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
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.
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.
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 (7.3.1.28-1+cuda10.0) over (7.3.1.28-1+cuda10.0) ...
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 (7.3.1.28-1+cuda10.0) over (7.3.1.28-1+cuda10.0) ...
Preparing to unpack libcudnn7-doc_7.3.1.28-1+cuda10.0_arm64.deb ...
Unpacking libcudnn7-doc (7.3.1.28-1+cuda10.0) over (7.3.1.28-1+cuda10.0) ...
Setting up cuda-repo-l4t-10-0-local-10.0.166 (1.0-1) ...
Setting up libcudnn7 (7.3.1.28-1+cuda10.0) ...
Setting up libcudnn7-dev (7.3.1.28-1+cuda10.0) ...
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 (7.3.1.28-1+cuda10.0) ...
Processing triggers for libc-bin (2.27-3ubuntu1) ...
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?
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.
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.