Update app.py

app.py

@@ -28,7 +28,7 @@ prompt_template="""<|begin_of_text|>Dibawah ini adalah percakapan antara dokter
 
 def output_inference(tanya, history):
     prompt = prompt_template.format(ask=tanya)
-
+    
     output = llm(
       prompt,
       stop=["<|end_of_text|>"],
@@ -39,9 +39,13 @@ def output_inference(tanya, history):
       min_p=0.05,
       typical_p=1.0,
       stream=False,
-
+    
     )
-    return output['choices'][0]['text']
+    inference = output['choices'][0]['text']
+    
+    for i in range(len(inference)):
+        yield inference[: i+1]
+    # return output['choices'][0]['text']
 
 
 gr.ChatInterface(