Instructions to use KexuanShi/Megatron-LM with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- NeMo
How to use KexuanShi/Megatron-LM with NeMo:
# tag did not correspond to a valid NeMo domain.
- Notebooks
- Google Colab
- Kaggle
How to use ?
Add these flags to enable optimizer cpu offload in MCore.
--optimizer-cpu-offload
--optimizer-offload-fraction 1.0
--use-precision-aware-optimizer
Configuration Recommendations
Gradient copy from GPU to CPU, CPU optimizer step, and subsequent parameter copy from CPU to GPU can be time-consuming operations, and it is recommended to use the flag --overlap-cpu-optimizer-d2h-h2d to execute them concurrently.