From e0cb3efce05388e21454556193d4069984b3154a Mon Sep 17 00:00:00 2001 From: Kai Hoang <56827800+hpdkhoa@users.noreply.github.com> Date: Sat, 10 Oct 2026 03:52:08 +0700 Subject: [PATCH] [Bugfix][Structured Output] Decode choice specs as JSON in outlines and LMFE (#60748) Signed-off-by: hpdkhoa Co-authored-by: Flora Feng <4florafeng@gmail.com> --- .../test_structured_output_manager.py | 18 +++++++++++++++++- .../backend_lm_format_enforcer.py | 3 +-- vllm/v1/structured_output/backend_outlines.py | 3 +-- 3 files changed, 19 insertions(+), 5 deletions(-) diff --git a/tests/v1/structured_output/test_structured_output_manager.py b/tests/v1/structured_output/test_structured_output_manager.py index fbceb3bd07..67c1582558 100644 --- a/tests/v1/structured_output/test_structured_output_manager.py +++ b/tests/v1/structured_output/test_structured_output_manager.py @@ -16,7 +16,8 @@ from vllm.parser.engine.parser_engine import ReasoningEnd from vllm.sampling_params import SamplingParams, StructuredOutputsParams from vllm.v1.request import Request from vllm.v1.structured_output import StructuredOutputManager -from vllm.v1.structured_output.backend_outlines import OutlinesGrammar +from vllm.v1.structured_output.backend_outlines import OutlinesBackend, OutlinesGrammar +from vllm.v1.structured_output.request import get_structured_output_key TOKENIZER = "gpt2" THINK_END = "\n" # reasoning-end marker (single GPT-2 token) @@ -560,3 +561,18 @@ def test_outlines_termination(tokenizer): assert grammar.validate_tokens([eos]) == [] assert grammar.accept_tokens(request.request_id, [one, eos, one]) assert grammar.is_terminated() + + +def test_outlines_choice_with_non_bmp_characters(tokenizer): + """The choice spec is JSON, which encodes emoji as surrogate pairs.""" + choices = ["😀 yes", "no"] + backend = OutlinesBackend( + VllmConfig(), tokenizer=tokenizer, vocab_size=len(tokenizer) + ) + request_type, grammar_spec = get_structured_output_key( + StructuredOutputsParams(choice=choices) + ) + + grammar = backend.compile_grammar(request_type, grammar_spec) + + assert grammar.accept_tokens("", tokenizer.encode(choices[0])) diff --git a/vllm/v1/structured_output/backend_lm_format_enforcer.py b/vllm/v1/structured_output/backend_lm_format_enforcer.py index 36eb5eb159..9a16ddc1fe 100644 --- a/vllm/v1/structured_output/backend_lm_format_enforcer.py +++ b/vllm/v1/structured_output/backend_lm_format_enforcer.py @@ -1,6 +1,5 @@ # SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project -import ast import json from dataclasses import dataclass, field from functools import lru_cache @@ -117,7 +116,7 @@ class LMFormatEnforcerBackend(StructuredOutputBackend): grammar_spec, ) elif request_type == StructuredOutputOptions.CHOICE: - choices = ast.literal_eval(grammar_spec) + choices = json.loads(grammar_spec) character_level_parser = lmformatenforcer.UnionParser( [lmformatenforcer.StringParser(choice) for choice in choices] ) diff --git a/vllm/v1/structured_output/backend_outlines.py b/vllm/v1/structured_output/backend_outlines.py index 1cfc6c8f5d..176d1a46e6 100644 --- a/vllm/v1/structured_output/backend_outlines.py +++ b/vllm/v1/structured_output/backend_outlines.py @@ -3,7 +3,6 @@ # SPDX-FileCopyrightText: Copyright contributors to the vLLM project from __future__ import annotations -import ast import json from dataclasses import dataclass, field from re import _constants, _parser # type: ignore[attr-defined] @@ -68,7 +67,7 @@ class OutlinesBackend(StructuredOutputBackend): elif request_type == StructuredOutputOptions.REGEX: regex = grammar_spec elif request_type == StructuredOutputOptions.CHOICE: - choices = ast.literal_eval(grammar_spec) + choices = json.loads(grammar_spec) choices = [regex_escape(c) for c in choices] regex = "(" + "|".join(choices) + ")" else: