Skip to content

Releases: AIKernel-NET/AIKernel.Cuda13.0

AIKernel.Cuda13.0 v0.1.0

08 Jun 18:27
@egumaru egumaru
v0.1.0
dc4d7fa
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Choose a tag to compare

View all tags
AIKernel.Cuda13.0 v0.1.0 Latest
Latest

June 9th, 2026 — Deterministic tensors enter the graph.
2026年6月9日──決定論的 Tensor がグラフへ流入した。

CUDA 13.0 bindings enter the graph: deterministic tensor ops join the semantic pipeline without leaking device complexity.
CUDA 13.0 バインディングがグラフへ参入──デバイスの複雑性を漏らさず、決定論的 Tensor 演算が意味パイプラインに合流する。

[CUDA] Device Binding: attached
[CUDA] Tensor Ops: deterministic
[CUDA] Device Complexity: suppressed

Today, device computation becomes a first‑class deterministic component.
今日、デバイス計算は決定論的な第一級コンポーネントとなりました。

AIKernel.Cuda13.0 v0.1.0

This release publishes the AIKernel.Cuda13.0 external Capability package for Windows win-x64, LibTorch 2.12.0, and CUDA 13.0.

Packages

  • AIKernel.Cuda13.0.Libtorch2.12.win-x64 0.1.0
  • aikernel-cuda13-libtorch2-12-win-x64 0.1.0

Changes

  • Switches AIKernel NuGet references to official published packages
  • Uses AIKernel contract packages at 0.1.0
  • Uses AIKernel Core runtime packages at 0.1.0.1
  • Updates Python package dependency to aikernel-net>=0.1.0.1
  • Aligns memory mapping integration with the official AIKernel.Core memory API
  • Builds the managed Capability package and Python wrapper package

Runtime Notes

The NuGet package contains the managed Capability descriptor/invoker, loader metadata, NuGet icon, and native bridge when available.

LibTorch 2.12.0 and CUDA 13.0 runtime binaries are distributed separately through GitHub Release archives. AIKernel.Core remains CUDA-free.

Verification

  • dotnet restore AIKernel.Cuda13.0.Libtorch2.12.win-x64.slnx --no-cache
  • dotnet test AIKernel.Cuda13.0.Libtorch2.12.win-x64.slnx -c Release --no-restore
  • py -m pytest
  • py -m build --wheel
  • py -m twine check dist\aikernel_cuda13_libtorch2_12_win_x64-0.1.0-py3-none-win_amd64.whl
Assets 2
Loading

AIKernel.Cuda13.0.Libtorch2.12.win-x64 0.0.5

06 Jun 15:03
@egumaru egumaru
v0.0.5
a73c5d6

Choose a tag to compare

View all tags
AIKernel.Cuda13.0.Libtorch2.12.win-x64 0.0.5

June 6th, 2026 — The GPU finally spoke.
2026年6月6日 — GPU はついに応答した。

As the Core achieved symmetry, the CUDA Capability returned its first external-world handshake.

Core が対称性を獲得したその瞬間、 CUDA Capability は初めて 外部世界としての応答 を返した。

On this symmetric day, CPU and GPU stood as two parallel execution domains.

この対称的な日に、 CPU と GPU は 二つの並列実行系 として立ち上がった。

Capability is not part of the OS. Capability is the outside world.

Capability は OS の内部ではない。 Capability は 外部世界そのもの だ。

コード

[CUDA] Capability v0.0.5 loaded  
[CUDA] Backend: CUDA 13.0 / LibTorch 2.12  
[CUDA] Target: win-x64  
[CUDA] Native ABI handshake: OK  
[CUDA] Ready to accelerate semantic execution

Today, the GPU became an extension of the semantic runtime.
今日、GPU はセマンティックランタイムの延長として動き始めた。

AIKernel.Cuda13.0.Libtorch2.12.win-x64 0.0.5

Initial CUDA Capability release for AIKernel.

This release targets:

  • Windows win-x64
  • CUDA 13.0
  • LibTorch 2.12.0
  • .NET 10

Package model

This release uses split distribution:

  • NuGet.org: lightweight C# package
  • PyPI: lightweight Python wrapper package
  • GitHub Release asset: full CUDA runtime archive

NuGet package

The NuGet package is intended for C# consumers and includes:

  • Managed Capability assembly
  • libtorch_bridge.dll
  • loader.json
  • Dynamic runtime loading logic

It does not include:

  • LibTorch
  • CUDA runtime
  • cuDNN
  • cuBLAS
  • other large runtime DLLs

Install:

dotnet add package AIKernel.Cuda13.0.Libtorch2.12.win-x64 --version 0.0.5

Python package

The Python wrapper is published separately through pip:

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

Import:

import aikernel_cuda13_libtorch2_12_win_x64 as cuda_capability

The Python wheel includes the lightweight Capability surface:

  • Managed Capability DLL
  • libtorch_bridge.dll
  • bundled loader.json
  • loader helper APIs

It does not include the large CUDA runtime payload.

Runtime archive

The GitHub Release runtime archive contains the large runtime payload:

  • LibTorch CUDA runtime
  • CUDA Runtime
  • cuDNN
  • cuBLAS
  • libtorch_bridge.dll
  • auto-configured loader.json
  • third-party notices

Extract the archive beside the consuming application, or set:

$env:AIKERNEL_CUDA13_LIBTORCH2_12_WIN_X64_LOADER="C:\path\to\loader.json"

Validation

  • .NET tests passed: 12/12
  • Python tests passed: 4/4
  • Python wheel passed twine check
  • NuGet package verified to exclude large CUDA runtime DLLs
Loading