Four ways to cut a model across a cluster

A large model does not fit on one chip, so it is cut along up to four dimensions at once. Set the degree of each split, then click a split to inspect what each chip holds and what has to talk to what. Every row of four chips is one node on a fast local fabric.

Inspecting
Tensor parallel
Degree
2 shards
Collective
all reduce
Cadence
every layer

Set each degree, then click a split to inspect its communication.

Pipeline utilisation, 89 percent compute, 11 percent bubble
Chips mapped 2 x 2 x 2 x 2 equals 16 of 16

The right combination is dictated by cluster topology and interconnect bandwidth, not by elegance. Tensor parallel wants a single node, data parallel spans the whole cluster, and pipeline depth is a trade against the bubble. Layout follows the wires.