flash-attention/training/configs/model/gpt2-hf.yaml
2022-11-28 17:34:40 -08:00

14 lines
444 B
YAML

defaults:
- _self_
- gpt2model: gpt2-small
_target_: transformers.models.gpt2.modeling_gpt2.GPT2LMHeadModel
_recursive_: True
config:
_target_: transformers.GPT2Config
# Mistral's config: https://github.com/stanford-crfm/mistral/blob/main/conf/models/gpt2-small.yaml
# However, reorder_and_upcast_attn slows things down
reorder_and_upcast_attn: false
scale_attn_by_inverse_layer_idx: true
n_positions: ${datamodule.max_length}