[REQUEST] Add optional "subsequence overlap" to UlyssesSPDataLoaderAdapter

For sequence parallel pipelines that have conv + downsampling layers prior to attention layers, it would be beneficial to allow for some overlap between subsequences sharded across GPUs, so as to not introduce convolution artifacts from zero-padded tensors.

I would like to see an argument I can provide to a DeepSpeed config file, like sequence_parallel_subsequence_overlap (or something less verbose), which is piped to UlyssesSPDataLoaderAdapter if present and used to adjust the sequence length for each micro-batch.

It could look something along the lines of:

class UlyssesSPDataLoaderAdapter:

    def __init__(
        self,
        dl: DataLoader,
        sp_rank: int,
        sp_group,
        sp_world_size,
        device,
        subsequence_overlap: int = 0    # <-- NEW ARG
    ):
        ...
        def refill(self):
            ...
            # update length of sharded tensor with user-defined subsequence_overlap
            # specified in DS config file
            chunk_len = seq_length // self.sp_world_size + subsequence_overlap

Locate UlyssesSPDataLoaderAdapter and the DeepSpeed config path that supplies its arguments. Trace how each micro-batch's sequence length is calculated and determine where the optional overlap should be propagated; done means the configured value adjusts the sharded subsequence length while the default preserves current behavior.