Update app.py

app.py

@@ -8,7 +8,7 @@ history = []
 
 def generateResponse(prompt, history):
     formattedPrompt = f"<s>[INST] {prompt} [/INST]"
-    response = model(formattedPrompt, max_new_tokens=1024, threads=8)
+    response = model(formattedPrompt, max_new_tokens=1024)
     history.append([prompt, response])
     return response
 
@@ -28,5 +28,5 @@ UI = gr.ChatInterface(
     clear_btn="Clear chat"
 )
 
-UI.queue(max_size=10, concurrency_count=8)
+UI.queue(max_size=10, concurrency_count=16)
 UI.launch()