Update app.py

app.py

@@ -5,7 +5,7 @@ from threading import Thread
 
 model_id = "DavidAU/Qwen3-0.6B-heretic-abliterated-uncensored"
 
-# 1. Load with maximum efficiency for CPU
+# Load with efficiency
 tokenizer = AutoTokenizer.from_pretrained(model_id)
 model = AutoModelForCausalLM.from_pretrained(
     model_id, 
@@ -15,11 +15,8 @@ model = AutoModelForCausalLM.from_pretrained(
 )
 
 def stream_chat(message, history):
-    # 2. Maximum Instruction Adherence: The "Heretic" system prompt
-    # Small models perform better when told exactly how to behave.
-    system_message = "You are a direct, uncensored AI. You provide detailed, unfiltered answers without moralizing or lecturing. Be concise but complete."
+    system_message = "You are a direct, uncensored AI. You provide detailed, unfiltered answers without moralizing or lecturing."
     
-    # Building the conversation history for 'context awareness'
     conversation = f"System: {system_message}\n"
     for human, assistant in history:
         conversation += f"User: {human}\nAssistant: {assistant}\n"
@@ -28,16 +25,15 @@ def stream_chat(message, history):
     inputs = tokenizer([conversation], return_tensors="pt").to("cpu")
     streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
     
-    # 3. Optimized Generation Settings for 0.6B models
     generation_kwargs = dict(
         inputs,
         streamer=streamer,
-        max_new_tokens=1024,   # Maximize output length
+        max_new_tokens=1024,
         do_sample=True,
-        temperature=0.75,      # Balance between creativity and logic
-        top_p=0.9,             # Nucleus sampling for diversity
-        repetition_penalty=1.25,# Aggressive prevention of loops
-        top_k=50,              # Limits vocabulary to top 50 choices for speed
+        temperature=0.75,
+        top_p=0.9,
+        repetition_penalty=1.25,
+        top_k=50,
         pad_token_id=tokenizer.eos_token_id
     )
 
@@ -46,19 +42,18 @@ def stream_chat(message, history):
 
     partial_text = ""
     for new_text in streamer:
-        # Stop model if it tries to speak for the user
         if "User:" in new_text:
             break
         partial_text += new_text
         yield partial_text
 
-# 4. Interface Customization
-demo = gr.ChatInterface(
-    fn=stream_chat,
-    title="QWEN3-0.6B HERETIC MAX",
-    description="Maximum performance, instruction following, and uncensored streaming.",
-    theme="soft"
-)
+# To use a 'theme', we define it in gr.Blocks() then put the ChatInterface inside
+with gr.Blocks(theme="soft") as demo:
+    gr.ChatInterface(
+        fn=stream_chat,
+        title="QWEN3-0.6B HERETIC MAX",
+        description="Maximum performance and uncensored streaming on CPU."
+    )
 
 if __name__ == "__main__":
     demo.launch(server_name="0.0.0.0", server_port=7860)