structured-outputs-auto-whitespace

Source: fixes/structured-outputs-auto-whitespace/

Allow disable_any_whitespace with the default auto structured-output backend, and make the outlines backend honour it.

Affects	servers that want compact JSON from structured outputs (disable_any_whitespace)
Releases	v0.26.0, v0.30.0
Patches	vllm/config/structured_outputs.py, vllm/v1/structured_output/backend_outlines.py
Upstream	vllm#63430 (open)
Test	none

Problem

--structured-outputs-config '{"disable_any_whitespace": true}' is rejected at startup with disable_any_whitespace is only supported for xgrammar and guidance backends unless the backend is named explicitly, because the validator checks the backend name and the default is auto. Naming xgrammar avoids the error but loses the fallback that auto provides: schemas xgrammar cannot compile (for example multipleOf, uniqueItems, patternProperties) then fail with HTTP 400 instead of falling back to guidance.

Symptom

The server refuses to start with the flag, or, with backend: xgrammar, some JSON schemas that work under auto start returning 400. Without the flag, JSON grammars allow arbitrary whitespace, and some models (reasoning models after a forced tool call in particular) then emit whitespace until max_tokens.

Who is affected

Anyone who wants compact JSON output while keeping the default backend selection.

The fix

xgrammar and guidance already read the flag from the engine config. The fix makes outlines pass an empty whitespace pattern to outlines_core when the flag is set, so every backend that auto can choose honours it, and relaxes the validator to reject only lm-format-enforcer. This matches the upstream pull request.

Verify

python3 fixes/structured-outputs-auto-whitespace/apply.py --check   # 0 = applied, 1 = not applied, 2 = source drifted

Start the server with --structured-outputs-config '{"disable_any_whitespace": true}' and no backend: it must start, and a response_format JSON schema request returns compact JSON.