]> git.djapps.eu Git - pkg/ggml/sources/whisper.cpp/commit
vulkan: Support asymmetric FA in coopmat2 path (llama/21753)
authorJeff Bolz <redacted>
Fri, 1 May 2026 13:28:32 +0000 (15:28 +0200)
committerGeorgi Gerganov <redacted>
Sat, 2 May 2026 12:02:42 +0000 (15:02 +0300)
commit95053f68e4c2b638b3b33c200cfd9f4dd96976b7
tree98a86e989ea8208e74adf58691e618171201200c
parent35cb6841299888d20ad320966ce2176c403ada7d
vulkan: Support asymmetric FA in coopmat2 path (llama/21753)

* vulkan: Support asymmetric FA in coopmat2 path

There has been some recent interest/experimentation with mixed quantization
types for FA. I had originally designed the cm2 FA shader with this in mind
(because I didn't realize it wasn't supported at the time!), this change
adds the missing pieces and enables it.

Also support Q1_0 since people have been trying that out (seems crazy, but
who knows).

We should be able to do similar things in the coopmat1/scalar path, but
there's another change open against the scalar path and I don't want to
conflict.

* reorder cases
ggml/src/ggml-vulkan/ggml-vulkan.cpp
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp