[jthomson04]

Summary

• add regression coverage for producer-supplied prompt token IDs, generation token IDs, and generation log probabilities
• verify the vLLM model proxy preserves those fields through response validation
• verify the existing compatibility /tokenize fallback is not awaited when native metadata is present

Why

NeMo-RL can return the exact token metadata already held by its vLLM engine. Gym already supports that response contract and skips its compatibility tokenization request, but the native fast path was not covered directly.

This is intentionally test-only; the existing fallback remains unchanged.

Testing

• uv run pytest tests/test_app.py -k preserves_upstream_token_metadata_without_tokenizing -q (1 passed)
• ruff check tests/test_app.py
• ruff format --check tests/test_app.py

Producer implementation: NVIDIA-NeMo/RL#3390

[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.