Skip to content

python-package-distribution.md

Latest commit

 

History

History
78 lines (56 loc) · 2.58 KB

python-package-distribution.md

File metadata and controls

78 lines (56 loc) · 2.58 KB

Python Package Distribution

日本語

This repository publishes a lightweight Python package for the CUDA Capability. Python distribution is independent from NuGet: NuGet packages are for C# consumers, while pip packages are for Python consumers.

The Python package is not embedded in the NuGet package. Release wheels carry the same lightweight runtime surface as the NuGet package: managed Capability DLL, native bridge DLL, and loader.json. They do not carry the CUDA runtime payload.

Stable Channel

Stable Python releases are published to PyPI.

Field Value
Distribution aikernel-cuda13-libtorch2-12-win-x64
Import name aikernel_cuda13_libtorch2_12_win_x64
Version line 0.1.0 -> ...
Contents Capability metadata, managed Capability DLL, libtorch_bridge.dll, bundled loader.json, loader helpers, and installation guidance

Install:

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

Use:

import aikernel_cuda13_libtorch2_12_win_x64 as cuda_capability

print(cuda_capability.capability_descriptor())
print(cuda_capability.install_instructions())

config = cuda_capability.load_loader_config()
print(config.resolved_runtime_search_paths())
print(cuda_capability.bundled_managed_assemblies())
print(cuda_capability.bundled_native_libraries())

Development Channel

Development Python wheels use the package identity aikernel-cuda13-libtorch2-12-win-x64-dev and version numbers such as 0.1.0.dev1.

GitHub Packages does not provide a PyPI registry. For pip users, development wheels are distributed as GitHub Release assets or installed directly from the repository:

pip install git+https://github.com/AIKernel-NET/AIKernel.Cuda13.0.git#subdirectory=python

Use development wheels only for CI/CD and compatibility testing. Breaking changes are allowed on the development channel.

Runtime Payload Boundary

The heavy executable CUDA runtime remains outside the Python package and outside the PyPI artifact:

  • LibTorch CUDA DLLs
  • CUDA runtime and cuDNN redistributables
  • Full GitHub Release runtime archive

Those assets are attached to GitHub Releases as the full runtime .zip. The PyPI package is a stable wrapper, discovery, and tooling package only. C# consumers use NuGet; Python consumers use pip.

The PyPI package does include a loader.json template so Python tooling can inspect and generate matching loader configuration. Release wheels also include the managed Capability DLL and native bridge DLL. They do not include LibTorch, CUDA, cuDNN, or cuBLAS runtime binaries.