Update embedded-devices example for the latest PyTorch and JetPack

[wwjang]

Describe the type of feature and its functionality.

Following @jafermarq's suggestion(#4381 , #4382 ), I am creating this new issue.

How would you feel about providing guidelines for the build script?
I’ve modified the build script (build_jetson_flower_client.sh) to use Flower (flwr) with JetPack 6.0 and successfully tested it. I believe it would be beneficial to provide guidelines for the build script. Utilizing the BASE_PYTORCH variable could simplify the process compared to reinstalling PyTorch from scratch.

Currently, the NVIDIA NGC Catalog offers base images supporting up to PyTorch v2.0.0. Additionally, with a bit more searching, we can find the dustynv base image for PyTorch v2.2.0, which is compatible with JetPack 6.0. I have also confirmed on the NVIDIA forums that the R36.2 container image is compatible with R36.3.

Let me know if this approach aligns with your plans or if there's anything else I can assist with. I'm happy to help further, whether it's refining the build script guidelines or contributing in other ways!

Describe step by step what files and adjustments are you planning to include.

build_jetson_flower_client.sh
`Update the build script to accept the BASE_PYTORCH and BASE_TF variables as input arguments.
Ensure that the script defaults to a standard base image if no argument is provided.

Dockerfile
`In my experience using the dusty-nv base image, I encountered an issue where libsndfile1 was missing. To resolve this, I added the following command to the Dockerfile:

RUN apt-get update && \
    apt-get install -y --no-install-recommends libsndfile1 && \
    rm -rf /var/lib/apt/lists/*

README.md
`Add a table that lists the available base images along with a brief description and reference links.

Is there something else you want to add?

No response

[jafermarq]

Hi @wwjang , thanks for opening the issue! There is some work done happening in this PR: #4381; but it still remains a TODO the jetson part

[jafermarq]

Hi @wwjang , I wanted to update you on this topic: the https://github.com/adap/flower/tree/main/examples/embedded-devices has been updated but we have left out NVIDIA Jetson devices for now. We'd like to include them in a future update. If you are still interested and have access to a new Jetson, would you want to update the guide?

[wwjang]

Hi @jafermarq, Sorry for the delayed reply.
I still have access to the Jetson device and can work on the update given some time. Should I refer to the current and previous README files and my suggestions when writing it? If you have any other methods or approaches in mind, please let me know.

[WilliamLindskog]

Hi @wwjang,

just checking in here. Did you have time to look and work at the updates? Is there anything we can do to help

[wwjang]

Hi @WilliamLindskog,

I’ve been quite busy, but I’ve just had a chance to test it. Based on the requirements in the pyproject.toml file, I created the Dockerfile and ran it. I can confirm that on my Jetson Orin NX device, the following steps in the current embedded-devices tutorial work fine: "Launching the Flower SuperLink", "Connecting Flower SuperNodes", and "Running the Flower App." However, I haven't had time to test the entire learning process yet.

I was thinking of creating a jetson_setup.md file, similar to the device_setup.md, and simply transferring the previous guide into it. Does that sound good to you? What are your thoughts?

[WilliamLindskog]

That sounds good! @wwjang

Feel free to create a PR where you see fit, happy to support you throughout.