vllm/tests/v1/structured_output/test_validation.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Request-time validation of structured output requests."""

import pytest

from vllm.config import StructuredOutputsConfig
from vllm.sampling_params import SamplingParams, StructuredOutputsParams

pytestmark = pytest.mark.cpu_test

JSON_SCHEMA = {
    "type": "object",
    "properties": {
        "invoice_id": {"type": "string"},
        "customer": {"type": "string"},
    },
    "required": ["invoice_id", "customer"],
    "additionalProperties": False,
}


class _StubModelConfig:
    def __init__(self, is_diffusion: bool):
        self.is_diffusion = is_diffusion


def test_structured_outputs_rejected_for_diffusion_models():
    """Diffusion LLMs denoise the canvas in parallel, which is incompatible
    with the token-by-token grammar FSM. The request must fail with a clear
    validation error instead of an FSM rejection mid-generation (#45436)."""
    params = SamplingParams(
        structured_outputs=StructuredOutputsParams(json=JSON_SCHEMA)
    )
    with pytest.raises(ValueError, match="not yet supported for diffusion"):
        params._validate_structured_outputs(
            _StubModelConfig(is_diffusion=True),
            StructuredOutputsConfig(),
            tokenizer=None,
        )


def test_plain_request_allowed_for_diffusion_models():
    """Requests without structured outputs are unaffected by the guard."""
    params = SamplingParams()
    params._validate_structured_outputs(
        _StubModelConfig(is_diffusion=True),
        StructuredOutputsConfig(),
        tokenizer=None,
    )