happyz
/
gemma.cpp
mirror of https://github.com/google/gemma.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
gemma.cpp/gemma/flash_attention.cc

1342 lines
61 KiB
C++
Raw Permalink Blame History

// Copyright 2025 Google LLC
// SPDX-License-Identifier: Apache-2.0
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// https://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include <stddef.h>
#include <stdint.h>
#include <algorithm>
#include <array>
#include <cmath>
#include <cstdlib>
#include <limits>
#include <vector>
#include "compression/types.h" // GEMMA_DISABLED_TARGETS
#include "gemma/flash_structs.h"
#include "gemma/kv_cache.h"
#include "gemma/query.h"
#include "util/basics.h"
#include "util/threading_context.h"
#include "util/zones.h"
#include "hwy/base.h"
#ifndef HWY_DISABLED_TARGETS
#define HWY_DISABLED_TARGETS GEMMA_DISABLED_TARGETS
#endif // HWY_DISABLED_TARGETS
#include "gemma/activations.h"
#include "gemma/configs.h" // kMaxQKVDim
#include "util/threading.h"
#include "hwy/profiler.h"
// Compiles this file for multiple architectures via "foreach_target.h", to
// which we pass the filename via macro 'argument'.
// clang-format off
#undef HWY_TARGET_INCLUDE
#define HWY_TARGET_INCLUDE "gemma/flash_attention.cc" // NOLINT
// clang-format on
#include "hwy/foreach_target.h" // IWYU pragma: keep
#include "hwy/highway.h"
// After highway.h
#include "compression/compress-inl.h"
#include "gemma/attention.h"
#include "ops/matmul-inl.h"
#include "ops/ops-inl.h"
HWY_BEFORE_NAMESPACE();
namespace gcpp {
namespace HWY_NAMESPACE {
static constexpr float kNegInf = -std::numeric_limits<float>::max() / 64.0f;
// Updates q in place for RMSNorm and positional encoding.
void RMSNormAndPositionalEncoding(const size_t num_tokens, const QBatch& qbatch,
MatPtrT<float>& q,
const MatPtr& query_norm_scale,
const size_t layer_idx,
const AttentionActivationsPtrs& activations,
ThreadingContext& ctx) {
const LayerConfig& layer_config = activations.config.layer_configs[layer_idx];
const float query_scale = activations.query_scale;
const hwy::Divisor div_qbatch(qbatch.Size());
const auto func = [&](const size_t task, size_t worker) HWY_ATTR {
GCPP_ZONE(ctx, worker, Zones::kFlashAttentionRmsNormAndPositionalEncoding);
size_t qi = div_qbatch.Remainder(task);
size_t batch_idx = div_qbatch.Divide(task);
for (size_t h = 0; h < layer_config.heads; ++h) {
const size_t tq_idx = qbatch.Size() * batch_idx + qi;
// Find the token position in the query and calculate
// the range of cache positions to attend to.
constexpr size_t offset = 0; // placeholder, do not remove
const size_t pos = qbatch.Pos(qi) + batch_idx + offset;
float* HWY_RESTRICT q_row = q.Row(tq_idx) + h * layer_config.qkv_dim;
// Apply rope and scaling to Q.
if (query_norm_scale.HasPtr()) {
CallUpcasted(&query_norm_scale, [&](const auto* weights_t) {
RMSNormInplace(weights_t->PackedScale1(), /*w_ofs=*/0, q_row,
layer_config.qkv_dim, ctx, worker);
});
}
PositionalEncodingQK(q_row, layer_idx, activations, ctx, worker, pos,
query_scale);
}
};
{
// kHierarchical is not worth the extra sync overhead because the tasks are
// very lightweight.
ParallelFor(Parallelism::kFlat, num_tokens * qbatch.Size(), ctx,
/*cluster_idx=*/0, Callers::kFlashRMSNormAndPositionalEncoding,
func);
}
}
// Zeroes out kVTileSize of the given vectors.
template <size_t kVTileSize, class DF, class VF = hn::Vec<DF>>
HWY_INLINE void ZeroResults(DF df, VF& sum0, VF& HWY_MAYBE_UNUSED sum1,
VF& HWY_MAYBE_UNUSED sum2,
VF& HWY_MAYBE_UNUSED sum3,
VF& HWY_MAYBE_UNUSED sum4,
VF& HWY_MAYBE_UNUSED sum5,
VF& HWY_MAYBE_UNUSED sum6,
VF& HWY_MAYBE_UNUSED sum7) {
sum0 = hn::Zero(df);
if constexpr (kVTileSize >= 4) {
sum1 = hn::Zero(df);
sum2 = hn::Zero(df);
sum3 = hn::Zero(df);
}
if constexpr (kVTileSize >= 8) {
sum4 = hn::Zero(df);
sum5 = hn::Zero(df);
sum6 = hn::Zero(df);
sum7 = hn::Zero(df);
}
}
// Returns a tile of 1, 4 or 8 Q rows by 2NF K Q.K dot products, in float32.
// K is always pre-transposed to shape:
// [seq_len / 2kNF, layers * kv_heads * qkv_dim/2 * 2kNF * 2], where the /2, *2
// represents that pairs of qkv_dim elements are kept together to make best use
// of BF16 dot product instructions.
// Note that this version assumes that Q is float32, and not transposed, and
// HWY_NATIVE_DOT_BF16 is false.
template <size_t kVTileSize, class DF, class VF = hn::Vec<DF>>
HWY_INLINE void QDotKTile148FloatNotNative(
DF df, const float* HWY_RESTRICT q, const uint32_t* HWY_RESTRICT q_offsets,
size_t half_cols, const MatPtrT<KV_t>& k, size_t pos, VF& sum00, VF& sum01,
VF& HWY_MAYBE_UNUSED sum10, VF& HWY_MAYBE_UNUSED sum11,
VF& HWY_MAYBE_UNUSED sum20, VF& HWY_MAYBE_UNUSED sum21,
VF& HWY_MAYBE_UNUSED sum30, VF& HWY_MAYBE_UNUSED sum31,
VF& HWY_MAYBE_UNUSED sum40, VF& HWY_MAYBE_UNUSED sum41,
VF& HWY_MAYBE_UNUSED sum50, VF& HWY_MAYBE_UNUSED sum51,
VF& HWY_MAYBE_UNUSED sum60, VF& HWY_MAYBE_UNUSED sum61,
VF& HWY_MAYBE_UNUSED sum70, VF& HWY_MAYBE_UNUSED sum71) {
ZeroResults<kVTileSize>(df, sum00, sum10, sum20, sum30, sum40, sum50, sum60,
sum70);
ZeroResults<kVTileSize>(df, sum01, sum11, sum21, sum31, sum41, sum51, sum61,
sum71);
using DBF = hn::ScalableTag<BF16>;
const DBF dbf;
using VBF = hn::Vec<DBF>;
const size_t kNF = hn::Lanes(df);
const float* HWY_RESTRICT q_base[kVTileSize];
for (size_t i = 0; i < kVTileSize; ++i) {
q_base[i] = q + q_offsets[i];
}
const BF16* HWY_RESTRICT k_base = k.Row(pos / (2 * kNF));
for (size_t i = 0; i < half_cols; ++i, k_base += kNF * 4) {
// TODO(rays): Replace with decompress2.
