llama.cpp
hexagon: FLASH_ATTN_EXT improvements for longer context and SWA models
#30731

Merged

hexagon: FLASH_ATTN_EXT improvements for longer context and SWA models #30731

max-krasnyansky merged 3 commits into ggml-org:master from qualcomm:hexagon-fa-tunning

hex-fa: support for masked block skipping to reduce DMA fetching

d4578c7b

hex-fa: finetuning the hmx->hvx crossover thresholds (boosting gemma-4)

92375241

hex-fa: update hmx gate to account for mdev.count

4fb18c99

max-krasnyansky requested a review 13 days ago

github-actions added ggml

github-actions added Hexagon

max-krasnyansky added merge ready

lhez approved these changes on 2026-10-17

jhen0409 approved these changes on 2026-10-19

max-krasnyansky merged 31288c9c into master 12 days ago

Reviewers

jhen0409

lhez

Assignees

No one assigned

Labels

ggml merge ready Hexagon

Milestone

No milestone