[Regression] Vulkan v1.52.0 does not use GPU/VRAM on gfx1151 (Ryzen AI MAX+ 395), x3 worse performance

[IgnatBeresnev]

Which version of LM Studio?
LM Studio 0.3.27

Which operating system?
I use arch btw

What is the bug?
Vulkan v1.50.2 uses VRAM to load models as expected. When loading gpt-oss 120b, I can clearly see that 60G of VRAM gets allocated, plus I can see that the GPU load is ~50% when the text is being generated (use btop).

Vulkan v1.52.0 doesn't use VRAM, it uses RAM + swap to load the model, and then it uses the CPU for compute. VRAM usage is at 1/96GB, GPU usage is at 1-2%.

This can be reproduced consistently just by switching the engine to 1.50.2 and back, you can check VRAM usage with btop, radeontop or amdgpu_top --smi.

I guess this bug is also present in v1.51.0 as per #1041

To Reproduce
Steps to reproduce the behavior:

1. Set the engine to Vulkan v1.50.2, load any model, make sure VRAM is used (commands above), generate some text and remember TPS.
2. Set the engine to Vulkan v1.52.0, load the same model, re-generate the response. Observe x3 worse performance, see that VRAM is not used and that the CPU load is higher than before.

TLDR fix

Set the engine to Vulkan v1.50.2

[yasutimpo]

私も同様のバグ？が発生しています。
LM Studioのどのバージョンですか?
LMスタジオ 0.3.28(Build2)

Vulcanはあなたと同じものです。

[hongcheng79]

I also experienced the same issue LLM 0.3.28(Build2) Linux with vulkan runtime v1.52.0.
I need to revert the runtime to v1.50.2 in order to use the GPU.

Both the chat and developer had this issue.

[MathieuVedana]

Same issue on my a395 128GB VRAM laptop, LM Studio 0.3.28 Build2 Vulkan 1.52.0

[glantucan]

Same here.
I was lucky and I had 1.50.2 which works (it's the one that came with the AppImage or dowloaded right after installation), but I couldn't find a way to download a previous vulkan runtime versions inside the app. If you downloaded the app after a bugged extension is released there is no way to select a working previous one :(

[DreamGallery]

Same issue, I have to switch back to v1.50.2

[brownt12684]

Can also confirm. Reverted back to 1.50.2 on AMD RYZEN AI MAX+ 395 w/ Radeon 8060S

[markinmaine]

Same issue. Affects both AI MAX+ 395 and AI HX 370. Reverting to v1.50,2 resolves the issue.

[yasutimpo]

My pc does not have Ryzen AI,but It has RYzen 7 pro 7840u and iGPU which has 4GB vrams and 32GB rams. I also resolved the issue by reverting to v1.50.2.

[nvdg]

Also was puzzled why performance dropped, also reverted back to v1.50.2!
I also noticed my model used only 10GB of memory instead of the usual 24GB

[jesusangelm]

Same here. using Vulkan v1.50.2 solved the issue.

[waltercool]

Same here. using Vulkan v1.50.2 solved the issue.

How do you download v1.50.2 at this point?

[jesusangelm]

Same here. using Vulkan v1.50.2 solved the issue.

How do you download v1.50.2 at this point?

I didn't need to download anything.
I just selected the necessary engine in:
Discover > Runtime > Engines > GGUF:

If you don't see the desired engine, click on “include older versions” in the upper right corner.

[SarcasticBaaka]

Same here. using Vulkan v1.50.2 solved the issue.

How do you download v1.50.2 at this point?

I didn't need to download anything. I just selected the necessary engine in: Discover > Runtime > Engines > GGUF:

If you don't see the desired engine, click on “include older versions” in the upper right corner.

I think that requires having downloaded v1.50.2 previously tho. I myself just installed the latest version of LMStudio for the first time and it came with v1.52.0 with no possible way to rollback to a previous version. Is it possible for you to share the files for that version so it could be manually loaded?