Update app.py

app.py

@@ -49,7 +49,7 @@ def model_inference(
         "Qwen/Qwen2-VL-2B-Instruct",
         attn_implementation="flash_attention_2", #doesn't work on zerogpu WTF?!
         trust_remote_code=True, 
-        torch_dtype="auto").cuda().eval()
+        torch_dtype=torch.bfloat16).to("cuda:0")
 
     # default processer
     min_pixels = 256*28*28