[zupengwang]

Overview

Add logprob_token_ids support to the Rust /v1/chat/completions and /v1/completions endpoints. Related to the selected-token item in #44280; scope claim.

Claims

• Forward non-empty token selections through the existing text/engine sampling path instead of silently dropping the field; preserve the Python frontend's empty-list fallback.
• Return the selected candidates and sampled token in streaming and non-streaming responses without applying the natural top-k cut again.
• Enforce the logprobs prerequisite and completion prompt-only echo restriction, reusing existing token-count and vocabulary validation.

Validation

Check	Baseline	With change
Conversion regressions	2 failures: selection is None	Passed
cargo test -p vllm-server -p vllm-text --locked (Rust 1.95)	—	543 passed, 1 upstream-ignored slow tokenizer test
cargo nextest run --workspace --all-features --locked --no-fail-fast (Rust 1.95, Python 3.12)	—	1923 passed, 13 upstream-ignored tests, 0 failures
Qwen2.5-0.5B-Instruct / RTX 4090, selected probabilities	Selected IDs omitted at all 15 observed generated positions	Python/Rust comparison across 45 positions: max absolute difference 4.87e-7
Ordinary requests without a selection	Original Rust binary	Same tokens, candidate maps and duplicate counts at all 15 positions

The Python/changed-Rust evaluation made 62 requests covering both endpoints, streaming/non-streaming, empty selections, sampled-token selections, invalid IDs/counts, missing logprobs, beam search, and prompt-only echo. Invalid requests returned 400, and each server remained healthy. One baseline run reproduced the omission and checked unchanged ordinary behavior. This is functional validation, not a throughput or accuracy benchmark.

The full Buildkite Rust style-clippy entrypoint passes, including fmt, Cargo ordering, dependency bans and workspace/all-targets/all-features Clippy with warnings denied. The frontend build and six applicable pre-commit hooks also pass.

The test bootstrap encountered external download failures. After installing nextest from locked source and preparing the CI Python environment, the identical full-workspace nextest command passed in one run. Initial missing tokenizer caches and first-run dependency timeouts were resolved without changing tests or adding skips. This reports local execution of the CI checks, not hosted CI status.

Reproduction commands:

cargo test --manifest-path rust/Cargo.toml -p vllm-server -p vllm-text --locked
bash .buildkite/scripts/run-rust-frontend-cargo-ci.sh style-clippy
cargo nextest run --manifest-path rust/Cargo.toml --workspace --all-features --locked --no-fail-fast

Example chat request used in validation:

{"model":"selected-logprobs-eval","messages":[{"role":"user","content":"The capital of France is"}],"temperature":0,"max_tokens":4,"logprobs":true,"top_logprobs":1,"logprob_token_ids":[0,1,2]}

Details

The shared sampling types already support selected-token logprobs, but the two OpenAI request types omit the field and conversion hard-codes it to None. The change clears natural top-k sampling for a non-empty selection and uses the existing all-candidates response option for chat. No sampler, engine protocol, GPU code or dependency changes are included.

Non-duplication: merged #56406 fixes gRPC conversion; #46427 adds /generative_scoring, #45569 addresses broader validation, and #57394/#49320 concern Python empty-list validation. None exposes selected IDs on these two Rust OpenAI routes. #59518 and #58711 also touch shared Rust chat candidate deduplication. This PR now dedupes chat top_logprobs by token ID before the top_logprobs cut, so a sampled token that is also selected (or the top-1 token) appears once, matching Python.

Evaluation used the exact Rust patch and Python source at cff08b461e08 with the same model snapshot, while reusing torch 2.13.0+cu132 and precompiled native extensions from the existing runtime. It was not a complete native rebuild. Rust 1.98 locally raises a newer lint in unchanged multimodal tensor code; pinned Rust 1.95 Clippy passes.

AI assistance was used for investigation, implementation and validation. The human submitter reviewed the patch and authorized this submission.

---

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).
• AI assistance was used during the creation of this PR.
• Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.
• Testing and Validation: Validation completed as reported above; human review confirmed before submission.
• Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.
• Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[BugenZhao]

[AUTOMATED] @codex review

[BugenZhao]

Thanks @zupengwang for filling this gap. I'll rebase the PR for you

[BugenZhao]

/ci run

[github-actions]

✅ Triggered Buildkite CI #94173 for commit 82985c9a8cb1.