Support native ARM64 in Microsoft.ML.LightGbm

[matouskozak]

Summary

Track native ARM64 support for Microsoft.ML.LightGbm on Linux, macOS, and Windows.

The managed package works on .NET, but its upstream LightGBM NuGet dependency currently contains only:

• linux-x64
• osx-x64
• win-x64

It does not contain native assets for linux-arm64, osx-arm64, or win-arm64. Consequently, native ARM64 applications cannot load lib_lightgbm without manually supplying a compatible binary or using architecture emulation where available.

Updated upstream context

This was previously reported for macOS in #7495 and closed as a platform limitation. However, the limitation appears to be in NuGet packaging rather than LightGBM itself:

• LightGBM publishes official macOS ARM64 Python wheels starting with 4.4.0: Provide binary python package for macOS M1 lightgbm-org/LightGBM#5328 (comment)
• The latest LightGBM 4.7.0 NuGet package still lacks ARM64 runtime assets for Linux, macOS, and Windows.
• An attempted Linux ARM64 NuGet change was closed without implementation, with an upstream maintainer explicitly requesting a feature request: [WIP] Add linux-arm64 support to NuGet package lightgbm-org/LightGBM#7085 (comment)
• Upstream is separately working on Windows ARM64 support: [ci] [python-package] support Windows on arm64 lightgbm-org/LightGBM#6967
• LightGBM.Fit() exits with 139 on osx-arm64 #6966 demonstrates that manually supplied macOS ARM64 binaries are possible, but version matching is brittle and unsupported.

Proposed outcome

1. Coordinate with LightGBM upstream to publish linux-arm64, osx-arm64, and win-arm64 native assets in its NuGet package.
2. Update ML.NET to the corresponding LightGBM package version.
3. Add native ARM64 training and inference coverage for each supported desktop OS.
4. Clearly document platform-specific limitations until each native asset is available.

Acceptance criteria

Native ARM64 applications on Linux, macOS, and Windows can install Microsoft.ML.LightGbm, train and run a LightGBM model without manually copying native libraries or using architecture emulation.

[github-actions]

🔍 No similar issues found with high confidence.

---

AI-powered similar issue detection