LARC optimizer missing 'state' attribute required by FP16_Optimizer

When I used LARC optimizer and fp16 optimizer as shown in the LARC comments, I got missing AttributeError. After I added optimizer.state to LARC's init() function, the error went away. Can you take a look at this issue? Thanks.

Error:

  File ".../apex/fp16_utils/fp16_optimizer.py", line 143, in __init__
    if param in self.optimizer.state:
AttributeError: 'LARC' object has no attribute 'state'

code I used:

    model = ...
    optim = torch.optim.Adam(model.parameters(), lr=...)
    optim = LARC(optim)
    optim = apex.fp16_utils.FP16_Optimizer(optim)

Start with apex/fp16_utils/fp16_optimizer.py at the reported line 143, then inspect LARC.init() and how the optimizer is wrapped in the issue example. Confirm that FP16_Optimizer can use a LARC-wrapped optimizer without the reported AttributeError, with coverage for this wrapping order.