The runtime Vulkan llama.cpp (Windows) v1.51.0 use CPU incorrectly

[SquareFong]

Which version of LM Studio?
LM Studio 0.3.27

Which operating system?
Windows 11

What is the bug?
I'm using AMD AI Max + 395, 16G was assigned as VRAM, running qwen3-30b-a3b through the runtime Vulkan llama.cpp. The older version 1.50.2 was pretty fast, nearly 50 tokens/s, but the testing version 1.51.0 is slow, it's only about 20 tokens/s. I open the Task Manager, the version 1.51.0 would load the model to RAM and computed by CPU, when 1.50.2 load the model to vram, and computed by GPU

[BerndAmend]

1.52.0 on Linux has the same issue.

[IgnatBeresnev]

See #1048

Vulkan engine v1.52.0 (and I guess v1.51.0) doesn't use the GPU/VRAM to do the work, it relies on the CPU, thus x3 worse performance.

Roll it back to Vulkan v1.50.2 and it will work as before and will utilize the GPU/VRAM. Check out the issue above for a reproducer and instructions on how to check it

[zzglitch]

My setup experiencing the same issue:

• AMD RYZEN AI MAX+ 395 w/ 8060S Graphics
• Linux 6.16.10-arch1-1
• LMStudio 0.3.28
• Vulkan 1.52.0 reverts to CPU
• ROCm 1.52 crashes

Vulkan and ROCm 1.50.2 work as expected.

[symphonic-navigator]

Yes, can confirm, I have the same issue here. 1.50.0 works, though.

Seems like a regression. I use a 780M (iGPU) on the mini PC here and can confirm that the iGPU, while not the fastest, is capable of running LLM inference here, even with at 24B model at a Q4_K_M quant (FireDolphin).

[yagil]

This issue is older, but the fix is tracked in #1048. Please Ctrl + Shift + R -> update your runtime