[Bugfix][Structured Output] Cap JSON schema nesting to prevent 500s and API-server stalls (#60036)

Signed-off-by: Mazukiri <maihoangduc04@gmail.com>
Signed-off-by: sfeng33 <4florafeng@gmail.com>
Co-authored-by: Flora Feng <4florafeng@gmail.com>
This commit is contained in:
Mai Hoàng Đức
2026-10-06 00:06:59 +00:00
committed by GitHub
co-authored by Flora Feng
parent fa423b614c
commit 5cd8a2626b
11 changed files with 259 additions and 8 deletions
@@ -1,6 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import json
from dataclasses import dataclass, field
from typing import Any
from unittest.mock import AsyncMock, MagicMock, patch
@@ -642,7 +643,15 @@ def test_batch_structural_tag_response_format_invalid(format_value):
)
@pytest.mark.parametrize("structural_tag", ["not json", ""])
@pytest.mark.parametrize(
"structural_tag",
[
"not json",
"",
# json.loads raises a plain ValueError, not JSONDecodeError
'{"type": "structural_tag", "x": ' + "1" * 5000 + "}",
],
)
def test_structured_outputs_structural_tag_invalid(structural_tag):
"""Malformed direct structured_outputs structural tags should be rejected."""
with pytest.raises(
@@ -656,6 +665,25 @@ def test_structured_outputs_structural_tag_invalid(structural_tag):
)
@pytest.mark.parametrize("field", ["response_format", "structured_outputs"])
def test_deeply_nested_structural_tag_rejected(field):
"""Structural tags are compiled during request parsing, so a deeply nested
one must be rejected there instead of blocking the server while converting,
and reported as too deep rather than malformed."""
schema = '{"type": "array", "items": ' * 200 + "{}" + "}" * 200
tag = (
'{"type": "structural_tag", "format": {"type": "json_schema", '
f'"json_schema": {schema}}}}}'
)
value = json.loads(tag) if field == "response_format" else {"structural_tag": tag}
with pytest.raises(VLLMValidationError, match="nested too deeply"):
ChatCompletionRequest(
model=MODEL_NAME,
messages=[{"role": "user", "content": "hello"}],
**{field: value},
)
@pytest.mark.parametrize("field_name", ["prompt_logprobs", "top_logprobs"])
def test_non_numeric_logprobs_rejected(field_name):
"""A non-numeric logprobs value must be a clean 400 validation error, not a
@@ -19,6 +19,7 @@ from vllm.entrypoints.generate.base.protocol import (
)
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
from vllm.exceptions import VLLMValidationError
from vllm.parser.cohere_command import (
CohereCommandParser,
CohereNormalizedTool,
@@ -339,6 +340,14 @@ class TestAdjustRequestFoldFromStructuredOutputs:
_make_chat_request(structured_outputs={"json": json_value}),
)
def test_deeply_nested_json_string_rejected(self, parser) -> None:
"""json.loads raises RecursionError on a string nested this deeply."""
schema = '{"type": "array", "items": ' * 20_000 + "{}" + "}" * 20_000
with pytest.raises(VLLMValidationError, match="nested too deeply"):
parser.adjust_request(
_make_chat_request(structured_outputs={"json": schema}),
)
@pytest.mark.parametrize(
"construct",
[
@@ -450,6 +459,15 @@ class TestCollectToolSchemaDefs:
self._tool_calls_json("get_weather", {"location": {"city": 1}}),
)
def test_deeply_nested_parameters_rejected(self) -> None:
"""Tool schemas are compiled here, before request validation, and end up
inside an EBNF grammar that the nesting check cannot see later."""
