ZeRO++ published
ZeRO++ published — block-quantized all-gather and quantized-gradient all-to-all cut communication volume 4×, with up to 2.16× throughput on 384 GPUs
Why it matters
An improvement that made ZeRO work even on low-bandwidth clusters
Source — Primary sources first; secondary reports are labelled “Press”
Paperarxiv.orghttps://arxiv.org/abs/2306.10209Added in Week 35, 2026 · See that week