import gradio as gr
import spaces
from transformers import AutoModelForCausalLM, AutoTokenizer

# --- Model Loading ---
model_name = "gitglubber/Qwen3-IWM"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# --- System Message ---
# Define the persona or instructions for the model
system_message = """"You are an expert at the terminal. If asked to perform a task - decorate the command with @command. Explaining why you would perform that task to complete the function. Then if there are follow up commands use @command2, etc. Be helpful and willing to correction."""

# --- Generation Function ---
@spaces.GPU(duration=120)
def generate_response(chat_history):
    # Prepare the model input from the chat history
    # The system message is the first entry
    messages = [{"role": "system", "content": system_message}]
    
    # Add previous user/assistant messages
    for user_msg, assistant_msg in chat_history:
        messages.append({"role": "user", "content": user_msg})
        messages.append({"role": "assistant", "content": assistant_msg})

    # Apply the chat template
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
    )
    model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

    # Generate text
    generated_ids = model.generate(
        **model_inputs,
        max_new_tokens=8192
    )
    output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
    
    content = tokenizer.decode(output_ids, skip_special_tokens=True)
    return content

# --- Gradio Interface ---
with gr.Blocks(fill_height=True) as demo:
    gr.Markdown("# IWM Chat Bot")
    # We use a state object to store the system message, though it's constant here
    chatbot = gr.Chatbot(scale=1)
    msg = gr.Textbox(label="Input", scale=0)
    clear = gr.Button("Clear")

    def respond(message, chat_history):
        if not message.strip(): # Check for empty or whitespace-only messages
            return "", chat_history
        
        # Append the new user message to the history
        chat_history.append((message, None))
        
        # Prepare history for the model (without the last empty spot)
        model_input_history = chat_history[:-1]
        model_input_history.append((message, None)) # Add current message for context
        
        # Flatten the history for the model function
        flat_history = []
        for user, assistant in chat_history:
            if user: flat_history.append((user, assistant))
            
        bot_response = generate_response(flat_history)
        
        # Update the last entry in chat_history with the bot's response
        chat_history[-1] = (message, bot_response)
        
        return "", chat_history

    msg.submit(respond, [msg, chatbot], [msg, chatbot])
    clear.click(lambda: None, None, chatbot, queue=False)

# Launch the app
demo.launch()