After an update, container recreate, or docker daemon restart, nomad_ollama's
HostConfig.DeviceRequests still lists the nvidia driver — but the NVIDIA
Container Toolkit binding inside the container is torn. `nvidia-smi` returns
"Failed to initialize NVML: Unknown Error" and Ollama silently falls back to
CPU inference. PR #208 detects this and shows a banner with a "Fix: Reinstall
AI Assistant" button. This change does that click automatically on admin boot.
New provider GpuPassthroughRemediationProvider runs once on web env boot:
1. Skip when KV `ai.autoFixGpuPassthrough = false` (default true).
2. Skip when Docker has no `nvidia` runtime registered (AMD-only and CPU-only
hosts unaffected).
3. Skip when nomad_ollama isn't running.
4. Exec `nvidia-smi --query-gpu=name --format=csv,noheader` inside the
container with an 8-second timeout. If the output matches
"Failed to initialize NVML", "Unknown Error", "TIMEOUT", or contains no
alphabetic characters, treat the passthrough as broken.
5. On broken: call DockerService.forceReinstall('nomad_ollama'). The existing
force-reinstall preserves the Ollama volume + installed models. Stamp
`gpu.autoRemediatedAt` on success.
6. On healthy: log and exit.
AMD passthrough_failed is intentionally not handled — its fix path is HSA
override handling (PR #804) rather than a simple service recreate, and false
positives during AMD startup log parsing would loop a recreate without fixing
anything. Left to a follow-up if it proves to be a recurring AMD issue.
Validated on NOMAD3 (RTX 5060, v1.32.0-rc.3 + this patch hot-applied):
- After admin restart with passthrough healthy: log line
"[GpuPassthroughRemediationProvider] NVIDIA passthrough healthy — no action
needed." Provider exits cleanly without touching the container.
- The broken-state branch hits the existing forceReinstall path, which was
manually invoked earlier in the same session to fix this exact box and
recovered GPU access in ~45s with model volume intact. No new failure mode
is introduced — the auto-trigger removes the user click but the underlying
operation is the same one the banner Fix button already calls.
Closes#755.
2997637ce0