Update app.py

app.py

@@ -5,6 +5,7 @@ from threading import Thread
 
 model_id = "DavidAU/Qwen3-0.6B-heretic-abliterated-uncensored"
 
+# Load model and tokenizer
 tokenizer = AutoTokenizer.from_pretrained(model_id)
 model = AutoModelForCausalLM.from_pretrained(
     model_id, 
@@ -14,17 +15,24 @@ model = AutoModelForCausalLM.from_pretrained(
 )
 
 def stream_chat(message, history):
-    # Strict prompt to stop reasoning/thinking
+    # Strict system prompt to kill the internal monologue
     system_message = "You are a direct assistant. Answer immediately. DO NOT explain your thinking. DO NOT show internal reasoning."
     
-    # Updated History Logic for Gradio 6.0
     conversation = f"System: {system_message}\n"
+    
+    # Robust history handling: Checks if history is list of lists or list of dicts
     for msg in history:
-        role = "User" if msg['role'] == 'user' else "Assistant"
-        content = msg['content']
+        if isinstance(msg, dict):
+            role = "User" if msg.get('role') == 'user' else "Assistant"
+            content = msg.get('content', '')
+        else:
+            # Fallback for older Gradio versions (list of lists)
+            role, content = "User", msg[0]
+            conversation += f"{role}: {content}\n"
+            role, content = "Assistant", msg[1]
+            
         conversation += f"{role}: {content}\n"
     
-    # Add current message
     conversation += f"User: {message}\nAssistant: "
 
     inputs = tokenizer([conversation], return_tensors="pt").to("cpu")
@@ -35,7 +43,7 @@ def stream_chat(message, history):
         streamer=streamer,
         max_new_tokens=512,
         do_sample=True,
-        temperature=0.3, # Low temp = No wandering into 'thinking'
+        temperature=0.3, # Low temp ensures the model doesn't drift into 'thinking'
         top_p=0.9,
         repetition_penalty=1.2,
         pad_token_id=tokenizer.eos_token_id
@@ -46,19 +54,20 @@ def stream_chat(message, history):
 
     partial_text = ""
     for new_text in streamer:
-        # Stop if the model starts hallucinating typical reasoning phrases
+        # Hard stop if model tries to start its 'Okay, let me think' routine
         if any(stop in new_text for stop in ["User:", "Okay,", "I think", "First"]):
             break
         partial_text += new_text
         yield partial_text
 
+# Removed 'type="messages"' to fix the TypeError
 with gr.Blocks() as demo:
     gr.ChatInterface(
         fn=stream_chat,
-        type="messages", # Explicitly tell Gradio to use the new message format
-        title="QWEN3-0.6B DIRECT",
-        description="Fixed history unpacking error and suppressed reasoning."
+        title="QWEN3-0.6B DIRECT MAX",
+        description="Reasoning suppressed. Fast, direct, uncensored responses."
     )
 
 if __name__ == "__main__":
+    # Passing theme here as per Gradio 6.0 logs
     demo.launch(server_name="0.0.0.0", server_port=7860, theme="soft")