Skip to content

Commit f64cd19

Browse files
tangzhiyi11tangzhiyi11
committed
code format
1 parent 77c7e1f commit f64cd19

File tree

2 files changed

+64
-71
lines changed

2 files changed

+64
-71
lines changed

‎dlinfer/graph/dicp/vendor/AtbGraph/codegen/runtime/ops/custom_ops/fused_lora_operation.cpp‎

Lines changed: 60 additions & 67 deletions
Original file line numberDiff line numberDiff line change
@@ -1,20 +1,20 @@
11
#include "fused_lora_operation.h"
22

3+
#include <atb/utils.h>
4+
5+
#include <algorithm>
36
#include <cstdint>
47
#include <unordered_set>
5-
#include <algorithm>
68

7-
#include "aclnnop/aclnn_mul.h"
89
#include "aclnnop/aclnn_grouped_matmul_v4.h"
10+
#include "aclnnop/aclnn_mul.h"
911
#include "aclnnop/aclnn_permute.h"
1012
#include "ops/operation_creator.h"
1113
#include "third_party/acl/inc/acl/acl_base.h"
1214
#include "utils/common.h"
1315
#include "utils/log.h"
1416
#include "utils/scalar.h"
1517

16-
#include <atb/utils.h>
17-
1818
namespace dicp {
1919

2020
const int NUM1 = 1;
@@ -72,10 +72,10 @@ int CustomFusedLoraOperation::CreateAclTensors(const atb::VariantPack& variantPa
7272

7373
const size_t inTensorCount = variantPack.inTensors.size();
7474
const size_t outTensorCount = variantPack.outTensors.size();
75-
75+
7676
aclInTensors_.resize(inTensorCount);
7777
aclOutTensors_.resize(outTensorCount);
78-
78+
7979
for (size_t i = 0; i < inTensorCount; ++i) {
8080
aclInTensors_[i] = CreateTensor(variantPack.inTensors.at(i));
8181
}
@@ -104,7 +104,7 @@ void CustomFusedLoraOperation::ClearInternal() {
104104
aclWeightATranspose_.clear();
105105
weightA_.clear();
106106
weightB_.clear();
107-
weightATranspose_.clear();
107+
weightATranspose_.clear();
108108

109109
aclScalingInput_.clear();
110110
scalingInput_.clear();
@@ -183,7 +183,7 @@ int CustomFusedLoraOperation::Setup(const atb::VariantPack& variantPack, uint64_
183183
const int64_t loraBDim = variantPack.inTensors.at(2).desc.shape.dims[1];
184184

185185
ClearInternal();
186-
186+
187187
// Pre-allocate vectors to avoid reallocations
188188
weightA_.reserve(adapterIdsVec.size());
189189
weightATranspose_.reserve(adapterIdsVec.size());
@@ -198,7 +198,6 @@ int CustomFusedLoraOperation::Setup(const atb::VariantPack& variantPack, uint64_
198198
aclScalingWeight_.reserve(adapterIdsVec.size());
199199
aclScalingInput_.reserve(adapterIdsVec.size());
200200

201-
202201
bool singleInfer = adapterIdsVec.size() == 1;
203202
int32_t totalRanks = 0;
204203

@@ -284,26 +283,22 @@ int CustomFusedLoraOperation::Setup(const atb::VariantPack& variantPack, uint64_
284283
} else {
285284
permuteDims = {1, 0};
286285
}
287-
aclIntArray *permuteDimsArray = aclCreateIntArray(permuteDims.data(), permuteDims.size());
286+
aclIntArray* permuteDimsArray = aclCreateIntArray(permuteDims.data(), permuteDims.size());
288287
for (const auto& [adapterId, weightATransposeIndex] : weightATransposeIdMap_) {
289288
aclWeightAPermuteExecutor_[adapterId] = nullptr;
290289
aclWeightAPermuteWorkspace_[adapterId] = 0;
291290

292291
auto& weightA = aclWeightA_[weightATransposeIndex];
293292
auto& weightATranspose = aclWeightATranspose_[weightATransposeIndex];
294293

295-
296-
int ret = aclnnPermuteGetWorkspaceSize(weightA.tensor,
297-
permuteDimsArray,
298-
weightATranspose.tensor,
299-
&aclWeightAPermuteWorkspace_[adapterId],
300-
&aclWeightAPermuteExecutor_[adapterId]);
294+
int ret = aclnnPermuteGetWorkspaceSize(
295+
weightA.tensor, permuteDimsArray, weightATranspose.tensor, &aclWeightAPermuteWorkspace_[adapterId], &aclWeightAPermuteExecutor_[adapterId]);
301296
DICP_LOG(INFO) << opName_ << " aclnnPermuteGetWorkspaceSize size[" << adapterId << "]: " << aclWeightAPermuteWorkspace_[adapterId] << ", ret: " << ret;
302297
}
303298

304299
// Setup grouped matrix multiplication
305300
DICP_LOG(INFO) << opName_ << " Setting up grouped matrix multiplication";
306-
301+
307302
// Create input tensor list
308303
std::vector<aclTensor*> xTmp;
309304
if (singleInfer) {
@@ -317,7 +312,7 @@ int CustomFusedLoraOperation::Setup(const atb::VariantPack& variantPack, uint64_
317312
slicedInput.desc.format = aclInTensors_.at(0).atbTensor.desc.format;
318313
slicedInput.desc.shape.dimNum = aclInTensors_.at(0).atbTensor.desc.shape.dimNum;
319314
slicedInput.desc.shape.dims[0] = seqLensVec[i];
320-
slicedInput.desc.shape.dims[1] = aclInTensors_.at(0).atbTensor.desc.shape.dims[1];
315+
slicedInput.desc.shape.dims[1] = aclInTensors_.at(0).atbTensor.desc.shape.dims[1];
321316
slicedInput.dataSize = atb::Utils::GetTensorSize(slicedInput.desc);
322317

323318
auto offset = CalculateWeightOffset(seqLensVec, i, slicedInput.dataSize / seqLensVec[i]);
@@ -338,14 +333,14 @@ int CustomFusedLoraOperation::Setup(const atb::VariantPack& variantPack, uint64_
338333
std::vector<aclTensor*> weightTmpB;
339334
weightTmpA.reserve(aclWeightATranspose_.size());
340335
weightTmpB.reserve(aclWeightB_.size());
341-
336+
342337
for (const auto& weight : aclWeightATranspose_) {
343338
weightTmpA.push_back(weight.tensor);
344339
}
345340
for (const auto& weight : aclWeightB_) {
346341
weightTmpB.push_back(weight.tensor);
347342
}
348-
343+
349344
aclTensorList* weightTensorListA = aclCreateTensorList(weightTmpA.data(), weightTmpA.size());
350345
aclTensorList* weightTensorListB = aclCreateTensorList(weightTmpB.data(), weightTmpB.size());
351346

@@ -363,7 +358,7 @@ int CustomFusedLoraOperation::Setup(const atb::VariantPack& variantPack, uint64_
363358
loraASliceOutput.desc.format = aclOutTensors_.at(1).atbTensor.desc.format;
364359
loraASliceOutput.desc.shape.dimNum = aclOutTensors_.at(1).atbTensor.desc.shape.dimNum;
365360
loraASliceOutput.desc.shape.dims[0] = aclOutTensors_.at(1).atbTensor.desc.shape.dims[0];
366-
loraASliceOutput.desc.shape.dims[1] = totalRanks / adapterIdsVec.size();
361+
loraASliceOutput.desc.shape.dims[1] = totalRanks / adapterIdsVec.size();
367362
loraASliceOutput.dataSize = atb::Utils::GetTensorSize(loraASliceOutput.desc);
368363
loraASliceOutput.deviceData = aclOutTensors_.at(1).atbTensor.deviceData;
369364
auto aclnnLoraASliceOutput = CreateTensor(loraASliceOutput);
@@ -378,7 +373,7 @@ int CustomFusedLoraOperation::Setup(const atb::VariantPack& variantPack, uint64_
378373
slicedOutput.desc.format = aclOutTensors_.at(1).atbTensor.desc.format;
379374
slicedOutput.desc.shape.dimNum = aclOutTensors_.at(1).atbTensor.desc.shape.dimNum;
380375
slicedOutput.desc.shape.dims[0] = seqLensVec[i];
381-
slicedOutput.desc.shape.dims[1] = ranksVec[adapterIdsVec[i]];
376+
slicedOutput.desc.shape.dims[1] = ranksVec[adapterIdsVec[i]];
382377
slicedOutput.dataSize = atb::Utils::GetTensorSize(slicedOutput.desc);
383378

