[REQUEST] being able to initialize stages 1, 2 engine w/ `optimizer=None`

Similar to how this was done for stage3, would it be possible to have a way to initialize stage 1/2 engine w/o optim+lr_scheduler?

We need it for integrating lm-eval harness into Megatron-Deepspeed which runs in an inference mode. We don't have a choice since the setup is for deepspeed so the easiest is to just load it as is but we don't want to load the optimizer files for inference.

I know unlike stage 3 this feature won't be useful to most users, but in an occasional special cases it will be very handy.

update: to clarify - this is not about not loading optim states from the checkpoint. It's about not allocating the optimizer and being able to pass to initialize optimizer=None, lr_scheduler=None. This has been implemented in Z3.

Thank you.

@jeffra

Start at the stage 1/2 engine initialize path and compare it with the stage 3 behavior referenced in the issue. Trace how the optimizer and lr_scheduler are allocated for inference. Done means stage 1/2 accepts optimizer=None, lr_scheduler=None without allocating optimizer state while preserving existing training behavior.