Problem

When CDI is enabled, the snapshot agent pod needs GPU access to run nvidia-smi. The existing --require-gpu flag solves this by requesting nvidia.com/gpu: 1, but this fails when all GPUs are already allocated to workloads.

Solution

Add a --runtime-class flag to aicr snapshot that:

1. Sets runtimeClassName on the agent Job's pod spec (e.g., nvidia)
2. Injects NVIDIA_VISIBLE_DEVICES=all environment variable into the container

This gives the agent access to nvidia-smi via the NVIDIA container runtime without consuming a GPU from the Device Plugin. The snapshot GPU collector only needs to run nvidia-smi -q -x — it does not need a dedicated GPU allocation.

Flags behavior

• --runtime-class and --require-gpu are mutually exclusive
• --runtime-class is the preferred approach; the error message when both are set recommends it
• Supports AICR_RUNTIME_CLASS environment variable

Acceptance criteria

• aicr snapshot --runtime-class nvidia sets runtimeClassName on the agent pod
• NVIDIA_VISIBLE_DEVICES=all is injected when --runtime-class is set
• --require-gpu and --runtime-class together produce a clear error recommending --runtime-class
• Table-driven tests cover the new flag in job_test.go
• make qualify passes