VBF k0_vec = hn::LoadU(dbf, k_base);
VBF k1_vec = hn::LoadU(dbf, k_base + kNF * 2);
VF k0_even = hn::PromoteEvenTo(df, k0_vec);
VF k0_odd = hn::PromoteOddTo(df, k0_vec);
VF k1_even = hn::PromoteEvenTo(df, k1_vec);
VF k1_odd = hn::PromoteOddTo(df, k1_vec);
VF q0_even = hn::Set(df, q_base[0][i * 2]);
VF q0_odd = hn::Set(df, q_base[0][i * 2 + 1]);
sum00 = hn::MulAdd(q0_even, k0_even, sum00);
sum01 = hn::MulAdd(q0_even, k1_even, sum01);
sum00 = hn::MulAdd(q0_odd, k0_odd, sum00);
sum01 = hn::MulAdd(q0_odd, k1_odd, sum01);
if constexpr (kVTileSize >= 4) {
VF q1_even = hn::Set(df, q_base[1][i * 2]);
VF q1_odd = hn::Set(df, q_base[1][i * 2 + 1]);
sum10 = hn::MulAdd(q1_even, k0_even, sum10);
sum11 = hn::MulAdd(q1_even, k1_even, sum11);
sum10 = hn::MulAdd(q1_odd, k0_odd, sum10);
sum11 = hn::MulAdd(q1_odd, k1_odd, sum11);
VF q2_even = hn::Set(df, q_base[2][i * 2]);
VF q2_odd = hn::Set(df, q_base[2][i * 2 + 1]);
sum20 = hn::MulAdd(q2_even, k0_even, sum20);
sum21 = hn::MulAdd(q2_even, k1_even, sum21);
sum20 = hn::MulAdd(q2_odd, k0_odd, sum20);
sum21 = hn::MulAdd(q2_odd, k1_odd, sum21);
VF q3_even = hn::Set(df, q_base[3][i * 2]);
VF q3_odd = hn::Set(df, q_base[3][i * 2 + 1]);
sum30 = hn::MulAdd(q3_even, k0_even, sum30);
sum31 = hn::MulAdd(q3_even, k1_even, sum31);
sum30 = hn::MulAdd(q3_odd, k0_odd, sum30);
sum31 = hn::MulAdd(q3_odd, k1_odd, sum31);
}
if constexpr (kVTileSize >= 8) {
VF q4_even = hn::Set(df, q_base[4][i * 2]);
VF q4_odd = hn::Set(df, q_base[4][i * 2 + 1]);
sum40 = hn::MulAdd(q4_even, k0_even, sum40);
sum41 = hn::MulAdd(q4_even, k1_even, sum41);
sum40 = hn::MulAdd(q4_odd, k0_odd, sum40);
sum41 = hn::MulAdd(q4_odd, k1_odd, sum41);
VF q5_even = hn::Set(df, q_base[5][i * 2]);
VF q5_odd = hn::Set(df, q_base[5][i * 2 + 1]);
sum50 = hn::MulAdd(q5_even, k0_even, sum50);
sum51 = hn::MulAdd(q5_even, k1_even, sum51);
sum50 = hn::MulAdd(q5_odd, k0_odd, sum50);
sum51 = hn::MulAdd(q5_odd, k1_odd, sum51);
VF q6_even = hn::Set(df, q_base[6][i * 2]);
VF q6_odd = hn::Set(df, q_base[6][i * 2 + 1]);
sum60 = hn::MulAdd(q6_even, k0_even, sum60);
sum61 = hn::MulAdd(q6_even, k1_even, sum61);
sum60 = hn::MulAdd(q6_odd, k0_odd, sum60);
sum61 = hn::MulAdd(q6_odd, k1_odd, sum61);
VF q7_even = hn::Set(df, q_base[7][i * 2]);
VF q7_odd = hn::Set(df, q_base[7][i * 2 + 1]);
sum70 = hn::MulAdd(q7_even, k0_even, sum70);
sum71 = hn::MulAdd(q7_even, k1_even, sum71);
sum70 = hn::MulAdd(q7_odd, k0_odd, sum70);
sum71 = hn::MulAdd(q7_odd, k1_odd, sum71);
}
}
}
// Loads an adjacent pair of floats, converts them to BF16, and broadcasts them
// across a vector of BF16 as alternating odd and even elements.
// hn::ReorderDemote2To(dbf, q_1_float, q_1_float); with q1_float containing
// alternating odd and even floats appears not to do this.
HWY_INLINE hn::Vec<hn::ScalableTag<BF16>> DemoteAndBroadcast2ToBF16(
const float* HWY_RESTRICT base) {
using DF = hn::ScalableTag<float>;
const DF df;
using VF = hn::Vec<DF>;
VF v_even = hn::Set(df, base[0]);
VF v_odd = hn::Set(df, base[1]);
VF interleaved = hn::OddEven(v_odd, v_even);
return hn::OrderedDemote2To(hn::ScalableTag<BF16>(), interleaved,
interleaved);
}
// Returns a tile of 1, 4 or 8 Q rows by 2NF K Q.K dot products, in float32.
// K is always pre-transposed to shape:
// [seq_len / 2kNF, layers * kv_heads * qkv_dim/2 * 2kNF * 2], where the /2, *2
// represents that pairs of qkv_dim elements are kept together to make best use
// of BF16 dot product instructions.
// Note that this version assumes that Q is float32, and not transposed, and
// HWY_NATIVE_DOT_BF16 is true.
template <size_t kVTileSize, class DF, class VF = hn::Vec<DF>>
HWY_INLINE void QDotKTile148FloatNative(
DF df, const float* HWY_RESTRICT q, const uint32_t* HWY_RESTRICT q_offsets,
size_t half_cols, const MatPtrT<KV_t>& k, size_t pos, VF& sum00, VF& sum01,
VF& HWY_MAYBE_UNUSED sum10, VF& HWY_MAYBE_UNUSED sum11,
VF& HWY_MAYBE_UNUSED sum20, VF& HWY_MAYBE_UNUSED sum21,
VF& HWY_MAYBE_UNUSED sum30, VF& HWY_MAYBE_UNUSED sum31,
VF& HWY_MAYBE_UNUSED sum40, VF& HWY_MAYBE_UNUSED sum41,
VF& HWY_MAYBE_UNUSED sum50, VF& HWY_MAYBE_UNUSED sum51,
VF& HWY_MAYBE_UNUSED sum60, VF& HWY_MAYBE_UNUSED sum61,
VF& HWY_MAYBE_UNUSED sum70, VF& HWY_MAYBE_UNUSED sum71) {
ZeroResults<kVTileSize>(df, sum00, sum10, sum20, sum30, sum40, sum50, sum60,
sum70);
ZeroResults<kVTileSize>(df, sum01, sum11, sum21, sum31, sum41, sum51, sum61,
sum71);
VF unused = hn::Zero(df);
using DBF = hn::ScalableTag<BF16>;
const DBF dbf;
using VBF = hn::Vec<DBF>;
const size_t kNF = hn::Lanes(df);
const float* HWY_RESTRICT q_base[kVTileSize];
for (size_t i = 0; i < kVTileSize; ++i) {
q_base[i] = q + q_offsets[i];
}
const BF16* HWY_RESTRICT k_base = k.Row(pos / (2 * kNF));
for (size_t i = 0; i < half_cols; ++i, k_base += kNF * 4) {
VBF kvec0 = hn::LoadU(dbf, k_base);
VBF kvec1 = hn::LoadU(dbf, k_base + kNF * 2);
VBF q0_bf16 = DemoteAndBroadcast2ToBF16(q_base[0] + i * 2);
sum00 = hn::ReorderWidenMulAccumulate(df, q0_bf16, kvec0, sum00, unused);
sum01 = hn::ReorderWidenMulAccumulate(df, q0_bf16, kvec1, sum01, unused);
if constexpr (kVTileSize >= 4) {
VBF q1_bf16 = DemoteAndBroadcast2ToBF16(q_base[1] + i * 2);
sum10 = hn::ReorderWidenMulAccumulate(df, q1_bf16, kvec0, sum10, unused);
sum11 = hn::ReorderWidenMulAccumulate(df, q1_bf16, kvec1, sum11, unused);
VBF q2_bf16 = DemoteAndBroadcast2ToBF16(q_base[2] + i * 2);
sum20 = hn::ReorderWidenMulAccumulate(df, q2_bf16, kvec0, sum20, unused);
sum21 = hn::ReorderWidenMulAccumulate(df, q2_bf16, kvec1, sum21, unused);
VBF q3_bf16 = DemoteAndBroadcast2ToBF16(q_base[3] + i * 2);
sum30 = hn::ReorderWidenMulAccumulate(df, q3_bf16, kvec0, sum30, unused);
sum31 = hn::ReorderWidenMulAccumulate(df, q3_bf16, kvec1, sum31, unused);
}
if constexpr (kVTileSize >= 8) {
VBF q4_bf16 = DemoteAndBroadcast2ToBF16(q_base[4] + i * 2);
sum40 = hn::ReorderWidenMulAccumulate(df, q4_bf16, kvec0, sum40, unused);
sum41 = hn::ReorderWidenMulAccumulate(df, q4_bf16, kvec1, sum41, unused);
VBF q5_bf16 = DemoteAndBroadcast2ToBF16(q_base[5] + i * 2);
sum50 = hn::ReorderWidenMulAccumulate(df, q5_bf16, kvec0, sum50, unused);
sum51 = hn::ReorderWidenMulAccumulate(df, q5_bf16, kvec1, sum51, unused);
VBF q6_bf16 = DemoteAndBroadcast2ToBF16(q_base[6] + i * 2);
sum60 = hn::ReorderWidenMulAccumulate(df, q6_bf16, kvec0, sum60, unused);
sum61 = hn::ReorderWidenMulAccumulate(df, q6_bf16, kvec1, sum61, unused);
VBF q7_bf16 = DemoteAndBroadcast2ToBF16(q_base[7] + i * 2);
sum70 = hn::ReorderWidenMulAccumulate(df, q7_bf16, kvec0, sum70, unused);
sum71 = hn::ReorderWidenMulAccumulate(df, q7_bf16, kvec1, sum71, unused);
}
}
}
// Returns a tile of 1, 4 or 8 Q rows by 2NF K Q.K dot products, in float32.
// K is always pre-transposed to shape:
// [seq_len / 2kNF, layers * kv_heads * qkv_dim/2 * 2kNF * 2], where the /2, *2
// represents that pairs of qkv_dim elements are kept together to make best use
// of BF16 dot product instructions.
