RX9070 (GFX1201) cannot run ROCm llama.cpp (Linux) properly in Arch Linux

[XiaoHaoJun]

Which version of LM Studio?
LM Studio 0.3.35

Which operating system?
Arch Linux (Kernel: Linux 6.17.9-arch1-1)
GPU: AMD Radeon RX 9070 (RADV GFX1201)

What is the bug?
ROCm llama.cpp (Linux) appears compatible in the download list, but after downloading it, an “Error surveying hardware” message appears, making it unusable. The Vulkan version functions normally.

Screenshots

Logs
No logs.

[kirito201711]

me too

[xiaoxihooo-source]

我是r9700,也是这样，没有gfx1201,LM Studio 的 ROCm 后端是预编译 + 架构白名单，需要等它把1201这个做出来。

[JSkier21]

I cannot reproduce this on ROCm runtime 2.27.1 — models load and infer correctly on gfx1201 at every size I tried, up to 32B. Looks like the bundled-runtime update you asked for has landed.

Environment

LM Studio	0.4.20 (Linux AppImage, AUR lmstudio-bin)
ROCm runtime ext	llama.cpp-linux-x86_64-amd-rocm-avx2@2.27.1
System ROCm	7.2.4
GPU	Radeon RX 9070 XT — Navi 48 [1002:7550], gfx1201, 15.92 GB VRAM
OS / kernel	CachyOS (Arch), 7.1.5-arch1-1
CPU	Ryzen 9 7950X
Mesa / RADV	26.1.5 (for the Vulkan comparison below)

rocminfo reports amdgcn-amd-amdhsa--gfx1201. No HSA_OVERRIDE_GFX_VERSION set—not needed.

Result

$ lms runtime select llama.cpp-linux-x86_64-amd-rocm-avx2@2.27.1
Selected llama.cpp-linux-x86_64-amd-rocm-avx2@2.27.1 for GGUF

$ lms load qwen3-32b
Model loaded successfully in 9.34s.

No SEGV, no crash. Loaded cleanly on ROCm at 0.6B, 14B, 32B, and two 30B MoE models. Inference over /v1/chat/completions returns coherent output, so it isn't merely surviving the load—the compute path is correct too.

Since your report used gpt-oss-20B specifically, I could not test that exact model, but for an unrelated reason—my copy is ollama's GGUF, which upstream llama.cpp rejects with unknown model architecture: 'gptoss'. That's a model-packaging mismatch, not a ROCm problem. qwen3-32b is larger than gpt-oss-20B and loads fine, so I don't think model size is the missing variable.

Incidental: Vulkan is still faster here

Same machine, same models, same runtime version, back-to-back. 3 runs after a warmup, 300 tokens, temp 0:

Model	Vulkan	ROCm	Vulkan advantage
qwen3-0.6b	302 tok/s	143 tok/s	2.11×
qwen3-14b	61.0 tok/s	50.4 tok/s	1.21×
qwen3-32b (spills to CPU)	6.5 tok/s	5.4 tok/s	1.20×

The 2.11× at 0.6B is overhead-dominated and not representative; at realistic sizes, it's a steady ~20%. Noting it only because ROCm being functional and ROCm being the fast path on RDNA4 are separate questions—and on a clean install here, the runtime auto-selected was Vulkan, with ROCm not installed by default.