feat(vllm): support NeMo-RL NCCL weight updates

[jthomson04]

Overview:

Details:

Where should the reviewer start?

Related Issues

⚠️ This section is required. Choose one path below and delete the other.

🔗 This PR is linked to an issue:

• Closes #XXXX

🚫 This PR is NOT linked to an issue:

• Confirmed — no related issue

Summary by CodeRabbit

• New Features

  • Weight updates can now be applied without pausing the worker when explicitly allowed.
  • Cache reset behavior is now configurable, letting updates optionally preserve existing prefix/KV cache state.

• Bug Fixes

  • Added stricter validation for new update options to prevent invalid requests.
  • Improved update handling so only the intended control fields are passed through during distributed updates.

• Tests

  • Added coverage for unpaused weight updates and default-safe error handling.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[coderabbitai]

Walkthrough

update_weights_from_distributed gains two boolean flags: allow_unpaused (default False) and reset_prefix_cache (default True). The pause guard is relaxed when allow_unpaused is True, both flags are stripped from RPC kwargs, and cache invalidation is now conditional. Two new unit tests cover async RL semantics and safe-default behavior.

Changes

Weight Update Flag Controls

Layer / File(s)	Summary
Handler: new flags, pause guard, and conditional cache reset components/src/dynamo/vllm/handlers.py	Parses and boolean-validates allow_unpaused and reset_prefix_cache; allows weight updates without pause when allow_unpaused is True; excludes both fields from rpc_kwargs; calls reset_prefix_cache() only when reset_prefix_cache is True.
Unit tests for new flag semantics components/src/dynamo/vllm/tests/test_vllm_worker_handler.py	Adds test_distributed_update_can_match_async_rl_semantics (verifies allow_unpaused path skips cache reset) and test_distributed_update_preserves_safe_defaults (verifies empty payload error and paused-path cache reset).

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

🚥 Pre-merge checks | ✅ 3 | ❌ 2

❌ Failed checks (2 warnings)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The PR description is only the template skeleton and omits the required content in Overview, Details, reviewer start, and issue linkage.	Fill in the required sections with a real summary, implementation details, reviewer guidance, and a valid linked issue or confirmed no-issue checkbox.
Docstring Coverage	⚠️ Warning	Docstring coverage is 20.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (3 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly summarizes the main change: NeMo-RL NCCL weight updates in vllm.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

---

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

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@components/src/dynamo/vllm/handlers.py`:
- Line 1644: The pause check in BaseWorkerHandler is using defensive getattr for
a known attribute, which can hide initialization regressions. Update the
condition in the handler logic to reference self._paused directly, and only
introduce a class-level default if __new__-based test doubles truly require it.
Keep the change scoped to the BaseWorkerHandler pause/allow_unpaused path and
follow the known-attribute guidance from the Python guidelines.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 99dfd0d4-12de-4c0a-8032-3936570e821f

📥 Commits

Reviewing files that changed from the base of the PR and between d672016 and f6d44eb.

📒 Files selected for processing (2)

• components/src/dynamo/vllm/handlers.py
• components/src/dynamo/vllm/tests/test_vllm_worker_handler.py

[jthomson04]

/ok to test 071684f

[jthomson04]

/ok to test 8db1596

[biswapanda]

lgtm

[jthomson04]

/ok to test 5ebd0d2