PyTorch opens the FSDP2 RFC
Opened the FSDP2 RFC — pointed out the limits of the flat parameter structure and proposed per-parameter sharding, a communication-free sharded state dict and mixed FP8 all-gather
Why it matters
A redesign of FSDP1. The precondition that made FP8 training, partial freezing and compiler optimization possible
Source — Primary sources first; secondary reports are labelled “Press”
Officialgithub.comhttps://github.com/pytorch/pytorch/issues/114299Added in Week 35, 2026 · See that week