask, bool varSeqlen,
bool useInt8ScaleMax)
: mLayerName(name)
, mS(0)
, mB(0)
, mHeadSize(hiddenSize / numHeads)
, mHiddenSize(hiddenSize)
, mNumHeads(numHeads)
, mType(type)
, mDqProbs(dqProbs)
, mHdim(HDIM)
{
mSM = getSMVersion();
mUseVarSeqlen = static_cast<int32_t>(varSeqlen);
mUseInt8ScaleMax = static_cast<int32_t>(useInt8ScaleMax);
mHasImask = static_cast<int32_t>(hasImask);
if (varSeqlen)
{
// variable sequence length is only supported with the fused MHA kernels
// we should not override mS!
bool isSMSupported = elem(mSM, {kSM_75, kSM_80, kSM_86, kSM_87, kSM_89, kSM_90, kSM_100, kSM_120});
PLUGIN_ASSERT(isSMSupported && (type == DataType::kINT8 || type == DataType::kHALF)
&& "requesting maxSeqlen not compatible with GPU arch");
// the layout changes: SxB will be a combined \sum_i s_i and hdim will be the 2nd dimension instead of the third
mHdim = 1;
}
}
QKVToContextVarSeqlenPlugin::QKVToContextVarSeqlenPlugin(std::string const name, int32_t const S, int32_t const B,
DataType const type, int32_t const hiddenSize, int32_t const numHeads, float const dqProbs, bool hasImask,
bool varSeqlen, bool useInt8ScaleMax, void const* runnerStateBuffer)
: mLayerName(name)
, mS(S)
, mB(B)
, mHeadSize(hiddenSize / numHeads)
, mHiddenSize(hiddenSize)
, mNumHeads(numHeads)
, mType(type)
, mDqProbs(dqProbs)
, mHdim(HDIM)
{
mSM = getSMVersion();
mUseVarSeqlen = static_cast<int32_t>(varSeqlen);
mUseInt8ScaleMax = static_cast<int32_t>(useInt8ScaleMax);
mHasImask = static_cast<int32_t>(hasImask);
if (varSeqlen)
{
// variable sequence length is only supported with the fused MHA kernels
// we should not override mS!
bool isSMSupported = elem(mSM, {kSM_75, kSM_80, kSM_86, kSM_87, kSM_89, kSM_90, kSM_100, kSM_120});
PLUGIN_ASSERT(isSMSupported && (type == DataType::kINT8 || type == DataType::kHALF)
&& "requesting maxSeqlen not compatible with GPU arch");
// the layout changes: SxB will be a combined \sum_i s_i and hdim will be the 2nd dimension instead of the third
mHdim = 1;
}
createMHARunner();
PLUGIN_ASSERT(runnerStateBuffer != nullptr);
auto length = mDispatcher->getSerializationSize();
mDispatcher->deserialize(runnerStateBuffer, length);
}
IPluginCapability* QKVToContextVarSeqlenPlugin::getCapabilityInterface(PluginCapabilityType type) noexcept
{
try
{
if (type == PluginCapabilityType::kBUILD)
{
return static_cast<IPluginV3OneBuild*>(this);
}
if (type == PluginCapabilityType::kRUNTIME)
{
return static_cast<IPluginV3OneRuntime*>(this);
}
PLUGIN_ASSERT(type == PluginCapabilityType::kCORE);
return static_cast<IPluginV3OneCore*>(this);
}
catch (std::exception const& e)
{
caughtError(e);
}
return nullptr;
}
void QKVToContextVarSeqlenPlugin::createMHARunner()
{
if (mDispatcher.get())
{
return;
}
if (mUseVarSeqlen)
{
PLUGIN_ASSERT(mHeadSize <= 64);
{
if (mHeadSize != 64)
{
mPatcher.reset(new QkvPaddingRunner(mType));
}
if (mType == DataType::kHALF)
{
mDispatcher.reset(new FusedMHARunnerFP16v2(mNumHeads, mSM));
}
else if (mType == DataType::kINT8)
{
mDispatcher.reset(new FusedMHARunnerInt8v2(mNumHeads, mSM, mDqProbs, mUseInt8ScaleMax));
}
}
}
else
{
PLUGIN_ASSERT(mType != DataType::kINT8);
mDispatcher.reset(new UnfusedMHARunner(mType, mNumHeads, mSM));
}
}
IPluginV3* QKVToContextVarSeqlenPlugin::clone() noexcept
{
BERT_DEBUG_MSG("QKV Clone");
QKVToContextVarSeqlenPlugin* ret = nullptr;
char* bufferData = nullptr;
// the workspacesize is 0 if we have not call setup the dispatcher yet.
