* Copyright (c) 2025 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
#include "flash_attention_score_grad.h"
#include "opdev/format_utils.h"
#include "opdev/make_op_executor.h"
#include "opdev/op_def.h"
#include "opdev/op_dfx.h"
#include "opdev/op_executor.h"
#include "opdev/op_log.h"
#include "opdev/shape_utils.h"
using namespace op;
#define CHECK_NULL(aclnTensor) do { if ((aclnTensor) == nullptr) { return {nullptr, nullptr, nullptr, nullptr};}} while (0)
namespace l0op {
OP_TYPE_REGISTER(FlashAttentionScoreGrad);
bool FakeArrayToTensor(const aclIntArray *inArray, aclTensor *&outTensor)
{
if (inArray != nullptr) {
int64_t size = static_cast<int64_t>(inArray->Size());
std::vector<int64_t> shape = {size};
outTensor = aclCreateTensor(shape.data(), shape.size(), aclDataType::ACL_INT64, nullptr,
0, ACL_FORMAT_ND, shape.data(), shape.size(), nullptr);
if (outTensor == nullptr) {
OP_LOGE(ACLNN_ERR_INNER_NULLPTR, "Try alloc tensor failed");
return false;
}
}
return true;
}
const std::array<const aclTensor *, MAX_FAG_OUTPUT_CNT> FlashAttentionScoreGrad(
const aclTensor *query, const aclTensor *key, const aclTensor *value, const aclTensor *dy,
const aclTensor *pseShiftOptional, const aclTensor *dropMaskOptional, const aclTensor *paddingMaskOptional,
const aclTensor *attenMaskOptional, const aclTensor *softmaxMaxOptional, const aclTensor *softmaxSumOptional,
const aclTensor *softmaxInOptional, const aclTensor *attentionInOptional, const aclIntArray *prefixOptional,
const aclIntArray *actualSeqQLenOptional, const aclIntArray *actualSeqKvLenOptional,
const aclIntArray *qStartIdxOptional, const aclIntArray *kvStartIdxOptional, const aclTensor *dScaleQOptional,
const aclTensor *dScaleKOptional, const aclTensor *dScaleVOptional, const aclTensor *dScaleDyOptional,
const aclTensor *dScaleOOptional, const aclTensor *dsScaleOptional, const aclTensor *pScaleOptional,
const aclTensor *queryRope, const aclTensor *keyRope, const aclTensor *sinkInOptional,
double scaleValueOptional, double keepProbOptional, int64_t preTockensOptional, int64_t nextTockensOptional, int64_t headNum,
char *inputLayout, int64_t innerPreciseOptional, int64_t sparseModeOptional, int64_t pseTypeOptional,
int64_t seed, int64_t offset, int64_t outDTypeOptional, char *softmaxInLayout, aclOpExecutor *executor,
bool isMaxWorkspace)
{
L0_DFX(FlashAttentionScoreGrad, query, key, value, dy, pseShiftOptional, dropMaskOptional, paddingMaskOptional,
attenMaskOptional, softmaxMaxOptional, softmaxSumOptional, softmaxInOptional, attentionInOptional,
prefixOptional, actualSeqQLenOptional, actualSeqKvLenOptional, qStartIdxOptional, kvStartIdxOptional,
dScaleQOptional, dScaleKOptional, dScaleVOptional, dScaleDyOptional, dScaleOOptional, dsScaleOptional, pScaleOptional,
queryRope, keyRope, scaleValueOptional, keepProbOptional, preTockensOptional, nextTockensOptional, headNum,
inputLayout, innerPreciseOptional, sparseModeOptional, pseTypeOptional, seed, offset, outDTypeOptional,softmaxInLayout, sinkInOptional);
DataType outputDtype = query->GetDataType();
if (outputDtype == DataType::DT_FLOAT8_E4M3FN || outputDtype == DataType::DT_FLOAT8_E5M2 || outputDtype == DataType::DT_HIFLOAT8) {
if (outDTypeOptional == 0) {
outputDtype = DataType::DT_FLOAT16;
} else {
outputDtype = DataType::DT_BF16;
