llama.cpp
metal: speed up Qwen3-VL image encoding on large images by ~11%
#21443
Open
Open
metal: speed up Qwen3-VL image encoding on large images by ~11% #21443
Avidanborisov
wants to merge
3 commits into
ggml-org:master
from
Avidanborisov:metal-img-encode-optim
metal: make flash attention support 16 queries per threadgroup
c0788eb3
metal: use 16 queries per threadgroup and 8 simdgroups in flash atten…
23bd7621
Merge branch 'ggml-org:master' into metal-img-encode-optim
7ef7cb05
Reviewers
No reviews
Assignees
No one assigned
Labels
ggml
Apple Metal
Milestone
No milestone
Login to write a write a comment.
Login via GitHub