#!/usr/bin/env python3
"""Simple NF4 quantization using BnB with device_map auto-distribution."""

import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig, AutoTokenizer


def quantize_model(model_path, output_path):
    print(f"Quantizing model from: {model_path}")

    bnb_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_compute_dtype=torch.bfloat16,
        bnb_4bit_use_double_quant=True,
    )

    print("Loading model with 4-bit quantization...")
    print("   Using CPU offloading to handle 70GB bf16 → 17GB 4-bit conversion\n")

    model = AutoModelForCausalLM.from_pretrained(
        model_path,
        quantization_config=bnb_config,
        device_map="auto",
        max_memory={0: "28GiB", 1: "28GiB", "cpu": "120GiB"},
        low_cpu_mem_usage=True,
        trust_remote_code=True,
    )

    print("\n✓ Model loaded and quantized")
    print(f"   GPU 0: {torch.cuda.memory_allocated(0) / 1e9:.2f} GB")
    print(f"   GPU 1: {torch.cuda.memory_allocated(1) / 1e9:.2f} GB")

    print(f"\nSaving quantized model to: {output_path}")
    model.save_pretrained(output_path)

    # Save tokenizer
    try:
        tokenizer = AutoTokenizer.from_pretrained(model_path)
        tokenizer.save_pretrained(output_path)
        print("✓ Tokenizer saved")
    except:
        print("⚠ No tokenizer found")

    print(f"\n✅ Quantized model saved to: {output_path}")


if __name__ == "__main__":
    import argparse
    parser = argparse.ArgumentParser()
    parser.add_argument("--model-path", type=str, default="/data/models/Ornith-1.0-35B")
    parser.add_argument("--output-path", type=str, default="/data/models/Ornith-1.0-35B-nf4")
    args = parser.parse_args()

    quantize_model(args.model_path, args.output_path)