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

Merge 7efeb4fe06 into f56d18dd68

This commit is contained in:
copybara-service[bot] 2026-03-25 12:02:06 +00:00 committed by GitHub
parent f56d18dd68 7efeb4fe06
commit 3cc9ec1177
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 21 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
gemma/tiled_attention.h
View File

@ -15,26 +15,27 @@
namespace gcpp {
// Passed to HWY_VISIT_TARGETS; declares for one target.
#define GEMMA_DECL_TILED_ATTENTION(TARGET, NAMESPACE) \
namespace NAMESPACE { \
void TiledAttention(AttentionImpl attention_impl, size_t num_tokens, \
size_t layer_idx, const LayerWeightsPtrs& layer, \
AttentionActivationsPtrs& activations, QBatch& qbatch, \
MatMulEnv& env, int flags); \
void TransposeStridedQueries(hwy::Span<float*> queries, int qkv_dim, \
hwy::Span<float> transposed_queries); \
void LocalAttentionForAllHeadsTokensAndBatch( \
AttentionImpl attention_impl, const size_t num_tokens, \
const size_t layer_idx, const LayerWeightsPtrs& layer, \
AttentionActivationsPtrs& activations, QBatch& qbatch, \
ThreadingContext& ctx); \
\
template <typename OutT> \
std::tuple<std::vector<OutT, hwy::AlignedAllocator<OutT>>, \
std::vector<OutT*>, AlignedFloatVector> \
TransposeQueriesToGroupsOfNBF16orInt16(hwy::Span<float*> queries_ptrs, \
int qkv_dim, size_t group_size); \
/* NOLINTNEXTLINE(google-readability-namespace-comments) */ \
#define GEMMA_DECL_TILED_ATTENTION(TARGET, NAMESPACE) \
namespace NAMESPACE { \
void TiledAttention(AttentionImpl attention_impl, size_t num_tokens, \
size_t layer_idx, const LayerWeightsPtrs& layer, \
AttentionActivationsPtrs& activations, QBatch& qbatch, \
MatMulEnv& env, int flags); \
void TransposeStridedQueries(hwy::Span<float*> queries, int qkv_dim, \
hwy::Span<float> transposed_queries); \
void LocalAttentionForAllHeadsTokensAndBatch( \
AttentionImpl attention_impl, const size_t num_tokens, \
const size_t layer_idx, const LayerWeightsPtrs& layer, \
AttentionActivationsPtrs& activations, QBatch& qbatch, \
ThreadingContext& ctx); \
\
template <typename OutT> \
std::tuple<std::vector<OutT, hwy::AlignedAllocator<OutT>>, \
std::vector<OutT*>, AlignedFloatVector> \
TransposeQueriesToGroupsOfNBF16orInt16(hwy::Span<float*> queries_ptrs, \
int qkv_dim, size_t group_size); \
\
/* NOLINTNEXTLINE(google-readability-namespace-comments) */ \
} // namespace NAMESPACE
// Function declarations for each SIMD target. Allows direct call from the