- # @package _global_
- defaults:
- - /experiment/pile/gpt3s-flash.yaml
- - override /model/gpt2model: gpt2-medium
- # Can enable mlp_checkpoint_lvl to fit batch_size 16 to A100 40GB
- # model:
- # config:
- # mlp_checkpoint_lvl: 1
- datamodule:
- batch_size: ${eval:"4 if ${train.gpu_mem} < 24 else (8 if ${train.gpu_mem} < 40 else (16 if ${train.gpu_mem} < 80 else 32))"}
- train:
- optimizer:
- lr: 3.0e-4
|
