Installation
Install witwin-radar 0.4.1 from PyPI with the Channel extra, on a CUDA 12.8 PyTorch, and check that the native library loads.
Radar is published on PyPI as witwin-radar. The wheel ships one prebuilt native CUDA library, so you do not compile anything. This page lists the requirements, gives the install commands, and shows how to check that the library loads on your GPU.
Requirements#
| Requirement | Supported |
|---|---|
| Operating system | Windows x86-64, or Linux x86-64 with glibc 2.28 or newer (manylinux_2_28) |
| Python | 3.11 for scene simulation (see the note below); the witwin-radar wheel itself accepts 3.10 to 3.14 |
| PyTorch | Exactly 2.10.0 built for CUDA 12.8 (torch==2.10.0+cu128) |
| GPU | NVIDIA GPU with a driver that supports CUDA 12.8. The library is built for compute capability 7.0, 7.5, 8.0, 8.6, 8.7, 8.9, 9.0, 10.0, 10.1 and 12.0 |
| Dependencies | witwin (Core) >=0.4.0,<0.5, numpy>=1.24, scipy>=1.10, and witwin-channel>=0.5.0,<0.6 through the channel extra |
Use Python 3.11
Every simulation goes through Channel, the propagation package. The witwin-channel 0.5.0 wheel on PyPI is built for CPython 3.11 only and pins torch==2.10.0 and witwin==0.4.0. On another Python version pip install "witwin-radar[channel]" cannot find a matching Channel wheel. Create a Python 3.11 environment.
Install#
Install the CUDA 12.8 build of PyTorch first, then Radar with the channel extra. If you install Radar first, pip may pull a CPU-only PyTorch from PyPI.
python3.11 -m venv .venv # Windows: py -3.11 -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install torch==2.10.0 --index-url https://download.pytorch.org/whl/cu128
pip install "witwin-radar[channel]==0.4.1"The install adds witwin 0.4.0 (Core), witwin-channel 0.5.0, NumPy and SciPy.
Optional extras#
| Extra | Adds | Needed for |
|---|---|---|
channel | witwin-channel>=0.5.0,<0.6 | Every Radar.trace, simulate and stream call. Install it unless you only use the processing functions |
smpl | smplpytorch, opencv-python | witwin.radar.smpl, the SMPL human body model. You also need the SMPL model files, which are not distributed with Radar |
pip install "witwin-radar[channel,smpl]==0.4.1"Verify the install#
Check that the native library loads and matches your PyTorch and CUDA runtime:
from witwin.radar.deployment import build_info, require_supported_runtime
print(build_info()["native_build"]["torch_version"])
report = require_supported_runtime()
print(report["package_version"], report["torch"], report["cuda_runtime"], report["sm_matrix_status"])On a supported machine this prints:
2.10.0
0.4.1 2.10.0+cu128 12.8 runtime_verifiedrequire_supported_runtime() raises RuntimeError with every failed check when the environment does not match, for example a CPU-only PyTorch or a different PyTorch version. runtime_diagnostics() returns the same report without raising. See Deployment for every field.
Note
sm_matrix_status is runtime_verified only on GPUs the release was tested on (compute capability 12.0 in 0.4.1). Other architectures in the build list report declared_unverified: the kernels are compiled for them, but the release was not run on them.
Install from source#
To build the native library yourself, clone witwin-ai/witwin-radar and follow its README. A source build needs the CUDA 12.8 toolkit and a C++ compiler. Most users should install the wheel.
Next step#
Run the Quickstart.