File size: 1,569 Bytes
0523608 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 | import argparse
import sys
import os
import torch
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from nanochat.gpt import GPT, GPTConfig
from scripts._sweep_utils import model_dims
def get_model_params(model_type, depth, gate_mode='mlp'):
n_layer, n_head, n_embd, max_seq_len = model_dims(depth)
config = GPTConfig(
sequence_len=max_seq_len,
vocab_size=100277,
n_layer=n_layer,
n_head=n_head,
n_embd=n_embd,
)
if model_type == 'dense':
config.use_remixed_linear = False
elif model_type == 'remixed':
config.use_remixed_linear = True
config.remix_basis_size = 256
config.remix_context_dim = 256
config.remixed_linear_kwargs = {
'use_basis_gate': True,
'use_output_gate': True,
'use_context': True,
'basis_gate_mode': gate_mode,
'output_gate_rank': 8,
}
model = GPT(config)
# also print parameter table to find the bloat
if depth == 12:
from scripts.base_train import print_params_table
print(f"\n--- {model_type} ({gate_mode}) ---")
print_params_table(model)
return sum(p.numel() for p in model.parameters() if p.requires_grad)
for d in [4, 12]:
print(f"Depth {d} (C={model_dims(d)[2]}):")
print(f" Dense: {get_model_params('dense', d):,}")
print(f" Remixed (MLP): {get_model_params('remixed', d, 'mlp'):,}")
print(f" Remixed (Lin): {get_model_params('remixed', d, 'linear'):,}")
|