Update app.py

app.py

@@ -13,7 +13,6 @@ pipe = pipeline(
     torch_dtype="auto",
     device_map="auto",
 )
-
 def format_conversation_history(chat_history):
     messages = []
     for item in chat_history:
@@ -23,14 +22,13 @@ def format_conversation_history(chat_history):
             content = content[0]["text"] if content and "text" in content[0] else str(content)
         messages.append({"role": role, "content": content})
     return messages
-
-@spaces.GPU(duration=60)
+    
+@spaces.GPU()
 def generate_response(input_data, chat_history, max_new_tokens, system_prompt, temperature, top_p, top_k, repetition_penalty):
     new_message = {"role": "user", "content": input_data}
     system_message = [{"role": "system", "content": system_prompt}] if system_prompt else []
     processed_history = format_conversation_history(chat_history)
     messages = system_message + processed_history + [new_message]
-    
     streamer = TextIteratorStreamer(pipe.tokenizer, skip_prompt=True, skip_special_tokens=True)
     generation_kwargs = {
         "max_new_tokens": max_new_tokens,
@@ -43,24 +41,25 @@ def generate_response(input_data, chat_history, max_new_tokens, system_prompt, t
     }
     thread = Thread(target=pipe, args=(messages,), kwargs=generation_kwargs)
     thread.start()
-    #streaming try #1
-    buffer = ""
-    full_response = ""
+    # simple formatting without harmony because of no tool usage etc. and experienced hf space problems with harmony 
+    thinking = ""
+    final = ""
+    started_final = False
     for chunk in streamer:
-        buffer += chunk
-        parts = re.split(r'(\s+)', buffer)
-        if re.match(r'\s+', parts[-1]) is not None:
-            to_append = ''.join(parts)
-            buffer = ""
+        if not started_final:
+            if "assistantfinal" in chunk.lower():
+                split_parts = re.split(r'assistantfinal', chunk, maxsplit=1)
+                thinking += split_parts[0]
+                final += split_parts[1]
+                started_final = True
+            else:
+                thinking += chunk
         else:
-            to_append = ''.join(parts[:-1])
-            buffer = parts[-1]
-        if to_append:
-            full_response += to_append
-            yield full_response
-    if buffer:
-        full_response += buffer
-        yield full_response
+            final += chunk
+        clean_thinking = re.sub(r'^analysis\s*', '', thinking).strip()
+        clean_final = final.strip()
+        formatted = f"<details open><summary>Click to view Thinking Process</summary>\n\n{clean_thinking}\n\n</details>\n\n{clean_final}"
+        yield formatted
 
 demo = gr.ChatInterface(
     fn=generate_response,
@@ -84,10 +83,8 @@ demo = gr.ChatInterface(
     ],
     cache_examples=False,
     type="messages",
-    description="""
-# gpt-oss-20b
-Wait couple of seconds initially. You can adjust reasoning level in the system prompt like "Reasoning: high.
-    """,
+    description="""# gpt-oss-20b
+Give it a couple of seconds to start. You can adjust reasoning level in the system prompt like "Reasoning: high. Click to view thinking process. Default is on""",
     fill_height=True,
     textbox=gr.Textbox(
         label="Query Input",