Update app.py

app.py

@@ -546,6 +546,9 @@ linear_proj = mm_backbone_mlp_model.mm_projector
 pretrained_llm_checkpoint = lazy_load(pretrained_llm_path)
 # 2. Load the fine-tuned LoRA weights
 lora_checkpoint = lazy_load(lora_path)
+for key, value in lora_checkpoint.items():
+    if isinstance(value, torch.Tensor):
+        lora_checkpoint[key] = value.to('cpu')
 # 3. merge the two checkpoints
 model_state_dict = {**pretrained_llm_checkpoint, **lora_checkpoint}
 model.load_state_dict(model_state_dict, strict=True)