Setting up CUDA/TensorRT manually on Jetson AGX Xavier

When I try to install Cuda/TensorRT/OpenCV through the Jetpack 4.1 installer it doesn’t actually install any of those packages.

This is almost exactly the problem I have but its for the TX2:

After asking me for the ssh password of nvidia@ its says copying cuda-repo-l4t-10-0-local-10.0.117_0-1_arm64.deb from the jetpack_download directory, then it asks for the nvidia@ ssh password again, and after that it asks again and just hangs, I’ve let it run for 2 hours with no change, and when i quit and check the Xavier nothing is installed.

And yes ssh is correctly set and I’ve confirmed that it works

Any Ideas?

Hi Sam, before installing the packages to Jetson, JetPack downloads the packages on the host to the jetpack_downloads/ directory. You can manually copy the CUDA, cuDNN, and TensorRT packages from this folder to your Jetson. You should then be able to install the packages on your Jetson similar to how the desktop versions of the packages are installed:

  • Alternatively, if you try installing the packages again with JetPack, it may work the second time around. Make sure you can ping your Jetson from your host PC first (and ideally test logging in with SSH from the host). In the JetPack Component Manager, you can select just the packages you wish to install again (CUDA, cuDNN, and TensorRT for the target) and it will not have to re-flash the board.

    Dear all,

    Is it necessary that I install CUDA 10.0, cuDNN, TensorRT 5.0 on my host ubuntu 18.04 amd 64 with GTX1080? I thought that Nvidia SDK Manager helps me to install all tools at my host and Xavier.

    I will try to install nv-tensorrt-repo-ubuntu1804-cuda10.0-trt5.1.5.0-ga-20190427_1-1_amd64.deb at my ubuntu host tomorrow.

    I still confuse which program should run on what platform.

    Thank you very much in advance.

    Warmest Regards,

    with sdkmanager where to get cuda packages for manual installation?
    located: /sdkm_downloads

    sudo dpkg -i *
    (Reading database ... 149017 files and directories currently installed.)
    Preparing to unpack cuda-repo-l4t-10-0-local-10.0.166_1.0-1_arm64.deb ...
    Unpacking cuda-repo-l4t-10-0-local-10.0.166 (1.0-1) over (1.0-1) ...
    Preparing to unpack libcudnn7_7.3.1.28-1+cuda10.0_arm64.deb ...
    Unpacking libcudnn7 ( over ( ...
    Preparing to unpack libcudnn7-dev_7.3.1.28-1+cuda10.0_arm64.deb ...
    update-alternatives: removing manually selected alternative - switching libcudnn to auto mode
    Unpacking libcudnn7-dev ( over ( ...
    Preparing to unpack libcudnn7-doc_7.3.1.28-1+cuda10.0_arm64.deb ...
    Unpacking libcudnn7-doc ( over ( ...
    Setting up cuda-repo-l4t-10-0-local-10.0.166 (1.0-1) ...
    Setting up libcudnn7 ( ...
    Setting up libcudnn7-dev ( ...
    update-alternatives: using /usr/include/aarch64-linux-gnu/cudnn_v7.h to provide /usr/include/cudnn.h (libcudnn) in auto mode
    Setting up libcudnn7-doc ( ...
    Processing triggers for libc-bin (2.27-3ubuntu1) ...
    sudo apt-key add /var/cuda-repo-10-0-local-10.0.166/
    sudo apt update
    Get:1 file:/var/cuda-repo-10-0-local-10.0.166  InRelease
    Ign:1 file:/var/cuda-repo-10-0-local-10.0.166  InRelease
    Get:2 file:/var/cuda-repo-10-0-local-10.0.166  Release [574 B]
    Get:2 file:/var/cuda-repo-10-0-local-10.0.166  Release [574 B]
    Get:3 file:/var/cuda-repo-10-0-local-10.0.166  Release.gpg [819 B]
    Get:3 file:/var/cuda-repo-10-0-local-10.0.166  Release.gpg [819 B]
    Get:4 file:/var/cuda-repo-10-0-local-10.0.166  Packages [10.9 kB]
    Hit:5 bionic InRelease
    Get:6 bionic-updates InRelease [88.7 kB]
    Hit:7 bionic-backports InRelease
    Get:8 bionic-security InRelease [88.7 kB]
    Get:9 bionic-updates/main arm64 Packages [528 kB]
    Get:10 bionic-updates/main Translation-en [247 kB]
    Get:11 bionic-updates/universe arm64 Packages [863 kB]
    Get:12 bionic-updates/universe Translation-en [285 kB]
    Fetched 2,102 kB in 4s (516 kB/s)                                 
    Reading package lists... Done
    Building dependency tree       
    Reading state information... Done
    4 packages can be upgraded. Run 'apt list --upgradable' to see them.
    sudo apt-get install cuda
    Reading package lists... Done
    Building dependency tree       
    Reading state information... Done
    E: Unable to locate package cuda

    seems resolved with

    sudo apt-get install cuda-toolkit-10-0
    1 Like

    It seems that my installation of the latest jetPack 4.2 did not get CUDA and CuDNN installed.

    I understand that it seems that I should should use sdkmanager to get CUDA/CuDNN installed however it wants to reflash the entire Xavier board to do it. I don’t want to reflash.

    I just want to be able to installed CUDA/CuDNN after the OS has been installed?

    How can I do that?

    Hi clemrom,

    No, you still need to upgrade the device to use the new version of CUDA/CuDNn due to driver dependency in low layer.

    Hi dusty,
    I’m trying to install jetpack on the Jetson AGX Xavier without flashing it.
    I downloaded the cuda, cuDNN and tensorRT packages with the sdkmanager.
    The board has already cuda and cuDNN but I have some issues with them, so I want to figure out if reinstalling them will solve the problem.

    Do I have to uninstall them first? Or I can follow the commands in the links you posted anyway?

    You probably want to uninstall the packages first if you are having issues with them.

    Also, SDK Manager can install CUDA/cuDNN/TensorRT without reflashing. You can de-select the flashing step and just choose to install the CUDA/cuDNN/TensorRT components.

    I want to install the Cuda manually but there are many different versions of them in the sdkm_downloads dir. Which one should I use for my Xavier ?

    dafaq - you buy a nvidia dev box and you’d expect it have already installed all the dev packages for cuda :(

    I found this which is very helpful for getting things going for opencv and cuda…

    Also you need to be aware of this issue:

    hi @kayccc I have Jetson nano running JetPack 4.2 LT 32.1.
    Using sdkmanager I have downloaded .deb packages of TensorRT,CUDNN,Cuda10.2 for Jetpack 4.4.1

    I want to install these deb packages directly on Jetson nano running Jetpack4.2, so that I can get latest version of TensorRT and Cuda on the board.

    It is not feasible for me to reflash the board as it is a location that is physically not reachable.
    Do you anticipate any issues in my plan?
    I have 500 such boards on JP4.2 which need latest TensorRt 7.1.3

    As I mentioned previously, each JatPack has it’s own corresponding low layer driver to work with CUDA, CuDNN and TensorRT…SDKs, so that will be many unknown issues if you do it in that way, and we can’t support to resolve it.