← All questions
Inference & ServingMedium· Sequence Parallelism· Communication· Efficiency

Why is sequence parallelism said to be free?

In a ring-based implementation, the AllReduce and the ReduceScatter + AllGather pair move the same number of bytes per device. Sequence parallelism rearranges when the scatter happens but does not increase the total volume of data transferred.

Code playground (Python)