fix: use BitsAndBytes INT4 for PEFT compatibility

training/scripts/train.py

@@ -32,15 +32,17 @@ def train(config_path):
     
     print(f"Loading model: {config['base_model']}")
     
-    # Load model - preserve NVFP4 quantization
+    # Load model with BitsAndBytes INT4 (PEFT supports this)
     print(f"Loading model: {config['base_model']}")
+    from transformers import BitsAndBytesConfig
+    
     model = AutoModelForCausalLM.from_pretrained(
         config["base_model"],
-        torch_dtype=torch.float16,  # Keep NVFP4 quantized
+        quantization_config=BitsAndBytesConfig(load_in_4bit=True),
         device_map="cpu",  # Load to CPU first
         trust_remote_code=True,
     )
-    print("Model loaded to CPU.")
+    print("Model loaded with INT4 quantization.")
     
     # Add LoRA
     lora_config = LoraConfig(