Set up RTX Spark for development, install CUDA-enabled PyTorch, and work around common dependency gaps.
For app compatibility and troubleshooting, see RTX Spark: Start Here. For porting, see the RTX Spark Developer Preview and porting guidance.
Before you start
Use the latest public Windows Arm64 tools and GPU-enabled framework builds where available.
-
Complete Windows Update and restart.
-
Install the latest CUDA Toolkit (currently 13.4.2) and NVIDIA driver for RTX Spark.
-
Use native Arm64 Visual Studio or Visual Studio Code. See Microsoft’s Windows on Arm developer overview.
-
Checkout the FAQ for more information
Match both OS and architecture: Linux Arm64 packages, including DGX Spark builds, cannot run in native Windows Python.
Set up your Python environment
Choose a Python version supported by the NVIDIA PyTorch wheels below. Create a fresh environment in Windows Command Prompt:
python -m venv .venv
.venv\\Scripts\\activate.bat
python -m pip install --upgrade pip
Select this environment in your IDE or notebook; use python -m pip to install packages into it.
Install conda
A native Windows ARM64 Miniconda installer is now available. We recommend installing it directly.
-
Download the Windows ARM64 Miniconda installer.
-
Run the installer, select Just Me, and leave Add Miniconda3 to PATH unchecked. Follow the installation guide.
-
Open Anaconda Prompt from the Start menu and verify the installation:
1. conda --version
4. python --version
For subsequent commands, use conda directly in Anaconda Prompt.
Install PyTorch with CUDA support
Use NVIDIA’s PyTorch Windows repository for CUDA-enabled Windows Arm64 builds.
Check that your Python version has matching PyTorch, TorchVision, and TorchAudio wheels.
Stable release
CUDA 13.4 builds:
python -m pip install "torch==2.14.0+cu134" "torchvision==0.29.0+cu134" "torchaudio==2.11.0+cu134" --extra-index-url https://pypi.nvidia.com/nvtorch_oot/
Nightly builds
Use a separate environment for nightly fixes:
python -m pip install --pre torch torchvision torchaudio --extra-index-url https://pypi.nvidia.com/nvtorch_oot_nightly/
--pre enables prereleases; --extra-index-url adds NVIDIA’s builds while retaining PyPI for other dependencies.
Verify the installation
python -c "import torch; print(torch.\__version_\_); print(torch.version.cuda); print(torch.cuda.is_available())"
Expected stable output: 2.14.0+cu134, 13.4, True.
Test a GPU operation before installing your app’s dependencies:
python -c "import torch; x = torch.ones(4, device='cuda'); print((x + x).cpu())"
If CUDA is unavailable, check your Python environment, PyTorch build, and NVIDIA driver. Recheck after installing app dependencies; version pins can replace PyTorch.
Python Package Installation Gaps
An app may support Windows Arm64 while a dependency lacks a compatible wheel (prebuilt package).
For “No matching distribution found,” check the package version, Python version, and architecture. If pip starts compiling, review the package’s source-build requirements.
We’re working with these projects on native Windows Arm64 support. Until compatible wheels are available, use these workarounds:
| Package | Workaround |
| cryptography | Follow the Windows source-build instructions using Rust, Arm64 C++ build tools, and Arm64 OpenSSL. |
| grpcio | Follow this community-tested recipe with the Arm64 C++ toolchain. Enable Windows long paths and use a short build folder. |
| PyArrow | Build and install a wheel with the Windows Arm64 build scripts. |
| tiktoken | Build an Arm64 wheel with Rust, C++ build tools, and the Windows Arm64 build configuration. |
| OpenCV Python | Build from source with the Windows Arm64 build configuration. |
Windows Arm64 support coming soon
-
Android Studio
-
Ostris AI Toolkit
Porting your application
See the porting overview and the RTX Spark porting guide.
Arm64 apps cannot directly load x64 DLLs or Python extensions. Arm64EC is a separate build target for x64 interoperability. Match native dependencies to your target; see Microsoft’s Arm64EC documentation.
Check app compatibility at Works on Windows on Arm and Python package progress in the CoreCollective Windows on Arm working group.
Get help
Search existing threads first. When reporting an issue, include:
-
RTX Spark model, Windows build, and NVIDIA driver version.
-
Tool/framework version, Python version, and architecture.
-
Install/build command and full error output.
-
A minimal reproducible example.
-
PyTorch verification output, if relevant.
Share what you’re building, what’s blocking you, and any working configuration with versions and steps.