]> git.djapps.eu Git - pkg/ggml/sources/whisper.cpp/commit
vulkan: coopmat2 mul_mat optimizations (llama/14934)
authorJeff Bolz <redacted>
Sat, 2 Aug 2025 09:21:37 +0000 (04:21 -0500)
committerGeorgi Gerganov <redacted>
Mon, 18 Aug 2025 17:30:45 +0000 (20:30 +0300)
commitb374fd6172bd5fcf94d4f07b0ad6a68b8799eb6f
treecfb58d984201ef280c80f99936f9095c8e9aa7aa
parent97341224b2c834e2419ee46301ced76bd22238b5
vulkan: coopmat2 mul_mat optimizations (llama/14934)

- Increase tile size for k-quants, to match non-k-quants
- Choose more carefully between large and medium tiles, considering how it
  interacts with split_k
- Allow larger/non-power of two split_k, and make the splits a multiple of 256
- Use split_k==3 to when >1/2 and <=2/3 of the SMs would hae been used
ggml/src/ggml-vulkan/ggml-vulkan.cpp