25 lines
538 B
YAML
25 lines
538 B
YAML
# @package _global_
|
|
defaults:
|
|
- /experiment/pile/gpt3s-flash.yaml
|
|
- override /optimizer: adamw-zero
|
|
|
|
model:
|
|
config:
|
|
n_embd: 1536
|
|
n_head: 16
|
|
n_layer: 24
|
|
# mlp_checkpoint_lvl: 1 # To fit batch_size 8
|
|
|
|
datamodule:
|
|
batch_size: ${eval:"2 if ${train.gpu_mem} < 24 else (4 if ${train.gpu_mem} < 40 else (8 if ${train.gpu_mem} < 80 else 16))"}
|
|
|
|
train:
|
|
optimizer:
|
|
lr: 2.5e-4
|
|
|
|
trainer:
|
|
strategy:
|
|
_target_: src.utils.ddp_zero1.DDPStrategyZero1
|
|
find_unused_parameters: False
|
|
gradient_as_bucket_view: True
|