Update app.py

app.py

@@ -29,44 +29,79 @@ h3 {
 }
 """
 
+class ConversationManager:
+    def __init__(self):
+        self.user_history = []  # For displaying to user (with markdown)
+        self.model_history = []  # For feeding back to model (with original tags)
+    
+    def add_exchange(self, user_message, assistant_response, formatted_response):
+        self.model_history.append((user_message, assistant_response))
+        self.user_history.append((user_message, formatted_response))
+    
+    def get_model_history(self):
+        return self.model_history
+    
+    def get_user_history(self):
+        return self.user_history
+
+conversation_manager = ConversationManager()
+
 device = "cuda"  # for GPU usage or "cpu" for CPU usage
 
 tokenizer = AutoTokenizer.from_pretrained(MODEL)
 model = AutoModelForCausalLM.from_pretrained(
     MODEL,
     torch_dtype=torch.bfloat16,
-    device_map="auto")
+    device_map="auto"
+)
 end_of_sentence = tokenizer.convert_tokens_to_ids("<|im_end|>")
 
+def format_response(response):
+    """Format the response for user display"""
+    if "<|end_reasoning|>" in response:
+        parts = response.split("<|end_reasoning|>")
+        reasoning = parts[0]
+        rest = parts[1] if len(parts) > 1 else ""
+        return f"<details><summary>Click to see reasoning</summary>\n\n{reasoning}\n\n</details>\n\n{rest}"
+    return response
+
 @spaces.GPU()
 def stream_chat(
-    message: str, 
+    message: str,
     history: list,
     system_prompt: str,
-    temperature: float = 0.2, 
-    max_new_tokens: int = 4096, 
-    top_p: float = 1.0, 
-    top_k: int = 1, 
+    temperature: float = 0.2,
+    max_new_tokens: int = 4096,
+    top_p: float = 1.0,
+    top_k: int = 1,
     penalty: float = 1.1,
 ):
-    print(f'message: {message}')
-    print(f'history: {history}')
-
+    model_history = conversation_manager.get_model_history()
+    
     conversation = []
-    for prompt, answer in history:
+    for prompt, answer in model_history:
         conversation.extend([
-            {"role": "user", "content": prompt}, 
+            {"role": "user", "content": prompt},
             {"role": "assistant", "content": answer},
         ])
-
+    
     conversation.append({"role": "user", "content": message})
-
-    input_ids = tokenizer.apply_chat_template(conversation, add_generation_prompt=True, return_tensors="pt").to(model.device)
     
-    streamer = TextIteratorStreamer(tokenizer, timeout=60.0, skip_prompt=True, skip_special_tokens=True)
+    input_ids = tokenizer.apply_chat_template(
+        conversation, 
+        add_generation_prompt=True, 
+        return_tensors="pt"
+    ).to(model.device)
+    
+    streamer = TextIteratorStreamer(
+        tokenizer, 
+        timeout=60.0, 
+        skip_prompt=True, 
+        skip_special_tokens=True
+    )
     
     generate_kwargs = dict(
-        input_ids=input_ids, 
+        input_ids=input_ids,
         max_new_tokens=max_new_tokens,
         do_sample=False if temperature == 0 else True,
         top_p=top_p,
@@ -76,43 +111,50 @@ def stream_chat(
         eos_token_id=[end_of_sentence],
         streamer=streamer,
     )
-
-    with torch.no_grad():
-        thread = Thread(target=model.generate, kwargs=generate_kwargs)
-        thread.start()
     
     buffer = ""
-    found_token = False
+    original_response = ""
     
-    for new_text in streamer:
-        buffer += new_text
+    with torch.no_grad():
+        thread = Thread(target=model.generate, kwargs=generate_kwargs)
+        thread.start()
         
-        if "<|end_reasoning|>" in buffer and not found_token:
-            # Split at the token
-            parts = buffer.split("<|end_reasoning|>")
-            reasoning = parts[0]
-            rest = parts[1] if len(parts) > 1 else ""
+        for new_text in streamer:
+            buffer += new_text
+            original_response += new_text
             
-            # Format with markdown and continue
-            buffer = f"<details><summary>Click to see reasoning</summary>\n\n{reasoning}\n\n</details>\n\n{rest}"
-            found_token = True
+            formatted_buffer = format_response(buffer)
             
-        yield buffer
+            if thread.is_alive() is False:
+                conversation_manager.add_exchange(
+                    message,
+                    original_response,  # Original for model
+                    formatted_buffer    # Formatted for user
+                )
+            
+            yield formatted_buffer
 
 chatbot = gr.Chatbot(height=600, placeholder=PLACEHOLDER)
 
 with gr.Blocks(css=CSS, theme="soft") as demo:
     gr.HTML(TITLE)
-    gr.DuplicateButton(value="Duplicate Space for private use", elem_classes="duplicate-button")
+    gr.DuplicateButton(
+        value="Duplicate Space for private use", 
+        elem_classes="duplicate-button"
+    )
     gr.ChatInterface(
         fn=stream_chat,
         chatbot=chatbot,
         fill_height=True,
-        additional_inputs_accordion=gr.Accordion(label="⚙️ Parameters", open=False, render=False),
+        additional_inputs_accordion=gr.Accordion(
+            label="⚙️ Parameters",
+            open=False,
+            render=False
+        ),
         additional_inputs=[
             gr.Textbox(
                 value="",
-                label="",
+                label="System Prompt",
                 render=False,
             ),
             gr.Slider(