# ONNX Runtime with the CUDA execution provider. Installed INSTEAD of # requirements-onnxruntime.txt (same `onnxruntime` module — both together shadow # each other; the deps manager uninstalls a CPU `onnxruntime` found in the venv # and reinstalls this file, and does the reverse when a node leaves the nvidia # image) on a node that has an NVIDIA card and an nvidia image: the wheel is # baked into $CAMSTACK_ORT_WHEELHOUSE (/opt/camstack/wheelhouse), matched to the # image's CUDA major (12.x: 1.21-1.26; nvidia-legacy, CUDA 11.8: 1.18.1). # pip expands ${CAMSTACK_ORT_WHEELHOUSE}; --no-index is NOT set so numpy etc. resolve. --find-links ${CAMSTACK_ORT_WHEELHOUSE} onnxruntime-gpu>=1.18,<1.27