[Core] Support structured outputs for beam search (#35022)

Signed-off-by: Guan-Ming (Wesley) Chiu <guanmingchiu@gmail.com>
Signed-off-by: Guan-Ming (Wesley) Chiu <105915352+guan404ming@users.noreply.github.com>
This commit is contained in:
Guan-Ming (Wesley) Chiu
2026-06-12 06:56:25 -07:00
committed by GitHub
parent fbc3a1907a
commit c7aa3d2630
3 changed files with 405 additions and 70 deletions
+63
View File
@@ -5,11 +5,16 @@
Run `pytest tests/samplers/test_beam_search.py`.
"""
import json
import jsonschema
import pytest
from transformers import AutoModelForSeq2SeqLM
from vllm.assets.audio import AudioAsset
from vllm.entrypoints.llm import LLM
from vllm.platforms import current_platform
from vllm.sampling_params import BeamSearchParams, StructuredOutputsParams
# Extra engine kwargs needed for numerically deterministic beam search.
# On ROCm, floating-point reductions in attention and GEMM kernels are
@@ -223,3 +228,61 @@ def test_beam_search_passes_multimodal_data(
# NOTE: encoder/decoder tests are currently located under
# tests/models/multimodal/generation/test_whisper.py
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("beam_width", BEAM_WIDTHS)
def test_beam_search_structured_output(
model: str,
dtype: str,
beam_width: int,
) -> None:
"""Ensure beam search with structured output produces valid JSON."""
json_schema = {
"type": "object",
"properties": {
"name": {"type": "string"},
"age": {"type": "integer"},
},
"required": ["name", "age"],
"additionalProperties": False,
}
llm = LLM(
model=model,
dtype=dtype,
max_model_len=512,
structured_outputs_config=dict(
backend="xgrammar",
disable_any_whitespace=True,
),
**(dict(enforce_eager=True) | EXTRA_ENGINE_KWARGS),
)
params = BeamSearchParams(
beam_width=beam_width,
max_tokens=64,
structured_outputs=StructuredOutputsParams(json=json_schema),
)
prompts = [
"Generate a JSON object for a person with name and age:",
]
outputs = llm.beam_search(prompts, params)
assert len(outputs) == len(prompts)
for output in outputs:
assert len(output.sequences) > 0
for seq in output.sequences:
assert seq.text is not None
print(f"Full text: {seq.text!r}")
# seq.text includes the prompt, extract generated JSON.
gen_start = seq.text.find("{")
assert gen_start != -1, f"No JSON found in output: {seq.text!r}"
generated = seq.text[gen_start:]
generated = generated.replace("</s>", "").strip()
print(f"Generated JSON: {generated!r}")
parsed = json.loads(generated)
jsonschema.validate(instance=parsed, schema=json_schema)