Why it matters

PP scales across many GPUs. Understanding shapes distributed design.

Advertisement

The architecture

Split layers into stages.

Micro-batches flow through.

Bubble minimization via scheduling.

Pipeline parallelism flowSplit layersinto stagesMicro-batchesflow throughSchedule1F1B, interleavedGPipe, PipeDream, 1F1B, Interleaved 1F1B are common schedules
PP.
Advertisement

How it works end to end

Schedules: GPipe (large bubbles), 1F1B (better), Interleaved (best).

Micro-batch size trade-off.

Cross-stage comm minimal.