app.py

import gradio as gr
import torch
import re
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
from duckduckgo_search import DDGS
from threading import Thread

# --- MODEL SETUP ---
MODEL_ID = "Qwen/Qwen3-0.6B" # Official HF Repo
print("Loading model and tokenizer...")

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True
)

# --- SEARCH FUNCTION ---
def web_search(query):
    try:
        with DDGS() as ddgs:
            results = list(ddgs.text(query, max_results=3))
            if not results: return ""
            blob = "\n\nSearch Results:\n"
            for r in results:
                blob += f"- {r['title']}: {r['body']}\n"
            return blob
    except:
        return ""

# --- UI HELPERS ---
CSS = """
.thought-box {
    background-color: rgba(255, 255, 255, 0.05);
    border-left: 4px solid #facc15;
    padding: 10px;
    margin: 10px 0;
    font-style: italic;
    color: #9ca3af;
}
details summary {
    cursor: pointer;
    color: #facc15;
    font-weight: bold;
}
"""

def parse_output(text):
    """Parses <think> tags into a clean UI format."""
    if "<think>" in text:
        parts = text.split("</think>")
        if len(parts) > 1:
            # Finished thinking
            thought = parts[0].replace("<think>", "").strip()
            answer = parts[1].strip()
            return f"<details open><summary>💭 Thought Process</summary><div class='thought-box'>{thought}</div></details>\n\n{answer}"
        else:
            # Still thinking
            thought = parts[0].replace("<think>", "").strip()
            return f"<details open><summary>🌀 Thinking...</summary><div class='thought-box'>{thought}</div></details>"
    return text

# --- GENERATION LOGIC ---
def chat(message, history, search_enabled, temperature, max_tokens):
    # 1. Handle Web Search
    search_context = ""
    if search_enabled:
        search_context = web_search(message)
    
    # 2. Build properly formatted prompt (Fixes AI talking to itself)
    # We use the official ChatML template
    conversation = []
    for user_msg, assistant_msg in history:
        conversation.append({"role": "user", "content": user_msg})
        if assistant_msg:
            # Remove UI formatting before feeding back to model
            clean_assistant = re.sub(r'<details.*?</details>', '', assistant_msg, flags=re.DOTALL).strip()
            conversation.append({"role": "assistant", "content": clean_assistant})
    
    user_content = message + search_context
    conversation.append({"role": "user", "content": user_content})
    
    input_ids = tokenizer.apply_chat_template(
        conversation, 
        add_generation_prompt=True, 
        return_tensors="pt"
    ).to(model.device)

    # 3. Streamer with stop criteria
    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
    
    generate_kwargs = dict(
        input_ids=input_ids,
        streamer=streamer,
        max_new_tokens=max_tokens,
        temperature=temperature,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id,
        # Stop generating once the model tries to start a new 'User' turn
        eos_token_id=[tokenizer.eos_token_id, tokenizer.convert_tokens_to_ids("<|im_end|>")]
    )

    thread = Thread(target=model.generate, kwargs=generate_kwargs)
    thread.start()

    buffer = ""
    for new_text in streamer:
        # Crucial Fix: If the model generates "User:" or "<|im_start|>", stop displaying
        if "User:" in new_text or "<|im_start|>" in new_text:
            break
            
        buffer += new_text
        yield parse_output(buffer)

# --- GRADIO UI ---
with gr.Blocks(css=CSS, theme=gr.themes.Soft()) as demo:
    gr.HTML("<h1>🧠 Qwen3 Reasoning Lab</h1>")
    
    with gr.Row():
        with gr.Column(scale=4):
            chat_box = gr.Chatbot(height=600, label="Qwen3-0.6B")
            msg_input = gr.Textbox(placeholder="Ask a logic question...", show_label=False)
        
        with gr.Column(scale=1):
            search_toggle = gr.Checkbox(label="🌐 Web Search (DDG)", value=False)
            temp_slider = gr.Slider(0.1, 1.0, 0.7, label="Temperature")
            token_slider = gr.Slider(512, 4096, 1024, label="Max Tokens")
            gr.Markdown("""
            ### Tips:
            - **Thinking:** This model is trained for Chain-of-Thought.
            - **Self-Talk Fix:** We use stop sequences to prevent the AI from acting as 'User'.
            """)
            clear_btn = gr.Button("🗑 Clear Chat")

    # Set up logic
    chat_event = msg_input.submit(
        lambda x, y: (x, y + [[x, None]]), 
        [msg_input, chat_box], 
        [msg_input, chat_box], 
        queue=False
    ).then(
        chat, 
        [msg_input, chat_box, search_toggle, temp_slider, token_slider], 
        chat_box
    )
    
    clear_btn.click(lambda: None, None, chat_box, queue=False)

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860)