uhhhh

app.py

@@ -135,7 +135,7 @@ def export_chat(history, system_prompt):
 
 def sanitize_chatbot_history(history):
     """Ensure each entry in the chatbot history is a tuple of two items."""
-    return [tuple(entry[:2]) if isinstance(entry, (list, tuple)) else (str(entry), None) for entry in history]
+    return [tuple(entry[:2]) for entry in history]
 
 with gr.Blocks(theme='gradio/monochrome') as demo:
     with gr.Row():
@@ -174,21 +174,17 @@ with gr.Blocks(theme='gradio/monochrome') as demo:
             return
         user_message = history[-1][0]
         bot_message = predict(user_message, history[:-1], system_prompt, temperature, top_p, top_k, frequency_penalty, presence_penalty, repetition_penalty, max_tokens)
-        history[-1] = (history[-1][0], "")
+        history[-1] = (history[-1][0], "")  # Ensure it's a tuple
         task_id = id(asyncio.current_task())
         active_tasks[task_id] = asyncio.current_task()
         try:
             async for chunk in bot_message:
                 if task_id not in active_tasks:
                     break
-                history[-1] = (history[-1][0], chunk)
+                history[-1] = (history[-1][0], chunk)  # Update as a tuple
                 yield history
         except asyncio.CancelledError:
-            print("Bot generation cancelled")
-        except GeneratorExit:
-            print("Generator exited")
-        except Exception as e:
-            print(f"Error in bot generation: {e}")
+            pass
         finally:
             if task_id in active_tasks:
                 del active_tasks[task_id]
@@ -197,16 +193,18 @@ with gr.Blocks(theme='gradio/monochrome') as demo:
             yield history
 
     async def regenerate_response(history, system_prompt, temperature, top_p, top_k, frequency_penalty, presence_penalty, repetition_penalty, max_tokens):
+        # Cancel any ongoing generation
         for task in list(active_tasks.values()):
             task.cancel()
         
+        # Wait for a short time to ensure cancellation is processed
         await asyncio.sleep(0.1)
         
         history = sanitize_chatbot_history(history or [])
         if history:
-            history[-1] = (history[-1][0], None)
+            history[-1] = (history[-1][0], None)  # Reset last response
             async for new_history in bot(history, system_prompt, temperature, top_p, top_k, frequency_penalty, presence_penalty, repetition_penalty, max_tokens):
-                yield sanitize_chatbot_history(new_history)
+                yield new_history
         else:
             yield []
 
@@ -216,7 +214,7 @@ with gr.Blocks(theme='gradio/monochrome') as demo:
 
     submit_event = msg.submit(user, [msg, chatbot], [msg, chatbot], queue=False).then(
         bot, [chatbot, system_prompt, temperature, top_p, top_k, frequency_penalty, presence_penalty, repetition_penalty, max_tokens], chatbot,
-        concurrency_limit=10
+        concurrency_limit=5
     )
 
     clear.click(lambda: [], None, chatbot, queue=False)
@@ -225,16 +223,16 @@ with gr.Blocks(theme='gradio/monochrome') as demo:
         regenerate_response,
         [chatbot, system_prompt, temperature, top_p, top_k, frequency_penalty, presence_penalty, repetition_penalty, max_tokens],
         chatbot,
-        concurrency_limit=10
+        concurrency_limit=5
     )
 
-    import_button.click(import_chat_wrapper, inputs=[import_textbox], outputs=[chatbot, system_prompt], concurrency_limit=10)
+    import_button.click(import_chat_wrapper, inputs=[import_textbox], outputs=[chatbot, system_prompt], concurrency_limit=5)
 
     export_button.click(
         export_chat,
         inputs=[chatbot, system_prompt],
         outputs=[import_textbox],
-        concurrency_limit=10
+        concurrency_limit=5
     )
 
     stop_btn.click(
@@ -246,4 +244,4 @@ with gr.Blocks(theme='gradio/monochrome') as demo:
     )
 
 if __name__ == "__main__":
-    demo.launch(debug=True, max_threads=40)
+    demo.launch(debug=True, max_threads=20)