Upload DPO-Training/quantize_dpo_model.py with huggingface_hub

DPO-Training/quantize_dpo_model.py

@@ -0,0 +1,127 @@
+#!/usr/bin/env python3
+"""
+Quantize DPO-trained Qwen3-0.6B model to GGUF format.
+
+Usage:
+    python quantize_dpo_model.py --model_path ./qwen3-0.6b-dpo-merged
+    python quantize_dpo_model.py --model_path ./qwen3-0.6b-dpo-merged --quantization Q4_K_M
+"""
+
+import argparse
+import subprocess
+import os
+import sys
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Quantize DPO model to GGUF")
+    parser.add_argument("--model_path", required=True, help="Path to merged DPO model")
+    parser.add_argument(
+        "--output_name",
+        default=None,
+        help="Output GGUF file name (auto-generated if not set)",
+    )
+    parser.add_argument(
+        "--quantization",
+        default="Q4_K_S",
+        choices=[
+            "Q2_K",
+            "Q3_K_S",
+            "Q3_K_M",
+            "Q3_K_L",
+            "Q4_K_S",
+            "Q4_K_M",
+            "Q5_K_S",
+            "Q5_K_M",
+            "Q6_K",
+            "Q8_0",
+        ],
+        help="Quantization type",
+    )
+    parser.add_argument(
+        "--convert_script",
+        default="/home/ma/prima.cpp/convert_hf_to_gguf.py",
+        help="Path to llama.cpp conversion script",
+    )
+
+    args = parser.parse_args()
+
+    # Generate output name if not provided
+    if args.output_name is None:
+        model_name = os.path.basename(args.model_path.rstrip("/"))
+        args.output_name = f"{model_name}-{args.quantization}.gguf"
+
+    print("=" * 60)
+    print("DPO Model Quantization")
+    print("=" * 60)
+    print(f"Model: {args.model_path}")
+    print(f"Quantization: {args.quantization}")
+    print(f"Output: {args.output_name}")
+    print("=" * 60)
+
+    # Check if model path exists
+    if not os.path.exists(args.model_path):
+        print(f"❌ Error: Model path not found: {args.model_path}")
+        sys.exit(1)
+
+    # Step 1: Convert to GGUF (FP16)
+    print("\n📥 Converting to GGUF format...")
+    temp_gguf = f"{args.output_name}.temp"
+
+    convert_cmd = [
+        sys.executable,
+        args.convert_script,
+        args.model_path,
+        "--outfile",
+        temp_gguf,
+        "--outtype",
+        "f16",
+    ]
+
+    print(f"Running: {' '.join(convert_cmd)}")
+    result = subprocess.run(convert_cmd, capture_output=True, text=True)
+
+    if result.returncode != 0:
+        print(f"❌ Conversion failed: {result.stderr}")
+        sys.exit(1)
+
+    print("✅ Conversion complete")
+
+    # Step 2: Quantize
+    print(f"\n🔧 Quantizing to {args.quantization}...")
+
+    quantize_cmd = [
+        "/home/ma/prima.cpp/llama-quantize",
+        temp_gguf,
+        args.output_name,
+        args.quantization,
+    ]
+
+    print(f"Running: {' '.join(quantize_cmd)}")
+    result = subprocess.run(quantize_cmd, capture_output=True, text=True)
+
+    if result.returncode != 0:
+        print(f"❌ Quantization failed: {result.stderr}")
+        sys.exit(1)
+
+    print("✅ Quantization complete")
+
+    # Clean up temp file
+    if os.path.exists(temp_gguf):
+        os.remove(temp_gguf)
+        print("🧹 Cleaned up temporary files")
+
+    # Get file size
+    if os.path.exists(args.output_name):
+        size_mb = os.path.getsize(args.output_name) / (1024 * 1024)
+        print(f"\n📦 Output file: {args.output_name} ({size_mb:.1f} MB)")
+
+    print("\n" + "=" * 60)
+    print("✅ Quantization Complete!")
+    print("=" * 60)
+    print(f"Quantized model: {args.output_name}")
+    print(f"Ready for deployment!")
+
+
+if __name__ == "__main__":
+    main()