// Note that this is optimized for the case where q and k are bf16, but there is
// no native_bf16 instruction.
template <size_t kVTileSize, class DF, class VF = hn::Vec<DF>>
HWY_INLINE void QDotKTile148BF16NotNative(
DF df, const BF16* HWY_RESTRICT q, const uint32_t* HWY_RESTRICT q_offsets,
size_t half_cols, const MatPtrT<KV_t>& k, size_t pos, VF& sum00, VF& sum01,
VF& HWY_MAYBE_UNUSED sum10, VF& HWY_MAYBE_UNUSED sum11,
VF& HWY_MAYBE_UNUSED sum20, VF& HWY_MAYBE_UNUSED sum21,
VF& HWY_MAYBE_UNUSED sum30, VF& HWY_MAYBE_UNUSED sum31,
VF& HWY_MAYBE_UNUSED sum40, VF& HWY_MAYBE_UNUSED sum41,
VF& HWY_MAYBE_UNUSED sum50, VF& HWY_MAYBE_UNUSED sum51,
VF& HWY_MAYBE_UNUSED sum60, VF& HWY_MAYBE_UNUSED sum61,
VF& HWY_MAYBE_UNUSED sum70, VF& HWY_MAYBE_UNUSED sum71) {
ZeroResults<kVTileSize>(df, sum00, sum10, sum20, sum30, sum40, sum50, sum60,
sum70);
ZeroResults<kVTileSize>(df, sum01, sum11, sum21, sum31, sum41, sum51, sum61,
sum71);
using DBF = hn::ScalableTag<BF16>;
const DBF dbf;
using VBF = hn::Vec<DBF>;
const size_t kNF = hn::Lanes(df);
const float* HWY_RESTRICT q_base[kVTileSize];
for (size_t i = 0; i < kVTileSize; ++i) {
q_base[i] = reinterpret_cast<const float*>(q + q_offsets[i]);
}
const BF16* HWY_RESTRICT k_base = k.Row(pos / (2 * kNF));
for (size_t i = 0; i < half_cols; ++i, k_base += kNF * 4) {
VBF kvec0 = hn::LoadU(dbf, k_base);
VBF kvec1 = hn::LoadU(dbf, k_base + kNF * 2);
VBF q0 = hn::BitCast(dbf, hn::Set(df, q_base[0][i]));
VF k0_even = hn::PromoteEvenTo(df, kvec0);
VF k0_odd = hn::PromoteOddTo(df, kvec0);
VF k1_even = hn::PromoteEvenTo(df, kvec1);
VF k1_odd = hn::PromoteOddTo(df, kvec1);
VF q0_even = hn::PromoteEvenTo(df, q0);
sum00 = hn::MulAdd(q0_even, k0_even, sum00);
sum01 = hn::MulAdd(q0_even, k1_even, sum01);
VF q0_odd = hn::PromoteOddTo(df, q0);
sum00 = hn::MulAdd(q0_odd, k0_odd, sum00);
sum01 = hn::MulAdd(q0_odd, k1_odd, sum01);
if constexpr (kVTileSize >= 4) {
VBF q1 = hn::BitCast(dbf, hn::Set(df, q_base[1][i]));
VF q1_even = hn::PromoteEvenTo(df, q1);
sum10 = hn::MulAdd(q1_even, k0_even, sum10);
sum11 = hn::MulAdd(q1_even, k1_even, sum11);
VF q1_odd = hn::PromoteOddTo(df, q1);
sum10 = hn::MulAdd(q1_odd, k0_odd, sum10);
sum11 = hn::MulAdd(q1_odd, k1_odd, sum11);
VBF q2 = hn::BitCast(dbf, hn::Set(df, q_base[2][i]));
VF q2_even = hn::PromoteEvenTo(df, q2);
sum20 = hn::MulAdd(q2_even, k0_even, sum20);
sum21 = hn::MulAdd(q2_even, k1_even, sum21);
VF q2_odd = hn::PromoteOddTo(df, q2);
sum20 = hn::MulAdd(q2_odd, k0_odd, sum20);
sum21 = hn::MulAdd(q2_odd, k1_odd, sum21);
VBF q3 = hn::BitCast(dbf, hn::Set(df, q_base[3][i]));
VF q3_even = hn::PromoteEvenTo(df, q3);
sum30 = hn::MulAdd(q3_even, k0_even, sum30);
sum31 = hn::MulAdd(q3_even, k1_even, sum31);
VF q3_odd = hn::PromoteOddTo(df, q3);
sum30 = hn::MulAdd(q3_odd, k0_odd, sum30);
sum31 = hn::MulAdd(q3_odd, k1_odd, sum31);
}
if constexpr (kVTileSize >= 8) {
VBF q4 = hn::BitCast(dbf, hn::Set(df, q_base[4][i]));
VF q4_even = hn::PromoteEvenTo(df, q4);
sum40 = hn::MulAdd(q4_even, k0_even, sum40);
sum41 = hn::MulAdd(q4_even, k1_even, sum41);
VF q4_odd = hn::PromoteOddTo(df, q4);
sum40 = hn::MulAdd(q4_odd, k0_odd, sum40);
sum41 = hn::MulAdd(q4_odd, k1_odd, sum41);
VBF q5 = hn::BitCast(dbf, hn::Set(df, q_base[5][i]));
VF q5_even = hn::PromoteEvenTo(df, q5);
sum50 = hn::MulAdd(q5_even, k0_even, sum50);
sum51 = hn::MulAdd(q5_even, k1_even, sum51);
VF q5_odd = hn::PromoteOddTo(df, q5);
sum50 = hn::MulAdd(q5_odd, k0_odd, sum50);
sum51 = hn::MulAdd(q5_odd, k1_odd, sum51);
VBF q6 = hn::BitCast(dbf, hn::Set(df, q_base[6][i]));
VF q6_even = hn::PromoteEvenTo(df, q6);
sum60 = hn::MulAdd(q6_even, k0_even, sum60);
sum61 = hn::MulAdd(q6_even, k1_even, sum61);
VF q6_odd = hn::PromoteOddTo(df, q6);
sum60 = hn::MulAdd(q6_odd, k0_odd, sum60);
sum61 = hn::MulAdd(q6_odd, k1_odd, sum61);
VBF q7 = hn::BitCast(dbf, hn::Set(df, q_base[7][i]));
VF q7_even = hn::PromoteEvenTo(df, q7);
sum70 = hn::MulAdd(q7_even, k0_even, sum70);
sum71 = hn::MulAdd(q7_even, k1_even, sum71);
VF q7_odd = hn::PromoteOddTo(df, q7);
sum70 = hn::MulAdd(q7_odd, k0_odd, sum70);
sum71 = hn::MulAdd(q7_odd, k1_odd, sum71);
}
}
}
// Returns a tile of 1, 4 or 8 Q rows by 2NF K Q.K dot products, in float32.
// K is always pre-transposed to shape:
// [seq_len / 2kNF, layers * kv_heads * qkv_dim/2 * 2kNF * 2], where the /2, *2
// represents that pairs of qkv_dim elements are kept together to make best use
// of BF16 dot product instructions.
// Note that this is optimized for the case where q and k are bf16, and there is
// a native_bf16 instruction.
template <size_t kVTileSize, class DF, class VF = hn::Vec<DF>>
HWY_INLINE void QDotKTile148BF16Native(
DF df, const BF16* HWY_RESTRICT q, const uint32_t* HWY_RESTRICT q_offsets,
size_t half_cols, const MatPtrT<KV_t>& k, size_t pos, VF& sum00, VF& sum01,
VF& HWY_MAYBE_UNUSED sum10, VF& HWY_MAYBE_UNUSED sum11,
VF& HWY_MAYBE_UNUSED sum20, VF& HWY_MAYBE_UNUSED sum21,
VF& HWY_MAYBE_UNUSED sum30, VF& HWY_MAYBE_UNUSED sum31,
VF& HWY_MAYBE_UNUSED sum40, VF& HWY_MAYBE_UNUSED sum41,
VF& HWY_MAYBE_UNUSED sum50, VF& HWY_MAYBE_UNUSED sum51,
VF& HWY_MAYBE_UNUSED sum60, VF& HWY_MAYBE_UNUSED sum61,
VF& HWY_MAYBE_UNUSED sum70, VF& HWY_MAYBE_UNUSED sum71) {
ZeroResults<kVTileSize>(df, sum00, sum10, sum20, sum30, sum40, sum50, sum60,
sum70);
ZeroResults<kVTileSize>(df, sum01, sum11, sum21, sum31, sum41, sum51, sum61,
sum71);
VF unused_sum1 = hn::Zero(df);
using DBF = hn::ScalableTag<BF16>;
const DBF dbf;
using VBF = hn::Vec<DBF>;
const size_t kNF = hn::Lanes(df);
const float* HWY_RESTRICT q_base[kVTileSize];
for (size_t i = 0; i < kVTileSize; ++i) {
q_base[i] = reinterpret_cast<const float*>(q + q_offsets[i]);
}
const BF16* HWY_RESTRICT k_base = k.Row(pos / (2 * kNF));
for (size_t i = 0; i < half_cols; ++i, k_base += kNF * 4) {
VBF k0_vec = hn::LoadU(dbf, k_base);
VBF k1_vec = hn::LoadU(dbf, k_base + kNF * 2);
VBF q0 = hn::BitCast(dbf, hn::Set(df, q_base[0][i]));
sum00 = hn::ReorderWidenMulAccumulate(df, q0, k0_vec, sum00, unused_sum1);
sum01 = hn::ReorderWidenMulAccumulate(df, q0, k1_vec, sum01, unused_sum1);
if constexpr (kVTileSize >= 4) {
VBF q1 = hn::BitCast(dbf, hn::Set(df, q_base[1][i]));
sum10 = hn::ReorderWidenMulAccumulate(df, q1, k0_vec, sum10, unused_sum1);
sum11 = hn::ReorderWidenMulAccumulate(df, q1, k1_vec, sum11, unused_sum1);
VBF q2 = hn::BitCast(dbf, hn::Set(df, q_base[2][i]));
sum20 = hn::ReorderWidenMulAccumulate(df, q2, k0_vec, sum20, unused_sum1);
sum21 = hn::ReorderWidenMulAccumulate(df, q2, k1_vec, sum21, unused_sum1);
VBF q3 = hn::BitCast(dbf, hn::Set(df, q_base[3][i]));
sum30 = hn::ReorderWidenMulAccumulate(df, q3, k0_vec, sum30, unused_sum1);
sum31 = hn::ReorderWidenMulAccumulate(df, q3, k1_vec, sum31, unused_sum1);
}
if constexpr (kVTileSize >= 8) {
VBF q4 = hn::BitCast(dbf, hn::Set(df, q_base[4][i]));
sum40 = hn::ReorderWidenMulAccumulate(df, q4, k0_vec, sum40, unused_sum1);
sum41 = hn::ReorderWidenMulAccumulate(df, q4, k1_vec, sum41, unused_sum1);
VBF q5 = hn::BitCast(dbf, hn::Set(df, q_base[5][i]));
sum50 = hn::ReorderWidenMulAccumulate(df, q5, k0_vec, sum50, unused_sum1);
sum51 = hn::ReorderWidenMulAccumulate(df, q5, k1_vec, sum51, unused_sum1);
VBF q6 = hn::BitCast(dbf, hn::Set(df, q_base[6][i]));
sum60 = hn::ReorderWidenMulAccumulate(df, q6, k0_vec, sum60, unused_sum1);
sum61 = hn::ReorderWidenMulAccumulate(df, q6, k1_vec, sum61, unused_sum1);
VBF q7 = hn::BitCast(dbf, hn::Set(df, q_base[7][i]));
sum70 = hn::ReorderWidenMulAccumulate(df, q7, k0_vec, sum70, unused_sum1);
sum71 = hn::ReorderWidenMulAccumulate(df, q7, k1_vec, sum71, unused_sum1);
}
}
}
// Handles NF v rows of flash attention for NF q.k dot products from one q row.
