Skip to content

Adjust work spliting to subtasks, about 1.5x faster prefill on zen5 machines, and +- neutral on older ones. Importantly to fully utilize this it's important to use high prefill token_batch like 4k #953

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

Merged
copybara-service merged 1 commit into from
Jul 15, 2026
+1 −1
Merged

Adjust work spliting to subtasks, about 1.5x faster prefill on zen5 machines, and +- neutral on older ones. Importantly to fully utilize this it's important to use high prefill token_batch like 4k #953

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: 1 addition & 1 deletion gemma/tiled_attention.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -605,7 +605,7 @@ void LocalAttentionForAllHeadsTokensAndBatch(
const size_t layer_idx, const LayerWeightsPtrs& layer,
AttentionActivationsPtrs& activations, QBatch& qbatch,
ThreadingContext& ctx) {
constexpr size_t kQueriesPerSubtask = 32;
constexpr size_t kQueriesPerSubtask = 128;
const size_t heads_per_kv_head =
layer.layer_config.heads / layer.layer_config.kv_heads;
const hwy::Divisor div_heads_per_kv_head(heads_per_kv_head);
Expand Down
Loading