AIKernel.Cuda13.0.Libtorch2.12.win-x64

日本語 README

Reference external AIKernel Capability module for one runtime combination: Windows win-x64, LibTorch 2.12.0, and CUDA 13.0. This repository owns the CUDA-specific managed invoker, native C ABI bridge, loader configuration, and runtime metadata that were intentionally separated from AIKernel.Core.

AIKernel.Core is CUDA-free. Install this package only on trusted GPU hosts that explicitly opt in to CUDA execution.

This repository participates in the AIKernel 0.1.0 prototype validation release line scheduled for 2026-06-09. It validates the external Capability split: Core remains CUDA-free, while this repository owns the Windows win-x64 CUDA 13.0 + LibTorch 2.12.0 runtime boundary.

For the full split-distribution rules, see docs/package-distribution.md.

Package Model

This Capability has two C# runtime artifacts and one Python wrapper artifact:

• Lightweight NuGet package: published to NuGet.org for C# consumers. It contains the managed assembly, libtorch_bridge.dll, loader.json, and dynamic loading logic. It does not contain LibTorch, CUDA, cuDNN, cuBLAS, or other large runtime DLLs.
• Full runtime archive: published as a GitHub Release .zip. It includes LibTorch CUDA, CUDA Runtime, cuDNN, cuBLAS, libtorch_bridge.dll, and an auto-configured loader.json.
• pip package: published to PyPI for Python consumers. It is not embedded in the NuGet package.

NuGet is the C# distribution channel. pip is the Python distribution channel. The GitHub Release archive carries the large CUDA runtime snapshot.

Install The Lightweight NuGet Package

For C# consumers:

dotnet add package AIKernel.Cuda13.0.Libtorch2.12.win-x64 --version 0.1.0

The managed package exposes:

• LibTorchCapabilityDescriptor.Create()
• LibTorchCapabilityInvoker
• C ABI operations: load_model, unload_model, forward

The NuGet package includes loader.json. Configure it by setting one of:

• AIKERNEL_LIBTORCH_PATH
• AIKERNEL_CUDA13_LIBTORCH2_12_WIN_X64_HOME
• AIKERNEL_CUDA13_LIBTORCH2_12_WIN_X64_LOADER

Install The Full Runtime Archive

Download the full runtime archive from:

https://github.com/AIKernel-NET/AIKernel.Cuda13.0/releases

Extract it beside the consuming application, or set AIKERNEL_CUDA13_LIBTORCH2_12_WIN_X64_LOADER to the extracted loader.json. The archive layout is designed so the default relative paths work after extraction:

loader.json
runtimes/win-x64/native/libtorch_bridge.dll
runtime/win-x64/libtorch/**

The public C ABI is stable. Do not expose LibTorch, CUDA, or C++ types across the ABI boundary.

Native Build

Windows/MSVC win-x64 is the only supported native build target in this repository revision. Do not add Linux, macOS, ARM64, or other CUDA/LibTorch combinations to this package.

Place LibTorch 2.12.0 + CUDA 13.0 under one of:

• runtime/win-x64/libtorch
• ref/libtorch-win-shared-with-deps-2.12.0+cu130/libtorch in the parent workspace
• AIKERNEL_LIBTORCH_PATH

native/CMakeLists.txt also reads ../ref/env.txt from the parent workspace for CUDA_PATH when present.

cmake -S native -B native/build/win-x64 -A x64
cmake --build native/build/win-x64 --config Release

The package project uses this default when packing the lightweight NuGet native bridge:

• AIKernelNativeBridgePath=native/build/win-x64/Release/libtorch_bridge.dll

Override this MSBuild property if your CI places the native bridge elsewhere.

