infrastructure
Muon's update size remains unchanged under fp16 loss scaling.
Muon's update size remains consistent across loss scales with this change.
1 sources · 1 verified claimsAI knowledge entity
Deepspeed
One evolving timeline of verified developments, without duplicating the underlying truth across pages.
3 verified events1 verified sourceUpdated 30 Sept 2026
Verified timeline
Developments connected to this entity
infrastructure
DeepSpeed v0.19.7 adds configurable dtype for ZeRO checkpoint export.
The v0.19.7 patch includes a new ARM SVE update kernel for CPU-Adam and fixes a Triton NFS detection crash.
1 sources · 1 verified claimsinfrastructure
DeepSpeed reduces NCCL backend timeout from 30 to 10 minutes.
The v0.19.6 patch includes fixes for checkpoint rank selection and staticmethod handling on Python 3.9.
1 sources · 1 verified claims