joestubbs
commented
3 months ago This seems to be fixed, thanks to Samuel's new automated image build. There is now an automated build that produces a new image, https://hub.docker.com/r/tapis/image_scoring_plugin_py_nano_3.8
Currently, the image scoring plugin Docker container image is based on the
tapis/camera_traps_pybase image, which is ultimately based on the official python image (e.g.,python:3.10). The iGPUs on edge devices, such as the Jetson Nano, appear to require separate CUDA drivers and recommend using theirl4tseries container images, for example:nvcr.io/nvidia/l4t-base:r36.2.0(base image)nvcr.io/nvidia/l4t-cuda:12.2.12-runtime(CUDA)nvcr.io/nvidia/l4t-tensorrt:r8.6.2-devel(TensorRT)nvcr.io/nvidia/l4t-pytorch:r35.2.1-pth2.0-py3(pytorch)Note that, for the CUDA image, the image tag above indicates the version of CUDA (i.e.,
12.2.12). We need to create new images for the image scoring plugin that bundle these CUDA drivers so that it can utilize the iGPU of the Nano. See [1] for background information and [2 - 5] for available images.[1] https://catalog.ngc.nvidia.com/orgs/nvidia/containers/l4t-base [2] https://catalog.ngc.nvidia.com/orgs/nvidia/containers/l4t-base/tags (Base) [3] https://catalog.ngc.nvidia.com/orgs/nvidia/containers/l4t-cuda/tags (CUDA) [4] https://catalog.ngc.nvidia.com/orgs/nvidia/containers/l4t-tensorrt/tags (TensorRT) [5] https://catalog.ngc.nvidia.com/orgs/nvidia/containers/l4t-pytorch/tags (pytorch)