app.py

import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

# Model and tokenizer loading (with error handling)
try:
    model_name = "google/gemma-3-1b-it"  # Correct model name
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        torch_dtype=torch.bfloat16,  # Use bfloat16 for efficiency, if supported
        device_map="auto",  # Automatically use GPU if available, otherwise CPU
    )
    # Create the pipeline
    pipe = pipeline(
        "text-generation",
        model=model,
        tokenizer=tokenizer,
        torch_dtype=torch.bfloat16, # Make sure pipeline also uses correct dtype
        device_map="auto", # and device mapping
        model_kwargs={"attn_implementation": "flash_attention_2"}  # Enable Flash Attention 2 if supported by your hardware and transformers version
    )

except Exception as e:
    error_message = f"Error loading model or tokenizer: {e}"
    print(error_message)  # Log the error to the console
    #  Provide a fallback, even if it's just displaying the error.
    def error_response(message, history):
        return f"Model loading failed.  Error: {error_message}"
    
    # Minimal Gradio interface to show the error
    with gr.Blocks() as demo:
         gr.ChatInterface(error_response)
    demo.launch()
    exit() # Important: exit to prevent running the rest of the (broken) code


# Chat template handling (important for correct prompting)
def apply_chat_template(messages, chat_template=None):
    """Applies the chat template to the message history.

    Args:
        messages: A list of dictionaries, where each dictionary has a "role"
            ("user" or "assistant") and "content" key.
        chat_template:  The chat template string (optional). If None,
        try to get from tokenizer.

    Returns:
        A single string representing the formatted conversation.
    """
    if chat_template is None:
        if hasattr(tokenizer, "chat_template") and tokenizer.chat_template:
            chat_template = tokenizer.chat_template
        else:
            # Fallback to a simple template if no chat template is found.  This is
            # *critical* to prevent the model from generating nonsensical output.
            chat_template = "{% for message in messages %}" \
                            "{{ '<start_of_turn>' + message['role'] + '\n' + message['content'] + '<end_of_turn>\n' }}" \
                            "{% endfor %}" \
                            "{% if add_generation_prompt %}{{ '<start_of_turn>model\n' }}{% endif %}"

    return tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True, chat_template=chat_template
    )

# Prediction function (modified for chat)
def predict(message, history):
    """Generates a response to the user's message.

    Args:
        message: The user's input message (string).
        history: A list of (user_message, bot_response) tuples representing
            the conversation history.

    Returns:
        The generated bot response (string).
    """
    # Build the conversation history in the required format.
    messages = []
    for user_msg, bot_response in history:
        messages.append({"role": "user", "content": user_msg})
        messages.append({"role": "model", "content": bot_response})
    messages.append({"role": "user", "content": message})

    # Apply the chat template.
    prompt = apply_chat_template(messages)

    # Generate the response using the pipeline (much cleaner).
    try:
        sequences = pipe(
            prompt,
            max_new_tokens=512,   # Limit response length
            do_sample=True,       # Use sampling for more diverse responses
            temperature=0.7,      # Control randomness (higher = more random)
            top_k=50,             # Top-k sampling
            top_p=0.95,            # Nucleus sampling
            repetition_penalty=1.2, # Reduce repetition
            pad_token_id=tokenizer.eos_token_id,  # Ensure padding is correct.

        )
        response = sequences[0]['generated_text'][len(prompt):].strip() # Extract *only* generated text
        return response

    except Exception as e:
        return f"An error occurred during generation: {e}"


# Gradio interface (using gr.ChatInterface for a chatbot UI)
with gr.Blocks() as demo:
    gr.ChatInterface(
        predict,
        chatbot=gr.Chatbot(height=500),  # Set a reasonable height
        textbox=gr.Textbox(placeholder="Ask me anything!", container=False, scale=7),
        title="Gemma-3-1b-it Chatbot",
        description="Chat with the Gemma-3-1b-it model.",
        retry_btn=None,   # Remove redundant buttons
        undo_btn=None,
        clear_btn=None,
    )

demo.launch(share=False) # Set share=True to create a publicly shareable link