[BUG] Recommended way to implement EMA

Describe the bug
A clear and concise description of what the bug is.

Hi deepspeed team, I have some code that uses exponential moving average (EMA) for training a UNet model, the code relies on the named_parameters() and parameters() of the model to store and update the params, the simplified impl is like below:

# Part 0 ====================================================================

class LitEma(nn.Module):

    def __init__(
        self,
        model:nn.Module,
        decay:float=0.9999,
        use_num_upates:bool=True,
    ):
        super().__init__()
        if decay < 0.0 or decay > 1.0:
            raise ValueError('Decay must be between 0 and 1')

        self.m_name2s_name = {}
        self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32))
        self.register_buffer(
            'num_updates',
            torch.tensor(0, dtype=torch.int) if use_num_upates else torch.tensor(-1, dtype=torch.int))

        for name, p in model.named_parameters():
            if p.requires_grad:
                #remove as '.'-character is not allowed in buffers
                s_name = name.replace('.','')
                self.m_name2s_name.update({ name: s_name })
                self.register_buffer(s_name, p.clone().detach().data)

        self.collected_params = []

    def forward(self, model:nn.Module):
        decay = self.decay

        if self.num_updates >= 0:
            self.num_updates += 1
            decay = min(self.decay, (1+self.num_updates)/(10+self.num_updates))

        one_minus_decay = 1.0 - decay

        with torch.no_grad():
            m_param = dict(model.named_parameters())
            shadow_params = dict(self.named_buffers())

            for key in m_param:
                if m_param[key].requires_grad:
                    sname = self.m_name2s_name[key]
                    shadow_params[sname] = shadow_params[sname].type_as(m_param[key])
                    shadow_params[sname].sub_(one_minus_decay*(shadow_params[sname]-m_param[key]))
                else:
                    assert not key in self.m_name2s_name

    def copy_to(self, model:nn.Module):
        m_param = dict(model.named_parameters())
        shadow_params = dict(self.named_buffers())
        for key in m_param:
            if m_param[key].requires_grad:
                m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data)
            else:
                assert not key in self.m_name2s_name

    def store(self, parameters:Iterable[nn.Parameter]):
        """
        Save the current parameters for restoring later.
        Args:
          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
            temporarily stored.
        """
        self.collected_params = [param.clone() for param in parameters]

    def restore(self, parameters:Iterable[nn.Parameter]):
        """
        Restore the parameters stored with the `store` method.
        Useful to validate the model with EMA parameters without affecting the
        original optimization process. Store the parameters before the
        `copy_to` method. After validation (or model saving), use this to
        restore the former parameters.
        Args:
          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
            updated with the stored parameters.
        """
        for c_param, param in zip(self.collected_params, parameters):
            param.data.copy_(c_param.data)

# Part 1 ====================================================================

    if self.use_ema:
        self.model_ema = LitEma(self.model)
        print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")

# Part 2 ====================================================================

    @contextmanager
    def ema_scope(self, context:str=None):
        if self.use_ema:
            self.model_ema.store(self.model.parameters())
            self.model_ema.copy_to(self.model)
            if context is not None:
                print(f"{context}: Switched to EMA weights")
        try:
            yield None
        finally:
            if self.use_ema:
                self.model_ema.restore(self.model.parameters())
                if context is not None:
                    print(f"{context}: Restored training weights")

# Part 3 ====================================================================

    def on_train_batch_end(self, *args, **kwargs):
        if self.use_ema:
            self.model_ema(self.model)

When using deepspeed, the .parameters() and .named_parameters() all returns empty, I'm wondering what is the recommended way of implementing the above LitEma class with deepspeed? Sorry if this seems to be a dumb question, but I'm new here and with offload and sharding it seems unclear to me how to implement it correctly.

To Reproduce
Steps to reproduce the behavior:

1. Go to '...'
2. Click on '....'
3. Scroll down to '....'
4. See error

Expected behavior
A clear and concise description of what you expected to happen.

ds_report output
Please run ds_report to give us details about your setup.

Screenshots
If applicable, add screenshots to help explain your problem.

System info (please complete the following information):

• OS: [e.g. Ubuntu 18.04]
• GPU count and types [e.g. two machines with x8 A100s each]
• Interconnects (if applicable) [e.g., two machines connected with 100 Gbps IB]
• Python version
• Any other relevant info about your setup

Launcher context
Are you launching your experiment with the deepspeed launcher, MPI, or something else?

Docker context
Are you using a specific docker image that you can share?

Additional context
Add any other context about the problem here.

Start with the LitEma implementation and its model.parameters() and model.named_parameters() calls described in the issue. Reproduce the empty-parameter behavior under the referenced offload and sharding setup, then establish and document a supported EMA approach with a clear validation case.