File size: 497 Bytes
d1396f0 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 |
# Optimizer
2 considerations:
1. AdamW - robust, but memory-hungry
2. Adafactor - more lean, but more difficult to figure out to converge - more likely to be used if the model is t5-like
## HF
default AdamW
## Deepspeed
default AdamW
## Megatron
Has `--optimizer adam` via `apex`
To add a new optimizer need to add a new option [here](https://github.com/NVIDIA/Megatron-LM/blob/aed2f75e209e525c842aec7c044af7acae2a4614/megatron/optimizer/__init__.py#L50) and import that new optimizer.
|