Update app.py

app.py

@@ -1,69 +1,49 @@
 import gradio as gr
-import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
-from threading import Thread
-
-model_id = "DavidAU/Qwen3-0.6B-heretic-abliterated-uncensored"
-
-tokenizer = AutoTokenizer.from_pretrained(model_id)
-model = AutoModelForCausalLM.from_pretrained(
-    model_id, 
-    dtype=torch.float32, 
-    device_map="cpu",
-    low_cpu_mem_usage=True
+from llama_cpp import Llama
+
+# Initialize the model
+# n_ctx is the context window; 2048 is plenty for fast chat
+llm = Llama(
+    model_path="llama3.2-1b-Uncensored-Q4_K_M.gguf",
+    n_ctx=2048,
+    n_threads=4  # Optimized for Hugging Face Free Tier CPU
 )
 
 def stream_chat(message, history):
-    # This system prompt is designed to override the 'quiz' persona
-    system_message = "You are a direct chat interface. Provide ONLY the spoken response. NEVER explain your logic. NEVER use 'The user says'. NEVER show multiple choice answers. Respond in English only."
-    
-    conversation = f"System: {system_message}\n"
+    # Llama 3 Prompt Format - Strict Directives
+    prompt = f"<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n" \
+             f"You are a direct AI. Respond immediately. No thinking. No internal monologue.<|eot_id|>"
     
     for msg in history:
-        role = "User" if msg['role'] == 'user' else "Assistant"
-        conversation += f"{role}: {msg['content']}\n"
+        role = msg['role']
+        content = msg['content']
+        prompt += f"<|start_header_id|>{role}<|end_header_id|>\n\n{content}<|eot_id|>"
     
-    conversation += f"User: {message}\nAssistant:"
-
-    inputs = tokenizer([conversation], return_tensors="pt").to("cpu")
-    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
-    
-    generation_kwargs = dict(
-        inputs,
-        streamer=streamer,
-        max_new_tokens=128,    # Shortened to force directness
-        do_sample=False,       # Greedy search: disables the 'creativity' that leads to thinking
-        repetition_penalty=1.5, # Very high to prevent looping phrases
-        pad_token_id=tokenizer.eos_token_id
+    prompt += f"<|start_header_id|>user<|end_header_id|>\n\n{message}<|eot_id|>" \
+              f"<|start_header_id|>assistant<|end_header_id|>\n\n"
+
+    # Streaming generation
+    stream = llm(
+        prompt,
+        max_tokens=512,
+        stop=["<|eot_id|>", "User:", "Assistant:"],
+        stream=True,
+        temperature=0,  # GREEDY: Absolute directness, no 'wandering' thoughts
+        repeat_penalty=1.2
     )
 
-    thread = Thread(target=model.generate, kwargs=generation_kwargs)
-    thread.start()
-
     partial_text = ""
-    # These are 'Hard Stop' keywords that trigger the moment the model starts thinking
-    stop_triggers = ["The user says", "Answer:", "Option", "A)", "B)", "Okay,", "In this case"]
-    
-    for new_text in streamer:
-        partial_text += new_text
-        
-        # Check if the model is drifting into reasoning
-        if any(trigger in partial_text for trigger in stop_triggers):
-            # Clean up the output to remove the trigger word itself
-            for trigger in stop_triggers:
-                if trigger in partial_text:
-                    partial_text = partial_text.split(trigger)[0]
-            yield partial_text.strip()
-            break
-            
-        yield partial_text.strip()
+    for output in stream:
+        token = output["choices"][0]["text"]
+        partial_text += token
+        yield partial_text
 
 with gr.Blocks() as demo:
     gr.ChatInterface(
         fn=stream_chat,
         type="messages",
-        title="QWEN3-0.6B ZERO-THINKING",
-        description="Greedy search and hard-coded stop triggers to ensure direct output."
+        title="LLAMA-3.2-1B UNCENSORED (GGUF)",
+        description="Running on Llama.cpp for maximum CPU speed. No-thinking mode active."
     )
 
 if __name__ == "__main__":