google/praxis

[Feature Request] Need ZeRo-1/2 to cooperate with PP+TP+DP. Which may more faster than FSDP sometimes.

MoFHeka opened this issue · 0 comments

MoFHeka commented

For example in single A100 machine. Llama2 13B training speed with TP2 DP 4 + Zero1 is more faster than FSDP.