app.py

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

model_name = 'yuntian-deng/gpt2-implicit-cot-multiplication'
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

def preprocess(num):
    num = str(num).strip().replace(' ', '')
    reversed_num = ' '.join(num[::-1])
    return reversed_num

def postprocess(raw_output):
    prediction = raw_output.replace(' ', '')[::-1]
    return prediction

@spaces.GPU
def predict_product(num1, num2):
    # Reverse input digits and add spaces
    input_text = f'{preprocess(num1)} * {preprocess(num2)} ='

    inputs = tokenizer(input_text, return_tensors='pt').to('cuda' if torch.cuda.is_available() else 'cpu')
    model.to('cuda' if torch.cuda.is_available() else 'cpu')

    # Generate output
    outputs = model.generate(**inputs, max_new_tokens=40)

    output = outputs[0][inputs['input_ids'].shape[-1]:]
    raw_output = tokenizer.decode(output, skip_special_tokens=True)
    prediction = postprocess(raw_output)

    # Evalaute the correctness of the result
    try:
        num1_int = int(num1)
        num2_int = int(num2)
        valid_input = True
    except ValueError:
        valid_input = False
    if valid_input:
        correct_product = num1_int * num2_int
        try:
            prediction_int = int(prediction)
            is_correct = (prediction_int == correct_product)
        except ValueError:
            is_correct = False
        result_color = "green" if is_correct else "red"
        result_message = "Correct!" if is_correct else f"Incorrect! The correct product is {correct_product}."
    else:
        result_color = "black"
        result_message = "Invalid input. Could not evaluate correctness."
    result_html = f"<div style='color: {result_color};'>{result_message}</div>"

    return input_text, raw_output, prediction, result_html

demo = gr.Interface(
    fn=predict_product,
    inputs=[
        gr.Textbox(label='First Number (up to 12 digits)', value='12345'),
        gr.Textbox(label='Second Number (up to 12 digits)', value='67890'),
    ],
    outputs=[
        gr.Textbox(label='Raw Input to GPT2 (reversed digits and added spaces)'),
        gr.Textbox(label='Raw Output from GPT2 (reversed digits and with spaces)'),
        gr.Textbox(label='Predicted Product'),
        gr.HTML(label='Result Message')
    ],
    title='GPT2 Direct Multiplication Calculator (Without Using Chain-of-Thought)',
    description='This demo uses GPT2 to directly predict the product of two numbers without using any intermediate reasoning steps. The GPT2 model has been fine-tuned to internalize chain-of-thought reasoning within its hidden states, following our stepwise internalization approach detailed in the paper linked at the bottom of this page.',
    article="""
    - [Paper: From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step](https://arxiv.org/pdf/2405.14838)
    - [Code Repository](https://github.com/da03/Internalize_CoT_Step_by_Step)
    - [Tweet Announcement](https://twitter.com/yuntiandeng/status/1795854740879774036)
    """,
    clear_btn=None,
    submit_btn="Multiply!",
    live=False
)

demo.launch()