Real-world uses

Created for this library

1. 1.

   A research engineer chooses a partitioning strategy that shards attention layers across devices while keeping embeddings replicated.

2. 2.

   An ML platform team configures a partitioning strategy across a 2D device mesh to combine model and data parallelism for very large training jobs.

3. 3.

   An ML platform team encodes its standard partitioning strategy as reusable configuration so research teams reuse a known-good setup.