Since proxmox-kernel-7.0.14-15-pve (the 7.0.14+15 CVE cherry-pick release) a node with an AMD Ryzen AI 9 HX 370 / Radeon 890M (Strix Point, 1002:150e rev c1) hangs its GPU about three minutes after boot as soon as an inference workload runs. The workload is llama.cpp and Ollama over ROCm/KFD inside an unprivileged LXC with /dev/dri and /dev/accel bind-mounted. On 7.0.14-14-pve the same node and workload ran for weeks without a single amdgpu error.
Sequence on -15 (dmesg, from boot): 09:34:03 SMU: No response msg_reg: 12 resp_reg: 0, then Failed to retrieve enabled ppfeatures! and Failed to export SMU metrics table! every few seconds; 09:34:29 Failed to disable gfxoff!, device coredump written, ring sdma0 timeout, signaled seq=24, emitted seq=26, Ring sdma0 reset failed, GPU reset begin!. Source: 1; from 09:35:17 task kworker/3:1:210 blocked for more than 122 seconds in kfd_process_wq_release [amdgpu], repeating until a hard reset. The container becomes unreachable, pvestatd blocks so the node reads “unknown” in the GUI, and a graceful systemctl reboot stalls in systemd-shutdown because the blocked kworker cannot be killed.
Pinning 7.0.14-14-pve with proxmox-boot-tool restores full stability. The two REG_WAIT optc35_disable_crtc lines at boot appear on both kernels and look unrelated.
Is there a known amdgpu/SMU change between 7.0.14-14 and -15 that affects Strix Point gfxoff or power management? Happy to provide the devcoredump and the full journal. BIOS 1.06 (Minisforum AI Series), pve-manager 9.2.11, ZFS 2.4.4.
Sequence on -15 (dmesg, from boot): 09:34:03 SMU: No response msg_reg: 12 resp_reg: 0, then Failed to retrieve enabled ppfeatures! and Failed to export SMU metrics table! every few seconds; 09:34:29 Failed to disable gfxoff!, device coredump written, ring sdma0 timeout, signaled seq=24, emitted seq=26, Ring sdma0 reset failed, GPU reset begin!. Source: 1; from 09:35:17 task kworker/3:1:210 blocked for more than 122 seconds in kfd_process_wq_release [amdgpu], repeating until a hard reset. The container becomes unreachable, pvestatd blocks so the node reads “unknown” in the GUI, and a graceful systemctl reboot stalls in systemd-shutdown because the blocked kworker cannot be killed.
Pinning 7.0.14-14-pve with proxmox-boot-tool restores full stability. The two REG_WAIT optc35_disable_crtc lines at boot appear on both kernels and look unrelated.
Is there a known amdgpu/SMU change between 7.0.14-14 and -15 that affects Strix Point gfxoff or power management? Happy to provide the devcoredump and the full journal. BIOS 1.06 (Minisforum AI Series), pve-manager 9.2.11, ZFS 2.4.4.