// Automatically handles masking for causal attention and different start_pos
// and last_pos values.
template <class DF, class VF = hn::Vec<DF>>
HWY_INLINE float SingleFlashAttentionRowVector(DF df, size_t start_pos,
size_t pos, size_t last_pos,
VF& x, float& old_max,
float& old_d) {
if (pos < start_pos) {
size_t mask_size = start_pos - pos;
const VF neg_inf = hn::Neg(hn::Inf(df));
x = hn::IfThenElse(hn::FirstN(df, mask_size), neg_inf, x);
}
if (pos + hn::Lanes(df) > last_pos) {
size_t mask_size = pos <= last_pos ? last_pos + 1 - pos : 0;
const VF neg_inf = hn::Neg(hn::Inf(df));
x = hn::IfThenElse(hn::FirstN(df, mask_size), x, neg_inf);
}
float m = hn::ReduceMax(df, x);
m = std::max(m, old_max);
x = hn::Exp(df, hn::Sub(x, hn::Set(df, m)));
float scale = old_d * std::exp(old_max - m);
old_d = hn::ReduceSum(df, x) + scale;
old_max = m;
if (old_d > 0.0f) {
const float one_over_d = 1.0f / old_d;
scale *= one_over_d;
x = hn::Mul(x, hn::Set(df, one_over_d));
} else {
scale = 0.0f;
x = hn::Zero(df);
}
return scale;
}
// Handles 2NF v rows of flash attention for 2NF q.k dot products from 1 q row.
// Automatically handles masking for causal attention and different start_pos
// and last_pos values.
template <class DF, class VF = hn::Vec<DF>>
HWY_INLINE float DoubleFlashAttentionRowVector(DF df, size_t start_pos,
size_t pos, size_t last_pos,
VF& x0, VF& x1, float& old_max,
float& old_d) {
const size_t kNF = hn::Lanes(df);
const VF neg_inf = hn::Neg(hn::Inf(df));
if (pos < start_pos) {
if (pos + kNF <= start_pos) {
x0 = neg_inf;
size_t mask_size = start_pos - (pos + kNF);
x1 = hn::IfThenElse(hn::FirstN(df, mask_size), neg_inf, x1);
} else {
size_t mask_size = start_pos - pos;
x0 = hn::IfThenElse(hn::FirstN(df, mask_size), neg_inf, x0);
}
}
if (pos + 2 * kNF > last_pos) {
if (pos + kNF > last_pos) {
size_t mask_size = pos <= last_pos ? last_pos + 1 - pos : 0;
x0 = hn::IfThenElse(hn::FirstN(df, mask_size), x0, neg_inf);
x1 = neg_inf;
} else {
size_t mask_size = last_pos + 1 - (pos + kNF);
x1 = hn::IfThenElse(hn::FirstN(df, mask_size), x1, neg_inf);
}
}
VF x_max = hn::Max(x0, x1);
float m = hn::ReduceMax(df, x_max);
m = std::max(m, old_max);
VF m_vec = hn::Set(df, m);
x0 = hn::Exp(df, hn::Sub(x0, m_vec));
x1 = hn::Exp(df, hn::Sub(x1, m_vec));
float scale = old_d * std::exp(old_max - m);
VF x_sum = hn::Add(x0, x1);
old_d = hn::ReduceSum(df, x_sum) + scale;
old_max = m;
if (old_d > 0.0f) {
const float one_over_d = 1.0f / old_d;
scale *= one_over_d;
VF one_over_d_vec = hn::Set(df, one_over_d);
x0 = hn::Mul(x0, one_over_d_vec);
x1 = hn::Mul(x1, one_over_d_vec);
} else {
scale = 0.0f;
x0 = hn::Zero(df);
x1 = hn::Zero(df);
}
return scale;
}
// Reduces each of x and stores in following lanes of max (tested with float32)
template <class DF, typename T = hn::TFromD<DF>,
class DF4 = hn::CappedTag<T, 4>, class VF4 = hn::Vec<DF4>,
class VF = hn::Vec<DF>, typename F>
static HWY_INLINE VF4 Reduce4(DF df, VF x_0, VF x_1, VF x_2, VF x_3,
F reducer) {
const DF4 df4;
constexpr size_t kMaxLanes = hn::MaxLanes(df);
HWY_LANES_CONSTEXPR size_t kLanes = hn::Lanes(df);
HWY_ALIGN T x_transposed[4 * kMaxLanes];
hn::StoreInterleaved4(x_0, x_1, x_2, x_3, df, x_transposed);
VF x01 =
reducer(hn::Load(df, x_transposed), hn::Load(df, x_transposed + kLanes));
VF x23 = reducer(hn::Load(df, x_transposed + 2 * kLanes),
hn::Load(df, x_transposed + 3 * kLanes));
VF x0123 = reducer(x01, x23);
hn::Store(x0123, df, x_transposed);
VF4 result = hn::Load(df4, x_transposed);
for (int i = 1; i < kLanes / 4; ++i) {
result = reducer(result, hn::Load(df4, x_transposed + i * 4));
}
return result;
}
// Handles Up to 4 Q rows by NF*2 timesteps of flash attention.
template <int kNumQueries, class DF, class VF = hn::Vec<DF>>
static HWY_INLINE void FlashAttentionTileStepAndApplySoftCap(
DF df, float att_cap, float one_over_att_cap, VF& x_0_p0, VF& x_0_p1,
VF& x_1_p0, VF& x_1_p1, VF& x_2_p0, VF& x_2_p1, VF& x_3_p0, VF& x_3_p1,
float* HWY_RESTRICT old_max, float* HWY_RESTRICT old_d,
float* HWY_RESTRICT scales) {
using DF4 = hn::CappedTag<float, 4>;
const DF4 df4;
using VF4 = hn::Vec<DF4>;
static_assert(kNumQueries >= 1 && kNumQueries <= 4);
VF4 new_max = hn::Set(df4, kNegInf);
VF max_0, max_1, max_2, max_3 = hn::Zero(df);
max_0 = hn::Max(x_0_p0, x_0_p1);
if constexpr (kNumQueries >= 2) {
max_1 = hn::Max(x_1_p0, x_1_p1);
}
if constexpr (kNumQueries >= 3) {
max_2 = hn::Max(x_2_p0, x_2_p1);
}
if constexpr (kNumQueries >= 4) {
max_3 = hn::Max(x_3_p0, x_3_p1);
}
if constexpr (kNumQueries == 1) {
new_max = hn::InsertLane(new_max, 0, hn::ReduceMax(df, max_0));
} else {
new_max = Reduce4(df, max_0, max_1, max_2, max_3,
[](auto a, auto b) HWY_ATTR { return hn::Max(a, b); });
}
if (att_cap > 0.0f) {
VF4 cap = hn::Set(df4, att_cap);
VF4 one_over_cap = hn::Set(df4, one_over_att_cap);
new_max = hn::Mul(cap, hn::Tanh(df4, hn::Mul(new_max, one_over_cap)));
}
VF4 old_max_vf = hn::Set(df4, kNegInf);
old_max_vf = hn::LoadU(df4, old_max);
new_max = hn::Max(new_max, old_max_vf);
auto changed_max = hn::Gt(new_max, hn::Set(df4, kNegInf));
// TODO figure out what was wrong with broadcasts and change to that.
