v0.8.16

EricLBuehler/mistral.rsv0.8.16Jun 20, 2026by EricLBuehler

AI Summary

Distributed training logic is optimized by restricting NCCL usage to multi-node setups, and a bug in the AFQ launch/index calculation for large MoE models is resolved.

Key Highlights

NCCL enabled only for world size >= 2
Fixed AFQ launch/index math for large MoE expert stacks

New Features

Distributed training optimization
AFQ launch/index math fix

Full Release Notes

## What's Changed
* feat(distributed): nccl on only for world size >= 2 by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/2257
* fix(afq): fix afq launch/index math for large MoE expert stacks by @EricLBuehler in https://github.com/EricLBuehler/mistral.rs/pull/2258


**Full Changelog**: https://github.com/EricLBuehler/mistral.rs/compare/v0.8.15...v0.8.16