https://github.com/microsoft/DeepSpeed/blob/937c5ceec1906e1a8f8e21d2dd165f5280050b26/deepspeed/runtime/fp16/onebit_adam.py#L120 here buffer_m with dimension same as parameters and always more than 1。
https://github.com/microsoft/DeepSpeed/blob/937c5ceec1906e1a8f8e21d2dd165f5280050b26/deepspeed/runtime/fp16/onebit_adam.py#L120
here buffer_m with dimension same as parameters and always more than 1。