hn::StoreU(new_max, df4, old_max);
if constexpr (kNumQueries >= 1) {
const VF new_max_0 = hn::Set(df, old_max[0]);
x_0_p0 = hn::Exp(df, hn::Sub(x_0_p0, new_max_0));
x_0_p1 = hn::Exp(df, hn::Sub(x_0_p1, new_max_0));
}
if constexpr (kNumQueries >= 2) {
const VF new_max_0 = hn::Set(df, old_max[1]);
x_1_p0 = hn::Exp(df, hn::Sub(x_1_p0, new_max_0));
x_1_p1 = hn::Exp(df, hn::Sub(x_1_p1, new_max_0));
}
if constexpr (kNumQueries >= 3) {
const VF new_max_0 = hn::Set(df, old_max[2]);
x_2_p0 = hn::Exp(df, hn::Sub(x_2_p0, new_max_0));
x_2_p1 = hn::Exp(df, hn::Sub(x_2_p1, new_max_0));
}
if constexpr (kNumQueries >= 4) {
const VF new_max_0 = hn::Set(df, old_max[3]);
x_3_p0 = hn::Exp(df, hn::Sub(x_3_p0, new_max_0));
x_3_p1 = hn::Exp(df, hn::Sub(x_3_p1, new_max_0));
}
VF4 old_d_vf = hn::Set(df4, 0.0f);
old_d_vf = hn::LoadU(df4, old_d);
VF4 scale = hn::Mul(old_d_vf, hn::Exp(df4, hn::Sub(old_max_vf, new_max)));
VF4 x_sum = hn::Zero(df4);
if constexpr (kNumQueries == 1) {
x_sum = hn::Set(df4, hn::ReduceSum(df, x_0_p0) + hn::ReduceSum(df, x_0_p1));
} else {
VF x_0_sum = hn::Add(x_0_p0, x_0_p1);
VF x_1_sum = hn::Add(x_1_p0, x_1_p1);
VF x_2_sum = hn::Add(x_2_p0, x_2_p1);
VF x_3_sum = hn::Add(x_3_p0, x_3_p1);
x_sum = Reduce4(df, x_0_sum, x_1_sum, x_2_sum, x_3_sum,
[](auto a, auto b) HWY_ATTR { return hn::Add(a, b); });
}
old_d_vf = hn::Add(scale, x_sum);
auto non_zero_mask = hn::Gt(old_d_vf, hn::Set(df4, 0.0f));
const VF zero = hn::Zero(df);
const VF4 zero4 = hn::Zero(df4);
const VF4 one_over_d =
hn::MaskedDivOr(zero4, non_zero_mask, hn::Set(df4, 1.0f), old_d_vf);
HWY_ALIGN float tmp_one_over_d[4];
hn::Store(one_over_d, df4, tmp_one_over_d);
hn::BlendedStore(old_d_vf, changed_max, df4, old_d);
scale = hn::Mul(scale, one_over_d);
hn::BlendedStore(scale, changed_max, df4, scales);
if (hn::ExtractLane(old_d_vf, 0) > 0.0f && scales[0] != 1.0f) {
const VF one_over_d_0 = hn::Set(df, tmp_one_over_d[0]);
x_0_p0 = hn::Mul(x_0_p0, one_over_d_0);
x_0_p1 = hn::Mul(x_0_p1, one_over_d_0);
} else {
x_0_p0 = zero;
x_0_p1 = zero;
}
if constexpr (kNumQueries >= 2) {
if (hn::ExtractLane(old_d_vf, 1) > 0.0f && scales[1] != 1.0f) {
const VF one_over_d_1 = hn::Set(df, tmp_one_over_d[1]);
x_1_p0 = hn::Mul(x_1_p0, one_over_d_1);
x_1_p1 = hn::Mul(x_1_p1, one_over_d_1);
} else {
x_1_p0 = zero;
x_1_p1 = zero;
}
}
if constexpr (kNumQueries >= 3) {
if (hn::ExtractLane(old_d_vf, 2) > 0.0f && scales[2] != 1.0f) {
const VF one_over_d_2 = hn::Set(df, tmp_one_over_d[2]);
x_2_p0 = hn::Mul(x_2_p0, one_over_d_2);
x_2_p1 = hn::Mul(x_2_p1, one_over_d_2);
} else {
x_2_p0 = zero;
x_2_p1 = zero;
}
}
if constexpr (kNumQueries >= 4) {
if (hn::ExtractLane(old_d_vf, 3) > 0.0f && scales[3] != 1.0f) {
const VF one_over_d_3 = hn::Set(df, tmp_one_over_d[3]);
x_3_p0 = hn::Mul(x_3_p0, one_over_d_3);
x_3_p1 = hn::Mul(x_3_p1, one_over_d_3);
} else {
x_3_p0 = zero;
x_3_p1 = zero;
}
}
}
// Implements flash attention for a strip of tiles of size 1, 4 or 8 query
// vectors by 2NF positions in K.
// It iterates through tiles in K from `params.min_start_pos / 2NF * 2NF` up to
// `params.max_last_pos` (rounded up to the nearest multiple of 2NF).
// Masking allows each row within a tile to have a different start and end
// position.
//
// @param params FlashAttentionParams containing the extent of the strip and
// size of the tiles.
// @param q The query matrix [batch_size * q_heads, qkv_dim] in BF16 format.
// @param k Key matrix from KV cache. K is always pre-transposed to shape:
// [seq_len / 2kNF, layers * kv_heads * qkv_dim/2 * 2kNF * 2],
// where the /2, *2 represents that pairs of qkv_dim elements are kept
// together to make best use of BF16 dot product instructions.
// @param v Value matrix [seq_len, qkv_dim] from KV cache.
// @param layer_idx The index of the current transformer layer.
// @param activations Attention configurations and buffers.
// @param att_out Output buffer for attention results.
// @param ctx Threading context.
// @param worker Worker thread index.
template <size_t kVTileSize, typename QType, typename KVType>
Tile4FlashState TileFlashAttention148(
const FlashAttentionParams& params, const MatPtrT<QType>& q,
const MatPtrT<KVType>& k, const MatPtrT<KVType>& v, const size_t layer_idx,
const AttentionActivationsPtrs& activations, MatPtrT<float>& att_out,
size_t qkv_dim, ThreadingContext& ctx, const size_t worker,
AttentionImpl attention_impl) {
constexpr Zones kZone =
kVTileSize == 8
? Zones::kFlashAttentionTileFlashAttention8
: (kVTileSize == 4 ? Zones::kFlashAttentionTileFlashAttention4
: Zones::kFlashAttentionTileFlashAttention1);
GCPP_ZONE(ctx, worker, kZone);
using DF = hn::ScalableTag<float>;
const DF df;
using VF = hn::Vec<DF>;
float att_cap = activations.config.att_cap;
float one_over_cap = att_cap > 0.0f ? 1.0f / att_cap : 0.0f;
const size_t kHTileSize = 2 * hn::Lanes(df);
float scales[kVTileSize];
for (size_t i = 0; i < kVTileSize; ++i) {
hwy::ZeroBytes(att_out.Row(0) + params.out_offsets[i],
qkv_dim * sizeof(att_out.Row(0)[0]));
}
Tile4FlashState state;
size_t position = params.min_start_pos / kHTileSize * kHTileSize;
while (position <= params.max_last_pos) {
// Each pair of vectors covers 2NF positions in K, with up to 8 pairs of
// vectors covering 1, 4 or 8 queries.
VF x00, x01;
VF HWY_MAYBE_UNUSED x10, x11;
VF HWY_MAYBE_UNUSED x20, x21;
VF HWY_MAYBE_UNUSED x30, x31;
VF HWY_MAYBE_UNUSED x40, x41;
VF HWY_MAYBE_UNUSED x50, x51;
VF HWY_MAYBE_UNUSED x60, x61;
VF HWY_MAYBE_UNUSED x70, x71;
constexpr size_t kMaxNF = hn::MaxLanes(df);
size_t v_pos[2 * kMaxNF];
for (size_t i = 0; i < kHTileSize; ++i) {
v_pos[i] = activations.div_seq_len.Remainder(position + i);
}
if constexpr (IsF32<QType>()) {
if constexpr (HWY_NATIVE_DOT_BF16) {
QDotKTile148FloatNative<kVTileSize>(df, q.Row(0), params.out_offsets,
qkv_dim / 2, k, position, x00, x01,
x10, x11, x20, x21, x30, x31, x40,
x41, x50, x51, x60, x61, x70, x71);
} else {
QDotKTile148FloatNotNative<kVTileSize>(
df, q.Row(0), params.out_offsets, qkv_dim / 2, k, position, x00,
x01, x10, x11, x20, x21, x30, x31, x40, x41, x50, x51, x60, x61,
x70, x71);
}
} else {
if constexpr (HWY_NATIVE_DOT_BF16) {
QDotKTile148BF16Native<kVTileSize>(df, q.Row(0), params.q_offsets,
qkv_dim / 2, k, position, x00, x01,
x10, x11, x20, x21, x30, x31, x40,
x41, x50, x51, x60, x61, x70, x71);
} else {
QDotKTile148BF16NotNative<kVTileSize>(
df, q.Row(0), params.q_offsets, qkv_dim / 2, k, position, x00, x01,
x10, x11, x20, x21, x30, x31, x40, x41, x50, x51, x60, x61, x70,
x71);
}
}
if (att_cap > 0.0f) {
// Compute tanh(x / cap) * cap, being LogitsSoftCap on the tile.
