feat(vllm): add Python RL serving parity for generate endpoint

[biswapanda]

Summary

This brings the Python dynamo.vllm worker to parity with Dynamo's RL path for aggregated vLLM deployments. It adapts Rust-frontend token-in/token-out Generate envelopes, preserves trainer replay metadata, publishes the topology and capabilities needed by RL orchestration, and serializes standard vLLM LoRA updates with active generation.

• adapt native Generate/TITO requests, including sampling controls, cache salts, and preprocessed image features
• fold engine streams into unary Generate responses while preserving routed-expert tensors and sampling masks
• expose RL administration routes, NCCL world size, native Generate support, and tower-LoRA capability metadata
• reject unsupported data-parallel and external-launcher RL topologies instead of publishing an incorrect receiver count
• keep versioned standard vLLM LoRA adapters loaded for the full request lifetime, draining active generation before hot-swap or unload
• add focused Python and Rust coverage for request adaptation, discovery metadata, replay metadata, topology validation, and LoRA lifecycle behavior

area breakdown by code owner

Changed area in PR	CODEOWNER team
components/src/dynamo/common/rl/admin.py	@ai-dynamo/dynamo-rl-codeowners
Other components/src/dynamo/common/... tests/support	@ai-dynamo/dynamo-runtime-codeowners
components/src/dynamo/vllm/... and its tests	@ai-dynamo/dynamo-backend-vllm-codeowners
lib/llm/src/protocols/openai/generate.rs	@ai-dynamo/dynamo-frontend-codeowners

Validation

• repository-configured pre-commit hooks passed after the Omni split
• cargo fmt --all -- --check
• cargo test -p dynamo-llm --no-default-features protocols::openai::generate::tests -- --nocapture — 32 passed
• test_vllm_legacy_lora.py — 17 passed with vLLM 0.29 after the split
• test_vllm_worker_handler.py and test_vllm_engine_generate.py — 119 passed, 5 dependency-gated skips after the split
• Qwen3-0.6B MathEnv dense NCCL run — 3/3 steps passed; mismatch KL 0.0004, 0.0004, 0.0003
• Qwen3-0.6B MathEnv filesystem-LoRA run after the lifecycle fix — 3/3 steps passed; mismatch KL 0.0003, 0.0004, 0.0006; immutable adapters v0-v3 loaded and drained successfully

Summary by CodeRabbit

• New Features

  • Added support for structured routed-expert metadata, including tensor payloads.
  • Generation responses can now include sampling-mask data.
  • Added native support for adapted vLLM generation requests, including multimodal prompts and routing metadata.
  • Route descriptions can report configured world size.

• Bug Fixes

  • Improved validation for malformed generation metadata, token limits, routing hashes, and sampling masks.
  • LoRA loading and unloading now safely wait for active requests to finish.
  • Added validation to prevent unsupported reinforcement-learning deployment topologies.

Post-review validation

• Exact vLLM 0.29 focused request-adaptation and LoRA suites: 28 passed.
• Worker-handler suite: 108 passed, 5 dependency-gated skips.
• Repository pre-commit hooks, Python compilation, and diff checks passed.

Where should the reviewer start?

• components/src/dynamo/vllm/engine_generate.py for native Generate request adaptation.
• components/src/dynamo/vllm/handlers.py and components/src/dynamo/vllm/lora_state.py for response metadata and LoRA request draining.
• components/src/dynamo/common/rl/admin.py for RL topology publication.

Related Issues

• None.

[github-actions]

👋 Hi biswapanda! Thank you for contributing to ai-dynamo/dynamo.

Just a reminder: The NVIDIA Test Github Validation CI runs an essential subset of the testing framework to quickly catch errors.Your PR reviewers may elect to test the changes comprehensively before approving your changes.

🚀

[dynamo-ops]

/ok to test 6190cd3

[dynamo-ops]

/ok to test e7b84bb

[dynamo-ops]

/ok to test 23061d1

[dynamo-ops]

/ok to test 613fd66

[dynamo-ops]

/ok to test 404be27

[dynamo-review-agent]

Previously reported defects still present:

• Original discussion: Still present: adapt_engine_generate_request rejects every sampling_params.stop request even though vLLM supports stop strings when detokenization is enabled, and test_tito_adapter_rejects_stop_strings_but_preserves_stop_token_ids now codifies that unsupported rejection instead of preserving stop-string handling.

[dynamo-ops]

/ok to test eee5e44

[dynamo-ops]

/ok to test 0ecc46b

[dynamo-ops]

/ok to test f168ae1

[dynamo-ops]

/ok to test d9a2532