I can't create a new file, the file system is reported as read-only

[land007]

Dockerfile FROM nvcr.io/nvidia/l4t-ml:r32.4.3-py3 Step 16/19 : RUN cd /usr/local/cuda/targets/aarch64-linux/lib && mkdir 123 ---> Running in d0f4ef0fa4c4 mkdir: cannot create directory '123': Read-only file system

[dusty-nv]

/usr/local/cuda is mounted from the host device in read-only mode, so you can't write to it from within the containers.

For more info, please see here: https://github.com/NVIDIA/nvidia-docker/wiki/NVIDIA-Container-Runtime-on-Jetson#mount-plugins

[land007]

Currently I have the following version environment -library: * CUDA: 10.2.89 * cuDNN: 8.0.0.145 * TensorRT: 7.1.0.16 I wish to make a docker image of the following environment -library: * CUDA: 10.0 * cuDNN: 7.5 * TensorRT: 5.1.6 If docker depends on the environment of the host, it is impossible to complete such an image

[dusty-nv]

It is currently not supported to use a different version of CUDA/ect than what comes with JetPack-L4T.

However you could try making your own base image and install the CUDA packages inside your container. What you should do is move/rename /etc/nvidia-container-runtime/host-files-for-container.d/cuda.csv on your host device first (and cudnn.csv, tensorrt.csv), so those CUDA files don't get mounted into your container.

[land007]

Thanks for your reply, it does not seem easy. The production of nvcr.io/nvidia/l4t-base does not seem to be in this project. I am not sure if the host cuda driver works for the low version cuda calls in Docker.

[land007]

I finally understand what you mean. After repeated attempts, I achieved the downgrade of CUDA.link