Hi @OscarSimetria
Welcome to the StereoLabs community.
GPU processing test passed! … 
Test suite failed: CURAND_STATUS_INITIALIZATION_FAILED
Thanks for the detailed log, it actually points straight at the cause. Notice that everything up to and including the grayscale conversion works: the camera opens, the image is retrieved on the GPU, and the simple processing test passes. Those steps run inside the ZED SDK’s own CUDA context and use plain CuPy array slicing, so they don’t exercise cuRAND. The failure appears only at the memory-allocation test, which is the first step that calls into cuRAND (curandCreateGenerator). That is the tell.
The root cause is the CuPy package you installed. pip install cupy-cuda12x pulls the generic PyPI wheel, which bundles its own set of CUDA libraries built for desktop/server x86_64 GPUs. On a ZED Box Mini (Jetson Orin NX, aarch64, L4T/JetPack), CUDA is provided by JetPack itself, and the bundled cuRAND in that wheel does not initialize correctly against the Tegra driver, hence CURAND_STATUS_INITIALIZATION_FAILED. CuPy 14.1.1 reporting “CUDA version: 12090” confirms it loaded the wheel’s bundled stack rather than your JetPack CUDA.
The fix is to use a CuPy that links against the JetPack CUDA on the device instead of the prebuilt wheel:
- Remove the wrong package (inside your
ZED venv):
pip uninstall cupy-cuda12x cupy
- Confirm your JetPack CUDA major version:
nvcc --version
On current Orin NX images this is CUDA 12.x.
- Install CuPy built from source against the on-device CUDA, so it links the JetPack cuRAND:
export CUDA_PATH=/usr/local/cuda
export PATH=$CUDA_PATH/bin:$PATH
pip install --no-cache-dir cupy
This compiles CuPy against the Tegra CUDA libraries. The build takes a while on Orin NX; if it is slow, increasing the active power mode (sudo nvpmodel -m 0 then sudo jetson_clocks) helps.
- Re-run and verify the right libraries are picked up:
python -c "import cupy; cupy.show_config()"
python hello_zed_gpu.py
show_config() should report your JetPack CUDA path, and the cuRAND step will then initialize.
Quick sanity check first, if you want to isolate it in one line:
python -c "import cupy as cp; print(cp.random.rand(4))"
If that throws the same cuRAND error, it confirms the wheel mismatch rather than anything ZED-side.
For reference: ZED SDK download (ZED SDK 5.4 - Download | Stereolabs) and the Python API install guide (Install the ZED Python API | StereoLabs). Product pages for your hardware: ZED Box Mini (ZED Box Mini | Stereolabs) and the ZED X camera (ZED X - AI Stereo Camera for Robotics | Stereolabs).
Let me know your exact JetPack/L4T version (cat /etc/nv_tegra_release) if rebuilding CuPy still gives trouble, and I’ll follow up.
Testing GPU image processing (basic grayscale conversion)…
Testing memory allocation strategies…