Skip to content

Release v0.104.5

Choose a tag to compare

View all tags
@github-actions github-actions released this 27 Jun 00:56
v0.104.5
26fb3f4
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Changes in v0.104.5

Performance Improvements

  • parallelize fused SSM/linear-attention scan kernels (Mamba, GLA, RWKV-7, GatedDeltaNet, xLSTM) (#703)
  • exceed OpenBLAS — JIT small-GEMM (1.5–6×) + scientific multi-thread routing/pinning wins on diffusion shapes (#700)
  • kill per-element indexer/numOps dispatch in sampling+distance kernels (GridSample ~50x +3) (#702)

Packages

  • AiDotNet.Native.CLBlast
  • AiDotNet.Native.CUDA
  • AiDotNet.Native.CUDA.Fragments
  • AiDotNet.Native.CUDA.Fragments
  • AiDotNet.Native.CUDA.Fragments
  • AiDotNet.Native.CUDA.Fragments
  • AiDotNet.Native.CUDA.Fragments
  • AiDotNet.Native.CUDA.Fragments
  • AiDotNet.Native.CUDA.Fragments
  • AiDotNet.Native.CUDA.Fragments
  • AiDotNet.Native.MoltenVK
  • AiDotNet.Native.OneDNN
  • AiDotNet.Native.OpenBLAS
  • AiDotNet.Native.ROCm
  • AiDotNet.Tensors

Note: AiDotNet.Native.CUDA bundles the NVIDIA CUDA 12 runtime (cudart / cuBLAS / cuBLASLt / nvRTC, ~770MB). Because that exceeds nuget.org's 250 MB per-package limit, it ships as a small meta package plus a set of AiDotNet.Native.CUDA.Fragments.* chunk packages that are reassembled into the full DLLs at your build (win-x64) — just dotnet add package AiDotNet.Native.CUDA alongside AiDotNet.Tensors on NVIDIA GPU machines; the fragments restore automatically.
Note: AiDotNet.Native.ROCm requires ROCm installed on the target machine for the kernel library (~496MB).

Installation

dotnet add package AiDotNet.Tensors
Assets 17
Loading