| # 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. | |