This PR fixes #49379 by making the Poolside reasoning parser follow Laguna’s default enable_thinking=true behavior during reasoning token ID initialization. It includes a focused ReasoningConfig regression test and validation with the real Laguna-S-2.1 tokenizer. DCO and all applicable local hooks pass.
I also checked for duplicates: #45874 only adds general parser tests and does not address the configuration initialization bug.
This is my first vLLM contribution, so the pre-run check requires a maintainer-applied verification/ready label. Thanks!
To trigger CI for [Bugfix] Fix Poolside reasoning token auto-initialization by nothiny · Pull Request #49434 · vllm-project/vllm · GitHub, a maintainer must add the ready or verified label, as first-time contributors cannot bypass the pre-run check themselves. Your PR addresses #49379 by ensuring the Poolside reasoning parser correctly inherits Laguna’s enable_thinking=true default for reasoning token ID initialization, and includes a targeted regression test. DCO and local checks are passing, and #45874 does not duplicate this fix—it only adds general tests, not the configuration bug fix.
Maintainers are encouraged to review and apply the ready or verified label to unblock CI, as described in the vLLM contribution guidelines and confirmed in PR review instructions for first-time contributors. Would you like more detail on the review or CI process?