scripts/validate_brain_gguf.py

"""Validate the GGUF brain returns the same valid 3-channel JSON as the Modal
backend, on the SAME inputs train/brain_serve.py was validated on.

    uv run --group train --with llama-cpp-python python scripts/validate_brain_gguf.py \
        --gguf data/gguf/yui-brain1-sft1.Q4_K_M.gguf
"""

from __future__ import annotations

import argparse
import json
import sys
from pathlib import Path

sys.path.insert(0, str(Path(__file__).resolve().parent.parent))

from voice.brain import LlamaCppBrain  # noqa: E402

# The exact devices_block from train/brain_serve.py main() — the Modal validation input.
DEVICES = (
    "light.kitchen 'Kitchen Light' = off\n"
    "light.living_room 'Living Room Light' = on;60%\n"
    "lock.front_door 'Front Door' = locked\n"
    "cover.garage 'Garage Door' = closed"
)
TIME = "3:45 PM on Tuesday, June 10 2026"  # fixed so output is reproducible
CASES = [
    ("turn on the kitchen light", "intents"),
    ("is the front door locked?", "response"),
]


def main() -> None:
    ap = argparse.ArgumentParser()
    ap.add_argument("--gguf", default="data/gguf/yui-brain1-sft1.Q4_K_M.gguf")
    args = ap.parse_args()

    brain = LlamaCppBrain(model_path=args.gguf, n_gpu_layers=0, verbose=False)
    for utt, expect_key in CASES:
        out = brain.infer(DEVICES, utt, TIME)
        ok = expect_key in out and not out.get("_parse_error")
        print(f"\nutterance : {utt}")
        print(f"expect    : top key '{expect_key}'")
        print(f"output    : {json.dumps(out, ensure_ascii=False)}")
        print(f"valid     : {ok}")


if __name__ == "__main__":
    main()