Fetching the paper…

SWARM Parallelism: Training Large Models Can Be Surprisingly Communication-Efficient · Around