AttributeError: 'FlatParameter' object has no attribute '_full_param_padded'

I use the nllb branch of fairseq for nllb training.Encountering a problem：

File "/media/walter/work/hui.li/fairscale/fairscale/nn/data_parallel/fully_sharded_data_parallel.py", line 2302, in _gather_optim_state
desired_buffer_size = non_shared_params[0]._full_param_padded.size()
AttributeError: 'FlatParameter' object has no attribute '_full_param_padded'
srun: error: localhost: task 0: Exited with exit code 1

Reproduce the NLLB training failure, then inspect fairscale/nn/data_parallel/fully_sharded_data_parallel.py at _gather_optim_state and the access to _full_param_padded. Done means optimizer-state gathering completes without the AttributeError; add or run a focused regression test if the repository provides one.

Written by the indexing model from the issue text.