What's Changed
- Update version post 0.19.8 release by @loadams in #8871
- Update to filter out version.txt from PR gates. by @loadams in #8872
- Skip the accelerator-only CPU optimizer tests on the cpu accelerator by @delock in #8836
- Keep Yuan's o_proj bias whole on every rank and describe it as replicated by @Achyuthan-S in #8857
- Expect AutoTP to keep a row-parallel output head tied to the embedding by @jinyouzhi in #8855
- Add Triton kernels for vocab-parallel logprobs, entropy and KL by @jinyouzhi in #8838
- Fix numerical stability in Triton masked softmax by @arcusbuilds in #8782
- Add RVV vectorized FP32 kernel to CPU Adam by @zhangjian29 in #8817
- [OPSD] Request last-position logits only in HybridEngine rollout prefill by @jinyouzhi in #8839
- docs: fix 30 dead links across README, docs pages, tutorials and posts by @pratikgx in #8688
- docs: fix remaining dead links in the Chinese README, tutorials and blog by @woodforestsen in #8894
- Bump modal-torch-latest CI Python from 3.10 to 3.12 by @delock in #8902
- [AffineIR] Plan expert transfers from the live runtime placement by @jinyouzhi in #8800
- Set TORCHINDUCTOR_COMPILE_THREADS=1 in modal CI sandbox to stop 300s test timeouts by @delock in #8914
- Fix scalar parameter byte-view broadcasts by @adenzhou1350 in #8642
- Honor fp16 hysteresis in the ZeRO stage 0 fp16 optimizers by @vineethsaivs in #8647
- Strip ANSI color codes before parsing FAILED lines in cpu-multi-triage by @delock in #8882
- Bump AWS capacity-fallback Python from 3.10.13 to 3.12.14 by @delock in #8921
- AutoEP: fused Triton kernel for the clamped expert activations by @pengdurice in #8652
- Keep fp32 masters on ZeRO-1 and ZeRO-2 when a resume skips optimizer state by @sfc-gh-truwase in #8891
- Skip accelerator sync for unprofiled comm ops by @guptaishaan in #8795
- Fix performance regression caused by comms logger and bug in comms_logger.enabled by @hwchen2017 in #8909
- Use nvtx's module-level API so NVTX domain ranges work on older nvtx releases by @hwchen2017 in #8900
- [AffineIR] Affine storage maps for live ZeRO parameters and optimizer states by @jinyouzhi in #8885
- Add a v1 multi-rank CPU suite (manual dispatch first) by @delock in #8926
- [AutoEP] Ensure consistent order of partition groups in all-gather when ZeRO-3 by @jinyouzhi in #8876
- Fix race between folded checkpoint load and metadata drop in AutoEP+AutoTP test by @jinyouzhi in #8943
- Check where column-split pieces write in a shard by @Achyuthan-S in #8922
New Contributors
- @arcusbuilds made their first contribution in #8782
- @zhangjian29 made their first contribution in #8817
- @pratikgx made their first contribution in #8688
- @woodforestsen made their first contribution in #8894
- @adenzhou1350 made their first contribution in #8642
- @guptaishaan made their first contribution in #8795
Full Changelog: v0.19.8...v0.19.9