disable cuda in main

llava/model/builder.py

@@ -138,7 +138,7 @@ def load_pretrained_model(
             from peft import PeftModel
 
             print("Loading LoRA weights...")
-            model = PeftModel.from_pretrained(model, model_path)
+            model = PeftModel.from_pretrained(model, model_path, device_map=device_map)
             print("Merging LoRA weights...")
             model = model.merge_and_unload()
             print("Model is loaded...")
@@ -200,7 +200,7 @@ def load_pretrained_model(
                 device_map=device_map,
             )
             print(f"Loading LoRA weights from {model_path}")
-            model = PeftModel.from_pretrained(model, model_path)
+            model = PeftModel.from_pretrained(model, model_path, device_map=device_map)
             print(f"Merging weights")
             model = model.merge_and_unload()
             print("Convert to BF16...")