Update app.py

made the stream = True in demo and added h11 req + changed gradio version

app.py

@@ -109,6 +109,7 @@ class ModelWrapper:
                 top_p=top_p,
                 min_p=0,
                 pad_token_id=self.tokenizer.pad_token_id
+                
             )
         output_text = self.tokenizer.decode(output_content[0], skip_special_tokens=True)
         sys_prompt_text = output_text.split("Brief explanation\n</explanation>")[0]
@@ -139,6 +140,7 @@ demo = gr.Interface(
     outputs=gr.Textbox(label="Compliance Output"),
     title="DynaGuard Compliance Checker",
     description="Paste your rules & transcript, then hit Submit."
+    streaming=True
 )
 
 if __name__ == "__main__":