Baike.dev
Log in
Back to tool/Back to issues
#1998·outlines

SGLangTypeAdapter silently drops whitespace_pattern from JsonSchema (vLLM backend handles it correctly)

Author: harsh4vardhanCreated Aug 13, 2026Updated Aug 13, 2026

Bug Description

SGLangTypeAdapter.format_output_type silently drops the whitespace_pattern field from a JsonSchema instance. When a user sets whitespace_pattern on their schema — to control whitespace in generated JSON — the SGLang backend ignores it entirely and passes no whitespace constraint to the model. The vLLM backend handles this correctly.

Affected file

src/outlines/models/sglang.py, lines 74–80:

python
elif isinstance(term, JsonSchema):
    return OpenAITypeAdapter().format_json_output_type(
        json.loads(term.schema)   # whitespace_pattern is never read
    )

Contrast with vLLM (correct behaviour)

src/outlines/models/vllm.py, lines 66–70:

python
structured_outputs = {"json": json.loads(term.schema)}
if term.whitespace_pattern:
    structured_outputs["whitespace_pattern"] = term.whitespace_pattern
return structured_outputs

Minimal reproducer

python
import json
from outlines.types.dsl import JsonSchema

js = JsonSchema('{"type": "string"}', whitespace_pattern=r"[ ]?")

# SGLang path (buggy):
sglang_result = {"json": json.loads(js.schema)}          # whitespace_pattern missing

# vLLM path (correct):
vllm_result = {"json": json.loads(js.schema)}
if js.whitespace_pattern:
    vllm_result["whitespace_pattern"] = js.whitespace_pattern

assert "whitespace_pattern" not in sglang_result   # passes — bug confirmed
assert "whitespace_pattern" in vllm_result          # passes — vLLM is correct

Impact

Users who rely on whitespace_pattern to control output whitespace will get silently wrong results when using the SGLang backend — no error is raised, but the constraint is not applied. This is a silent correctness failure.

Fix

Forward whitespace_pattern in SGLangTypeAdapter.format_output_type, mirroring the vLLM implementation:

python
elif isinstance(term, JsonSchema):
    structured_outputs = {"json": json.loads(term.schema)}
    if term.whitespace_pattern:
        structured_outputs["whitespace_pattern"] = term.whitespace_pattern
    return structured_outputs

Environment

Reproduced on Python 3.12 with Outlines main branch (2026-08-13).

Source: dottxt-ai/outlines

View original on GitHubView discussion on GitHub