Update app.py

app.py

@@ -53,7 +53,7 @@ def infer_fin_pali(image, question):
     device = "cuda:0" if torch.cuda.is_available() else "cpu"
     torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
     
-    model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch_dtype, trust_remote_code=True, token=access_token).to(device)
+    model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch_dtype, trust_remote_code=True, quantization_config=bnb_config,token=access_token).to(device)
     processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True, token=access_token)
 ###