immich-ml SIGSEGV with Intel Arc A310 (OpenVINO)
🚧Setup
Usage❓Question
Usage❓QuestionContext:
* Environment: Proxmox 9 Host -> Unprivileged LXC (Debian 13 Trixie).
* Installation: Proxmox Helper Scripts (https://helper-scripts.com/) (Bare-metal systemd services, not Docker).
* Hardware: AMD Ryzen 7 3700X, Intel Arc A310 GPU (Passthrough /dev/dri/renderD128).
* Storage: NFS Bind-mount from TrueNAS.
The Problem:
The immich-ml service enters a crash loop immediately upon loading the CLIP model (ViT-B-32openai). The worker is sent a SIGSEGV (Segmentation Fault) by
the kernel.
Error Log Snippet (ml.log / dmesg):
1 [immich-ml] INFO Loading visual model 'ViT-B-32openai' to memory
2 [immich-ml] INFO Setting execution providers to ['OpenVINOExecutionProvider', 'CPUExecutionProvider']
3 [immich-ml] WARNING Worker (pid:XXXX) was sent SIGSEGV!
Host dmesg shows the crash occurs in:
python[XXXX]: segfault at 0 ip 00007a1f1249346b sp 00007a1f36ffab00 error 6 in libopenvino.so.2541
Troubleshooting Steps Taken:
1. Permissions: Verified GID/UID mapping. User immich has full access to /dev/dri/*. Crash persists even when running as root.
2. Drivers: Updated intel-opencl-icd and libze-intel-gpu1 to 26.05.37020.3 (Latest Intel Compute Runtime).
3. Backend Workaround: Attempted forcing OV_GPU_BACKEND=OPENCL via .env, but the segmentation fault occurs before the provider choice is fully respected.
4. Hardware isolation: Removed GPU devices from LXC config. Interestingly, libopenvino.so still segfaults during library initialization even without
physical GPU access.
Current Workaround:
Manually patched /opt/immich/app/machine-learning/immich_ml/sessions/ort.py to filter out OpenVINOExecutionProvider from the default providers list. Smart
Search is now stable on CPU.
Checklist:
* [x] Verified on latest release (release version).
* [x] Tried local IP.
* [x] Checked hardware compatibility (Intel Arc/DG2).
* Environment: Proxmox 9 Host -> Unprivileged LXC (Debian 13 Trixie).
* Installation: Proxmox Helper Scripts (https://helper-scripts.com/) (Bare-metal systemd services, not Docker).
* Hardware: AMD Ryzen 7 3700X, Intel Arc A310 GPU (Passthrough /dev/dri/renderD128).
* Storage: NFS Bind-mount from TrueNAS.
The Problem:
The immich-ml service enters a crash loop immediately upon loading the CLIP model (ViT-B-32openai). The worker is sent a SIGSEGV (Segmentation Fault) by
the kernel.
Error Log Snippet (ml.log / dmesg):
1 [immich-ml] INFO Loading visual model 'ViT-B-32openai' to memory
2 [immich-ml] INFO Setting execution providers to ['OpenVINOExecutionProvider', 'CPUExecutionProvider']
3 [immich-ml] WARNING Worker (pid:XXXX) was sent SIGSEGV!
Host dmesg shows the crash occurs in:
python[XXXX]: segfault at 0 ip 00007a1f1249346b sp 00007a1f36ffab00 error 6 in libopenvino.so.2541
Troubleshooting Steps Taken:
1. Permissions: Verified GID/UID mapping. User immich has full access to /dev/dri/*. Crash persists even when running as root.
2. Drivers: Updated intel-opencl-icd and libze-intel-gpu1 to 26.05.37020.3 (Latest Intel Compute Runtime).
3. Backend Workaround: Attempted forcing OV_GPU_BACKEND=OPENCL via .env, but the segmentation fault occurs before the provider choice is fully respected.
4. Hardware isolation: Removed GPU devices from LXC config. Interestingly, libopenvino.so still segfaults during library initialization even without
physical GPU access.
Current Workaround:
Manually patched /opt/immich/app/machine-learning/immich_ml/sessions/ort.py to filter out OpenVINOExecutionProvider from the default providers list. Smart
Search is now stable on CPU.
Checklist:
* [x] Verified on latest release (release version).
* [x] Tried local IP.
* [x] Checked hardware compatibility (Intel Arc/DG2).
