fix zero gpu

app.py

@@ -44,9 +44,10 @@ tokenizer, model, data_loader = load_model_and_dataloader(
     load_4bit=load_4bit,
     load_bf16=load_bf16,
     scene_to_obj_mapping=scene_to_obj_mapping,
-)
+    device_map='cpu',
+)  # Huggingface Zero-GPU has to use .to(device) to set the device, otherwise it will fail
+model.to("cuda")  # Huggingface Zero-GPU requires explicit device placement
 
-@spaces.GPU
 def get_chatbot_response(user_chat_input, scene_id):
     # Get the response from the model
     prompt, response = get_model_response(

model.py

@@ -12,8 +12,7 @@ from llava.mm_utils import get_model_name_from_path
 from llava.model.builder import load_pretrained_model
 from llava.utils import disable_torch_init
 
-@spaces.GPU
-def load_model_and_dataloader(model_path, model_base, scene_to_obj_mapping, obj_context_feature_type="text", load_8bit=False, load_4bit=False, load_bf16=False):
+def load_model_and_dataloader(model_path, model_base, scene_to_obj_mapping, obj_context_feature_type="text", load_8bit=False, load_4bit=False, load_bf16=False, device_map='auto'):
 
     model_name = get_model_name_from_path(model_path)
 
@@ -24,6 +23,7 @@ def load_model_and_dataloader(model_path, model_base, scene_to_obj_mapping, obj_
         load_8bit=load_8bit,
         load_4bit=load_4bit,
         load_bf16=load_bf16,
+        device_map=device_map,
     )
 
     dataset = ObjIdentifierDataset(
@@ -41,7 +41,6 @@ def load_model_and_dataloader(model_path, model_base, scene_to_obj_mapping, obj_
     return tokenizer, model, data_loader
 
 
-@spaces.GPU
 def get_model_response(model, tokenizer, data_loader, scene_id, user_input, max_new_tokens=50, temperature=0.2, top_p=0.9):
     input_data = [
         {