[Core] Add RL-Kernel mode controls

[inaniloquentee]

Summary

• add public RL-Kernel controls --rlk-fast {off,auto,strict} and --rlk-consistency {off,audit,strict}
• add env aliases VIME_RLK_FAST and VIME_RLK_CONSISTENCY
• preserve legacy compatibility flags: --enable-rl-kernel, --rl-kernel-strict, and --rl-kernel-ops
• centralize resolved values in args.rlk_mode_config

Fixes #7

Tests

• pre-commit run --all-files --show-diff-on-failure --color=always
• python -m pytest tests/test_megatron_argument_validation.py -m unit
• python -m pytest tests/utils/test_megatron_server_arguments.py tests/test_megatron_argument_validation.py -m unit
• git diff --check

CI note

• python -m pytest tests -m unit was attempted locally, but collection fails before running this change's tests due to existing local environment/import issues on Windows: dotted test module names such as test_qwen2.5... are imported as packages, and ray.util is unavailable in the local stubbed environment.

AI assistance

This PR includes AI-assisted code changes generated with OpenAI Codex and reviewed before submission.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: feb8d6a2-e61e-4720-ac0f-f97db4eefd8b

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feature/issue7-rlk-mode-parsing

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[Flink-ddd]

Overall, the code is extremely solid and fully meets the Phase 1 acceptance criteria. LGTM!

[Flink-ddd]

Two Sources of Truth: Writing resolved values to both the root args (e.g., args.rl_kernel_ops) and args.rlk_mode_config is fine for backward compatibility. However, consider adding an inline deprecation comment to the root attributes to strictly guide future development toward the immutable rlk_mode_config.

[inaniloquentee]

Two Sources of Truth: Writing resolved values to both the root args (e.g., args.rl_kernel_ops) and args.rlk_mode_config is fine for backward compatibility. However, consider adding an inline deprecation comment to the root attributes to strictly guide future development toward the immutable rlk_mode_config.

Thanks, addressed. I added an inline compatibility comment explaining that the root attributes are legacy aliases and new code should prefer the immutable args.rlk_mode_config.

[Flink-ddd]

Ambiguous Injection Point: Hooking add_rl_kernel_arguments inside add_ci_arguments is slightly misleading in its naming. Just ensure this specific entry point is actually triggered during standard production training, not just in CI environments.

[inaniloquentee]

Ambiguous Injection Point: Hooking add_rl_kernel_arguments inside add_ci_arguments is slightly misleading in its naming. Just ensure this specific entry point is actually triggered during standard production training, not just in CI environments.

Good point. I clarified the production parser entry point with an inline comment and added a regression test that builds the standard vime parser and parses the RL-Kernel flags.

[Flink-ddd]

LGTM, Thank.

[Flink-ddd]

cc @zhangj1an @maxiaosong1124 @a-kaa PTAL, if no further questions, we will merge this PR first. Thanks.