Fetching the paper…

Communication-Efficient Sparsely-Activated Model Training via Sequence Migration and Token Condensation · Around