fix: unset CI in ray.sub and virtual_cluster to prevent vLLM aliasing errors

[kajalj22]

Summary

• PyTorch sets `error_on_custom_op_aliasing=True` when `CI=true` (via `bool(os.getenv("CI"))` in `torch/_functorch/config.py`), causing `RuntimeError` in vLLM custom ops such as `sequence_parallel_chunk_impl`
• `CI` propagates to Ray workers via the Ray daemon environment (started by sbatch before any test script runs), so per-recipe `vllm_cfg.env_vars` or `unset CI` in test scripts alone are insufficient
• Fix at two levels so all launchers are covered:
  • `ray.sub`: `unset CI` in both `head_cmd` and `worker_cmd` before `ray start`, ensuring Ray daemons and all workers they spawn never see `CI=true`
  • `virtual_cluster.py`: strip `CI` from the `runtime_env` env_vars passed to `ray.init` and emit `UserWarning` so non-ray.sub launchers (e.g. k8s, local) also get the fix with visibility

Test plan

• Nightly audiomcq tests pass without per-recipe `vllm_cfg.env_vars: {CI: ""}` workaround
• `UserWarning` appears in logs when `CI` is set at launch time

🤖 Generated with Claude Code

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.