method summary please?

Could you possibly post a brief summary of the used approach? Or a link to a paper/article that discusses your approach?

I'm documenting the different approaches to training huge models here: https://github.com/huggingface/transformers/issues/9766

Would love to add yours if you could share the approach you use.

Thank you very much!