I am using Clara AGX with dGPU mode. And I am trying to install pytorch with GPU on it. I tried the following:
- Could not find a pytorch installer with GPU for arm64 on Pytorch’s website
- Installers from PyTorch for Jetson - Jetson & Embedded Systems / Jetson Nano - NVIDIA Developer Forums also doesn’t work. After switching to dGPU mode, the cuda version is 11.6. None of the installers in this website support CUDA 11.6. I guess they only works for iGPU modes
Could you please provide some help about how to install Pytorch with GPU?
Hello, there is a NGC Pytorch container which has support for both arm64 and amd64. This should support the devkit with dGPU. For determining which version of the container to pull, we can reference Frameworks Support Matrix - NVIDIA Docs and the dGPU base image for the Holoscan Docker image (in v0.6)
nvcr.io/nvidia/tensorrt:22.03-py3, so you could try
Thanks a lot for your help. This looks great, I will try it tomorrow.
I dont have Clara AGX on hand so I could not try what my thoughts. If I want to include Pytorch with dGPU in the holoscan docker image, should I replace the base image for it from
nvcr.io/nvidia/pytorch:22.03-py3. Or should I add a new a line of
FROM nvcr.io/nvidia/pytorch:22.03-py3? If I add
FROM nvcr.io/nvidia/pytorch:22.03-py3 in the dockerfile, should I copy something by
COPY ***? I am new to docker but trying to learn about it. Thanks in advance for your help.
There are several ways to install the Holoscan SDK, we could see SDK Installation - NVIDIA Docs for the different ways.
Only if you want to build the SDK from source, you could replace the base image in the Dockerfile from
FROM nvcr.io/nvidia/tensorrt:22.03-py3 to
FROM nvcr.io/nvidia/pytorch:22.03-py3. However, please see section “Need more control over the SDK?” and “Attention” on when building from source is appropriate. If you want to build from source, you could reference this Dockerfile modified for v0.6, then please see GitHub instructions after changing the Dockerfile.
If not building the SDK from source, you could run the PyTorch container and install Holoscan SDK inside of it via the deb package or pip install, described in here. Please see the Overview on the NGC page for the
docker run command. Note that the option
Automatically remove the container when it exits in docker run | Docker Documentation.
Hi, thank you so much for your help. Now it’s working :).
Have a nice day
Awesome to hear it’s working for you! Have a great day too.
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.