[Bug] "Exit code: null" on Strix Halo (gfx1151) despite successful ROCm 7.2 / Kernel 6.18 ID and model load #494
Description
LM Studio Version: 0.4.6 build 1
ROCm runtime (LM Studio): v2.5.1
ROCm Version: 7.2.0
Operating System: Linux Mint 22.1 (Ubuntu 24.04 Noble base)
Kernel: 6.18.14-061814-generic (Mainline)
GPU/APU: AMD Strix Halo (gfx1151) - 128GB LPDDR5x Unified Memory - AMD 395 + Radeon 8060S
What is the bug?
OS, ROCm, ROCm runtime and LM Studio identify the hardware properly, show no errors when operating. But just before the model loads, LM Studio throws a generic error.
Steps to Reproduce
Launch LM Studio
Try to load any model with any settings
Error -> cause: '(Exit code: null). Please check settings and try loading the model again.'
Logs:
2026-03-05.1 (copy).log
cli log.log
Extended information:
Tried to fix the issue with the typical steps for the rig and OS:
- Configure system with 128GB GTT memory via GRUB (ttm.pages_limit=33554432).
- Library Swap: Manually symlinked system ROCm 7.2 libhipblaslt.so and librocblas.so into the LM Studio linux-llama-rocm-vendor-v3 folder to ensure gfx1151 compatibility—error persists.
- Launch LM Studio with:
"export HSA_OVERRIDE_GFX_VERSION=11.5.1
export HSA_ENABLE_SDMA=0
export LD_LIBRARY_PATH=/opt/rocm/lib:$LD_LIBRARY_PATH
/opt/LM-Studio.AppImage"
Kernel & Driver: Verified amdgpu initialized correctly. dmesg confirms: 131072M of GTT memory ready.
Permissions: User is in render and video groups.
Isolated Benchmark: Running llama-bench (built against ROCm 7.2) works perfectly on this hardware, achieving expected T/s.
While llama.cpp (standalone) handles this architecture when forced via HSA_OVERRIDE, the LM Studio NodeProcessForker appears to receive a null exit code before the inference engine can fully initialize the memory map. Since the hardware is verified as "Ready" at the OS level, the failure point seems to be in the handshake between the Harmony parser and the ROCm runtime specifically for the GFX1151 (Strix Halo) target.