Jeff Bolz cea22b3075 vulkan: For coopmat2 FA, use fp16 accumulators for the final result (llama/19376) ...

The cpu and cuda backends use fp16 for the VKQ accumulator type, this change
does the same for vulkan. This helps particularly with large head sizes which
are very register-limited.

I tried this for the coopmat1 path and it slowed down a bit. I didn't try for
scalar.

I applied the softmax bias that the cuda backend uses to avoid overflow,
although I was not able to reproduce the original bug without it.

2026-02-08 09:29:10 +02:00

..

cmake

cmake : remove unused file (ggml/1419)

2026-02-08 09:29:10 +02:00

include

ggml-virtgpu: make the code thread safe (llama/19204)

2026-02-08 09:29:10 +02:00

src

vulkan: For coopmat2 FA, use fp16 accumulators for the final result (llama/19376)

2026-02-08 09:29:10 +02:00

.gitignore

…

CMakeLists.txt

Bump cmake max version (needed for Windows on Snapdragon builds) (llama/19188)

2026-02-08 09:29:10 +02:00