384379
auto offset = CalculateWeightOffset(seqLensVec, i, slicedOutput.dataSize / seqLensVec[i]);
@@ -398,59 +393,59 @@ int CustomFusedLoraOperation::Setup(const atb::VariantPack& variantPack, uint64_
398393
DICP_LOG(ERROR) << opName_ << " Failed to create output tensor lists";
399394
return -1;
400395
}
401-
396+
402397
// Setup LoRA A grouped matrix multiplication
403-
int ret = aclnnGroupedMatmulV4GetWorkspaceSize(xTensorList, // x
404-
weightTensorListA, // weight
405-
nullptr, // biasOptional
406-
nullptr, // scaleOptional
407-
nullptr, // offsetOptional
408-
nullptr, // antiquantScaleOptional
409-
nullptr, // antiquantOffsetOptional
410-
nullptr, // perTokenScaleOptional
411-
singleInfer ? aclInTensors_.at(5).tensor : nullptr, // groupListOptional
412-
nullptr, // activationInputOptional
413-
nullptr, // activationQuantScaleOptional
414-
nullptr, // activationQuantOffsetOptional
415-
singleInfer ? 2 : 0, // splitItem
416-
singleInfer ? 0 : -1, // groupType
417-
1, // groupListType
418-
0, // actType
419-
loraAOutTensorList, // out
420-
nullptr, // activationFeatureOutOptional
421-
nullptr, // dynQuantScaleOutOptional
398+
int ret = aclnnGroupedMatmulV4GetWorkspaceSize(xTensorList, // x
399+
weightTensorListA, // weight
400+
nullptr, // biasOptional
401+
nullptr, // scaleOptional
402+
nullptr, // offsetOptional
403+
nullptr, // antiquantScaleOptional
404+
nullptr, // antiquantOffsetOptional
405+
nullptr, // perTokenScaleOptional
406+
singleInfer ? aclInTensors_.at(5).tensor : nullptr, // groupListOptional
407+
nullptr, // activationInputOptional
408+
nullptr, // activationQuantScaleOptional
409+
nullptr, // activationQuantOffsetOptional
410+
singleInfer ? 2 : 0, // splitItem
411+
singleInfer ? 0 : -1, // groupType
412+
1, // groupListType
413+
0, // actType
414+
loraAOutTensorList, // out
415+
nullptr, // activationFeatureOutOptional
416+
nullptr, // dynQuantScaleOutOptional
422417
&loraAGroupedGemmWorkspace_,
423418
&aclLoraAGroupedGemmExecutor_);
424419
DICP_LOG(INFO) << opName_ << " LoRA A grouped matmul workspace size: " << loraAGroupedGemmWorkspace_ << ", ret: " << ret;
425420