parameters = json.loads('{"type": "array", "items": ' * 200 + "{}" + "}" * 200)
with pytest.raises(VLLMValidationError, match="nested too deeply"):
collect_tool_schema(
[CohereNormalizedTool(name="deep", parameters=parameters)]
)
def test_input_parameters_not_mutated(self) -> None:
params = json.loads(json.dumps(NESTED_DEFS_PARAMETERS))
collect_tool_schema(
+14
View File
@@ -35,6 +35,7 @@ from vllm.entrypoints.generate.base.protocol import (
from vllm.entrypoints.openai.chat_completion.protocol import (
ChatCompletionRequest,
)
from vllm.exceptions import VLLMValidationError
from vllm.parser.engine.events import EventType
from vllm.parser.engine.streaming_parser_engine import StreamingParserEngine
from vllm.parser.mistral import _DEFAULT_JSON_SCHEMA, MistralParser, mistral_config
@@ -1890,6 +1891,19 @@ def test_adjust_request_structured_outputs_generates_grammar(
assert len(result.structured_outputs.grammar) > 0
def test_adjust_request_deeply_nested_json_string_rejected(
mistral_tool_parser: MistralToolParser,
) -> None:
"""The schema string is parsed here, before request validation, and
json.loads raises RecursionError on a string nested this deeply."""
schema = '{"type": "array", "items": ' * 20_000 + "{}" + "}" * 20_000
request = _make_request(
structured_outputs=StructuredOutputsParams(json=schema),
)
with pytest.raises(VLLMValidationError, match="nested too deeply"):
mistral_tool_parser.adjust_request(request)
@pytest.mark.parametrize(
"response_format_kwargs,expected_json_schema",
[
+17 -1
View File
@@ -13,7 +13,8 @@ from vllm.entrypoints.openai.chat_completion.protocol import (
ChatCompletionRequest,
ChatCompletionToolsParam,
)
from vllm.parser.abstract_parser import DelegatingParser
from vllm.exceptions import VLLMValidationError
from vllm.parser.abstract_parser import DelegatingParser, structured_outputs_to_format
from vllm.sampling_params import StructuredOutputsParams
from vllm.tool_parsers.abstract_tool_parser import ToolParser
from vllm.tool_parsers.qwen3_engine_tool_parser import Qwen3EngineToolParser
@@ -303,3 +304,18 @@ class TestToolChoice_Plus_ResponseFormat:
assert _is_grammar_accept_string(grammar, self._qwen_tool_call())
assert not _is_grammar_accept_string(grammar, '{"text": "hi"}')
mock_warn.assert_called_once()
@pytest.mark.parametrize("field", ["json", "structural_tag"])
def test_structured_outputs_to_format_rejects_deeply_nested_string(field):
"""Parsers convert the constraint before request validation, and json.loads
raises RecursionError on a string nested this deeply."""
schema = '{"type": "array", "items": ' * 20_000 + "{}" + "}" * 20_000
value = schema
if field == "structural_tag":
value = (
'{"type": "structural_tag", "format": {"type": "json_schema", '
f'"json_schema": {schema}}}}}'
)
with pytest.raises(VLLMValidationError, match="nested too deeply"):
structured_outputs_to_format(StructuredOutputsParams(**{field: value}))
+100 -1
View File
@@ -2,11 +2,18 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Request-time validation of structured output requests."""
import json
import pytest
from vllm.config import StructuredOutputsConfig
from vllm.exceptions import VLLMClientError, VLLMValidationError
from vllm.sampling_params import SamplingParams, StructuredOutputsParams
from vllm.sampling_params import (
MAX_STRUCTURED_OUTPUT_JSON_NESTING,
STRUCTURAL_TAG_WRAPPER_NESTING,
SamplingParams,
StructuredOutputsParams,
)
pytestmark = pytest.mark.cpu_test
@@ -110,6 +117,98 @@ def test_regex_with_nul_byte_rejected(regex):
validate_xgrammar_grammar(params)
def _nested_array_schema(levels: int) -> str:
