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)
MAX_PRODUCT_DIGITS = 100

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):
    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')
    eos_token_id = tokenizer.eos_token_id

    input_ids = inputs['input_ids']
    input_len = input_ids.shape[-1]
    prediction = ""
    correct_product = ""
    valid_input = True

    try:
        num1_int = int(num1)
        num2_int = int(num2)
        correct_product = str(num1_int * num2_int)
    except ValueError:
        valid_input = False

    generated_ids = inputs['input_ids']
    past_key_values = None
    for _ in range(MAX_PRODUCT_DIGITS):  # Set a maximum limit to prevent infinite loops
        outputs = model(
            input_ids=generated_ids,
            past_key_values=past_key_values,
            use_cache=True
        )
        logits = outputs.logits
        
        next_token_id = torch.argmax(logits[:, -1, :], dim=-1)
        generated_ids = torch.cat((generated_ids, next_token_id.view(1,-1)), dim=-1)
        
        if next_token_id.item() == eos_token_id:
            break
        past_key_values = outputs.past_key_values
        
        output_text = tokenizer.decode(generated_ids[0, input_len:], skip_special_tokens=True)
        #prediction = postprocess(output_text)
        predicted_digits_reversed = output_text.strip().split(' ')
        correct_digits_reversed = ' '.join(correct_product)[::-1]
        
        # Create the diff for HighlightedText
        diff = []
        correct_digits = []
        is_correct_sofar = True
        for i in range(len(predicted_digits_reversed)):
            predicted_digit = predicted_digits_reversed[i]
            correct_digit = correct_digits_reversed[i]
            correct_digits.append((correct_digit, None))
            if i >= len(correct_digits_reversed):
                if predicted_digit == '0' and is_correct_sofar:
                    is_correct_digit = True
                else:
                    is_correct_digit = True
            else:
                if predicted_digit == correct_digit:
                    is_correct_digit = True
                else:
                    is_correct_digit = False
            if not is_correct_digit:
                is_correct_sofar = False
            if is_correct_digit:
                diff.append((correct_product[i], "-"))
            else:
                diff.append((predicted_digit, "+"))
        diff = diff[::-1]
        correct_digits = correct_digits[::-1]

        yield correct_digits, diff, ""

    #if valid_input:
    #    is_correct = prediction == correct_product
    #    result_message = "Correct!" if is_correct else f"Incorrect! The correct product is {correct_product}."
    #else:
    #    result_message = "Invalid input. Could not evaluate correctness."

    ## Final diff for the complete prediction
    #final_diff = []
    #for i in range(max(len(prediction), len(correct_product))):
    #    if i < len(prediction) and i < len(correct_product) and prediction[i] == correct_product[i]:
    #        final_diff.append((prediction[i], None))  # No highlight for correct digits
    #    elif i < len(prediction) and (i >= len(correct_product) or prediction[i] != correct_product[i]):
    #        final_diff.append((prediction[i], "+"))  # Highlight incorrect digits in red
    #    if i < len(correct_product) and (i >= len(prediction) or prediction[i] != correct_product[i]):
    #        final_diff.append((correct_product[i], "-"))  # Highlight missing/incorrect digits in green

    #yield final_diff, result_message

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='Ground Truth Product'),
        gr.HighlightedText(label='Predicted Product', combine_adjacent=False, show_legend=False, color_map={"-": "green", "+": "red"}),
        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()