In plain English
The amdgpu driver declares the GPU hung and issues a MODE1 (SMU) reset. VRAM is lost across the reset. Cited reports: an amd-gfx thread (July 2026) on gfx12 HSA selftests after MES failed to suspend compute gangs; ROCm #6127 on an R9700 where `gst-plugin-scan` / VA-API probing ended in MODE1; and a gfx1201 llama.cpp HIP hang (legacy-rocm-build #6630) where Vulkan on the same card passed. #6127 was later closed for inactivity, not marked fixed.
Impact on consumers
A hang can take the desktop with it. Reinstalling userspace will not help if the kernel/MES path is wedged. Fine-tune GPU-resets are a different class — do not apply training workarounds to inference blindly.
Workaround
Capture `dmesg | grep -iE 'amdgpu|MODE1|MES'` before you reboot if you can. If the hang started after a kernel upgrade, bisect kernel vs ROCm userspace. For #6127-style login resets, try disabling `libgstvaapi.so`. For inference, switch to llama.cpp Vulkan and see if the hang follows HIP. AMD asked one reporter for nightlies after later gfx1201 page-fault fixes — that is a test request, not a measured fix on our rig.
Upstream
- amd-gfx — GPU reset / MODE1 on gfx12 HSA tests (July 2026) (amd-gfx)
- ROCm (ROCm/ROCm)
- legacy-rocm-build (ROCm/legacy-rocm-build)