src/examples/run_model.py

#!/usr/bin/env python3
"""
Run inference using SFT local model + GRPO model from HuggingFace.
"""

import os
import sys
import logging
from pathlib import Path
import torch

# ============================
# Fix PYTHONPATH
# ============================
PROJECT_ROOT = Path(__file__).resolve().parents[2]
sys.path.append(str(PROJECT_ROOT))

# ============================
# Logging setup
# ============================
logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s - %(levelname)s - %(message)s"
)
logger = logging.getLogger("model_inference")

# ============================
# Import AFTER fixing path
# ============================
try:
    from src.utils.inference import GRPOModelInference
except Exception as e:
    print("❌ Import src.utils.inference FAILED!")
    print("PROJECT_ROOT =", PROJECT_ROOT)
    print("Error =", e)
    sys.exit(1)


def main():
    print("🚀 Starting inference...")

    # ============================
    # Load model
    # ============================
    try:
        model_inference = GRPOModelInference(
            sft_model_path="models/sft/checkpoint-100",
            grpo_model_path="Dat1710/countdown-grpo-qwen2",
            base_model_id="Qwen/Qwen2.5-Math-1.5B",
            device="auto",
            dtype=torch.float16,
        )
        print("✅ Models loaded successfully!")
    except Exception as e:
        print("❌ Failed to load models!")
        print(e)
        return

    # ============================
    # Run inference
    # ============================
    problem = (
        "Your task: Use 53, 3, 47, and 36 exactly once each with "
        "only +, -, *, and / operators to create an expression equal to 133."
    )

    print("\n📌 Problem:")
    print(problem)

    response, extracted_answer, is_valid = model_inference.solve_problem(
        problem_description=problem,
        max_new_tokens=512,
        temperature=0.8,
    )

    # ============================
    # Print output clearly
    # ============================
    print("\n================= MODEL OUTPUT =================")
    print(response)
    print("================================================\n")

    print("📘 Extracted Answer:", extracted_answer)
    print("🔍 Valid format:", is_valid)


if __name__ == "__main__":
    main()