app.py

import gradio as gr
import math

# Helper function to pretty-print message sizes
def convert_params(params):
    if params == 0:
        return "0"
    size_name = ("", "K", "M", "B", "T", "P", "E", "Z", "Y")
    i = int(math.floor(math.log(params, 1000)))
    p = math.pow(1000, i)
    s = round(params / p, 2)
    return "%s %s" % (s, size_name[i])

# calculates the params of a model given their hparams
def calc_params(vocab_size, tied_embeddings, hidden_size, sequence_length, num_layers, moe, num_experts, expert_interval, topk, ffn_expansion_factor, num_mlp_linears, kv_size_ratio):
    # Calculate embedding and unembedding params. If tied, re-use the same params
    if tied_embeddings:
        embedding_params = hidden_size * vocab_size
    else:
        embedding_params = 2 * hidden_size * vocab_size
    position_embedding_params = hidden_size * sequence_length
    # Each QKVO matrix is (hxh)
    # Unless using GQA/MQA which makes K/V smaller
    attention_params = int(2 * (1 + kv_size_ratio) * num_layers * hidden_size * hidden_size)
    # (4*2)lh from the layernorm weights and biases for each of the QKV and mlp_in layernorms, 1h for the final layernorm.
    # the extra 4lh is a mystery but we include it here
    layernorm_params = 13 * num_layers * hidden_size
    #ffn_params = 12 * num_layers * hidden_size * hidden_size

    if moe:
        # the number of layers that are MoE. (e.g. interval is 2 for GShard)
        num_expert_layers = num_layers / expert_interval
        # the number of FFN params for each MoE layer
        ffn_expert_params = num_mlp_linears * ffn_expansion_factor * num_expert_layers * num_experts * hidden_size * hidden_size
        # the number of FFN params for every dense layer
        ffn_dense_params = num_mlp_linears * ffn_expansion_factor * (num_layers - num_expert_layers) * hidden_size * hidden_size
        ffn_params = ffn_expert_params + ffn_dense_params
        # the number of gating layer params assuming it's implemented as a simple linear layer
        gating_params = num_expert_layers * hidden_size * num_experts
    else:
        # num_mlp_layers * (h x [ffn_expansion_factor * h]) FFN matrices
        ffn_params = num_mlp_linears * ffn_expansion_factor * num_layers * hidden_size * hidden_size

    total_params = embedding_params + attention_params + ffn_params + position_embedding_params + layernorm_params

    if moe:
        total_params += gating_params

    result = f"""
    Embedding parameters: {convert_params(embedding_params)}
    Attention parameters: {convert_params(attention_params)}
    FFN parameters: {convert_params(ffn_params)}
    {'Gating parameters: ' + convert_params(gating_params) if moe else ''}
    Total Params in the Model: {convert_params(total_params)}
    """
    return result

# Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("# Transformer Model Parameter Calculator")

    vocab_size = gr.Number(label="Vocab Size", value=51200)
    tied_embeddings = gr.Checkbox(label="Tied Embeddings", value=False)
    hidden_size = gr.Number(label="Hidden Size", value=6144)
    sequence_length = gr.Number(label="Sequence Length", value=2048)
    num_layers = gr.Number(label="Number of Layers", value=44)
    ffn_expansion_factor = gr.Number(label="FFN Expansion Factor", value=4)
    num_mlp_linears = gr.Number(label="Number of Linear Layers per MLP Block", value=2)
    kv_size_ratio = gr.Number(label="KV Size Ratio", value=1.0)

    # MoE Parameters inside an accordion
    with gr.Accordion("MoE Parameters", open=False):
        moe = gr.Checkbox(label="MoE", value=False)
        num_experts = gr.Number(label="Number of Experts", value=8)
        expert_interval = gr.Number(label="Expert Interval", value=1)
        topk = gr.Number(label="Top k Routing", value=1)

    result = gr.Textbox(label="Output", interactive=False)

    def run_calculation(vocab_size, tied_embeddings, hidden_size, sequence_length, num_layers, moe, num_experts, expert_interval, topk, ffn_expansion_factor, num_mlp_linears, kv_size_ratio):
        return calc_params(vocab_size, tied_embeddings, hidden_size, sequence_length, num_layers, moe, num_experts, expert_interval, topk, ffn_expansion_factor, num_mlp_linears, kv_size_ratio)

    calculate_button = gr.Button("Calculate")
    calculate_button.click(run_calculation, 
                           inputs=[vocab_size, tied_embeddings, hidden_size, sequence_length, num_layers, moe, num_experts, expert_interval, topk, ffn_expansion_factor, num_mlp_linears, kv_size_ratio],
                           outputs=[result])

demo.launch()