[TOC]
LLM参数量
e.g. GPT2-small 参数量计算
{
"activation_function": "gelu_new",
"architectures": [
"GPT2LMHeadModel"
],
"attn_pdrop": 0.1,
"bos_token_id": 50256,
"embd_pdrop": 0.1,
"eos_token_id": 50256,
"initializer_range": 0.02,
"layer_norm_epsilon": 1e-05,
"model_type": "gpt2",
"n_ctx": 1024,
"n_embd": 768,
"n_head": 12,
"n_layer": 12,
"n_positions": 1024,
"resid_pdrop": 0.1,
"summary_activation": null,
"summary_first_dropout": 0.1,
"summary_proj_to_labels": true,
"summary_type": "cls_index",
"summary_use_proj": true,
"task_specific_params": {
"text-generation": {
"do_sample": true,
"max_length": 50
}
},
"vocab_size": 50257
}
- input embedding
词表embedding 50257*768 = 38597376
(不计入，不可学习) 位置编码embedding 1024*768 = 786432
- transformer block * 12
attention模块权重矩阵加偏置：768 * 768 * 3 + 768 * 3 = 1771776
attention结尾的线性变换：768 * 768 + 768 = 590592
第一层全连接：768 * 768 * 4 + 768 * 4 = 2362368
第二层全连接：768 * 768 * 4 + 768 = 2360064
2个Layer Normalization (其中参数量只涉及 $\gamma$ 和 $\beta$): 2 * (768 + 768) = 3072
- output embedding
一般和input embedding共享参数，一般bert、GPT都是采用这样的策略
相加上述数据，得到总的参数量为0.12365184B，与官方公布的124M一致
hidden_dim = 768
vocal_size = 50257
n_layers = 12
hidden_dim_2 = 4*hidden_dim
res = 0
res += hidden_dim * vocal_size
for i in range(n_layers):
res += 4*(hidden_dim * hidden_dim + hidden_dim)
res += 2*(2*hidden_dim)
res += hidden_dim*hidden_dim_2+hidden_dim_2
res += hidden_dim_2*hidden_dim+hidden_dim
print(f"{res/10**9}B")
