← All questions
Inference & ServingMedium· Parallelism Strategies· Distributed Training· Networking

When would you choose pipeline parallelism over FSDP?

FSDP’s AllGather-per-layer pattern costs 3|θ| per step, which can stall on low-bandwidth links like 25 GB/s InfiniBand. Pipeline parallelism’s activation handoffs are one to two orders of magnitude smaller than FSDP’s weight AllGathers, making it more viable for large models in bandwidth-constrained environments.

Code playground (Python)