[Draft] Newton-Schulz via cuSOLVERMp#2706

Open

vcherepanov-nv wants to merge 30 commits intoNVIDIA:mainfrom

vcherepanov-nv:newton-schulz

Commits on Feb 9, 2026

[PyTorch] Fix NCCL comm extraction and pass global dims to Newton-Schulz

vcherepanov-nv
and
claude
committed

Commits on Feb 18, 2026

[Common] Cache cuSolverMp handle and grid in Newton-Schulz context

vcherepanov-nv
and
claude
committed
[Common] Create dedicated CUDA stream in Newton-Schulz context

vcherepanov-nv
and
claude
committed
[Common] Fix Newton-Schulz zero output with event-based stream sync

vcherepanov-nv
and
claude
committed
[Common] Fix Newton-Schulz NaNs by keeping host workspace alive

vcherepanov-nv
and
claude
committed
[Common] Cache CUDA event in Newton-Schulz context

vcherepanov-nv
and
claude
committed
[Common] Use separate in/out events for Newton-Schulz stream sync

vcherepanov-nv
and
claude
committed
Correct coefficients
vcherepanov-nv
committed
No stream synchronize
vcherepanov-nv
committed
[Test] Verify Newton-Schulz result with XAX=I identity check

vcherepanov-nv
and
claude
committed

Commits on Feb 19, 2026