diff --git a/tests/v1/structured_output/test_backend_guidance.py b/tests/v1/structured_output/test_backend_guidance.py index 055ac9c364..59d1abf468 100644 --- a/tests/v1/structured_output/test_backend_guidance.py +++ b/tests/v1/structured_output/test_backend_guidance.py @@ -1,5 +1,6 @@ # SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project +import json import time from concurrent.futures import Future @@ -14,7 +15,10 @@ from vllm.sampling_params import SamplingParams, StructuredOutputsParams from vllm.tokenizers import get_tokenizer from vllm.v1.request import Request from vllm.v1.structured_output import StructuredOutputManager -from vllm.v1.structured_output.backend_guidance import GuidanceBackend +from vllm.v1.structured_output.backend_guidance import ( + GuidanceBackend, + process_for_additional_properties, +) from vllm.v1.structured_output.backend_types import StructuredOutputOptions TOKENIZER = "openai-community/gpt2" @@ -236,3 +240,82 @@ def test_mistral_tokenizer_compile_grammar( grammar = backend.compile_grammar(request_type, grammar_spec) assert grammar is not None assert not grammar.is_terminated() + + +def test_process_for_additional_properties_preserves_literals() -> None: + """Literal values under const/enum/default/examples must not be rewritten + when they contain a 'properties' or 'patternProperties' key (#58695).""" + feature = { + "type": "Feature", + "geometry": None, + "properties": {"name": "HQ"}, + "patternProperties": {"^x": "y"}, + } + schema = { + "type": "object", + "properties": { + "const_field": {"const": feature}, + "enum_field": {"enum": [feature]}, + "default_field": {"type": "object", "default": feature}, + "examples_field": {"type": "object", "examples": [feature]}, + "properties": {"type": "string"}, + "const": { + "type": "object", + "properties": {"id": {"type": "integer"}}, + }, + }, + "$defs": { + "item": { + "type": "object", + "properties": {"label": {"type": "string"}}, + } + }, + } + + processed = process_for_additional_properties(schema) + + assert processed["additionalProperties"] is False + assert processed["$defs"]["item"]["additionalProperties"] is False + assert processed["properties"]["const"]["additionalProperties"] is False + assert "additionalProperties" not in processed["properties"] + assert processed["properties"]["const_field"]["const"] == feature + assert processed["properties"]["enum_field"]["enum"] == [feature] + assert processed["properties"]["default_field"]["default"] == feature + assert processed["properties"]["examples_field"]["examples"] == [feature] + + +@pytest.mark.parametrize("keyword", ["const", "enum"]) +def test_disable_additional_properties_accepts_literal_with_properties_key( + keyword: str, +) -> None: + """GuidanceBackend with disable_additional_properties=True must accept the + original literal value under const/enum and reject injected + additionalProperties (#58695).""" + feature = {"type": "Feature", "geometry": None, "properties": {"name": "HQ"}} + rewritten = {**feature, "additionalProperties": False} + schema = {keyword: feature if keyword == "const" else [feature]} + + vllm_config = VllmConfig( + structured_outputs_config=StructuredOutputsConfig( + backend="guidance", + disable_additional_properties=True, + ) + ) + tokenizer = AutoTokenizer.from_pretrained(TOKENIZER) + backend = GuidanceBackend( + vllm_config, + tokenizer=tokenizer, + vocab_size=50257, + ) + grammar = backend.compile_grammar( + StructuredOutputOptions.JSON, + json.dumps(schema), + ) + + valid_tokens = tokenizer.encode(json.dumps(feature)) + [tokenizer.eos_token_id] + rewritten_tokens = tokenizer.encode(json.dumps(rewritten)) + [ + tokenizer.eos_token_id + ] + + assert grammar.validate_tokens(valid_tokens) == valid_tokens + assert grammar.validate_tokens(rewritten_tokens) != rewritten_tokens diff --git a/vllm/v1/structured_output/backend_guidance.py b/vllm/v1/structured_output/backend_guidance.py index 5f43c5a890..6dc2dc91ff 100644 --- a/vllm/v1/structured_output/backend_guidance.py +++ b/vllm/v1/structured_output/backend_guidance.py @@ -34,17 +34,58 @@ else: logger = init_logger(__name__) +_SCHEMA_MAP_KEYWORDS = ( + "properties", + "patternProperties", + "$defs", + "definitions", + "dependentSchemas", + "dependencies", +) + +_SUBSCHEMA_KEYWORDS = ( + "additionalProperties", + "unevaluatedProperties", + "propertyNames", + "contains", + "additionalItems", + "unevaluatedItems", + "not", + "if", + "then", + "else", + "contentSchema", + "items", + "prefixItems", + "allOf", + "anyOf", + "oneOf", +) + + def _walk_json_for_additional_properties(data: object): if isinstance(data, dict): - for value in data.values(): - _walk_json_for_additional_properties(value) + for key in _SCHEMA_MAP_KEYWORDS: + value = data.get(key) + if isinstance(value, dict): + for subschema in value.values(): + if isinstance(subschema, dict): + _walk_json_for_additional_properties(subschema) + + for key in _SUBSCHEMA_KEYWORDS: + value = data.get(key) + if isinstance(value, (dict, list)): + _walk_json_for_additional_properties(value) + if "additionalProperties" not in data and ( - "properties" in data or "patternProperties" in data + isinstance(data.get("properties"), dict) + or isinstance(data.get("patternProperties"), dict) ): data["additionalProperties"] = False elif isinstance(data, list): for item in data: - _walk_json_for_additional_properties(item) + if isinstance(item, dict): + _walk_json_for_additional_properties(item) def has_guidance_unsupported_json_features(schema: dict[str, Any]) -> bool: