Update app.py

app.py

@@ -98,6 +98,17 @@ def run_huggingface_model(model, messages, max_tokens, temperature, top_p):
             result = response.content
         return result
 
+def run_huggingface_model_alt(model, messages, max_tokens, temperature, top_p):
+    client = InferenceClient(model)
+    response = client.chat_completion(
+        messages,
+        max_tokens=max_tokens,
+        stream=False,
+        temperature=temperature,
+        top_p=top_p,
+    )
+    return response.choices[0].message.content
+
 def run_together_model(model, messages, max_tokens, temperature, top_p):
     client = Together()
     response = client.chat.completions.create(
@@ -193,7 +204,7 @@ with demo:
     )
 
     huggingface_run_button.click(
-        run_huggingface_model,
+        run_huggingface_model_alt,
         inputs=[huggingface_model_input, messages_state, max_tokens_slider, temperature_slider, top_p_slider],
         outputs=[response_output],
     )