fix nvidia runtime edge case

[luke-lombardi]

Summary by cubic

Preserve NVIDIA_VISIBLE_DEVICES during CDI generation to prevent void from hiding GPUs from the worker. Adds GPU discovery edge-case tests and improves allocation error messages.

• Bug Fixes

  • Capture and restore NVIDIA_VISIBLE_DEVICES around nvidia-ctk cdi generate so the worker’s env isn’t overwritten with void.
  • Enriched “not enough GPUs” error with requested, allocable, visible, configured, already_allocated, and current NVIDIA_VISIBLE_DEVICES.
  • Added tests for AvailableGPUDevices: mismatched visibility, missing /proc entries, procfs errors, and query failures.

• Dependencies

  • Bumped github.com/VictoriaMetrics/metrics and github.com/beam-cloud/blobcache-v2.
  • Moved github.com/opencontainers/runc and github.com/containerd/console to indirect; removed sdk/uv.lock.

^{Written for commit 5b67b27. Summary will update on new commits.}

[cubic-dev-ai]

1 issue found across 5 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="pkg/worker/nvidia.go">

<violation number="1" location="pkg/worker/nvidia.go:47">
P2: Preserve whether `NVIDIA_VISIBLE_DEVICES` was originally unset; using `Getenv` + unconditional `Setenv` can incorrectly create an empty env var.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}