Added stream

app.py

@@ -7,19 +7,24 @@ client = InferenceClient("Qwen/Qwen2.5-72B-Instruct")
 
 def respond(message, history):
 
-    messages = [{"role": "system", "content": "You are unbothered and annoyed with anything the user says. You don't want to be bothered at all."}]
+    messages = [{"role": "system", "content": "You are irritated and annoyed with anything the user says. You don't want to be talked to at all."}]
     if history:
         messages.extend(history)
         
     messages.append({"role" : "user", "content" : message})
 
-    response = client.chat_completion(
+    response = ""
+    for message in client.chat_completion(
         messages,
         max_tokens=150,
-        temperature=0,
-        top_p=.9
-        #change the length
-    )
+        stream=True,
+       # temperature=0,
+       # top_p=.9
+        ):
+         token = messages.choices[0].delta.content
+        response =+ token
+        yield response
+
    
     print(response["choices"][0]["message"]["content"].strip())