What's Changed
- Update patch version after 0.16.8 release by @loadams in #7296
- Avoid graph break by removing another redundant requires grad false by @deepcharm in #7263
- Add qwen3 meta loading for AutoTP by @delock in #7293
- Modernize system executable detection across components by @emmanuel-ferdman in #7290
- Enable ZeRO set/get APIs for NVMe offload by @tjruwase in #7046
- Add qwen3moe meta loading for AutoTP by @ranzhejiang in #7297
- disable license check until the new license situation has been sorted… by @stas00 in #7301
- Fix extra_repr_str when weight is None / in zero-3 by @HollowMan6 in #7254
- [XPU] Support XCCL on deepspeed side by @ys950902 in #7299
New Contributors
- @emmanuel-ferdman made their first contribution in #7290
Full Changelog: v0.16.8...v0.16.9