Release Verification

Before publishing, verify both managed and native surfaces:

dotnet test AIKernel.Cuda13.0.Libtorch2.12.win-x64.slnx -c Release --no-restore
dotnet pack AIKernel.Cuda13.0.Libtorch2.12.win-x64.slnx -c Release --no-restore
cmake --build native/build/win-x64 --config Release

The NuGet.org lightweight .nupkg should contain:

• aikernel-logo128x128.png
• lib/net10.0/AIKernel.Cuda13.0.Libtorch2.12.win-x64.dll
• runtimes/win-x64/native/libtorch_bridge.dll
• loader.json
• LICENSE
• README.md

It must not contain LibTorch, CUDA, cuDNN, cuBLAS, or other large runtime DLLs.

The GitHub Release runtime .zip should contain:

• runtimes/win-x64/native/libtorch_bridge.dll
• runtime/win-x64/libtorch/**
• loader.json
• LICENSE
• README.md
• RELEASE_NOTES.md
• LICENSE-THIRD-PARTY/*

Release Workflow

Use .github/workflows/release.yml for the split publication flow:

1. Build and test the managed Capability.
2. Build libtorch_bridge.dll with Windows MSVC.
3. Pack the lightweight NuGet .nupkg.
4. Push the lightweight NuGet package to NuGet.org.
5. Build the full runtime .zip.
6. Upload the full runtime .zip to GitHub Release assets.

The full package job intentionally runs on a self-hosted Windows runner labeled Windows, X64, and CUDA13. That runner must have:

• Visual Studio/MSVC C++ tools
• CMake
• CUDA Toolkit 13.0.x
• LibTorch 2.12.0 + CUDA 13.0 available at the workflow libtorch_path input, or another path passed through AIKERNEL_LIBTORCH_PATH

See docs/package-distribution.md for the publisher and consumer checklist.

Python / pip

Python distribution is independent from NuGet. NuGet packages are for C# consumers, while Python wrappers are published through pip.

The Python package follows the same channel policy as AIKernel.Core:

• Stable packages are published to PyPI.
• Development packages are attached to GitHub Releases for CI/CD and compatibility testing.

The stable Python package for this Capability is:

pip install aikernel-cuda13-libtorch2-12-win-x64

Import it as:

import aikernel_cuda13_libtorch2_12_win_x64 as cuda_capability

This Python package is intentionally lightweight. It carries the CUDA Capability identity, descriptor metadata, managed Capability DLL, libtorch_bridge.dll, loader.json, and installation guidance. It does not include LibTorch, CUDA runtime DLLs, cuDNN, or cuBLAS; those runtime assets remain in the full GitHub Release runtime archive.

Development wheels use aikernel-cuda13-libtorch2-12-win-x64-dev and versions such as 0.1.0.dev1. GitHub Packages does not provide a PyPI registry, so dev pip artifacts are distributed as GitHub Release assets or installed from the repository with pip install git+...#subdirectory=python.

See docs/python-package-distribution.md for the Python package policy.

Third-Party Notices

LibTorch/PyTorch runtime binaries are redistributed under the BSD 3-Clause License and Additional Terms. CUDA Runtime and cuDNN DLLs included in the LibTorch Windows CUDA package are NVIDIA redistributables.

The lightweight NuGet package does not redistribute LibTorch, CUDA, cuDNN, or cuBLAS binaries.

The GitHub Release runtime archive includes:

• LICENSE-THIRD-PARTY/pytorch-LICENSE.txt
• LICENSE-THIRD-PARTY/pytorch-NOTICE.txt

Keep those files in the runtime archive whenever LibTorch CUDA binaries are included.

Fork Model

Create separate Capability repositories for other GPU, OS, RID, or runtime targets. This repository remains the Windows win-x64 CUDA 13.0 + LibTorch 2.12.0 package only.

• AIKernel.Cuda13.0.Libtorch2.12.win-arm64
• AIKernel.Cuda13.0.Libtorch2.12.linux-x64
• AIKernel.Cuda12.4.Libtorch2.3.win-x64
• AIKernel.ROCm6.Libtorch2.12.linux-x64
• AIKernel.DirectML.win-x64
• AIKernel.Vulkan.linux-x64

Fork this repository and update the runtime version metadata, native build settings, and descriptor metadata while preserving AIKernel Capability contracts.

License

Apache License 2.0.