no cuda

app.py

@@ -51,7 +51,7 @@ def respond(
         tokenize = True,
         add_generation_prompt = True, # Must add for generation
         return_tensors = "pt",
-    ).to("cuda")
+    )
 
     outputs = model.generate(input_ids = inputs, max_new_tokens = max_tokens, use_cache = True,
                             temperature = 1.5, min_p = 0.1)