Release Candidate Gate

Current Baseline

As of 2026-06-12 the release candidate gate starts from a zero-missing interface baseline:

  • TensorRT interface coverage: 0 missing rows.
  • CUDA runtime interface coverage: 0 missing rows.
  • Manifest inventory: 3271 API records across 102 manifests.
  • Deferred APIs are explicit manifest/native boundaries and must not be treated as safe public wrappers.

The release candidate gate is split into two paths:

  • Hosted gate: source build, runtime manifest validation, binding generator determinism, workflow contracts, DocFX, managed package content, project quality tests, and report exports.
  • Self-hosted runtime gate: native build, runtime asset collection, runtime nupkg, package consumer validation, optional GPU smoke, and publish readiness for the selected runtime key.

Current primary Windows runtime keys:

  • win-x64-trt8.6-cuda11.8-cudnn8.9
  • win-x64-trt8.6-cuda12.1-cudnn8.9
  • win-x64-trt10.11-cuda11.8-cudnn8.9
  • win-x64-trt10.11-cuda12.9-cudnn9.22
  • win-x64-trt11.0-cuda12.9-cudnn9.22
  • win-x64-trt11.0-cuda13.2-cudnn9.22

At least one locally validated runtime key should pass consumer smoke before a release candidate is considered usable. CUDA 12.9 and TensorRT 11 Windows keys are now eligible only when their exact CUDA/TensorRT/cuDNN dependency chain and package consumer smoke evidence are present.

Current package-consumer evidence on 2026-06-12:

  • win-x64-trt10.11-cuda11.8-cudnn8.9: restore/build/native-copy/smoke passed with 16/16 native assets; probe output TensorRT 10.11.0, CUDA 11.8.
  • win-x64-trt10.11-cuda12.9-cudnn9.22: restore/build/native-copy/smoke passed with 19/19 native asset patterns; probe output TensorRT 10.11.0, CUDA 12.9.
  • win-x64-trt11.0-cuda12.9-cudnn9.22: restore/build/native-copy/smoke passed with 19/19 native asset patterns; probe output TensorRT 11.0.0, CUDA 12.9.
  • win-x64-trt11.0-cuda13.2-cudnn9.22: 2026-06-14 full split package set and collection package packed locally; restore/build/native-copy passed with 19/19 native asset patterns, but package consumer smoke remains pending and publish readiness is blocked until CUDA 13 runtime/builder validation is available.

Required Local Quality Gate

Run this gate before considering any release-candidate tag:

powershell -ExecutionPolicy Bypass -File .\eng\Export-InterfaceCoverageMatrix.ps1
dotnet restore .\TensorRtSharp.sln
dotnet build .\TensorRtSharp.sln -c Debug --no-restore
powershell -ExecutionPolicy Bypass -File .\eng\Test-PublicApiBilingualDocumentation.ps1
dotnet test .\tests\JYPPX.ProjectQuality.Tests\JYPPX.ProjectQuality.Tests.csproj -c Debug --no-build
dotnet docfx .\docs\docfx.json

If native, manifest, or generated binding files changed, also run:

powershell -ExecutionPolicy Bypass -File .\eng\Generate-Bindings.ps1
powershell -ExecutionPolicy Bypass -File .\eng\Test-BindingGeneratorOutputs.ps1
cmake --preset win-x64-trt11-cuda13-release
cmake --build --preset win-x64-trt11-cuda13-release --parallel

Expected local gate result:

  • TensorRT missing rows remain 0.
  • CUDA runtime missing rows remain 0.
  • dotnet build reports 0 warnings and 0 errors.
  • Public API XML comments contain both English and Chinese text.
  • JYPPX.ProjectQuality.Tests passes.
  • DocFX reports 0 warnings and 0 errors.

After the local quality gate, run validation-oriented smoke runners from smoke/README.md in this order:

  1. CudaSmokeRunner
  2. MultiStream
  3. TensorRtSmokeRunner
  4. LifecycleSmokeRunner
  5. OnnxToEngineSmokeRunner
  6. DynamicShape
  7. InferenceBindings
  8. NetworkBuilderSmokeRunner
  9. Layer-specific network runners

Asset-dependent sample projects are executable but require user-provided model metadata:

  • Classification
  • YoloDet

Those projects are not release blockers by themselves as long as their README files describe required external assets and runnable substitutes. CUDA custom-kernel preprocessing remains a documentation roadmap item until safe public CudaModule / CudaKernel wrappers are available.

Runtime Package Gate

For any runtime package selected for release:

  • Validate explicit TensorRT, CUDA, and cuDNN roots before asset collection.
  • Collect runtime assets from the matching CMake preset output.
  • Pack the managed package and matching runtime package.
  • Run package consumer validation for restore/build/native asset copy.
  • Run package consumer smoke when the machine has a compatible driver/GPU/runtime stack.
  • Require local-validated before private-feed or split-delivery readiness can be treated as ready.
  • Record WDAC/application-control blockers separately from package layout failures.

Remote Publication Prerequisites

Before using the remote publication lanes:

  • package-managed.yml with publish_to_nuget=true requires NUGET_API_KEY to be a plain-text ASCII nuget.org API key. Encrypted local credential blobs and machine-generated exports are not valid substitutes.
  • runtime-windows.yml requires the Windows self-hosted runner to stay online with the labels self-hosted, windows, and x64.
  • runtime-linux.yml requires a separate Linux x64 self-hosted runner. If no Linux runner is available yet, leave linux_runtime_keys empty in release-bundle.yml so the Linux module cleanly no-ops.
Edit this page