How to use only fairscale's OSS

[toriving]

Hi,

I want to use only fairscale Optimizer State Sharded (OSS) similar to Deepspeed's zero1, is there any way?
Currently, ddp_sharded is in the form of zero2.
Strangely, in the case of ddp_sharded(zero2), there is an issue that the learning speed is very slow, so I want to use only OSS.