* 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.
*/
* \file llama_def.h
* \brief
*/
#pragma once
#ifndef LLAMA_DEF_H
#define LLAMA_DEF_H
#include "interface/inner/pre_def.h"
#include "tilefwk/tilefwk.h"
#include "interface/inner/tilefwk.h"
#define LLAMA_FUNCTION(n, ...) FUNCTION(#n, ##__VA_ARGS__)
#define LLAMA_PROGRAM(n, ...) PROGRAM(#n, ##__VA_ARGS__)
namespace npu::tile_fwk {
struct AttentionDims {
int b;
int n;
int s;
int d;
int singleM;
int singleN;
};
struct AttentionVecTileConfig {
int defaultVecTileX;
int defaultVecTileY;
int softmaxTileX;
int softmaxTileY;
int updateTileX;
int updateTileY;
int castTileX;
int castTileY;
};
struct AttentionCubeTileConfig {
int c1L1M;
int c1L1K;
int c1L1N;
int c2L1M;
int c2L1K;
int c2L1N;
int c1L0 = 128;
int c2L0 = 128;
};
struct KeyConfig {
int max;
int min;
int dbType;
int nBuffer;
int isPartitionCv;
int cTileX;
int cTileY;
};
constexpr AttentionVecTileConfig DFS_VEC_CFG = {128, 128, 16, 128, 16, 128, 32, 128};
constexpr AttentionVecTileConfig SMALL_DFS_VEC_CFG = {64, 128, 16, 128, 16, 128, 32, 128};
constexpr AttentionCubeTileConfig DFS_CUBE_CFG = {128, 128, 128, 128, 128, 128};
constexpr AttentionVecTileConfig OOO_VEC_CFG = {64, 128, 16, 512, 32, 128, 32, 128};
constexpr AttentionCubeTileConfig OOO_CUBE_CFG = {128, 128, 512, 64, 256, 64, 128, 64};
constexpr KeyConfig DFT_BASIC_CFG = {8192, 1024, 1, 1, 0, 128, 128};
constexpr int DFT_SINGLE_M = 128;
constexpr int DFT_SINGLE_N = 128;
Tensor LlamaLayer(
Tensor hiddenStates, const Tensor& attnWight, const Tensor& denseWeight, const Tensor& ffnWeight,
const AttentionDims& atDims, const AttentionVecTileConfig& vecCfg, const AttentionCubeTileConfig& cubeCfg);
Tensor FlashAttention(
const Tensor& q, const Tensor& k, const Tensor& v, const Tensor& m, const Tensor& l, const AttentionDims& atDims,
const AttentionVecTileConfig& vecCfg, const AttentionCubeTileConfig& cubeCfg);
Tensor FlashAttentionNew(
const Tensor& q, const Tensor& k, const Tensor& v, const Tensor& m, const Tensor& l, const AttentionDims& atDims);
static inline void SetC1CubeConfig(const AttentionCubeTileConfig& cubeCfg)
{
TileShape::Current().SetCubeTile(
{cubeCfg.c1L0, cubeCfg.c1L1M}, {cubeCfg.c1L0, cubeCfg.c1L1K}, {cubeCfg.c1L0, cubeCfg.c1L1N});
}
static inline void SetC2CubeConfig(const AttentionCubeTileConfig& cubeCfg)
{
TileShape::Current().SetCubeTile(
{cubeCfg.c2L0, cubeCfg.c2L1M}, {cubeCfg.c2L0, cubeCfg.c2L1K}, {cubeCfg.c2L0, cubeCfg.c2L1N});
}
}
#endif