Build guide

This guide covers everything needed to build the TensorRT RTX Execution Provider from source on Windows, Windows ARM64, and Linux, including the Python wheel.

Note: The repository also provides convenience scripts build.bat (Windows) and build.sh (Linux) that wrap the CMake steps below and add path validation, vcpkg bootstrapping, and wheel packaging. The plain CMake commands documented here are sufficient for all standard builds.

Prerequisites

Software requirements

Dependency	Minimum version	Platform	Notes
CMake	3.20	All	Must be on system `PATH`
Visual Studio	2022 (17.0+)	Windows	“Desktop development with C++” workload required. VS 2026 also supported.
GCC / Clang	GCC 11+ / Clang 14+	Linux
Python 3	3.8+	All	Required during CMake configuration for bundled ONNX dependency
CUDA Toolkit	12.9 (cu12) or 13.x (cu13)	All	See CUDA versions below. Download from the CUDA Toolkit page.
ONNX Runtime	1.24.0	All	Prebuilt SDKs at ORT releases.
TensorRT RTX	1.4.0	All	Download from TensorRT RTX.

CUDA versions

Two CUDA variants are supported:

cu12 — CUDA 12.9 (pinned)
cu13 — any CUDA 13.x release (Windows ARM64 requires 13.4+)

CUDA and TensorRT RTX must target the same CUDA major version.

Note: If you already have protobuf installed via winget or another system package manager, it will conflict with CMake’s FetchContent and fail configuration. Remove it or use -DUSE_VCPKG=ON.

Build: Windows

Open a Developer PowerShell for VS 2022 (or VS 2026) so that MSVC and the Windows SDK are on the path, then:

cmake -S . -B build -G "Visual Studio 17 2022" -A x64 `
  -DCUDAToolkit_ROOT="C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9" `
  -DONNXRUNTIME_ROOT="C:\SDK\onnxruntime-win-x64-<version>" `
  -DTRT_RTX_ROOT="C:\SDK\TensorRT-RTX-<version>"

cmake --build build --config Release --parallel

Build: Linux

cmake -S . -B build -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DCUDAToolkit_ROOT=/usr/local/cuda \
  -DPython3_EXECUTABLE="$(command -v python3)" \
  -DONNXRUNTIME_ROOT=/opt/onnxruntime-<version> \
  -DTRT_RTX_ROOT=/opt/TensorRT-RTX-<version>

cmake --build build --parallel

Build: Windows ARM64

Windows ARM64 (WoA) cross-compilation runs on an x64 Windows host. Open a Developer PowerShell for VS 2022 and pass -A ARM64:

cmake -S . -B build -G "Visual Studio 17 2022" -A ARM64 `
  -DCUDAToolkit_ROOT="C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" `
  -DONNXRUNTIME_ROOT="C:\SDK\onnxruntime-win-arm64-<version>" `
  -DTRT_RTX_ROOT="C:\SDK\TensorRT-RTX-<version>-arm64" `
  -DUSE_PRECOMPILED_HOST_PROTOC=ON

cmake --build build --config Release --parallel

Note: Windows ARM64 requires CUDA 13.4+. The cross-compiled output runs on ARM64 devices but the build itself executes on an x64 Windows host.

Python wheel

Building the wheel is a three-step process:

Step 1 — Build the C++ library (as documented above)

Step 2 — Stage runtime DLLs into the Python package directory (Windows):

cd python
python scripts/stage_windows_dlls.py `
  --ep-dll "..\build\Release\onnxruntime_providers_nv_tensorrt_rtx.dll" `
  --trt-lib-dir "C:\SDK\TensorRT-RTX-<version>\bin" `
  --cuda-bin "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\bin\x64" `
  --trt-doc-dir "C:\SDK\TensorRT-RTX-<version>\doc"

For Linux use python scripts/stage_linux_so.py with equivalent arguments.

Step 3 — Build the wheel:

Windows:

$env:TRT_RTX_EP_VERSION = "<M.m.p>"
python -m build --wheel --no-isolation --outdir dist

Linux:

TRT_RTX_EP_VERSION="<M.m.p>" python -m build --wheel --no-isolation --outdir dist

Outputs (under python\dist\ / python/dist/):

onnxruntime_ep_nv_tensorrt_rtx_cu12-<M.m.p>-py3-none-win_amd64.whl — CUDA 12 variant
onnxruntime_ep_nv_tensorrt_rtx-<M.m.p>-py3-none-any.whl — meta wheel

The meta wheel (onnxruntime-ep-nv-tensorrt-rtx) defaults to the cu13 variant. CUDA 12 users must install onnxruntime-ep-nv-tensorrt-rtx-cu12 explicitly.

Note: build.bat --build_wheel / build.sh --build_wheel automate all three steps above in a single command.

CMake options reference

Option	Description
`CUDAToolkit_ROOT`	CUDA Toolkit installation root
`ONNXRUNTIME_ROOT`	Root of an extracted ONNX Runtime. If omitted, CMake downloads it.
`ONNXRUNTIME_VERSION`	ORT version to download when `ONNXRUNTIME_ROOT` is not set
`TRT_RTX_ROOT`	Root of an extracted TensorRT RTX
`BUILD_TESTS`	Build the unit test suite (default: `ON`)
`BUILD_EXAMPLES`	Build the C++ examples (default: `OFF`)
`TRT_RTX_EP_PRODUCTION_BUILD`	Enable production-build checks and signature verification
`TRT_RTX_EP_VERSION`	Version string embedded in production artifacts

Run tests

Most runtime tests require an NVIDIA GPU. Building the tests does not.

Windows:

.\build\tests\Release\unittests.exe

Linux:

./build/tests/unittests

Troubleshooting

CMake cannot find Python

Set Python3_EXECUTABLE explicitly:

cmake ... -DPython3_EXECUTABLE="C:\Python311\python.exe" ...

`TRT_RTX_ROOT must be set`

Same as above for TensorRT RTX:

Test-Path "C:\SDK\TensorRT-RTX-<version>"

`Could not find CUDAToolkit`

Provide the path explicitly and confirm CUDA is installed:

nvcc --version
cmake ... -DCUDAToolkit_ROOT="C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9" ...

CUDA and TRT RTX major version mismatch

CUDA and TensorRT RTX must target the same CUDA major version — both cu12 or both cu13. Mixing versions causes linker errors.

protobuf conflict

If protobuf is installed system-wide (e.g. via winget install protobuf), it conflicts with CMake’s FetchContent. Uninstall it before configuring.

Stale build cache

After changing SDK roots, generator, or CUDA version, configure a fresh build directory:

# Windows
Remove-Item -Recurse -Force build

# Linux
rm -rf build

C++20 compiler errors

Ensure Visual Studio 2022 17.0+ is installed (_MSC_VER >= 1930). On Linux, GCC 11+ or Clang 14+ are required.

See the integration guide for provider integration and runtime usage.