Update app.py

app.py

@@ -94,7 +94,7 @@ def bot_streaming(message, history):
 
     image_tensor = image_tensor.unsqueeze(0)
     image_tensor = image_tensor.to(dtype=torch.float16, device='cuda', non_blocking=True)
-    inputs.to(device='cuda', non_blocking=True)
+    inputs = inputs.to(device='cuda', non_blocking=True)
 
     print('image', image_tensor.shape)
     print('inputs', inputs.shape)
@@ -104,11 +104,14 @@ def bot_streaming(message, history):
     
     generation_kwargs = dict(inputs=inputs, images=image_tensor, streamer=streamer, max_new_tokens=1024, do_sample=False, eos_token_id = terminators)
 
-    print('here?')
 
     thread = Thread(target=model.generate, kwargs=generation_kwargs)
+
+    print('here?')
+    
     thread.start()
 
+    print('start')
 
     buffer = ""
     # time.sleep(0.5)