Add support for structured outputs

[pminervini]

As in https://docs.vllm.ai/en/latest/features/structured_outputs/#online-serving-openai-api and https://developers.openai.com/api/docs/guides/structured-outputs

I tried it with ds4 using the latest code from main but it doesn't seem to be working fine (both DGX Spark and m3 max) -- here's the test:

#!/usr/bin/env python3

import argparse
import json
import os
import sys
from pathlib import Path

from openai import APIConnectionError, APIStatusError, OpenAI, OpenAIError
from pydantic import BaseModel, ConfigDict, ValidationError


DEFAULT_BASE_URL = "http://127.0.0.1:8000/v1"
CONFIG_FILE = Path(__file__).resolve().parent / "config.json"


class CalendarEvent(BaseModel):
    model_config = ConfigDict(extra="forbid")

    name: str
    date: str
    participants: list[str]


def default_model() -> str:
    if os.getenv("SO_TEST_MODEL"):
        return os.environ["SO_TEST_MODEL"]
    if CONFIG_FILE.exists():
        try:
            config = json.loads(CONFIG_FILE.read_text(encoding="utf-8"))
        except json.JSONDecodeError:
            config = {}
        model = config.get("llm_model")
        if isinstance(model, str) and model:
            return model
    return "default"


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(
        description="Test Responses API structured output parsing against a local endpoint.",
    )
    parser.add_argument("--base-url", default=DEFAULT_BASE_URL)
    parser.add_argument("--model", default=default_model())
    parser.add_argument(
        "--api-key",
        default=os.getenv("OPENAI_API_KEY", "not-needed"),
        help="Dummy values are fine for local OpenAI-compatible endpoints.",
    )
    return parser.parse_args()


def main() -> int:
    args = parse_args()
    client = OpenAI(base_url=args.base_url, api_key=args.api_key)

    print(f"[so-test] base_url={args.base_url}")
    print(f"[so-test] model={args.model}")
    print("[so-test] calling client.responses.parse(..., text_format=CalendarEvent)")

    try:
        response = client.responses.parse(
            model=args.model,
            input=[
                {"role": "system", "content": "Extract the event information."},
                {
                    "role": "user",
                    "content": "Alice and Bob are going to a science fair on Friday.",
                },
            ],
            text_format=CalendarEvent,
        )
    except APIStatusError as exc:
        print(f"[so-test] API status error: HTTP {exc.status_code}", file=sys.stderr)
        print(str(exc), file=sys.stderr)
        return 1
    except APIConnectionError as exc:
        print(f"[so-test] connection error: {exc}", file=sys.stderr)
        return 1
    except OpenAIError as exc:
        print(f"[so-test] OpenAI SDK error: {exc}", file=sys.stderr)
        return 1
    except ValidationError as exc:
        print("[so-test] validation error while parsing structured output", file=sys.stderr)
        print(exc, file=sys.stderr)
        print(
            "[so-test] interpretation: the endpoint responded to /responses, "
            "but the returned text did not conform to CalendarEvent.",
            file=sys.stderr,
        )
        return 2

    event = response.output_parsed
    print("[so-test] parsed object:")
    print(event.model_dump_json(indent=2))
    print("[so-test] raw response id:", getattr(response, "id", None))
    return 0


if __name__ == "__main__":
    raise SystemExit(main())

Output:

$ python3 so-test-cli.py 
[so-test] base_url=http://127.0.0.1:8000/v1
[so-test] model=default
[so-test] calling client.responses.parse(..., text_format=CalendarEvent)
[so-test] validation error while parsing structured output
2 validation errors for CalendarEvent
name
  Field required [type=missing, input_value={'event': 'science fair',...Bob'], 'date': 'Friday'}, input_type=dict]
    For further information visit https://errors.pydantic.dev/2.12/v/missing
event
  Extra inputs are not permitted [type=extra_forbidden, input_value='science fair', input_type=str]
    For further information visit https://errors.pydantic.dev/2.12/v/extra_forbidden
[so-test] interpretation: the endpoint responded to /responses, but the returned text did not conform to CalendarEvent.

A bit swamped but happy to look into this at some point