'''
 Description  : 
 Version      : 1.0
 Author       : Chaofan Tao
 Mail         : tcftrees@gmail.com
 Github       : https://github.com/sail-sg/scaling-with-vocab
 Date         : 2024-08-09 00:25
 Copyright (C) 2024 Chaofan Tao. All rights reserved.
'''
import gradio as gr
from utils import approach1_isoflops, approach2_derivative, approach3_isoloss


def compute_optimal_vocab(Nnv, flops):
    try:
        Nnv = float(eval(Nnv))
    except ValueError:
        return "Invalid input for Non-vocabulary Parameters."

    if flops:
        try:
            flops = float(eval(flops))
        except ValueError:
            return "Invalid input for FLOPs."

    if flops is None or flops == "":
        Vopt_app1 = approach1_isoflops(Nnv)
        Vopt_app2 = approach2_derivative(Nnv)
        Vopt_app3 = approach3_isoloss(Nnv)
    else:
        Vopt_app1, Vopt_app2 = None, None
        Vopt_app3 = approach3_isoloss(Nnv, flops)

    results = f"The optimal vocabulary size is:\nApproach 1: {Vopt_app1}\nApproach 2: {Vopt_app2}\nApproach 3: {Vopt_app3}"
    return results

with gr.Blocks() as demo:
    with gr.Column():
        gr.Markdown(
            """<h1>The Optimal Vocabulary Size Predictor</h1>

            This repo is the offical demo space for [Scaling Laws with Vocabulary: Larger Models Deserve Larger Vocabularies](https://huggingface.co/papers/2407.13623). In summary, we show that when scaling up model size, increase vocabulary size too, but at a slower rate than other parameters.
            
            <img src="https://huggingface.co/spaces/sail/scaling-with-vocab-demo/resolve/main/figures/vocabulary_demo.png" style="max-width:720px; display:block; margin-left:auto; margin-right:auto;">

            This tool is used to predict the optimal vocabulary size given the non-vocabulary parameters. We provide 3 ways for prediction:
            
            - **Approach 1: Build the relationship between studied attributes and FLOPs**: Build the relationship between the optimal data points (the points that reach the lowest loss under the same FLOPs budget) and the FLOPs.
            - **Approach 2: Derivative-Based Estimation**: Fast calculation method using the derivative of FLOPs with respect to the vocabulary size.
            - **Approach 3: Parametric Fit of Loss Formula**: Design a loss formula that considers the effect of vocabulary size and utilizes the loss to make prediction.
            
            Approach 1 and 2 can only be used to compute the optimal vocabulary size when the compute is optimally allocated to non-vocabulary parameters, vocabulary parameters and data jointly. Approach 3 will not only consider the case above, but also consider the case when the amount of data does not satisfy the optimal compute allocation, and can calculate the optimal vocabulary size with specified FLOPs.
            
            **Thanks for trying** 🌟🌟🌟!
            """)
        

        with gr.Row():
            Nnv = gr.Textbox(label="Non-vocabulary Parameters (1e9=1B)", placeholder="Required (e.g. 3.0e9)")
            flops = gr.Textbox(label="FLOPs", placeholder="Optional (e.g. 7.05e21)")
            output_text = gr.Textbox(label="Prediction")
        with gr.Row():
            btn = gr.Button("Press it to compute the optimal vocabulary size")

        btn.click(
            compute_optimal_vocab,
            inputs=[Nnv, flops],
            outputs=output_text
        )
demo.launch()