Upload folder using huggingface_hub

scripts/run_autoround.py

@@ -0,0 +1,95 @@
+#!/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()

scripts/run_reap.py

@@ -0,0 +1,115 @@
+#!/usr/bin/env python3
+"""
+REAP (Router-Experts Activation Pruning) for GLM-4.7 MoE
+
+This script prunes MoE experts from GLM-4.7 using the REAP methodology from Cerebras.
+Requires: https://github.com/Cerebras/reap (or fork with GLM support)
+
+Usage:
+    python run_reap.py --compression-ratio 0.50 --model-path /path/to/GLM-4.7
+
+For observation reuse (instant pruning at different ratios):
+    python run_reap.py --compression-ratio 0.35 --reuse-observations observations_1360_angular-seed_42.pt
+"""
+
+import argparse
+import subprocess
+import sys
+from pathlib import Path
+
+
+def main():
+    parser = argparse.ArgumentParser(description="REAP pruning for GLM-4.7")
+    parser.add_argument("--model-path", type=str, required=True,
+                        help="Path to GLM-4.7 model")
+    parser.add_argument("--compression-ratio", type=float, required=True,
+                        help="Compression ratio (0.30 = keep 70%, 0.50 = keep 50%)")
+    parser.add_argument("--output-dir", type=str, default=None,
+                        help="Output directory (default: auto-generated)")
+    parser.add_argument("--dataset", type=str,
+                        default="0xSero/glm47-reap-calibration-v2",
+                        help="Calibration dataset")
+    parser.add_argument("--samples", type=int, default=1360,
+                        help="Number of calibration samples")
+    parser.add_argument("--seed", type=int, default=42,
+                        help="Random seed")
+    parser.add_argument("--distance", type=str, default="angular",
+                        choices=["angular", "cosine", "euclidean"],
+                        help="Distance measure for expert clustering")
+    parser.add_argument("--reuse-observations", type=str, default=None,
+                        help="Path to pre-computed observations file for instant pruning")
+    parser.add_argument("--reap-repo", type=str, default="./reap",
+                        help="Path to REAP repository")
+
+    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)
+
+    reap_script = Path(args.reap_repo) / "src" / "reap" / "prune.py"
+    if not reap_script.exists():
+        print(f"ERROR: REAP prune.py not found at: {reap_script}")
+        print("Clone the REAP repo: git clone https://github.com/Cerebras/reap")
+        sys.exit(1)
+
+    # Build output directory name
+    if args.output_dir is None:
+        ratio_pct = int(args.compression_ratio * 100)
+        args.output_dir = f"./GLM-4.7-REAP-{ratio_pct}"
+
+    Path(args.output_dir).mkdir(parents=True, exist_ok=True)
+
+    # Build command
+    cmd = [
+        sys.executable, str(reap_script),
+        "--model-name", args.model_path,
+        "--dataset-name", args.dataset,
+        "--compression-ratio", str(args.compression_ratio),
+        "--prune-method", "reap",
+        "--seed", str(args.seed),
+        "--do-eval", "false",
+        "--profile", "false",
+        "--samples_per_category", str(args.samples),
+        "--model_max_length", "2048",
+        "--distance_measure", args.distance,
+        "--record_pruning_metrics_only", "true",
+        "--output_file_name", f"observations_{args.samples}_{args.distance}-seed_{args.seed}.pt",
+    ]
+
+    if args.reuse_observations:
+        cmd.extend(["--load_observations", args.reuse_observations])
+        print(f"Reusing observations from: {args.reuse_observations}")
+        print("This enables instant pruning without re-running calibration!")
+
+    print("=" * 60)
+    print(f"REAP Pruning: GLM-4.7 @ {args.compression_ratio*100:.0f}% compression")
+    print("=" * 60)
+    print(f"Model: {args.model_path}")
+    print(f"Output: {args.output_dir}")
+    print(f"Dataset: {args.dataset} ({args.samples} samples)")
+    print(f"Distance: {args.distance}")
+    print("=" * 60)
+
+    # Run REAP
+    env = {
+        **dict(__import__('os').environ),
+        "CUDA_VISIBLE_DEVICES": "0,1,2,3,4,5,6,7",
+        "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True",
+    }
+
+    result = subprocess.run(cmd, env=env)
+
+    if result.returncode == 0:
+        print("\n" + "=" * 60)
+        print("REAP pruning complete!")
+        print(f"Pruned model saved to: {args.output_dir}")
+        print("=" * 60)
+    else:
+        print(f"\nERROR: REAP failed with code {result.returncode}")
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()