if (mDispatcher.get())
{
mRunnerStateBuffer.resize(mDispatcher->getSerializationSize());
mDispatcher->serialize(mRunnerStateBuffer.data());
bufferData = mRunnerStateBuffer.data();
ret = new QKVToContextVarSeqlenPlugin(mLayerName, mS, mB, mType, mHiddenSize, mNumHeads, mDqProbs, mHasImask,
mUseVarSeqlen, mUseInt8ScaleMax, static_cast<void const*>(bufferData));
}
else
{
// dispatcher not setup yet, use type 1 constructor
ret = new QKVToContextVarSeqlenPlugin(
mLayerName, mType, mHiddenSize, mNumHeads, mDqProbs, mHasImask, mUseVarSeqlen, mUseInt8ScaleMax);
}
ret->setPluginNamespace(mNamespace.c_str());
BERT_DEBUG_MSG("QKV Clone done");
return ret;
}
int32_t QKVToContextVarSeqlenPlugin::getOutputShapes(DimsExprs const* inputs, int32_t nbInputs,
DimsExprs const* shapeInputs, int32_t nbShapeInputs, DimsExprs* outputs, int32_t nbOutputs,
IExprBuilder& exprBuilder) noexcept
{
try
{
PLUGIN_ASSERT(inputs != nullptr);
PLUGIN_ASSERT(nbInputs == 1 + mHasImask + 2 * mUseVarSeqlen);
PLUGIN_ASSERT(nbShapeInputs == 0);
PLUGIN_ASSERT(outputs != nullptr);
PLUGIN_ASSERT(nbOutputs == 1);
// Input is BxSx3*N*H, output should be BxSxN*H
// Copy over everything
outputs[kIIDX] = inputs[kIIDX];
// Divide last dim by three
auto const* three = exprBuilder.constant(3);
// mHdim is 2 for fixed seqlen and is 1 for varseqlen
outputs[kIIDX].d[mHdim]
= exprBuilder.operation(DimensionOperation::kFLOOR_DIV, *inputs[kIIDX].d[mHdim], *three);
return pluginStatus_t::STATUS_SUCCESS;
}
catch (std::exception const& e)
{
caughtError(e);
}
return pluginStatus_t::STATUS_FAILURE;
}
bool QKVToContextVarSeqlenPlugin::supportsFormatCombination(
int32_t pos, DynamicPluginTensorDesc const* inOut, int32_t nbInputs, int32_t nbOutputs) noexcept
{
// we only support variable sequence and int8 IO in fused mha runner, and we only support fused mha runner on
// Turing, Ampere, Hopper and Blackwell
bool const hasV2Kernels = elem(mSM, {kSM_75, kSM_80, kSM_86, kSM_87, kSM_89, kSM_90, kSM_100, kSM_120});
PLUGIN_ASSERT((mType != DataType::kINT8 || hasV2Kernels)
&& "INT8 IO is only supported on Xavier, Turing, Ampere, Hopper and Blackwell!");
PLUGIN_ASSERT((!mUseVarSeqlen || hasV2Kernels)
&& "Variable sequence is only supported on Xavier, Turing, Ampere, Hopper and Blackwell!");
PLUGIN_ASSERT(pos >= 0);
PLUGIN_ASSERT(pos < 2 + mHasImask + 2 * mUseVarSeqlen);
PLUGIN_ASSERT(nbInputs == 1 + mHasImask + 2 * mUseVarSeqlen);
PLUGIN_ASSERT(nbOutputs == 1);
auto const* in = inOut;
auto const* out = inOut + nbInputs;
if (mUseVarSeqlen)
{
PLUGIN_ASSERT((mType == DataType::kHALF || mType == DataType::kINT8)
&& "Conditions for variable seqlen support not fulfilled");
// qkv, mask, cu_seqlens, dummy
PLUGIN_ASSERT(nbInputs == 4 && "for varseqlen, expected 4 inputs");
}
auto const inDims = in->desc.dims;
auto const outDims = out->desc.dims;
auto supportedFormat = TensorFormat::kLINEAR;
if (mType == DataType::kINT8)
{
supportedFormat = (inDims.d[mHdim] % 32U == 0) ? TensorFormat::kCHW32 : TensorFormat::kCHW4;
}
int32_t supportedNbDims = 5;
if (mUseVarSeqlen)
{
supportedNbDims = 4;
}
bool supportedHdim = (pos == 0) ? (inDims.d[mHdim] % 3U == 0) : (inDims.d[mHdim] / 3 == outDims.d[mHdim]);
if (pos == 0 || pos == nbInputs)
{ // check input and output
auto const& desc = inOut[pos].desc;
return (desc.type == mType) && // check type
(desc.format == supportedFormat) && // check format
(desc.dims.nbDims == supportedNbDims) && // check dims:
(supportedHdim) && // - hidden dims multiple of 3 for qkv
(desc.dims.d[mHdim + 1] == 1) && // - dummy 1 or h
(desc.dims.d[mHdim + 2] == 1) // - dummy 1 or w
;