app.py

import gradio as gr 
import matplotlib.pyplot as plt
import numpy as np

### CHINCHILLA PARAMS:
E = 1.62
A = 406.4
B = 410.7
alpha = 0.336
beta = 0.283

Bn = 10**9

G = ((alpha*A)/(beta*B))**(1/(alpha+beta)) 
###

def to_flops(N, D):
    return 6 * N * D

def n_opt(C):
    return G * ((C/6) ** (beta / (alpha+beta)))

def d_opt(C):
    return (1/G) * ((C/6) ** (alpha / (alpha+beta)))

def compute_kd(kn):
    frac = (A/B)*(G**(-alpha-beta))    
    kd = (1-((kn**-alpha -1)*frac))**(1/(-beta))
    return kd

def compute_overhead(kn, kd):
    return kn*kd - 1

### PRECOMPUTE CURVE:
kn_min = 0.2
kn_max = 2

kns = np.linspace(0.05, 2, 100)
overheads = []
for kn in np.linspace(0.2, 2, 100):
    kd = compute_kd(kn)
    overheads.append(compute_overhead(kn, kd)*100)

def plot_curve(kn, kd):
    fig = plt.figure()
    plt.plot(kns, overheads)
    plt.scatter([kn], [kd])
    plt.xlabel("Fraction of compute optimal model size")
    plt.ylabel("Compute overhead (%)")
    return fig


def compute(N, D):
    
    C = to_flops(N * Bn, D * Bn)
    N_opt = n_opt(C)
    D_opt = d_opt(C)

    kn = N/N_opt
    kd = compute_kd(kn)

    print(N, D, N_opt, D_opt, kn, kd)
    
    fig = plot_curve(kn, kd)

    text = f"""Compute budget (TFLOPs): {C:.2E}\n\nTraining compute overhead (%): {100*compute_overhead(kn, kd):.2f}\n\nInference cost fraction (%): {kn*100:.2f}"""
    return text, fig

with gr.Blocks() as demo:
    N = gr.Number(value=1, label="Model size (in B parameters)")
    D = gr.Number(value=100, label="Dataset size (in B tokens")
    button = gr.Button("Compute!")
    
    plot = gr.Plot(value=plt)
    md = gr.Markdown("")

    button.click(fn=compute, inputs=[N, D], outputs=[md, plot])

demo.launch()