fix: remove PEFT prepare (causes OOM)

quantize_to_bnb.py

@@ -25,14 +25,8 @@ def quantize_model(model_path, output_path):
     total_params = sum(p.numel() for p in model.parameters())
     print(f"   Total parameters: {total_params / 1e9:.2f}B")
     
-    # Apply PEFT prepare for k-bit training
-    print("\nApplying PEFT prepare_model_for_kbit_training...")
-    from peft import prepare_model_for_kbit_training
-    model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=False)
-    print("✓ Model prepared for k-bit training")
-    
     # Quantize using BnB's actual API
-    print("Quantizing with BnB 4-bit...")
+    print("\nQuantizing with BnB 4-bit...")
     from bitsandbytes.nn import Linear4bit
     from torch import nn