Installation Layout

CUDA root:

  • standard NVIDIA Toolkit install under %ProgramFiles%\NVIDIA GPU Computing Toolkit\CUDA

Example CUDA folders validated in this repository:

  • v11.6
  • v11.8
  • v12.1
  • v12.9
  • v13.2

TensorRT root:

  • <repo-root>\third_party\nvidia

Example TensorRT package folders:

  • TensorRT-10.11.0.33-cuda 11.8
  • TensorRT-10.11.0.33-cuda 12.9
  • TensorRT-8.6.1.6-cuda 11.8
  • TensorRT-8.6.1.6-cuda 12.0

Matching rules

  • TensorRT-10.11.0.33-cuda 11.8 matches CUDA 11.0 to 11.8
  • TensorRT-10.11.0.33-cuda 12.9 matches CUDA 12.0 to 12.9
  • TensorRT-8.6.1.6-cuda 11.8 matches CUDA 11.0 to 11.8
  • TensorRT-8.6.1.6-cuda 12.0 matches CUDA 12.0 to 12.1

Important limitation

When multiple CUDA and TensorRT versions exist on the same machine, build and packaging steps must explicitly select the intended combination. Do not rely on implicit discovery alone.

Machine-specific Windows roots should be stored in:

  • pack/runtime/runtime-packages.local.json

This file is ignored by Git. Use pack/runtime/runtime-packages.local.example.json as the starting point. The public runtime-packages.manifest.json must not contain machine-specific local drive roots.

Managed runtime loading is production-first:

  • normal probing checks the application base directory and runtimes/<rid>/native
  • explicit native bridge loading uses JYPPX_NATIVE_BRIDGE_PATH
  • explicit vendor roots use JYPPX_TENSORRT_ROOT and JYPPX_CUDA_ROOT
  • local development scanning of build-out and third_party requires JYPPX_ENABLE_DEVELOPMENT_PROBING=1

CMake presets now refresh selected CUDA toolkit cache variables when a CUDA root is resolved, so stale CUDAToolkit_NVCC_EXECUTABLE values from another CUDA line should not survive reconfiguration.

Validated combinations

  • TensorRT package: TensorRT-10.11.0.33-cuda 11.8
  • CUDA root: %CUDA_PATH_V11_8%
  • CMake preset: win-x64-trt10-cuda11-release

Additional Windows combinations validated in this repository:

  • TensorRT-10.11.0.33-cuda 12.9 + CUDA v12.9
  • TensorRT-8.6.1.6-cuda 11.8 + CUDA v11.8
  • TensorRT-8.6.1.6-cuda 12.0 + CUDA v12.1
  • TensorRT-11.0.0.114-cuda 12.9 + CUDA v12.9

TensorRT 11 status

TensorRT 11 is now part of the Windows build and runtime matrix. The trt11.0-cuda12.9-cudnn9.22 path has native adapter, runtime package, package consumer, and smoke validation. The trt11.0-cuda13.2-cudnn9.22 path compiles, but runtime smoke remains pending until a CUDA 13-capable driver/runtime environment is available.

Linux expectation

The repository assumes Linux runtime packaging will be driven from explicit roots on self-hosted Linux x64 runners rather than auto-discovery from a Windows development machine.

Edit this page