* Copyright (c) 2026 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_attn.h"
#include "opdev/make_op_executor.h"
#include "opdev/op_dfx.h"
using namespace op;
namespace l0op {
OP_TYPE_REGISTER(FlashAttn);
const std::array<const aclTensor *, 2> FlashAttn(
const aclTensor *q,
const aclTensor *k,
const aclTensor *v,
const aclTensor *blockTableOptional,
const aclTensor *cuSeqlensQOptional,
const aclTensor *cuSeqlensKvOptional,
const aclTensor *sequsedQOptional,
const aclTensor *sequsedKvOptional,
const aclTensor *sinksOptional,
const aclTensor *attnMaskOptional,
const aclTensor *metadataOptional,
double softmaxScale,
int32_t maskMode,
int32_t winLeft,
int32_t winRight,
int32_t maxSeqlenQ,
int32_t maxSeqlenKV,
const char *layoutQ,
const char *layoutKv,
const char *layoutOut,
int32_t returnSoftmaxLse,
aclOpExecutor *executor)
{
L0_DFX(FlashAttn, q, k, v, blockTableOptional, cuSeqlensQOptional, cuSeqlensKvOptional,
sequsedQOptional, sequsedKvOptional, sinksOptional, attnMaskOptional, metadataOptional,
softmaxScale, maskMode, winLeft, winRight, maxSeqlenQ, maxSeqlenKV, layoutQ, layoutKv, layoutOut,
returnSoftmaxLse);
if (blockTableOptional == nullptr) {
blockTableOptional = executor->AllocTensor(DataType::DT_INT32, Format::FORMAT_ND, Format::FORMAT_ND);
}
if (cuSeqlensQOptional == nullptr) {
cuSeqlensQOptional = executor->AllocTensor(DataType::DT_INT32, Format::FORMAT_ND, Format::FORMAT_ND);
}
if (cuSeqlensKvOptional == nullptr) {
cuSeqlensKvOptional = executor->AllocTensor(DataType::DT_INT32, Format::FORMAT_ND, Format::FORMAT_ND);
}
if (sequsedQOptional == nullptr) {
sequsedQOptional = executor->AllocTensor(DataType::DT_INT32, Format::FORMAT_ND, Format::FORMAT_ND);
}
if (sequsedKvOptional == nullptr) {
sequsedKvOptional = executor->AllocTensor(DataType::DT_INT32, Format::FORMAT_ND, Format::FORMAT_ND);
}
if (sinksOptional == nullptr) {
sinksOptional = executor->AllocTensor(DataType::DT_FLOAT, Format::FORMAT_ND, Format::FORMAT_ND);
}
if (attnMaskOptional == nullptr) {
attnMaskOptional = executor->AllocTensor(DataType::DT_INT8, Format::FORMAT_ND, Format::FORMAT_ND);
}
if (metadataOptional == nullptr) {
metadataOptional = executor->AllocTensor(DataType::DT_INT32, Format::FORMAT_ND, Format::FORMAT_ND);
}
auto attentionOutAlloc = executor->AllocTensor(q->GetDataType(), Format::FORMAT_ND, Format::FORMAT_ND);
auto softmaxLseAlloc = executor->AllocTensor(DataType::DT_FLOAT, Format::FORMAT_ND, Format::FORMAT_ND);
auto ret = INFER_SHAPE(FlashAttn,
OP_INPUT(q, k, v, blockTableOptional, cuSeqlensQOptional, cuSeqlensKvOptional,
sequsedQOptional, sequsedKvOptional, sinksOptional, attnMaskOptional, metadataOptional),
OP_OUTPUT(attentionOutAlloc, softmaxLseAlloc),
OP_ATTR(softmaxScale, maskMode, winLeft, winRight, maxSeqlenQ, maxSeqlenKV,
layoutQ, layoutKv, layoutOut, returnSoftmaxLse));
if (ret != ACLNN_SUCCESS) {
OP_LOGE(ACLNN_ERR_PARAM_INVALID, "FlashAttn InferShape failed.");
return {nullptr, nullptr};
}
ret = ADD_TO_LAUNCHER_LIST_AICORE(
FlashAttn,
OP_INPUT(q, k, v, blockTableOptional, cuSeqlensQOptional, cuSeqlensKvOptional,
sequsedQOptional, sequsedKvOptional, sinksOptional, attnMaskOptional, metadataOptional),
OP_OUTPUT(attentionOutAlloc, softmaxLseAlloc),
OP_ATTR(softmaxScale, maskMode, winLeft, winRight, maxSeqlenQ, maxSeqlenKV,
layoutQ, layoutKv, layoutOut, returnSoftmaxLse));
if (ret != ACLNN_SUCCESS) {
OP_LOGE(ACLNN_ERR_PARAM_INVALID, "FlashAttn launch kernel failed.");
return {nullptr, nullptr};
}
return {attentionOutAlloc, softmaxLseAlloc};
}
}