Useful fine-tuning framework comparison fixes a key EP vs CP misconception

StasBekman · x · 2026-07-26

A useful comparison of popular fine-tuning frameworks, with one important correction: EP does not have to be mutually exclusive with CP.

The post notes that ALST/SP works fine with EP, and because SP overlaps with DP, you can run EP=8 plus SP=8 on a single node and still reach a 2M sequence length even with a 30B MoE model. The main critique is that the article often quotes “max” numbers without specifying whether they assume one GPU or many, or what GPU size is being used.

Related event: Fine-Tuning Framework Comparison Sparks Parallelism Debate(3 posts)→

Original post →

More from Infra

Infra channel →