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

Merge 267dbe00cb into 5a05857deb

This commit is contained in:
copybara-service[bot] 2025-10-28 04:22:08 +00:00 committed by GitHub
parent 5a05857deb 267dbe00cb
commit 877abf4da5
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
5 changed files with 14 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
gemma/activations.h
View File

@ -203,6 +203,12 @@ struct Activations {
ffw_out.OverrideRows(batch_size);
attention_storage.SetBatchSize(batch_size);
attention.q = attention_storage.q;
attention.q_T = attention_storage.q_T;
attention.pre_att_rms_out = attention_storage.pre_att_rms_out;
attention.att = attention_storage.att;
attention.att_out = attention_storage.att_out;
attention.att_sums = attention_storage.att_sums;
}
const LayerConfig& layer_config;

4
gemma/attention.cc
View File

@ -130,7 +130,7 @@ static HWY_INLINE void WeightedSumV(
void SingleDotSoftmaxWeightedSum(
const size_t pos, const size_t start_pos, const size_t last_pos,
float* HWY_RESTRICT q, const MatPtrT<KV_t>& k, const MatPtrT<KV_t>& v,
const MatPtrT<float>& query_norm_scale, const size_t layer_idx,
const MatPtr& query_norm_scale, const size_t layer_idx,
const AttentionActivationsPtrs& activations, float* HWY_RESTRICT att,
float* HWY_RESTRICT att_out, ThreadingContext& ctx, const size_t worker) {
const float att_cap = activations.config.att_cap;
@ -169,7 +169,7 @@ size_t StartPos(size_t pos, const ModelConfig& config, size_t layer_idx) {
}
void DotSoftmaxWeightedSum(const size_t num_tokens, const size_t layer_idx,
const MatPtrT<float>& query_norm_scale,
const MatPtr& query_norm_scale,
AttentionActivationsPtrs& activations,
QBatch& qbatch, ThreadingContext& ctx) {
GCPP_ZONE(ctx, 0, Zones::kGenAttentionDotSoftmaxWeightedSumInclusive);

4
gemma/attention.h
View File

@ -38,12 +38,12 @@ namespace gcpp {
void SingleDotSoftmaxWeightedSum( \
const size_t pos, const size_t start_pos, const size_t last_pos, \
float* HWY_RESTRICT q, const MatPtrT<KV_t>& k, const MatPtrT<KV_t>& v, \
const MatPtrT<float>& query_norm_scale, size_t layer_idx, \
const MatPtr& query_norm_scale, size_t layer_idx, \
const AttentionActivationsPtrs& activations, float* HWY_RESTRICT att, \
float* HWY_RESTRICT att_out, ThreadingContext& ctx, size_t worker); \
\
void DotSoftmaxWeightedSum(const size_t num_tokens, size_t layer_idx, \
const MatPtrT<float>& query_norm_scale, \
const MatPtr& query_norm_scale, \
AttentionActivationsPtrs& activations, \
QBatch& qbatch, ThreadingContext& ctx); \
\

5
gemma/flash_attention.cc
View File

@ -91,7 +91,7 @@ static void TransposeQ(const MatPtrT<float>& q, MatPtrT<float>& q_t,
// Updates q in place for RMSNorm and positional encoding.
void RMSNormAndPositionalEncoding(const size_t num_tokens, const QBatch& qbatch,
MatPtrT<float>& q,
const MatPtrT<float>& query_norm_scale,
const MatPtr& query_norm_scale,
const size_t layer_idx,
const AttentionActivationsPtrs& activations,
ThreadingContext& ctx) {
@ -592,8 +592,7 @@ size_t GetVTileSize(size_t kNF, size_t num_head_groups, size_t num_tokens,
// grouped together so that mode 1 or 2 can be used, and choosing which of the
// 3 modes to use for best efficiency.
void FlashAttention(const size_t num_tokens, const size_t target_parallelism,
const size_t layer_idx,
const MatPtrT<float>& query_norm_scale,
const size_t layer_idx, const MatPtr& query_norm_scale,
AttentionActivationsPtrs& activations, QBatch& qbatch,
ThreadingContext& ctx) {
GCPP_ZONE(ctx, 0, Zones::kFlashAttentionInclusive);

5
gemma/flash_attention.h
View File

@ -30,7 +30,7 @@ namespace gcpp {
namespace NAMESPACE { \
void RMSNormAndPositionalEncoding( \
size_t num_tokens, const QBatch& qbatch, MatPtrT<float>& q, \
const MatPtrT<float>& query_norm_scale, size_t layer_idx, \
const MatPtr& query_norm_scale, size_t layer_idx, \
const AttentionActivationsPtrs& activations, ThreadingContext& ctx); \
\
void SingleFlashAttention(size_t start_pos, size_t last_pos, \
@ -45,8 +45,7 @@ namespace gcpp {
size_t total_tasks, size_t target_parallelism); \
\
void FlashAttention(size_t num_tokens, size_t target_parallelism, \
size_t layer_idx, \
const MatPtrT<float>& query_norm_scale, \
size_t layer_idx, const MatPtr& query_norm_scale, \
AttentionActivationsPtrs& activations, QBatch& qbatch, \
ThreadingContext& ctx); \
/* NOLINTNEXTLINE(google-readability-namespace-comments) */ \