426421
// Setup LoRA B grouped matrix multiplication
427-
ret = aclnnGroupedMatmulV4GetWorkspaceSize(loraAOutTensorList, // x
428-
weightTensorListB, // weight
429-
nullptr, // biasOptional
430-
nullptr, // scaleOptional
431-
nullptr, // offsetOptional
432-
nullptr, // antiquantScaleOptional
433-
nullptr, // antiquantOffsetOptional
434-
nullptr, // perTokenScaleOptional
435-
aclInTensors_.at(5).tensor, // groupListOptional
436-
nullptr, // activationInputOptional
437-
nullptr, // activationQuantScaleOptional
438-
nullptr, // activationQuantOffsetOptional
439-
2, // splitItem
440-
0, // groupType
441-
1, // groupListType
442-
0, // actType
443-
loraBOutTensorList, // out
444-
nullptr, // activationFeatureOutOptional
445-
nullptr, // dynQuantScaleOutOptional
422+
ret = aclnnGroupedMatmulV4GetWorkspaceSize(loraAOutTensorList, // x
423+
weightTensorListB, // weight
424+
nullptr, // biasOptional
425+
nullptr, // scaleOptional
426+
nullptr, // offsetOptional
427+
nullptr, // antiquantScaleOptional
428+
nullptr, // antiquantOffsetOptional
429+
nullptr, // perTokenScaleOptional
430+
aclInTensors_.at(5).tensor, // groupListOptional
431+
nullptr, // activationInputOptional
432+
nullptr, // activationQuantScaleOptional
433+
nullptr, // activationQuantOffsetOptional
434+
2, // splitItem
435+
0, // groupType
436+
1, // groupListType
437+
0, // actType
438+
loraBOutTensorList, // out
439+
nullptr, // activationFeatureOutOptional
440+
nullptr, // dynQuantScaleOutOptional
446441
&loraBGroupedGemmWorkspace_,
447442
&aclLoraBGroupedGemmExecutor_);
448443
DICP_LOG(INFO) << opName_ << " LoRA B grouped matmul workspace size: " << loraBGroupedGemmWorkspace_ << ", ret: " << ret;
449-
444+
450445
// Setup scaling operations
451446
aclScalingWorkspace_.resize(adapterIdsVec.size());
452447
aclScalingExecutor_.resize(adapterIdsVec.size());
453-
448+
454449
for (size_t i = 0; i < adapterIdsVec.size(); ++i) {
455450
const int32_t adapterId = adapterIdsVec[i];
456451
const auto& inputAtbTensor = aclOutTensors_.at(0).atbTensor;
@@ -494,10 +489,8 @@ int CustomFusedLoraOperation::Setup(const atb::VariantPack& variantPack, uint64_
494489
aclnnScalingWeight.CreateTensor(opName_);
495490
aclScalingWeight_.push_back(aclnnScalingWeight);
496491

497-
ret = aclnnInplaceMulGetWorkspaceSize(aclScalingInput_.back().tensor,
498-
aclScalingWeight_.back().tensor,
499-
&aclScalingWorkspace_[i],
500-
&aclScalingExecutor_[i]);
492+
ret =
493+
aclnnInplaceMulGetWorkspaceSize(aclScalingInput_.back().tensor, aclScalingWeight_.back().tensor, &aclScalingWorkspace_[i], &aclScalingExecutor_[i]);
501494
DICP_LOG(INFO) << opName_ << " Scaling workspace size[" << i << "]: " << aclScalingWorkspace_[i] << ", ret: " << ret;
502495
}
503496

‎dlinfer/graph/dicp/vendor/AtbGraph/codegen/runtime/ops/custom_ops/fused_lora_operation.h‎

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -1,8 +1,9 @@
11
#pragma once
22

3-
#include "ops/aclnn_ops/acl_nn_operation.h"
4-
#include <vector>
53
#include <unordered_map>
4+
#include <vector>
5+
6+
#include "ops/aclnn_ops/acl_nn_operation.h"
67

78
namespace dicp {
89

@@ -26,7 +27,7 @@ class CustomFusedLoraOperation : public atb::Operation {
2627
int CreateAclTensors(const atb::VariantPack& variantPack);
2728
void ClearAclScalrs();
2829
void ClearInternal();
29-
30+
3031
// Helper functions for weight tensor creation and offset calculation
3132
atb::Tensor CreateWeightTensor(const atb::Tensor& baseTensor, int64_t rank, int64_t dim, uint64_t offset);
3233
uint64_t CalculateWeightOffset(const std::vector<int32_t>& ranksVec, size_t adapterId, uint64_t tensorSizePerRank);
@@ -62,7 +63,6 @@ class CustomFusedLoraOperation : public atb::Operation {
6263

6364
std::vector<uint64_t> aclScalingWorkspace_;
6465
std::vector<aclOpExecutor*> aclScalingExecutor_;
65-
6666
};
6767

6868
} // namespace dicp

0 commit comments

Comments
 (0)