feat: detect pre-quantized model and skip re-quantization

train.py

@@ -33,7 +33,23 @@ def train(config_path):
     
     print(f"Loading model: {config['base_model']}")
     
-    # Load bf16 to CPU with BnB 4-bit quantization
+    # Check if model is quantized (has model-XXXXX-of-XXXXX files)
+    import glob
+    safetensor_files = glob.glob(f"{config['base_model']}/*.safetensors")
+    is_quantized = any("of-" in f for f in safetensor_files)
+    
+    if is_quantized:
+        print(f"\n[INFO] Loading pre-quantized BnB 4-bit model (no re-quantization)...")
+        model = AutoModelForCausalLM.from_pretrained(
+            config["base_model"],
+            device_map="cpu",
+            torch_dtype=torch.float16,
+            trust_remote_code=True,
+            low_cpu_mem_usage=True,
+        )
+        print("✓ Model loaded to CPU (BnB 4-bit)")
+    else:
+        # Load bf16 with BnB 4-bit quantization
         print(f"\n[INFO] Loading {config['base_model']} with BnB 4-bit to CPU...")
         from transformers import BitsAndBytesConfig
         bnb_config = BitsAndBytesConfig(