radv vs amdvlk which one is koboldcpp using?

[1506Eden]

I recently saw a comparison that while radv is better for games when it comes to llms amdvlk is 40% faster in prompt processing compared to radv with similar tg.

I wanted to know which koboldcpp is using and if it might be worth it to do a manual install with custom vulkan sdk if it's using radv.

and test the difference.

[1506Eden]

amdvlk will soon no longer be supported by amd so i understand having it as an option is not worth it still wanted to know how to get it running for testing purposes

[wbruna]

Short answer is: it uses the system's default.

With both installed, AMDVLK is usually the default. You can check it on the log at model loading time; here I get:

Using Vulkan backend
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon RX 7600 XT (RADV NAVI33) (radv) | uma: 0 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat

so I'm using radv.

You can choose between them with environment variables, like AMD_VULKAN_ICD and VK_DRIVER_FILES. They are well documented elsewhere, eg. on the Arch wiki.

[1506Eden]

I see thanks