Update app.py

app.py

@@ -41,7 +41,7 @@ user
 {user_prompt}
 assistant
 '''
-    input_ids = tokenizer(prompt_template, return_tensors='pt').input_ids
+    input_ids = tokenizer(prompt_template, return_tensors='pt').input_ids.cuda()
     
     # Gerar a saída
     output = model.generate(
@@ -119,4 +119,4 @@ with gr.Blocks(css=CSS) as demo:
     )
     stop.click(fn=None, inputs=None, outputs=None, cancels=[submit_click_event], queue=False)
 
-demo.queue(max_size=128, concurrency_count=48).launch(debug=True, server_name="0.0.0.0", server_port=7860)
+demo.queue(max_size=128, concurrency_count=2)