scripts/run_autoround.py

#!/usr/bin/env python3
"""
AutoRound W4A16 Quantization for GLM-4.7 REAP models

This script quantizes a REAP-pruned GLM-4.7 model to INT4 weights using Intel's AutoRound.
Reduces model size by ~4x while maintaining quality.

Requirements:
    pip install auto-round

Usage:
    python run_autoround.py --model-path ./GLM-4.7-REAP-50 --output-dir ./GLM-4.7-REAP-50-W4A16
"""

import argparse
import subprocess
import sys
from pathlib import Path


def main():
    parser = argparse.ArgumentParser(description="AutoRound W4A16 quantization")
    parser.add_argument("--model-path", type=str, required=True,
                        help="Path to REAP-pruned model")
    parser.add_argument("--output-dir", type=str, default=None,
                        help="Output directory (default: {model-path}-W4A16)")
    parser.add_argument("--bits", type=int, default=4,
                        help="Weight bit width (default: 4)")
    parser.add_argument("--group-size", type=int, default=128,
                        help="Quantization group size (default: 128)")
    parser.add_argument("--format", type=str, default="auto_gptq",
                        choices=["auto_gptq", "auto_awq", "auto_round"],
                        help="Output format (default: auto_gptq)")
    parser.add_argument("--iters", type=int, default=200,
                        help="Optimization iterations (default: 200)")

    args = parser.parse_args()

    # Validate
    if not Path(args.model_path).exists():
        print(f"ERROR: Model path not found: {args.model_path}")
        sys.exit(1)

    # Build output directory
    if args.output_dir is None:
        args.output_dir = f"{args.model_path}-W{args.bits}A16"

    Path(args.output_dir).mkdir(parents=True, exist_ok=True)

    # Get model size info
    model_size_gb = sum(f.stat().st_size for f in Path(args.model_path).rglob("*.safetensors")) / (1024**3)
    expected_output_gb = model_size_gb / 4  # ~4x compression for W4

    print("=" * 60)
    print(f"AutoRound W{args.bits}A16 Quantization")
    print("=" * 60)
    print(f"Input Model: {args.model_path}")
    print(f"Input Size: {model_size_gb:.1f} GB")
    print(f"Output: {args.output_dir}")
    print(f"Expected Output Size: ~{expected_output_gb:.1f} GB")
    print(f"Config: {args.bits}-bit, group_size={args.group_size}, format={args.format}")
    print("=" * 60)
    print("\nThis will take ~2-3 hours for a 92-layer MoE model...")
    print()

    # Build command
    cmd = [
        "auto-round",
        "--model", args.model_path,
        "--bits", str(args.bits),
        "--group_size", str(args.group_size),
        "--format", args.format,
        "--output_dir", args.output_dir,
        "--iters", str(args.iters),
    ]

    result = subprocess.run(cmd)

    if result.returncode == 0:
        # Calculate actual output size
        output_size_gb = sum(f.stat().st_size for f in Path(args.output_dir).rglob("*.safetensors")) / (1024**3)
        compression = model_size_gb / output_size_gb if output_size_gb > 0 else 0

        print("\n" + "=" * 60)
        print("AutoRound quantization complete!")
        print(f"Output: {args.output_dir}")
        print(f"Output Size: {output_size_gb:.1f} GB ({compression:.1f}x compression)")
        print("=" * 60)
    else:
        print(f"\nERROR: AutoRound failed with code {result.returncode}")
        sys.exit(1)


if __name__ == "__main__":
    main()