# Built as text: json.dumps on a deeply nested dict would itself recurse.
wrappers = levels - 1
return (
'{"type": "array", "items": ' * wrappers + '{"type": "string"}' + "}" * wrappers
)
def _structural_tag(schema: str) -> str:
return (
'{"type": "structural_tag", "format": {"type": "json_schema", '
f'"json_schema": {schema}}}}}'
)
@pytest.mark.parametrize(
"structured_outputs",
[
pytest.param(
{"json": _nested_array_schema(MAX_STRUCTURED_OUTPUT_JSON_NESTING + 1)},
id="json-str-over-limit",
),
pytest.param(
{"json": json.loads(_nested_array_schema(2_000))},
id="json-dict-past-recursion-limit",
),
pytest.param(
{"json": _nested_array_schema(200_000)},
id="json-str-past-json-loads-recursion-limit",
),
pytest.param(
{
"structural_tag": _structural_tag(
_nested_array_schema(
MAX_STRUCTURED_OUTPUT_JSON_NESTING
+ STRUCTURAL_TAG_WRAPPER_NESTING
)
)
},
id="structural-tag-over-limit",
),
],
)
@pytest.mark.parametrize(
"backend", ["auto", "xgrammar", "guidance", "outlines", "lm-format-enforcer"]
)
def test_deeply_nested_schema_rejected(structured_outputs, backend):
"""Converting a schema to a grammar gets steeply more expensive with nesting
and runs on the API server; past the recursion limit it used to escape as a
RecursionError (HTTP 500 with an empty message). Deep nesting must be a
validation error before any backend touches the schema."""
params = SamplingParams(
structured_outputs=StructuredOutputsParams(**structured_outputs)
)
with pytest.raises(VLLMValidationError, match="nested too deeply"):
params._validate_structured_outputs(
_StubModelConfig(is_diffusion=False),
StructuredOutputsConfig(backend=backend),
tokenizer=object(),
)
@pytest.mark.parametrize("field", ["json", "structural_tag"])
def test_xgrammar_validation_rejects_deeply_nested_schema(field):
"""Request parsing and tool parsers call validate_xgrammar_grammar directly,
before _validate_structured_outputs, so it must check nesting itself."""
from vllm.v1.structured_output.backend_xgrammar import validate_xgrammar_grammar
schema = _nested_array_schema(2_000)
value = _structural_tag(schema) if field == "structural_tag" else schema
params = SamplingParams(
structured_outputs=StructuredOutputsParams(**{field: value})
)
with pytest.raises(VLLMValidationError, match="nested too deeply"):
validate_xgrammar_grammar(params)
@pytest.mark.parametrize("field", ["json", "structural_tag"])
def test_schema_at_nesting_limit_accepted(field):
"""A structural tag wraps its schema in a few more levels (more for the
tags tool parsers generate), which must not count against the schema."""
