| # Handy info utils for gpt2 | |
| ## Calculate model size | |
| Calculate the number of params in the model: | |
| - h = hidden size | |
| - l = num_layers | |
| - s = sequence length | |
| - v = vocabulary size | |
| ``` | |
| $ python -c "h=1024; l=24; s=1024; v=50257; print(f'{l*(12*h**2 + 13*h) + v*h + s*h + 2*h >> 20}M')" | |
| 338M | |
| ``` | |
| For our scripts where we only care for Billions: | |
| ``` | |
| NHIDDEN=4096 | |
| NLAYERS=36 | |
| SEQ_LEN=512 | |
| VOCAB_SIZE=50257 | |
| python -c "h=$NHIDDEN; l=$NLAYERS; s=$SEQ_LEN; v=$VOCAB_SIZE; print(f'Model size: {(l*(12*h**2 + 13*h) + v*h + s*h + 2*h) / 10**9 :.0f}B')" | |
| ``` | |
| Full math for the above final formula: (num_heads is not really part of the calculations) | |
| ```# Let h = hidden size, n = num_layers, k = num_heads, s = sequence length, v = vocabulary size | |
| # Compute Embedding Parameters (Vocab + Position) | |
| emb_params = (v * h) + (s * h) | |
| # Compute Parameters per Transformer Block | |
| head_dim = h / k | |
| qkv_params_w = k * (3 * (h * (h / k))) = 3 * h * h # 3h^2 | |
| mh_reduce_w = (k * ((h / k)) * h = h * h # h^2 | |
| qkv_params_b = k * (3 * (h / k)) = 3 * h # 3h | |
| mh_reduce_b = h # h | |
| pos_ff_exp_w = h * (4 * h) # 4h^2 | |
| pos_ff_con_w = (4 * h) * h # 4h^2 | |
| pos_ff_exp_b = 4 * h # 4h | |
| pos_ff_con_b = h # h | |
| layer_norm1 = 2 * h # 2h | |
| layer_norm2 = 2 * h # 2h | |
| # Magic Formula: | |
| total_params = n * (12h^2 + 13h) + (v * h) + (s * h) + 2*h | |
| ``` | |
| credits: Sidd Karamcheti | |
| An estimate variations of this for large hidden size and number of layers (seq and vocab size have very small contribution) | |
| ``` | |
| NHIDDEN=4096 | |
| NLAYERS=36 | |
| python -c "h=$NHIDDEN; l=$NLAYERS; print(f'Model size: {(12*l*h**2) / 10**9 :.0f}B')" | |
| ``` | |
| credits: Mohammad Shoeybi | |
| Can calculate the same on a given `model` object (counts shared params once): | |
| ``` | |
| sum(dict((p.data_ptr(), p.numel()) for p in model.parameters()).values()) | |
| ``` | |