* 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 incre_flash_attention.cpp
* \brief
*/
#include "interface/operation/operation_impl.h"
#include "tilefwk/tilefwk.h"
#include "interface/inner/tilefwk.h"
using namespace npu::tile_fwk;
namespace npu::tile_fwk {
void IncreFlashAttention(
Tensor& qNope, Tensor& kNopeCache, Tensor& vNopeCache, Tensor& qRope, Tensor& kRopeCache,
std::vector<std::vector<int>>& blockTable, std::vector<int>& actSeqs, float softmaxScale, Tensor& attentionOut,
IfaTileShapeConfig& tileConfig)
{
auto batchSize = blockTable.size();
ASSERT(batchSize == actSeqs.size());
int dN = qNope.GetShape()[1];
int dR = qRope.GetShape()[1];
int nQ = qNope.GetShape()[0] / batchSize;
int nTile = tileConfig.headNumQTile;
int blockSize = tileConfig.blockSize;
int nLoop = CeilDiv(nQ, nTile);
auto v0Tile = tileConfig.v0TileShape;
auto c1Tile = tileConfig.c1TileShape;
auto v1Tile = tileConfig.v1TileShape;
auto c2Tile = tileConfig.c2TileShape;
auto v2Tile = tileConfig.v2TileShape;
std::vector<std::pair<Tensor, std::vector<int>>> aggregation;
std::vector<Tensor> tiledOut;
for (size_t bIdx = 0; bIdx < batchSize; bIdx++) {
int curSeq = actSeqs[bIdx];
int bnPerBatch = CeilDiv(curSeq, blockSize);
for (int nIdx = 0; nIdx < nLoop; nIdx++) {
Tensor oiUpdate;
Tensor liUpdate;
Tensor miUpdate;
auto nTileCur = Min(nTile, nQ - nIdx * nTile);
auto curOffset = bIdx * nQ + nIdx * nTile;
TileShape::Current().SetVecTile(v0Tile[0], v0Tile[1]);
auto qn = View(qNope, {nTileCur, dN}, {static_cast<int>(curOffset), 0});
auto qr = View(qRope, {nTileCur, dR}, {static_cast<int>(curOffset), 0});
auto qi = Assemble({{qn, {0, 0}}, {qr, {0, dN}}});
for (int bn = 0; bn < bnPerBatch; bn++) {
auto curBlockIdx = blockTable[bIdx][bn];
auto s2TileCur = Min(blockSize, curSeq - bn * blockSize);
TileShape::Current().SetVecTile(v0Tile[0], v0Tile[1]);
auto kn = View(kNopeCache, {s2TileCur, dN}, {curBlockIdx * blockSize, 0});
auto kr = View(kRopeCache, {s2TileCur, dR}, {curBlockIdx * blockSize, 0});
auto kj = Assemble({{kn, {0, 0}}, {kr, {0, dN}}});
auto vj = View(vNopeCache, {s2TileCur, dN}, {curBlockIdx * blockSize, 0});
TileShape::Current().SetCubeTile(
{c1Tile[0], c1Tile[1]}, {c1Tile[2], c1Tile[3]}, {c1Tile[4], c1Tile[5]});
TileShape::Current().SetMatrixSize({qi.GetShape()[0], 0, kj.GetShape()[0]});
auto sij = Matrix::Matmul(DataType::DT_FP32, qi, kj, false, true);
TileShape::Current().SetVecTile(v1Tile[0], v1Tile[1]);
auto sijScale = Mul(sij, Element(DataType::DT_FP32, softmaxScale));
auto tildaMij = Amax(sijScale, -1, true);
auto tsub = Sub(sijScale, tildaMij);
auto tildaPij = Exp(tsub);
auto tildaPijF16 = Cast(tildaPij, DataType::DT_BF16);
auto tildaLij = Sum(tildaPij, -1, true);
if (bn == 0) {
TileShape::Current().SetCubeTile(
{c2Tile[0], c2Tile[1]}, {c2Tile[2], c2Tile[3]}, {c2Tile[4], c2Tile[5]});
TileShape::Current().SetMatrixSize(
{tildaPijF16.GetShape()[0], tildaPijF16.GetShape()[1], vj.GetShape()[1]});
auto oiTmp = Matrix::Matmul(DataType::DT_FP32, tildaPijF16, vj, false, false);
TileShape::Current().SetVecTile(v2Tile[0], v2Tile[1]);
oiUpdate = (bnPerBatch == 1 ? Div(oiTmp, tildaLij) : oiTmp);
liUpdate = tildaLij;
miUpdate = tildaMij;
continue;
}
auto oi = oiUpdate;
auto li = liUpdate;
auto mi = miUpdate;
auto miNew = Maximum(mi, tildaMij);
auto t1 = Sub(mi, miNew);
auto t2 = Exp(t1);
auto t3 = Sub(tildaMij, miNew);
auto t4 = Exp(t3);
auto t5 = Mul(t4, tildaLij);
auto t6 = Mul(t2, li);
auto liNew = Add(t6, t5);
auto q3 = Mul(oi, t2);
TileShape::Current().SetCubeTile(
{c2Tile[0], c2Tile[1]}, {c2Tile[2], c2Tile[3]}, {c2Tile[4], c2Tile[5]});
TileShape::Current().SetMatrixSize(
{tildaPijF16.GetShape()[0], tildaPijF16.GetShape()[1], vj.GetShape()[1]});
auto q1 = Matrix::Matmul(DataType::DT_FP32, tildaPijF16, vj, false, false);
TileShape::Current().SetVecTile(v2Tile[0], v2Tile[1]);
auto q2 = Mul(q1, t4);
auto oiTmp = Add(q3, q2);
oiUpdate = (bn == bnPerBatch - 1 ? Div(oiTmp, liNew) : oiTmp);
liUpdate = liNew;
miUpdate = miNew;
}
tiledOut.push_back(oiUpdate);
}
}
attentionOut = Cat(tiledOut, 0);
}
}