schema = _nested_array_schema(MAX_STRUCTURED_OUTPUT_JSON_NESTING)
value = _structural_tag(schema) if field == "structural_tag" else schema
structured_outputs = StructuredOutputsParams(**{field: value})
SamplingParams(structured_outputs=structured_outputs)._validate_structured_outputs(
_StubModelConfig(is_diffusion=False),
StructuredOutputsConfig(),
tokenizer=object(),
)
assert structured_outputs._backend == "xgrammar"
INVALID_JSON_SCHEMA = {"type": "object", "properties": {"name": {"type": "str"}}}
+8 -1
View File
@@ -230,7 +230,11 @@ def validate_structural_tag_response_format(
def validate_structural_tag_payload(payload: Any, *, parameter: str) -> None:
from vllm.sampling_params import SamplingParams, StructuredOutputsParams
from vllm.sampling_params import (
SamplingParams,
StructuredOutputsParams,
check_json_nesting,
)
from vllm.v1.structured_output.backend_xgrammar import validate_xgrammar_grammar
if isinstance(payload, str) and not payload:
@@ -239,6 +243,9 @@ def validate_structural_tag_payload(payload: Any, *, parameter: str) -> None:
parameter=parameter,
)
if isinstance(payload, str):
# Raised here so the error is not reported as a malformed tag below
check_json_nesting(payload, structural_tag=True)
try:
validate_xgrammar_grammar(
SamplingParams(
+7 -1
View File
@@ -42,7 +42,11 @@ from vllm.parser.engine.adapters import ParserEngineToolAdapter
from vllm.parser.metrics import record_tool_parser_invocation
from vllm.parser.utils import count_history_tool_calls
from vllm.reasoning.abs_reasoning_parsers import ReasoningParser
from vllm.sampling_params import SamplingParams, StructuredOutputsParams
from vllm.sampling_params import (
SamplingParams,
StructuredOutputsParams,
check_json_nesting,
)
from vllm.tokenizers import TokenizerLike
from vllm.tool_parsers.abstract_tool_parser import Tool, ToolParser
from vllm.tool_parsers.streaming import (
@@ -274,6 +278,7 @@ def structured_outputs_to_format(params: StructuredOutputsParams) -> Format | No
if params.json is not None:
schema = params.json
if isinstance(schema, str):
check_json_nesting(schema)
schema = json.loads(schema)
return JSONSchemaFormat(json_schema=schema)
if params.regex is not None:
@@ -295,6 +300,7 @@ def structured_outputs_to_format(params: StructuredOutputsParams) -> Format | No
raise VLLMValidationError("Invalid grammar specification.") from e
return GrammarFormat(grammar=grammar)
if params.structural_tag is not None:
check_json_nesting(params.structural_tag, structural_tag=True)
s_tag = json.loads(params.structural_tag)
if "structures" in s_tag:
# LegacyStructuralTagResponseFormat
+3 -1
View File
@@ -45,7 +45,7 @@ from vllm.reasoning.cohere_command_reasoning_parser import (
BaseCohereCommandReasoningParser,
)
from vllm.renderers.cohere import POSITION_TO_SOURCE_KEY
from vllm.sampling_params import StructuredOutputsParams
from vllm.sampling_params import StructuredOutputsParams, check_json_nesting
from vllm.tool_parsers.cohere_command_tool_parser import BaseCohereCommandToolParser
if TYPE_CHECKING:
@@ -127,6 +127,7 @@ def collect_tool_schema(tool_schema: list[CohereNormalizedTool]) -> str:
tool_dictionary: dict[str, str] = {}
for tool in tool_schema:
tool_name = tool["name"]
check_json_nesting(tool["parameters"])
tool_parameters = dict(tool["parameters"])
json_schema: dict[str, Any] = {
"type": "object",
@@ -404,6 +405,7 @@ def _schema_dict_from_structured_outputs(
if isinstance(raw, str):
if not raw.strip():
raise ValueError("structured_outputs.json cannot be empty.")
check_json_nesting(raw)
try:
raw = json.loads(raw)
except json.JSONDecodeError as e:
+2 -1
View File
@@ -48,7 +48,7 @@ from vllm.parser.engine.parser_engine_config import (
ParserState,
Transition,
)
from vllm.sampling_params import StructuredOutputsParams
from vllm.sampling_params import StructuredOutputsParams, check_json_nesting
from vllm.utils.mistral import is_mistral_tokenizer
if TYPE_CHECKING:
@@ -381,6 +381,7 @@ class MistralParser(ParserEngine):
json_schema = _DEFAULT_JSON_SCHEMA
elif request.structured_outputs.json is not None:
if isinstance(request.structured_outputs.json, str):
check_json_nesting(request.structured_outputs.json)
json_schema = json.loads(request.structured_outputs.json)
else:
json_schema = request.structured_outputs.json
+58
View File
@@ -33,6 +33,15 @@ MAX_LOGPROB_TOKEN_IDS = 128
"""Upper bound on `SamplingParams.logprob_token_ids` list length. Must match
the per-request row width allocated by the sampler's `LogprobTokenIdsState`."""