ApplySoftCap<kVTileSize>(df, att_cap, one_over_cap, x00, x10, x20, x30,
x40, x50, x60, x70);
ApplySoftCap<kVTileSize>(df, att_cap, one_over_cap, x01, x11, x21, x31,
x41, x51, x61, x71);
}
scales[0] = DoubleFlashAttentionRowVector(
df, params.start_pos[0], position, params.last_pos[0], x00, x01,
state.row_states[0].max, state.row_states[0].d);
if constexpr (kVTileSize >= 4) {
scales[1] = DoubleFlashAttentionRowVector(
df, params.start_pos[1], position, params.last_pos[1], x10, x11,
state.row_states[1].max, state.row_states[1].d);
scales[2] = DoubleFlashAttentionRowVector(
df, params.start_pos[2], position, params.last_pos[2], x20, x21,
state.row_states[2].max, state.row_states[2].d);
scales[3] = DoubleFlashAttentionRowVector(
df, params.start_pos[3], position, params.last_pos[3], x30, x31,
state.row_states[3].max, state.row_states[3].d);
MulByConstAndAddVT4Mem(df, scales, x00, x01, x10, x11, x20, x21, x30, x31,
v, v_pos, params.max_last_pos + 1 - position,
att_out.Row(0), params.out_offsets, qkv_dim);
} else {
MulByConstAndAddVT1Mem(df, scales, x00, x01, v, v_pos,
params.max_last_pos + 1 - position, att_out.Row(0),
params.out_offsets, qkv_dim);
}
if constexpr (kVTileSize >= 8) {
scales[4] = DoubleFlashAttentionRowVector(
df, params.start_pos[4], position, params.last_pos[4], x40, x41,
state.row_states[4].max, state.row_states[4].d);
scales[5] = DoubleFlashAttentionRowVector(
df, params.start_pos[5], position, params.last_pos[5], x50, x51,
state.row_states[5].max, state.row_states[5].d);
scales[6] = DoubleFlashAttentionRowVector(
df, params.start_pos[6], position, params.last_pos[6], x60, x61,
state.row_states[6].max, state.row_states[6].d);
scales[7] = DoubleFlashAttentionRowVector(
df, params.start_pos[7], position, params.last_pos[7], x70, x71,
state.row_states[7].max, state.row_states[7].d);
MulByConstAndAddVT4Mem(df, scales + 4, x40, x41, x50, x51, x60, x61, x70,
x71, v, v_pos, params.max_last_pos + 1 - position,
att_out.Row(0), params.out_offsets + 4, qkv_dim);
}
position += kHTileSize;
}
return state;
}
// The vertical tile size is determined by the ability to use tiling and the
// target_parallelism. In practice the possible tile sizes in order of
// preference for efficiency are 8, 4, 1. The final tile size is chosen to be
// the largest possible that allows for target_parallelism parallel tasks.
size_t GetVTileSize(size_t kNF, size_t num_head_groups, size_t num_tokens,
size_t total_tasks, size_t target_parallelism) {
const size_t kMaxEqualK = num_head_groups * num_tokens;
if (total_tasks / k8xNFVTileSize >= target_parallelism &&
kMaxEqualK >= k8xNFVTileSize && kNF >= k8xNFVTileSize) {
return k8xNFVTileSize;
}
if (total_tasks / k4xNFVTileSize >= target_parallelism &&
kMaxEqualK >= k4xNFVTileSize && kNF >= k4xNFVTileSize) {
return k4xNFVTileSize;
}
return 1;
}
// Clears and fills the params vector with FlashAttentionParams for the given
// num_tokens, target_parallelism, and layer_idx. Computes tile sizes and
// offsets for each tile to achieve target_parallelism.
void ComputeFlashParams(size_t num_tokens, const size_t target_parallelism,
size_t layer_idx, AttentionActivationsPtrs& activations,
QBatch& qbatch, AttentionImpl attention_impl,
std::vector<FlashAttentionParams>& params) {
const LayerConfig& layer_config = activations.config.layer_configs[layer_idx];
const hwy::Divisor div_qbatch(qbatch.Size());
const size_t qkv_dim = layer_config.qkv_dim;
using DF = hn::ScalableTag<float>;
const DF df;
const size_t kNF = hn::Lanes(df);
// A "head group" in the context of GQA refers to a collection of query
// heads that share the same key and value heads.
const size_t kHeadGroups = layer_config.heads / layer_config.kv_heads;
const size_t cache_layer_size = layer_config.CacheLayerSize();
const size_t token_batch = num_tokens * div_qbatch.GetDivisor();
const size_t total_tasks = token_batch * layer_config.heads;
size_t kVTileSize = GetVTileSize(kNF, kHeadGroups, num_tokens, total_tasks,
target_parallelism);
// All layers should have the same number of heads.
HWY_DASSERT(activations.div_heads.GetDivisor() == layer_config.heads);
// To maximize adjacent tasks with the same kv matrices, task index is encoded
// thus: [qi][kv_head][batch_idx][head_group]. Note that the head index is
// split into kv_head and head_group, since the head_group does not affect
// the KV matrices, and kv_head does. batch_idx does not affect the KV
// matrices, but does affect the last position in the sequence. qi affects
// everything.
params.clear();
for (uint32_t qi = 0; qi < div_qbatch.GetDivisor(); ++qi) {
for (uint32_t kv_head = 0; kv_head < layer_config.kv_heads; ++kv_head) {
const size_t head_offset = kv_head * qkv_dim * 2;
const uint32_t kv_offset = layer_idx * cache_layer_size + head_offset;
params.push_back(FlashAttentionParams{
.qi_index = qi,
.kv_offset = kv_offset,
});
for (uint32_t batch_idx = 0; batch_idx < num_tokens; ++batch_idx) {
const size_t pos = qbatch.Pos(qi) + batch_idx;
const size_t start_pos = StartPos(pos, activations.config, layer_idx);
size_t last = pos;
const size_t prefix_end = qbatch.PrefixEnd(qi);
if (prefix_end > 0 && prefix_end - 1 > last) {
// last_pos is inclusive.
last = prefix_end - 1;
}
for (size_t head_group = 0; head_group < kHeadGroups; ++head_group) {
size_t tasks_remaining = kHeadGroups - head_group +
kHeadGroups * (num_tokens - 1 - batch_idx);
// We want to fill a tile of size kVTileSize or k4xNFVTileSize if
// smaller, otherwise everything is singles to the next head group.
size_t tasks_required = params.back().v_tile_size < k4xNFVTileSize
? k4xNFVTileSize
: kVTileSize;
if (params.back().v_tile_size + tasks_remaining < tasks_required ||
params.back().v_tile_size == kVTileSize) {
// We don't have enough tasks remaining to fill a tile, or the
// current tile is full so start new tile.
params.push_back(FlashAttentionParams{
.qi_index = qi,
.kv_offset = kv_offset,
});
}
const size_t head = head_group + kHeadGroups * kv_head;
const size_t tq_idx = div_qbatch.GetDivisor() * batch_idx + qi;
auto& param = params.back();
size_t offset = param.v_tile_size;
param.q_offsets[offset] = activations.q_bf.Row(tq_idx) +
head * qkv_dim - activations.q_bf.Row(0);
param.out_offsets[offset] = activations.att_out.Row(tq_idx) +
head * qkv_dim -
activations.att_out.Row(0);
param.tq_idx[offset] = tq_idx;
param.start_pos[offset] = start_pos;
param.min_start_pos = HWY_MIN(param.min_start_pos, start_pos);
param.last_pos[offset] = last;
param.max_last_pos = HWY_MAX(param.max_last_pos, last);
++param.v_tile_size;
}
}
}
}
}
// Returns the maximum number of tiles needed for any query in the batch.
size_t GetMaxTiles(const std::vector<FlashAttentionParams>& params,
const size_t kHTileSize) {
size_t max_tiles = 0;
for (const auto& param : params) {
size_t start = param.min_start_pos / kHTileSize;
size_t last = param.max_last_pos / kHTileSize;
max_tiles = HWY_MAX(last + 1 - start, max_tiles);
}
return max_tiles;
}
// Splits params into smaller k-strips to allow for more parallelism.
// The strips are of size num_tiles_per_task * kHTileSize.
// split_params is cleared and filled with the split tasks.
void SplitTasksByKPos(std::vector<FlashAttentionParams>& params,
const size_t kHTileSize, const size_t num_tiles_per_task,
const size_t out_stride,
std::vector<FlashAttentionParams>& split_params) {
split_params.clear();
for (auto& param : params) {
param.split_index = split_params.size();
size_t start = param.min_start_pos / kHTileSize;
size_t last = param.max_last_pos / kHTileSize;
for (size_t tile_pos = start; tile_pos <= last;
tile_pos += num_tiles_per_task) {
auto& split_param = split_params.emplace_back(param);
split_param.i_of_n = (tile_pos - start) / num_tiles_per_task;
uint32_t tile_last = (tile_pos + num_tiles_per_task) * kHTileSize - 1;
if (tile_last < param.max_last_pos) {
split_param.max_last_pos = tile_last;
for (auto& last_pos : split_param.last_pos) {
last_pos = std::min(last_pos, tile_last);
}
}
uint32_t tile_start = tile_pos * kHTileSize;
if (tile_start > param.min_start_pos) {
split_param.min_start_pos = tile_start;
for (auto& start_pos : split_param.start_pos) {
start_pos = std::max(start_pos, tile_start);
}
}
if (split_param.i_of_n > 0) {
for (size_t i = 0; i < split_param.v_tile_size; ++i) {
split_param.tq_idx[i] =
param.tq_idx[i] * AttentionActivations::kThreadReplicationFactor +
split_param.i_of_n - 1;
split_param.out_offsets[i] =
param.out_offsets[i] +
(split_param.tq_idx[i] - param.tq_idx[i]) * out_stride;
}
}
}
}
}
// Clears and fills activations.flash_params with FlashAttentionParams for the
// given num_tokens, target_parallelism, and layer_idx. Computes tile sizes and
// offsets for each tile to achieve target_parallelism.
