llama.cpp
ggml-webgpu: improve flastAttention performance by software pipelining
#19151
Merged
Merged
ggml-webgpu: improve flastAttention performance by software pipelining #19151
reeselevine
merged
15 commits into
ggml-org:master
from
ArberSephirotheca:zheyuan-fa
webgpu : pipeline flash_attn Q/K loads in WGSL
84ceaca6
ggml-webgpu: unroll Q*K accumlation inner loop
01db8b60
ggml-webgpu: vectorization
57887189
ggml-webgpu: unrolling
e302f1f1
ggml-webgpu: remove redundant unrolling
cfdce2a0
ggml-webgpu: restore the config
eaa26c63
ggml-webgpu: remove redundant comments
8f2daee1
ggml-webgpu: formatting
2bd304ff
ggml-webgpu: formatting and remove vectorization
17eee16d
ggml-webgpu: remove unnecessary constants
faa9a76c
reeselevine
commented
on 2026-01-28
ggml-webgpu: change QKV buffer to read_write to pass validation
178f85c4
ggml-webgpu: add explanation for the additional bracket around Q K ac…
243a299d
Indentation and for -> if for tail
29f0b88b
reeselevine
approved these changes
on 2026-01-29
Merge remote-tracking branch 'upstream/master' into zheyuan-fa
e6f15ca4
Kick off CI on wgsl only commits
a150993a
Login to write a write a comment.
Login via GitHub