}
}
auto dqOut = executor->AllocTensor(outputDtype, op::Format::FORMAT_ND, op::Format::FORMAT_ND);
auto dkOut = executor->AllocTensor(outputDtype, op::Format::FORMAT_ND, op::Format::FORMAT_ND);
auto dvOut = executor->AllocTensor(outputDtype, op::Format::FORMAT_ND, op::Format::FORMAT_ND);
auto dpseOut = executor->AllocTensor(outputDtype, op::Format::FORMAT_ND, op::Format::FORMAT_ND);
auto dqRopeOut = executor->AllocTensor(outputDtype, op::Format::FORMAT_ND, op::Format::FORMAT_ND);
auto dkRopeOut = executor->AllocTensor(outputDtype, op::Format::FORMAT_ND, op::Format::FORMAT_ND);
auto dsinkOut = executor->AllocTensor(DataType::DT_FLOAT, op::Format::FORMAT_ND, op::Format::FORMAT_ND);
const aclTensor *prefix = nullptr;
if (prefixOptional) {
if (!isMaxWorkspace) {
prefix = executor->ConvertToTensor(prefixOptional, op::DataType::DT_INT64);
CHECK_NULL(prefix);
const_cast<aclTensor *>(prefix)->SetStorageFormat(op::Format::FORMAT_ND);
const_cast<aclTensor *>(prefix)->SetViewFormat(op::Format::FORMAT_ND);
const_cast<aclTensor *>(prefix)->SetOriginalFormat(op::Format::FORMAT_ND);
} else {
FakeArrayToTensor(prefixOptional, const_cast<aclTensor *&>(prefix));
}
}
const aclTensor *actualSeqQLen = nullptr;
if (actualSeqQLenOptional) {
if (!isMaxWorkspace) {
actualSeqQLen = executor->ConvertToTensor(actualSeqQLenOptional, op::DataType::DT_INT64);
CHECK_NULL(actualSeqQLen);
const_cast<aclTensor *>(actualSeqQLen)->SetStorageFormat(op::Format::FORMAT_ND);
const_cast<aclTensor *>(actualSeqQLen)->SetViewFormat(op::Format::FORMAT_ND);
const_cast<aclTensor *>(actualSeqQLen)->SetOriginalFormat(op::Format::FORMAT_ND);
} else {
FakeArrayToTensor(actualSeqQLenOptional, const_cast<aclTensor *&>(actualSeqQLen));
}
}
const aclTensor *actualSeqKvLen = nullptr;
if (actualSeqKvLenOptional) {
if (!isMaxWorkspace) {
actualSeqKvLen = executor->ConvertToTensor(actualSeqKvLenOptional, op::DataType::DT_INT64);
CHECK_NULL(actualSeqKvLen);
const_cast<aclTensor *>(actualSeqKvLen)->SetStorageFormat(op::Format::FORMAT_ND);
const_cast<aclTensor *>(actualSeqKvLen)->SetViewFormat(op::Format::FORMAT_ND);
const_cast<aclTensor *>(actualSeqKvLen)->SetOriginalFormat(op::Format::FORMAT_ND);
} else {
FakeArrayToTensor(actualSeqKvLenOptional, const_cast<aclTensor *&>(actualSeqKvLen));
}
}
const aclTensor *qStartIdxOptionalTensor = nullptr;
if (qStartIdxOptional) {
if (!isMaxWorkspace) {
qStartIdxOptionalTensor = executor->ConvertToTensor(qStartIdxOptional, DataType::DT_INT64);
CHECK_NULL(qStartIdxOptionalTensor);
const_cast<aclTensor *>(qStartIdxOptionalTensor)->SetStorageFormat(Format::FORMAT_ND);
const_cast<aclTensor *>(qStartIdxOptionalTensor)->SetViewFormat(Format::FORMAT_ND);
const_cast<aclTensor *>(qStartIdxOptionalTensor)->SetOriginalFormat(Format::FORMAT_ND);
} else {
FakeArrayToTensor(qStartIdxOptional, const_cast<aclTensor *&>(qStartIdxOptionalTensor));
}
}
const aclTensor *kvStartIdxOptionalTensor = nullptr;
if (kvStartIdxOptional) {
if (!isMaxWorkspace) {
kvStartIdxOptionalTensor = executor->ConvertToTensor(kvStartIdxOptional, DataType::DT_INT64);
CHECK_NULL(kvStartIdxOptionalTensor);
const_cast<aclTensor *>(kvStartIdxOptionalTensor)->SetStorageFormat(Format::FORMAT_ND);
const_cast<aclTensor *>(kvStartIdxOptionalTensor)->SetViewFormat(Format::FORMAT_ND);
