DeepSpeed v0.18.6 补丁发布：修复竞态条件与梯度问题

AI 摘要

DeepSpeed 发布了 v0.18.6 补丁版本，包含多项修复和改进。主要修复了叶模块竞态条件、ZeRO stage 0 的 bf16 梯度范数发散等问题，并支持 AutoTP 自定义分区模式及 Python 3.14 注解处理。此外，将 torch.jit.script 替换为 torch.compile，并新增了两位贡献者。

正文节选

## What's Changed * Update version.txt to 0.18.6 after latest release by @loadams in https://github.com/deepspeedai/DeepSpeed/pull/7826 * Fix leaf module race condition by @tohtana in https://github.com/deepspeedai/DeepSpeed/pull/7825 * Skip sequence parallel operations during eval by @jp1924 in https://github.com/deepspeedai/DeepSpeed/pull/7821 * Support custom partitioning patterns for AutoTP by @tohtana in https://github.com/deepspeedai/DeepSpeed/pull/7806 * Fix gradient is ready with z2

发布时间：2026-02-13 07:21

抓取时间：2026-08-02 00:31

来源机构：Microsoft