[AndreasKaratzas]

Summary

• Make native_ci an explicit runtime switch: true selects the native AMD Kubernetes pod, while false or omission preserves legacy DinD. ci-infra does not infer runtime mode from the device family or queue name.
• Complete the native pod contract with the per-commit ci_base image, explicit GPU/no-GPU quantities, 16 GiB /dev/shm, isolated workspace, IPC_LOCK, restricted security context, image pull policy, artifact/checksum parameters, job-local runtime settings, and Docker-hook suppression.
• Reject non-boolean native_ci, native multi-node jobs, and native no_plugin jobs during pipeline generation.
• Preserve the legacy runner for every AMD queue unless its configuration explicitly opts into native execution.
• Preserve upstream AMD mirror soft-fail behavior, per-step timeouts, and the latest upstream main.

This is paired with vllm-project/vllm#48387. The current vLLM configuration explicitly opts MI300 jobs into native execution while MI250, MI325, and MI355 remain on DinD. Future AMD queues can choose either mode without an infra code change.

[khluu]

This would add to many things into buildkite_step.py. Can you make sure to utilize all other files like constants.py or k8s_plugin.py?

[AndreasKaratzas]

CI runs with vllm-project/vllm#48387:

• https://buildkite.com/vllm/ci/builds/77920/canvas
• https://buildkite.com/vllm/amd-ci/builds/10821/canvas