calc_params.py

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])

# Parameter Calculation function
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

    return 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)}
    """