#!/usr/bin/env python3
"""Quantize bf16 model to BnB 4-bit using transformers' built-in mechanism."""

import argparse
import gc
import torch
from pathlib import Path
from transformers import AutoModelForCausalLM, BitsAndBytesConfig


def quantize_model(model_path, output_path):
    """Load bf16 model, quantize to BnB 4-bit, save."""
    
    print(f"Loading model from: {model_path}")
    
    # Load with BnB quantization config and device_map="cpu"
    bnb_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_compute_dtype=torch.float16,
    )
    
    print("Loading with BnB 4-bit quantization to CPU...")
    model = AutoModelForCausalLM.from_pretrained(
        model_path,
        quantization_config=bnb_config,
        device_map="cpu",
        torch_dtype=torch.float16,
        trust_remote_code=True,
        low_cpu_mem_usage=True,
    )
    print("✓ Model loaded with BnB 4-bit to CPU")
    
    # Check if model is actually quantized
    bnb_modules = sum(
        1 for m in model.modules() 
        if hasattr(m, 'weight') and hasattr(m.weight, 'quant_state')
    )
    print(f"   BnB quantized modules: {bnb_modules}")
    
    # Save model
    print(f"\nSaving to: {output_path}")
    model.save_pretrained(output_path)
    print("✓ Model saved")
    
    # Free memory
    del model
    gc.collect()
    torch.cuda.empty_cache()
    
    print("\nDone! Model is ready for QLoRA training.")
    print(f"Save location: {output_path}")


def main():
    parser = argparse.ArgumentParser(description="Quantize model to BnB 4-bit")
    parser.add_argument("--model-path", type=str, 
                       default="/data/models/Ornith-1.0-35B",
                       help="Path to bf16 model")
    parser.add_argument("--output-path", type=str,
                       default="/data/models/Ornith-1.0-35B-bnb-4bit",
                       help="Output path for quantized model")
    args = parser.parse_args()
    
    Path(args.output_path).mkdir(parents=True, exist_ok=True)
    quantize_model(args.model_path, args.output_path)


if __name__ == "__main__":
    main()