fix(agent): detect NVIDIA GPUs on distros without the unversioned NVML soname

On Debian and derivatives the NVIDIA driver ships only libnvidia-ml.so.1
(the unversioned symlink belongs to the dev package), and nvml-wrapper's
default init dlopens the unversioned name — so gfxinfo reported 'No GPU
found' on a fully functional RTX A2000 host while nvidia-smi worked
fine. Arch-family distros ship the symlink, which is why the desktop
never showed this.

The GPU worker now falls back to initializing NVML directly with the
versioned soname when gfxinfo's probe fails, collecting name/util/vram
through the same handle-caching path. nvml-wrapper was already in the
tree via gfxinfo — same version, no new build cost.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

This commit is contained in:

jasonwitty

2026-08-21 11:16:41 -07:00

parent 8e0effe361

commit 2a7951cf65

3 changed files with 61 additions and 18 deletions

Cargo.lock

Generated

View File

@@ -2441,6 +2441,7 @@ dependencies = [
  "futures-util",
  "gfxinfo",
  "hostname",
  "nvml-wrapper",
  "once_cell",
  "prost",
  "prost-build",