Could TransformerEngine work with Deepspeed Zero w/ offloading?

[leiwen83]

Hi,

Since it is common to use with deepspeed zero w/ offloading when training large LLM, does TE currently support in this mode?

Currently deepspeed support is just unittest as refered by TE's readme: deepspeedai/DeepSpeed#3731

Thx~

[sbhavani]

@leiwen83 I'd recommend using https://github.com/huggingface/accelerate/tree/main/benchmarks/fp8 which has an example with DS ZeRO 1-3 support. Please let us know if it's missing any features.