Add wheel support for Newton-Schulz method via cuSolverMp

[ksivaman]

Description

#2706 added distributed Newton-Schulz matrix orthogonalization API via cuSolverMp, this PR brings the support for the same via published wheels.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

• Enable NVTE_WITH_CUSOLVERMP TE build via PyPI wheel.

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[greptile-apps]

Greptile Summary

This PR updates the wheel build infrastructure to include cuSolverMP as an optional feature by installing the system packages in both Dockerfiles, creating a canonical /opt/nvidia/cusolvermp layout with symlinked include/lib directories, and exporting NVTE_WITH_CUSOLVERMP=1 in the build script.

• Dockerfiles (x86 + aarch64): libcusolvermp0-cuda-${CUDA_MAJOR} and its devel package are installed via dnf, symlinks create a unified /opt/nvidia/cusolvermp/{include,lib} tree, ldconfig is invoked in the same layer, and CUSOLVERMP_HOME is exported — all addressing prior review feedback.
• build_wheels.sh: Only NVTE_WITH_CUSOLVERMP=1 is exported; the three other flags cited in the PR description (NVTE_WITH_CUBLASMP, NVTE_ENABLE_NVSHMEM, NVTE_UB_WITH_MPI) have no corresponding exports or dependency installs anywhere in the changed files.

Confidence Score: 4/5

Safe to merge for cuSolverMP support, but the three other features promised in the PR description are not implemented.

The cuSolverMP wiring — package install, symlink tree, ldconfig, ENV exports — is self-consistent and correct in both Dockerfiles. The gap is between what the PR claims to deliver (four optional build flags) and what is actually wired up (only one). Users relying on the PR description to know the wheel now includes cuBLASMP, NVSHMEM, or MPI support will be wrong.

build_tools/wheel_utils/build_wheels.sh — the three missing flag exports (and the Dockerfiles that would need matching dependency installs) are the unfinished portion of this change.

Important Files Changed

Filename	Overview
build_tools/wheel_utils/build_wheels.sh	Adds NVTE_WITH_CUSOLVERMP=1 export before the build loop, but NVTE_WITH_CUBLASMP, NVTE_ENABLE_NVSHMEM, and NVTE_UB_WITH_MPI — all claimed in the PR description — are absent.
build_tools/wheel_utils/Dockerfile.x86	Installs libcusolvermp from the CUDA repo, creates /opt/nvidia/cusolvermp with include/lib symlinks, calls ldconfig in the same layer, and sets CUSOLVERMP_HOME. Only cuSolverMP is set up; no cuBLASMP/NVSHMEM/OpenMPI packages are installed.
build_tools/wheel_utils/Dockerfile.aarch	Mirror of Dockerfile.x86 for aarch64 — same cuSolverMP install/symlink/ldconfig/ENV additions, same missing cuBLASMP/NVSHMEM/MPI setup.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Docker build\nDockerfile.x86 / .aarch] --> B[dnf install libcusolvermp0-cuda-CUDA_MAJOR\n+ devel package]
    B --> C[Create /opt/nvidia/cusolvermp\nwith include + lib symlinks]
    C --> D[echo lib path to ld.so.conf.d\n+ ldconfig]
    D --> E[ENV CUSOLVERMP_HOME=/opt/nvidia/cusolvermp\nENV LD_LIBRARY_PATH += cusolvermp/lib]
    E --> F[Container starts\nbuild_wheels.sh]
    F --> G[export NVTE_WITH_CUSOLVERMP=1]
    G --> H[python setup.py bdist_wheel\ncommon lib]
    H --> I[Wheel includes cuSolverMP]
    F -.->|NOT exported| J[NVTE_WITH_CUBLASMP\nNVTE_ENABLE_NVSHMEM\nNVTE_UB_WITH_MPI]
    J -.->|NOT installed| K[cuBLASMP / NVSHMEM / OpenMPI\ndependencies]
    style J fill:#ffcccc,stroke:#cc0000
    style K fill:#ffcccc,stroke:#cc0000

Loading

_{Reviews (2): Last reviewed commit: "Add NS via cusolvermp to wheel build" | Re-trigger Greptile}

[greptile-apps]

Likely incorrect CUSOLVERMP_HOME path

The path ${SITE_PACKAGES}/nvidia/cu${CUDA_MAJOR} is missing the package-name segment. Every other NVIDIA Python package follows the layout site-packages/nvidia/<package-name>/cu<ver>/ — for example, nvidia-cublasmp-cu12 installs under nvidia/cublasmp/cu12/, so nvidia-cusolvermp-cu12 should install under nvidia/cusolvermp/cu12/. With the current path the .so symlink loop silently skips cuSolverMP's lib/ directory ([ -d "$lib_dir" ] || continue), no unversioned .so stubs are created, and the linker will not find cuSolverMP at build time even though NVTE_WITH_CUSOLVERMP=1 is exported.

[greptile-apps]

Three of the four advertised flags never get exported

The PR description and title claim to enable NVTE_WITH_CUSOLVERMP, NVTE_WITH_CUBLASMP, NVTE_ENABLE_NVSHMEM, and NVTE_UB_WITH_MPI in the wheel build. Only NVTE_WITH_CUSOLVERMP is exported here. Neither NVTE_WITH_CUBLASMP, NVTE_ENABLE_NVSHMEM, nor NVTE_UB_WITH_MPI are exported in build_wheels.sh, and no corresponding packages (cuBLASMP, NVSHMEM, OpenMPI) are installed in either Dockerfile. Wheels built from this script will silently omit those three features.