]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
overview / pkg / ggml / sources / llama.cpp / commit
summary | shortlog | log | commit | commitdiff | tree
(parent: ad12647)
vulkan: Fix FA coopmat1 invalid array indexing (#16365)
authorJeff Bolz <redacted>
Fri, 3 Oct 2025 09:52:46 +0000 (04:52 -0500)
committerGitHub <redacted>
Fri, 3 Oct 2025 09:52:46 +0000 (11:52 +0200)
commit0e1f83855609d73beaf05d818640b6cfd39d287b
treefba1d34acc846c75f1bcb8ccd5ee17bd8db7f0f8tree
parentad126479c25cf983a0f994a08ba0911cf49ed62bcommit | diff
vulkan: Fix FA coopmat1 invalid array indexing (#16365)

When computing sinks, the cm1 shader was looping r from 0 to Br rather than
to rows_per_thread. I must have copied this from the scalar path (where it is
correct), and somehow it wasn't causing failures on current drivers.
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp diff | blob | history
Packaging of ggml-org/llama.cpp