// If the parallelism is insufficient for this processor type, and the sequence
// length is sufficient, the tiles are upgraded to k4xNFVTileSize and the tasks
// are split along the k positions to achieve the desired parallelism.
// If splitting was required, returns that factor by which the tiles were
// upgraded, k4xNFVTileSize, otherwise returns 0.
uint32_t ComputeAndSplitFlashParams(const size_t kNF, const size_t num_tokens,
const size_t target_parallelism,
size_t layer_idx,
AttentionActivationsPtrs& activations,
QBatch& qbatch, ThreadingContext& ctx,
AttentionImpl attention_impl) {
ComputeFlashParams(num_tokens, target_parallelism, layer_idx, activations,
qbatch, attention_impl, activations.flash_params);
if (activations.flash_params.size() < ctx.pools.MaxWorkers()) {
// Insufficient parallelism for this processor type. Try splitting along the
// k positions.
size_t max_tiles = GetMaxTiles(activations.flash_params, kNF);
size_t desired_tiles_per_task = hwy::DivCeil(
activations.flash_params.size() * max_tiles, ctx.pools.MaxWorkers());
// The cost of combining split tasks is significant, so we want a minimum
// number of tiles per task, and we want to use k4xNFVTileSize if possible.
constexpr size_t kMinTilesPerTask = 4;
if (desired_tiles_per_task >= k4xNFVTileSize * kMinTilesPerTask) {
// We can afford to use k4xNFVTileSize vertically, so recompute params.
ComputeFlashParams(num_tokens,
activations.flash_params.size() / k4xNFVTileSize,
layer_idx, activations, qbatch, attention_impl,
activations.flash_params);
desired_tiles_per_task =
hwy::DivCeil(desired_tiles_per_task, k4xNFVTileSize);
SplitTasksByKPos(activations.flash_params, kNF, desired_tiles_per_task,
activations.att_out_reps.Stride(),
activations.split_flash_params);
return k4xNFVTileSize;
}
}
return 0;
}
// Combines results from split tasks, processing kNumNF * NF qkv values where
// kNumNF can be 1 4 or 16. This enables the intermediate results to be held in
// registers, which speeds up the combination step significantly.
template <size_t kNumNF>
void CombineSplitTasks1416(hwy::Span<const FlashAttentionParams> params,
size_t tile_pos, size_t qkv_offset,
AttentionActivationsPtrs& activations) {
using DF = hn::ScalableTag<float>;
const DF df;
using VF = hn::Vec<DF>;
const size_t kNF = hn::Lanes(df);
float overall_m = params[0].end_state.row_states[tile_pos].max;
float overall_d = params[0].end_state.row_states[tile_pos].d;
float* HWY_RESTRICT att_out =
activations.att_out.Row(0) + params[0].out_offsets[tile_pos] + qkv_offset;
VF result_0 = hn::Load(df, att_out);
VF result_1, result_2, result_3, result_4, result_5, result_6, result_7;
VF result_8, result_9, result_10, result_11, result_12, result_13, result_14;
VF result_15;
if constexpr (kNumNF > 1) {
result_1 = hn::Load(df, att_out + kNF);
result_2 = hn::Load(df, att_out + 2 * kNF);
result_3 = hn::Load(df, att_out + 3 * kNF);
}
if constexpr (kNumNF == 16) {
result_4 = hn::Load(df, att_out + 4 * kNF);
result_5 = hn::Load(df, att_out + 5 * kNF);
result_6 = hn::Load(df, att_out + 6 * kNF);
result_7 = hn::Load(df, att_out + 7 * kNF);
result_8 = hn::Load(df, att_out + 8 * kNF);
result_9 = hn::Load(df, att_out + 9 * kNF);
result_10 = hn::Load(df, att_out + 10 * kNF);
result_11 = hn::Load(df, att_out + 11 * kNF);
result_12 = hn::Load(df, att_out + 12 * kNF);
result_13 = hn::Load(df, att_out + 13 * kNF);
result_14 = hn::Load(df, att_out + 14 * kNF);
result_15 = hn::Load(df, att_out + 15 * kNF);
}
for (size_t i = 1; i < params.size() && params[i].i_of_n > 0; ++i) {
float m = params[i].end_state.row_states[tile_pos].max;
float d = params[i].end_state.row_states[tile_pos].d;
float new_m = std::max(overall_m, m);
// Scale factor for existing total given the change in max.
float old_scale = overall_d * std::exp(overall_m - new_m);
// Scale factor for new group to add.
float new_scale = d * std::exp(m - new_m);
float new_d = old_scale + new_scale;
float one_over_d = 1.0f / new_d;
old_scale *= one_over_d;
new_scale *= one_over_d;
overall_m = new_m;
overall_d = new_d;
float* HWY_RESTRICT att_in = activations.att_out_reps.Row(0) +
params[i].out_offsets[tile_pos] + qkv_offset;
VF old_scale_vec = hn::Set(df, old_scale);
VF new_scale_vec = hn::Set(df, new_scale);
result_0 = hn::Mul(result_0, old_scale_vec);
result_0 = hn::MulAdd(hn::Load(df, att_in), new_scale_vec, result_0);
if constexpr (kNumNF > 1) {
result_1 = hn::Mul(result_1, old_scale_vec);
result_2 = hn::Mul(result_2, old_scale_vec);
result_3 = hn::Mul(result_3, old_scale_vec);
result_1 =
hn::MulAdd(hn::Load(df, att_in + kNF), new_scale_vec, result_1);
result_2 =
hn::MulAdd(hn::Load(df, att_in + 2 * kNF), new_scale_vec, result_2);
result_3 =
hn::MulAdd(hn::Load(df, att_in + 3 * kNF), new_scale_vec, result_3);
}
if constexpr (kNumNF == 16) {
result_4 = hn::Mul(result_4, old_scale_vec);
result_5 = hn::Mul(result_5, old_scale_vec);
result_6 = hn::Mul(result_6, old_scale_vec);
result_7 = hn::Mul(result_7, old_scale_vec);
result_8 = hn::Mul(result_8, old_scale_vec);
result_9 = hn::Mul(result_9, old_scale_vec);
result_10 = hn::Mul(result_10, old_scale_vec);
result_11 = hn::Mul(result_11, old_scale_vec);
result_12 = hn::Mul(result_12, old_scale_vec);
result_13 = hn::Mul(result_13, old_scale_vec);
result_14 = hn::Mul(result_14, old_scale_vec);
result_15 = hn::Mul(result_15, old_scale_vec);
result_4 =
hn::MulAdd(hn::Load(df, att_in + 4 * kNF), new_scale_vec, result_4);
result_5 =
hn::MulAdd(hn::Load(df, att_in + 5 * kNF), new_scale_vec, result_5);
result_6 =
hn::MulAdd(hn::Load(df, att_in + 6 * kNF), new_scale_vec, result_6);
result_7 =
hn::MulAdd(hn::Load(df, att_in + 7 * kNF), new_scale_vec, result_7);
result_8 =
hn::MulAdd(hn::Load(df, att_in + 8 * kNF), new_scale_vec, result_8);
result_9 =
hn::MulAdd(hn::Load(df, att_in + 9 * kNF), new_scale_vec, result_9);
result_10 =
hn::MulAdd(hn::Load(df, att_in + 10 * kNF), new_scale_vec, result_10);
result_11 =
hn::MulAdd(hn::Load(df, att_in + 11 * kNF), new_scale_vec, result_11);
result_12 =
hn::MulAdd(hn::Load(df, att_in + 12 * kNF), new_scale_vec, result_12);
result_13 =
hn::MulAdd(hn::Load(df, att_in + 13 * kNF), new_scale_vec, result_13);
result_14 =
hn::MulAdd(hn::Load(df, att_in + 14 * kNF), new_scale_vec, result_14);
result_15 =
hn::MulAdd(hn::Load(df, att_in + 15 * kNF), new_scale_vec, result_15);
}
}
hn::Store(result_0, df, att_out);
if constexpr (kNumNF > 1) {
hn::Store(result_1, df, att_out + kNF);
hn::Store(result_2, df, att_out + 2 * kNF);
hn::Store(result_3, df, att_out + 3 * kNF);
}
if constexpr (kNumNF == 16) {
hn::Store(result_4, df, att_out + 4 * kNF);
hn::Store(result_5, df, att_out + 5 * kNF);
hn::Store(result_6, df, att_out + 6 * kNF);
hn::Store(result_7, df, att_out + 7 * kNF);
hn::Store(result_8, df, att_out + 8 * kNF);
hn::Store(result_9, df, att_out + 9 * kNF);
hn::Store(result_10, df, att_out + 10 * kNF);
hn::Store(result_11, df, att_out + 11 * kNF);
hn::Store(result_12, df, att_out + 12 * kNF);
hn::Store(result_13, df, att_out + 13 * kNF);
hn::Store(result_14, df, att_out + 14 * kNF);
hn::Store(result_15, df, att_out + 15 * kNF);
}
}
// Recombines results from split tasks, activations.att_out_reps ->
// activations.att_out. Instead of repeatedly calling MultiplyByConstAndAdd,
// which reads/writes the sum each time, the result is kept entirely in
// registers, and the task is split into 16NF, 4NF, and NF chunks, so that there
// are enough registers to hold the intermediate results.
