Skip to content

perf(megatron): reduce NPU gradient, RMSNorm and MoE permutation overhead - #10513

Draft
addsubmuldiv wants to merge 8 commits into
modelscope:mainfrom
addsubmuldiv:npu-gradient-performance
Draft

perf(megatron): reduce NPU gradient, RMSNorm and MoE permutation overhead#10513
addsubmuldiv wants to merge 8 commits into
modelscope:mainfrom
addsubmuldiv:npu-gradient-performance

Commits

  1. Commits on Nov 4, 2026

  2. Commits on Nov 6, 2026

  3. Commits on Nov 9, 2026