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])

# ---- Transformer Parameter Calculation ---- #
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):
    if tied_embeddings:
        embedding_params = hidden_size * vocab_size
    else:
        embedding_params = 2 * hidden_size * vocab_size
    position_embedding_params = hidden_size * sequence_length
    attention_params = int(2 * (1 + kv_size_ratio) * num_layers * hidden_size * hidden_size)
    layernorm_params = 13 * num_layers * hidden_size

    if moe:
        num_expert_layers = num_layers / expert_interval
        ffn_expert_params = num_mlp_linears * ffn_expansion_factor * num_expert_layers * num_experts * hidden_size * hidden_size
        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
        gating_params = num_expert_layers * hidden_size * num_experts
    else:
        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

# ---- Memory Calculation Code (from the second script) ---- #
def calc_mem(args):
    dp_degree = args.num_gpus / (args.tensor_parallel_size * args.pipeline_parallel_size)
    embed_params = 2 * args.vocab_size * args.hidden_size
    positional_params = args.hidden_size * args.sequence_length
    ln_params = 8 * args.hidden_size * args.num_layers + (2 * args.hidden_size)
    attention_params = int(2 * (1 + args.kv_size_ratio) * args.num_layers * args.hidden_size * args.hidden_size)
    mlp_params = args.num_mlp_linears * args.num_layers * args.hidden_size * args.ffn_expansion_factor * args.hidden_size
    total_params = embed_params + positional_params + ln_params + attention_params + mlp_params

    bytes_per_param = args.low_prec_bytes_per_val if args.is_mixed_precision else args.high_prec_bytes_per_val
    model_mem = total_params * bytes_per_param
    per_gpu_model_mem = model_mem / (args.tensor_parallel_size * args.pipeline_parallel_size)
    per_gpu_mem_gib = per_gpu_model_mem / 1024**3 + args.misc_mem_gib

    return f"Per-GPU Memory Required for Training: {per_gpu_mem_gib:.2f} GiB"

# Gradio Interface
with gr.Blocks() as demo:
    with gr.Tabs():
        with gr.TabItem("Parameter Calculation"):
            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)

            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)
            calculate_button = gr.Button("Calculate")
            calculate_button.click(calc_params, 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)

        with gr.TabItem("Memory Calculation"):
            hf_model_name_or_path = gr.Textbox(label="HuggingFace Model Name or Path", value="")
            num_gpus = gr.Number(label="Number of GPUs", value=1)
            tensor_parallel_size = gr.Number(label="Tensor Parallel Size", value=1)
            pipeline_parallel_size = gr.Number(label="Pipeline Parallel Size", value=1)
            batch_size_per_gpu = gr.Number(label="Batch Size per GPU", value=8)
            sequence_length = gr.Number(label="Sequence Length", value=2048)
            vocab_size = gr.Number(label="Vocab Size", value=51200)
            hidden_size = gr.Number(label="Hidden Size", value=6144)
            num_attention_heads = gr.Number(label="Number of Attention Heads", value=64)
            num_layers = gr.Number(label="Number of Layers", value=44)
            ffn_expansion_factor = gr.Number(label="FFN Expansion Factor", value=4)
            is_mixed_precision = gr.Checkbox(label="Mixed Precision", value=True)
            misc_mem_gib = gr.Number(label="Misc Memory Overhead (GiB)", value=5)

            memory_result = gr.Textbox(label="Memory Calculation Result", interactive=False)
            calc_memory_button = gr.Button("Calculate Memory")
            calc_memory_button.click(calc_mem, inputs=[num_gpus, tensor_parallel_size, pipeline_parallel_size, batch_size_per_gpu, sequence_length, vocab_size, hidden_size, num_attention_heads, num_layers, ffn_expansion_factor, is_mixed_precision, misc_mem_gib], outputs=memory_result)

demo.launch()