void CombineSplitTasks(size_t qkv_dim, uint32_t tile_factor,
AttentionActivationsPtrs& activations,
ThreadingContext& ctx) {
GCPP_ZONE(ctx, 0, Zones::kFlashAttentionCombineSplit);
using DF = hn::ScalableTag<float>;
const DF df;
const size_t kNF = hn::Lanes(df);
uint32_t num_16 = qkv_dim / (16 * kNF);
uint32_t num_4 = (qkv_dim - kNF * 16 * num_16) / (4 * kNF);
uint32_t num_1 = hwy::DivCeil(qkv_dim - kNF * (16 * num_16 + 4 * num_4), kNF);
uint32_t tasks_per_qkv = num_16 + num_4 + num_1;
ParallelFor(
Parallelism::kFlat,
activations.flash_params.size() * tasks_per_qkv * tile_factor, ctx,
/*cluster_idx=*/0, Callers::kFlashAttention,
[&](size_t p, size_t worker) {
uint32_t tile = p / tasks_per_qkv;
uint32_t p_idx =
activations.flash_params[tile / tile_factor].split_index;
const auto& param = activations.split_flash_params[p_idx];
size_t remaining_params = activations.split_flash_params.size() - p_idx;
tile %= tile_factor;
if (tile >= param.v_tile_size) return;
int32_t qkv_task = p % tasks_per_qkv;
if (qkv_task < num_16) {
uint32_t qkv_offset = qkv_task * 16 * kNF;
CombineSplitTasks1416<16>(
hwy::Span<const FlashAttentionParams>(&param, remaining_params),
tile, qkv_offset, activations);
} else if (qkv_task < num_16 + num_4) {
uint32_t qkv_offset = (num_16 * 16 + (qkv_task - num_16) * 4) * kNF;
CombineSplitTasks1416<4>(
hwy::Span<const FlashAttentionParams>(&param, remaining_params),
tile, qkv_offset, activations);
} else {
uint32_t qkv_offset =
(num_16 * 16 + num_4 * 4 + (qkv_task - num_16 - num_4)) * kNF;
CombineSplitTasks1416<1>(
hwy::Span<const FlashAttentionParams>(&param, remaining_params),
tile, qkv_offset, activations);
}
});
}
// The nominal aim of attention is to combine 3 inputs Q[L,D], K[L,D], V[L,D]
// into a single output O[L,D].
// Conventional attention first computes A[L,L] = Q . KT
// followed by A = softmax(A) (over invididual rows).
// Then A is multiplied by V to get O[L,D].
// For each row of O, this takes a read of one row of Q L times, all of K,
// 3 write/reads of one row of A, read all of V, and read/write the one row of O
// L times. Ignoring the computation for now, and focusing just on memory,
// the one row of O takes L(4D+3) reads and L(D+3) writes.
// For the whole of Q, this is L^2(4D+3) reads and L^2(D+3) writes.
//
// Flash attention fuses these operations together, and operates on tiles of
// n Q rows x NF K positions, accumulated in n registers, where n is in
// {1, 4, 8} and NF is the number of float lanes in a register, being 16 for
// AVX3. This reduces the number of reads of Q by NF and reads of K by n. The
// softmax is converted to streaming form using the algorithm from:
// https://courses.cs.washington.edu/courses/cse599m/23sp/notes/flashattn.pdf,
// which eliminates the need to store A to memory. The accumulated Q.KT result
// is passed via the streaming softmax directly to the A.V step.
// To make the dot product computation more efficient, Q, K, and V are stored
// as BF16 and K is transposed to shape:
// [seq_len / NF, layers * kv_heads * qkv_dim/2 * NF * 2], where the /2, *2
// represents that pairs of qkv_dim elements are kept together to make best
// use of BF16 dot product instructions, where each pair of adjacent BF16
// values from Q and K are mul-added into a single F32 result.
//
// A further complication is that real attention is not as simple as documented
// in the paper and above. There are multiple query heads, differing KV, and
// different sequence lengths, and the difference between prefill and decode,
// so a lot of the work in FlashAttention is making sure that a collection of q
// rows with the same KV and sequence length are grouped together so that the
// largest possible tiles can be used. This is dealt with by the
// ComputeAndSplitFlashParams() function.
void FlashAttention(const size_t num_tokens, const size_t target_parallelism,
const size_t layer_idx, const MatPtr& query_norm_scale,
AttentionActivationsPtrs& activations, QBatch& qbatch,
ThreadingContext& ctx, AttentionImpl attention_impl) {
GCPP_ZONE(ctx, 0, Zones::kFlashAttentionInclusive);
RMSNormAndPositionalEncoding(num_tokens, qbatch, activations.q,
query_norm_scale, layer_idx, activations, ctx);
const LayerConfig& layer_config = activations.config.layer_configs[layer_idx];
const size_t qkv_dim = layer_config.qkv_dim;
const size_t seq_len =
static_cast<size_t>(activations.div_seq_len.GetDivisor());
using DF = hn::ScalableTag<float>;
const DF df;
const size_t kNF = hn::Lanes(df);
// Compress q to q_bf.
// TODO(rays): Move this into RMSNormAndPositionalEncoding().
ParallelFor(
Parallelism::kWithinCluster, activations.q.Rows(), ctx,
/*cluster_idx=*/0, Callers::kFlashAttention,
[&](size_t row, size_t worker) {
CompressPerThread tls;
const hn::ScalableTag<float> df;
CompressTraits<BF16>::Compress(
df, activations.q.Row(row), activations.q.Cols(), tls,
MakeSpan(activations.q_bf.Row(row), activations.q_bf.Cols()), 0);
});
int tile_factor =
ComputeAndSplitFlashParams(kNF, num_tokens, target_parallelism, layer_idx,
activations, qbatch, ctx, attention_impl);
auto& params = tile_factor >= 1 ? activations.split_flash_params
: activations.flash_params;
size_t num_tasks = params.size();
// For each head/token/query, compute fused flash Q.K, softmax and weighted V.
const auto func = [&](const size_t task, size_t worker) HWY_ATTR {
GCPP_ZONE(ctx, worker, Zones::kFlashAttentionFlashAttention);
auto& param = params[task];
auto& kv_cache = qbatch.KV(param.qi_index).kv_cache;
auto& kT_cache = qbatch.KV(param.qi_index).k_cache;
MatPtrT<KV_t> kT("k_T_view", Extents2D(hwy::DivCeil(seq_len, 2 * kNF),
qkv_dim * 2 * kNF));
kT.SetPtr(kT_cache.Row(0) + param.kv_offset * kNF, kT_cache.Stride());
MatPtrT<KV_t> v("v_view", Extents2D(seq_len, qkv_dim));
v.SetPtr(kv_cache.Row(0) + param.kv_offset + qkv_dim, kv_cache.Stride());
auto& vT_cache = qbatch.KV(param.qi_index).v_cache;
MatPtrT<KV_t> vT("v_T_view", Extents2D(hwy::DivCeil(seq_len, 2 * kNF),
qkv_dim * 2 * kNF));
vT.SetPtr(vT_cache.Row(0) + param.kv_offset * kNF, vT_cache.Stride());
MatPtrT<float>& att_out =
param.i_of_n == 0 ? activations.att_out : activations.att_out_reps;
if (param.v_tile_size == k8xNFVTileSize) {
param.end_state = TileFlashAttention148<k8xNFVTileSize>(
param, activations.q_bf, kT, vT, layer_idx, activations, att_out,
qkv_dim, ctx, worker, attention_impl);
} else if (param.v_tile_size == k4xNFVTileSize) {
param.end_state = TileFlashAttention148<k4xNFVTileSize>(
param, activations.q_bf, kT, vT, layer_idx, activations, att_out,
qkv_dim, ctx, worker, attention_impl);
} else {
param.end_state = TileFlashAttention148<1>(
param, activations.q_bf, kT, vT, layer_idx, activations, att_out,
qkv_dim, ctx, worker, attention_impl);
}
};
{
PROFILER_ZONE("Gen.FlashAttention.ForkJoin");
// Full parallelism is helpful, SmallParallelFor is insufficient.
HierarchicalParallelFor(num_tasks, ctx, Callers::kFlashAttention, func);
}
if (tile_factor >= 1) {
// Run the flash attention correction on the partial outputs.
CombineSplitTasks(qkv_dim, tile_factor, activations, ctx);
}
}
// NOLINTNEXTLINE(google-readability-namespace-comments)
} // namespace HWY_NAMESPACE
} // namespace gcpp
HWY_AFTER_NAMESPACE();
Reference in New Issue View Git Blame Copy Permalink