Use ARC Kubernetes mode to run benchmark
ARC Kubernetes mode has the highest security level among all runner modes. We can keep using CPU-only instance for docker building, but all GPU instances should use ARC Kubernetes mode.
Reference: https://www.stepsecurity.io/blog/how-to-use-docker-in-actions-runner-controller-runners-securelly
On PyTorch, we decided to stop ARC migration. Is this still relevant for TorchBench?
Yes, it is still relevant for TorchBench, TB uses ARC runners on GCP K8s cluster for PR and nightly testing: https://github.com/pytorch/benchmark/tree/main/docker/infra
If we migrate all OSS workflows to Internal, or if pytorch CI provides capacity for GCP A100 runners, we can also deprecate the ARC runner setup.