NuGet packaging and consumption

LLamaSharp-style model: one pure managed package + per-RID native runtime packages.

Package Contents TFM
LiteRtLmSharp Managed assembly only (lib/net10.0/LiteRtLmSharp.dll). No natives. net10.0
LiteRtLmSharp.Extensions.AI IChatClient and IEmbeddingGenerator connectors (Microsoft.Extensions.AI). Depends on LiteRtLmSharp (same version). net10.0
LiteRtLmSharp.SemanticKernel IChatCompletionService connector and embedding generator registration, built on LiteRtLmSharp.Extensions.AI (same version). net10.0
LiteRtLmSharp.runtime.win-x64 runtimes/win-x64/native/: the official LiteRtLm.dll (static CRT) + the DirectX Shader Compiler runtime (dxcompiler.dll, dxil.dll) the GPU backend needs. No lib. (native-only)
LiteRtLmSharp.runtime.linux-x64 runtimes/linux-x64/native/libLiteRtLm.so (one official library). No lib. (native-only)
LiteRtLmSharp.runtime.linux-arm64 runtimes/linux-arm64/native/libLiteRtLm.so (one official library). No lib. (native-only)
LiteRtLmSharp.runtime.android-arm64 runtimes/android-arm64/native/libLiteRtLm.so (one official library, accelerators and samplers embedded; packed into the APK as lib/arm64-v8a/). (native-only)
LiteRtLmSharp.runtime.android-x64 runtimes/android-x64/native/libLiteRtLm.so (the same library for the x86_64 Android emulator; packed into the APK as lib/x86_64/). (native-only)
LiteRtLmSharp.runtime.osx-arm64 runtimes/osx-arm64/native/libLiteRtLm.dylib (one official library, Apple Silicon). (native-only)
LiteRtLmSharp.runtime.ios-arm64 Google's official CLiteRTLM.xcframework (device + simulator slices) injected via NativeReference (buildTransitive .targets); CPU backend. No lib. (native-only)

iOS (ios-arm64). Unlike desktop, a nupkg runtimes/ folder is not auto-embedded on iOS; the native library ships as Google's official CLiteRTLM.xcframework, consumed via NativeReference Kind=Framework from a buildTransitive .targets conditioned on the net10.0-ios TFM — embedded and code-signed into the app bundle and loaded at runtime by NativeLibraryResolver as Frameworks/CLiteRTLM.framework/CLiteRTLM (see the iOS linking decision in docs/roadmap.md). The xcframework carries device and simulator slices. GPU (Metal) needs companion libraries the xcframework only references by name and the package does not ship, so the CPU backend is the supported path. Build/link is validated in CI (ios-package-check.yml, which also asserts the package layout); on-device runtime is pending hardware validation (see docs/roadmap.md). Every package ships README.md, LICENSE.txt, NOTICE and THIRD-PARTY-NOTICES.md at its root; the runtime packages add upstream's THIRD_PARTY_NOTICES.litert-lm.txt for the binaries they carry.

All packages share one version per release (enforced by the single Version in Directory.Build.props) and that version is independent of the LiteRT-LM native version, which is pinned via LiteRtLmVersion and surfaced in the package release notes and the README compatibility table. Install the managed and runtime packages with the same version number.

Consumption (including MAUI)

<PackageReference Include="LiteRtLmSharp" Version="1.2.0" />
<PackageReference Include="LiteRtLmSharp.runtime.win-x64" Version="1.2.0" />
<!-- and/or linux-x64 / linux-arm64 / android-arm64 / android-x64 (emulator) / osx-arm64, per target -->

The SDK copies runtimes/<rid>/native/* into the consumer's output; NativeLibraryResolver resolves them. The managed package does not depend on any runtime package (each consumer picks its RID, like LLamaSharp).

Validated end-to-end: managed + runtime.win-x64 from a local feed → a consumer app resolved the natives and ran inference.

Producing the packages

  1. Native release (native-release.yml) pinned to the upstream tag (v0.18.0, the current pin) with publish_release → downloads Google's official prebuilts (the C API zip release asset, or the litert-lm-api wheels on PyPI when the release carries no zip), verifies each download against the digest its host publishes, inspects each library and publishes the native-v0.18.0 release with the litertlm-*.tar.gz assets plus THIRD_PARTY_NOTICES.litert-lm.txt.
  2. Pack (pack-nuget.yml): downloads those assets, lays them into runtimes/<rid>/native/, runs dotnet pack with the given package_version, uploads the .nupkgs as an artifact, and (optionally, push=true) publishes to nuget.org via Trusted Publishing (the GitHub OIDC token is exchanged for a short-lived API key with NuGet/login — no stored secret).

Locally: dotnet pack LiteRtLmSharp/LiteRtLmSharp.csproj -c Release -o .nupkgs (managed) and the projects under packaging/ (these need the natives already present in runtimes/<rid>/native/).

Notes / pending

  • No Vulkan loader needed on Linux for the CPU backend: the v0.18.0 libraries carry no hard dependency on libvulkan.so.1 (the v0.16.0 one did, which is why 1.2.0 documented libvulkan1). The GPU backend runs on Vulkan and needs the GPU's Vulkan driver plus the Vulkan loader.
  • No VC++ Redistributable on Windows since 1.2.0: the official LiteRtLm.dll links the CRT statically.
  • Optional future: a LiteRtLmSharp.Backend.Desktop meta-package depending on the win/linux runtime packages.
  • Android GPU: consumers must declare <uses-native-library> in their manifest (see the README and docs/android.md).
Edit this page