* Copyright (c) 2024 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.
*/
#ifndef ATBOPS_PARAMS_UNPAD_FLASH_ATTENTION_H
#define ATBOPS_PARAMS_UNPAD_FLASH_ATTENTION_H
#include <cstdint>
#include <string>
#include <sstream>
#include <vector>
#include "mki/tensor.h"
#include "mki/utils/compare/compare.h"
namespace AtbOps {
namespace OpParam {
struct UnpadFlashAttention {
enum Type {
UNPAD_FLASH_ATTENTION_ND = 1,
UNPAD_DYNAMIC_BATCH_FLASH_ATTENTION = 4,
UNPAD_FLASH_ATTENTION_ENCODER_ND = 10,
UNPAD_ALIBI_FLASH_ATTENTION_ND = 11,
UNPAD_FLASH_ATTENTION_DECODER_ND = 12,
UNPAD_DYNAMIC_BATCH_FLASH_ATTENTION_DECODER = 13,
UNPAD_FLASH_ATTENTION_FP32_ND = 2001,
MULTI_LATENT_ATTENTION_COMBINE_CACHE = 2008,
MULTI_LATENT_ATTENTION_HIGH_PRECISION_COMBINE_CACHE = 2010,
UNPAD_FLASH_ATTENTION_ENCODER_PREFIX_CACHE_ND = 2012,
UNPAD_FLASH_ATTENTION_RAZOR_FUSION = 2014,
RELAY_ATTENTION_DECODER_ND = 14
};
enum ScaleType {
SCALE_TOR = 0,
SCALE_LOGN = 1,
SCALE_LOGN_FP32 = 2
};
Type type;
enum QuantType {
TYPE_QUANT_UNDEFINED = 0,
TYPE_DEQUANT_FUSION,
TYPE_QUANT_QKV_OFFLINE,
TYPE_QUANT_QKV_ONLINE
};
QuantType quantType = TYPE_QUANT_UNDEFINED;
Mki::TensorDType outDataType = Mki::TensorDType::TENSOR_DTYPE_FLOAT16;
ScaleType scaleType = SCALE_TOR;
int32_t headSize = 0;
int32_t headDimV = 0;
int32_t razorLen = 0;
int32_t preTokens = 0;
int32_t nextTokens = 0;
int32_t tileQ = 0;
int32_t tileKv = 0;
int32_t textQLen = 0;
int32_t textKvLen = 0;
std::vector<int32_t> qSeqLen;
std::vector<int32_t> kvSeqLen;
std::vector<Mki::Tensor> kTensorList;
std::vector<Mki::Tensor> vTensorList;
std::vector<Mki::Tensor> kShareTensorList;
std::vector<Mki::Tensor> vShareTensorList;
std::vector<int32_t> kvShareMap;
std::vector<int32_t> kvShareLen;
float tor = 0;
int32_t kvHead = 0;
std::vector<int32_t> batchRunStatus;
int32_t isClamp = 0;
float clampMin = 0;
float clampMax = 0;
uint32_t windowSize = 0;
uint32_t isTriuMask = 0;
enum MaskType {
MASK_TYPE_NONE = 0,
MASK_TYPE_NORM = 1,
MASK_TYPE_ALIBI = 2,
MASK_TYPE_LOOK_AHEAD = 3,
MASK_TYPE_SWA_NORM = 4,
MASK_TYPE_SWA_COMPRESS = 5,
MASK_TYPE_ALIBI_COMPRESS = 6,
MASK_TYPE_ALIBI_COMPRESS_SQRT = 7,
MASK_TYPE_ALIBI_COMPRESS_LEFT_ALIGN = 8,
MASK_TYPE_ALIBI_COMPRESS_128 = 9,
MASK_TYPE_CAUSAL_MASK = 10
};
MaskType maskType = MASK_TYPE_NORM;
enum CacheType {
CACHE_TYPE_NORM = 0,
CACHE_TYPE_CYCLE = 1
};
CacheType cacheType = CACHE_TYPE_NORM;
uint32_t isAlibiMaskSqrt = 0;
uint32_t alibiLeftAlign = 0;
bool compressHead = false;
enum DataShapeType : int {
TYPE_BSND = 0,
TYPE_BNSD
};
DataShapeType dataShapeType = TYPE_BSND;
bool operator==(const UnpadFlashAttention &other) const
{
return this->type == other.type && this->headSize == other.headSize && this->qSeqLen == other.qSeqLen &&
this->kvSeqLen == other.kvSeqLen && Mki::Utils::Compare<float>::IsEqual(this->tor, other.tor) &&
this->kvHead == other.kvHead && this->batchRunStatus == other.batchRunStatus &&
this->isClamp == other.isClamp && Mki::Utils::Compare<float>::IsEqual(this->clampMin, other.clampMin) &&
Mki::Utils::Compare<float>::IsEqual(this->clampMax, other.clampMax) &&
this->windowSize == other.windowSize && this->cacheType == other.cacheType &&
this->isTriuMask == other.isTriuMask && this->maskType == other.maskType &&
this->isAlibiMaskSqrt == other.isAlibiMaskSqrt && this->alibiLeftAlign == other.alibiLeftAlign &&
this->compressHead == other.compressHead && this->dataShapeType == other.dataShapeType &&
this->quantType == other.quantType && this->outDataType == other.outDataType &&
this->scaleType == other.scaleType && this->razorLen == other.razorLen &&
this->preTokens == other.preTokens && this->nextTokens == other.nextTokens &&
this->tileQ == other.tileQ && this->tileKv == other.tileKv &&
this->textQLen == other.textQLen && this->textKvLen == other.textKvLen;
}
};
}
}
#endif