MAX_STRUCTURED_OUTPUT_JSON_NESTING = 64
"""Upper bound on nested objects and arrays in a structured-output JSON schema
or structural tag. Grammar conversion runs on the API server and its cost grows
steeply with nesting. OpenAI allows 10 levels of schema nesting."""
STRUCTURAL_TAG_WRAPPER_NESTING = 16
"""Extra nesting allowed for a structural tag, which wraps its JSON schemas in
tag structure (up to 15 levels for tool-call tags generated by tool parsers)."""
def _verify_num_sequences(value: int, parameter_name: str) -> None:
if not isinstance(value, int):
@@ -202,6 +211,52 @@ class RequestOutputKind(Enum):
FINAL_ONLY = 2
def _exceeds_json_nesting(spec: str | dict[str, Any], limit: int) -> bool:
"""Whether a JSON document nests objects and arrays deeper than `limit`."""
if isinstance(spec, str):
try:
spec = json_mod.loads(spec)
except RecursionError:
return True
except ValueError:
# Malformed JSON is reported by the backend validation.
return False
level: list[Any] = [spec]
for _ in range(limit):
level = [
child
for node in level
if isinstance(node, (dict, list))
for child in (node.values() if isinstance(node, dict) else node)
if isinstance(child, (dict, list))
]
if not level:
return False
return True
def check_json_nesting(
spec: str | dict[str, Any] | None, *, structural_tag: bool = False
) -> None:
"""Reject a JSON schema nested more than `MAX_STRUCTURED_OUTPUT_JSON_NESTING`
levels deep, or a structural tag nested deeper than that plus
`STRUCTURAL_TAG_WRAPPER_NESTING`.
Raises:
VLLMValidationError: If `spec` is nested too deeply.
"""
limit = MAX_STRUCTURED_OUTPUT_JSON_NESTING
if structural_tag:
limit += STRUCTURAL_TAG_WRAPPER_NESTING
if spec and _exceeds_json_nesting(spec, limit):
raise VLLMValidationError(
"The JSON schema is nested too deeply: objects and arrays may nest at "
f"most {MAX_STRUCTURED_OUTPUT_JSON_NESTING} levels"
)
def _is_non_tekken_mistral(tokenizer: TokenizerLike) -> bool:
return is_mistral_tokenizer(tokenizer) and not tokenizer.is_tekken
@@ -1230,6 +1285,9 @@ class SamplingParams(
raise VLLMValidationError(
"structured_outputs.json cannot be an empty string"
)
# Reject deep nesting before any backend walks or converts the schema
check_json_nesting(self.structured_outputs.json)
check_json_nesting(self.structured_outputs.structural_tag, structural_tag=True)
# Reject json_object=False early to avoid engine-side crashes
if self.structured_outputs.json_object is False:
raise VLLMValidationError(
@@ -10,7 +10,7 @@ import torch
import vllm.envs
from vllm.exceptions import VLLMValidationError
from vllm.logger import init_logger
from vllm.sampling_params import SamplingParams
from vllm.sampling_params import SamplingParams, check_json_nesting
from vllm.utils.import_utils import LazyLoader
from vllm.utils.mistral import is_mistral_tokenizer
from vllm.v1.structured_output.backend_types import (
@@ -392,6 +392,7 @@ def validate_xgrammar_grammar(sampling_params: SamplingParams) -> None:
return
if so_params.json:
check_json_nesting(so_params.json)
if isinstance(so_params.json, str):
try:
schema = json.loads(so_params.json)
@@ -426,6 +427,7 @@ def validate_xgrammar_grammar(sampling_params: SamplingParams) -> None:
return
if so_params.structural_tag:
check_json_nesting(so_params.structural_tag, structural_tag=True)
try:
s_tag = json.loads(so_params.structural_tag)