[Refactor] Break down single-controller training support into focused PRs

[garrett4wade]

Checklist

• This refactor maintains backward compatibility with all user-facing APIs.
• For large-scale refactors, I've prepared a phased implementation plan.

Current Limitations

PR #528 is too large to review and merge as a single unit.

Proposed Refactor Plan

The following changes will be implemented across separate, focused pull requests:

• Refactor InferenceEngine API to accept importable strings as workflow arguments: refactor: allow passing string paths and init kwargs as rollout workflows #525
• Refactor InferenceEngine API to support server process launching Refactor InferenceEngine API to support server process launching #530
• Modify the wait method of InferenceEngine to allow silent TimeoutError handling [Refactor] Modify the wait method of InferenceEngine to allow silent TimeoutError handling #531
• Consolidate sglang/vllm tests: refactor: merge unit tests of sglang and vllm engines  #514
• Refactor training statistics collection to use inline stats_tracker extensively and disregard returned statistics [Refactor] Refactor training statistics collection to use inline stats_tracker extensively and disregard returned statistics #532
• Implement Flask-based RPC server with serialization and corresponding tests [Refactor] Implement Flask-based RPC server with serialization and corresponding tests #533
• Implement LocalScheduler with corresponding tests
• Implement TrainController with an SFT example
• Implement RolloutController with an eval-only example The implementation of RolloutController #469
• Add an end-to-end single-controller RL training script

[daihaowz]

glad to help. i will implement LocalScheduler and TrainController.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity within the last 14 days.

If this issue is still relevant, please add a comment to keep it open. Otherwise, it will be automatically closed in 16 days.

Thank you for your contribution!

[garrett4wade]

FYI the single-controller mode has been broken down into #610 #612 #614

Pending review and merge.