Skip to content

Internal changes #782

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 1 commit into
base: dev
Choose a base branch
from
Open

Internal changes #782

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions gemma/flash_attention.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,9 @@
#include <stdint.h>

#include <algorithm>
#include <array>
#include <cmath>
#include <cstdlib>
#include <limits>

#include "compression/types.h" // GEMMA_DISABLED_TARGETS
Expand Down
65 changes: 33 additions & 32 deletions gemma/flash_attention.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,38 +29,39 @@
namespace gcpp {

// Passed to HWY_VISIT_TARGETS; declares for one target.
#define GEMMA_DECL_FLASH_ATTENTION(TARGET, NAMESPACE) \
namespace NAMESPACE { \
void RMSNormAndPositionalEncoding( \
size_t num_tokens, const QBatch& qbatch, MatPtrT<float>& q, \
const MatPtr& query_norm_scale, size_t layer_idx, \
const AttentionActivationsPtrs& activations, ThreadingContext& ctx); \
\
void SingleFlashAttention(size_t start_pos, size_t last_pos, \
const float* HWY_RESTRICT q, \
const MatPtrT<KV_t>& k, const MatPtrT<KV_t>& v, \
size_t layer_idx, \
const AttentionActivationsPtrs& activations, \
float* HWY_RESTRICT att_out, \
ThreadingContext& ctx, size_t worker); \
\
Tile4FlashState TileFlashAttention4( \
const MatPtrT<BF16>& q, const uint32_t* HWY_RESTRICT q_offsets, \
const MatPtrT<KV_t>& k, size_t start_pos, \
const uint32_t* HWY_RESTRICT last_pos, size_t min_last_pos, \
size_t max_last_pos, const MatPtrT<KV_t>& v, size_t layer_idx, \
const LayerWeightsPtrs& layer, const AttentionActivations& activations, \
MatPtrT<float>& att_out, const uint32_t* HWY_RESTRICT out_offsets, \
ThreadingContext& ctx, const size_t worker); \
\
size_t GetVTileSize(size_t kNF, size_t num_head_groups, size_t num_tokens, \
size_t total_tasks, size_t target_parallelism); \
\
void FlashAttention(size_t num_tokens, size_t target_parallelism, \
size_t layer_idx, const MatPtr& query_norm_scale, \
AttentionActivationsPtrs& activations, QBatch& qbatch, \
ThreadingContext& ctx); \
/* NOLINTNEXTLINE(google-readability-namespace-comments) */ \
#define GEMMA_DECL_FLASH_ATTENTION(TARGET, NAMESPACE) \
namespace NAMESPACE { \
void RMSNormAndPositionalEncoding( \
size_t num_tokens, const QBatch& qbatch, MatPtrT<float>& q, \
const MatPtr& query_norm_scale, size_t layer_idx, \
const AttentionActivationsPtrs& activations, ThreadingContext& ctx); \
\
void SingleFlashAttention(size_t start_pos, size_t last_pos, \
const BF16* HWY_RESTRICT q, \
const MatPtrT<KV_t>& k, const MatPtrT<KV_t>& v, \
size_t layer_idx, \
const AttentionActivationsPtrs& activations, \
float* HWY_RESTRICT att_out, \
ThreadingContext& ctx, size_t worker); \
\
Tile4FlashState TileFlashAttention4( \
const MatPtrT<BF16>& q, const uint32_t* HWY_RESTRICT q_offsets, \
const MatPtrT<KV_t>& k, size_t start_pos, \
const uint32_t* HWY_RESTRICT last_pos, size_t min_last_pos, \
size_t max_last_pos, const MatPtrT<KV_t>& v, size_t layer_idx, \
const LayerWeightsPtrs& layer, const AttentionActivations& activations, \
MatPtrT<float>& att_out, const uint32_t* HWY_RESTRICT out_offsets, \
ThreadingContext& ctx, const size_t worker); \
\
size_t GetVTileSize(size_t kNF, size_t num_head_groups, size_t num_tokens, \
size_t total_tasks, size_t target_parallelism); \
\
void FlashAttention(size_t num_tokens, size_t target_parallelism, \
size_t layer_idx, const MatPtr& query_norm_scale, \
AttentionActivationsPtrs& activations, QBatch& qbatch, \
ThreadingContext& ctx); \
\
/* NOLINTNEXTLINE(google-readability-namespace-comments) */ \
} // namespace NAMESPACE

// Function declarations for each SIMD target. Allows direct call from the
Expand Down
1 change: 1 addition & 0 deletions ops/ops-inl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,6 +25,7 @@
#include <cstdint>
#include <random>
#include <type_traits> // std::enable_if_t
#include <utility>
#include <vector>

#include "ops/matmul.h"
Expand Down
Loading