const_cast<aclTensor *>(kvStartIdxOptionalTensor)->SetOriginalFormat(Format::FORMAT_ND);
} else {
FakeArrayToTensor(kvStartIdxOptional, const_cast<aclTensor *&>(kvStartIdxOptionalTensor));
}
}
if (dScaleQOptional == nullptr) {
dScaleQOptional = executor->AllocTensor(DataType::DT_FLOAT, op::Format::FORMAT_ND, op::Format::FORMAT_ND);
}
if (dScaleKOptional == nullptr) {
dScaleKOptional = executor->AllocTensor(DataType::DT_FLOAT, op::Format::FORMAT_ND, op::Format::FORMAT_ND);
}
if (dScaleVOptional == nullptr) {
dScaleVOptional = executor->AllocTensor(DataType::DT_FLOAT, op::Format::FORMAT_ND, op::Format::FORMAT_ND);
}
if (dScaleDyOptional == nullptr) {
dScaleDyOptional = executor->AllocTensor(DataType::DT_FLOAT, op::Format::FORMAT_ND, op::Format::FORMAT_ND);
}
if (dScaleOOptional == nullptr) {
dScaleOOptional = executor->AllocTensor(DataType::DT_FLOAT, op::Format::FORMAT_ND, op::Format::FORMAT_ND);
}
if (dsScaleOptional == nullptr) {
dsScaleOptional = executor->AllocTensor(DataType::DT_FLOAT, op::Format::FORMAT_ND, op::Format::FORMAT_ND);
}
if (pScaleOptional == nullptr) {
pScaleOptional = executor->AllocTensor(DataType::DT_FLOAT, op::Format::FORMAT_ND, op::Format::FORMAT_ND);
}
auto ret = INFER_SHAPE(FlashAttentionScoreGrad,
OP_INPUT(query, key, value, dy, pseShiftOptional, dropMaskOptional, paddingMaskOptional,
attenMaskOptional, softmaxMaxOptional, softmaxSumOptional, softmaxInOptional,
attentionInOptional, prefix, actualSeqQLen, actualSeqKvLen,
qStartIdxOptionalTensor, kvStartIdxOptionalTensor,
dScaleQOptional, dScaleKOptional, dScaleVOptional, dScaleDyOptional, dScaleOOptional,
queryRope, keyRope, sinkInOptional, dsScaleOptional, pScaleOptional),
OP_OUTPUT(dqOut, dkOut, dvOut, dpseOut, dqRopeOut, dkRopeOut, dsinkOut),
OP_ATTR(static_cast<float>(scaleValueOptional), static_cast<float>(keepProbOptional),
preTockensOptional, nextTockensOptional, headNum, inputLayout, innerPreciseOptional,
sparseModeOptional, pseTypeOptional, seed, offset, outDTypeOptional, softmaxInLayout));
if (ret != ACLNN_SUCCESS) {
OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Fag InferShape failed.");
return {nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr};
}
ret = ADD_TO_LAUNCHER_LIST_AICORE(
FlashAttentionScoreGrad,
OP_INPUT(query, key, value, dy, pseShiftOptional, dropMaskOptional, paddingMaskOptional, attenMaskOptional,
softmaxMaxOptional, softmaxSumOptional, softmaxInOptional, attentionInOptional, prefix, actualSeqQLen,
actualSeqKvLen, qStartIdxOptionalTensor, kvStartIdxOptionalTensor,
dScaleQOptional, dScaleKOptional, dScaleVOptional, dScaleDyOptional, dScaleOOptional,
queryRope, keyRope, sinkInOptional, dsScaleOptional, pScaleOptional),
OP_OUTPUT(dqOut, dkOut, dvOut, dpseOut, dqRopeOut, dkRopeOut,dsinkOut),
OP_ATTR(static_cast<float>(scaleValueOptional), static_cast<float>(keepProbOptional), preTockensOptional,
nextTockensOptional, headNum, inputLayout, innerPreciseOptional, sparseModeOptional, pseTypeOptional,
seed, offset, outDTypeOptional, softmaxInLayout));
if (ret != ACLNN_SUCCESS) {
OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Fag launch kernel failed.");
return {nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr};
}
return {dqOut, dkOut, dvOut, dpseOut, dqRopeOut, dkRopeOut, dsinkOut};
}
}