[FEA] Better support for CONDA on Embedded GPUs

[boonedoggle]

Can we test and potentially make this conda environment the official version for cuSignal?

https://github.com/conda-forge/miniforge.

[datametrician]

@kkraus14 @mike-wendt thoughts for RAPIDS on Arm in general, can someone add Jason as well (@awthomp)

[awthomp]

@JasonAtNvidia

[awthomp]

This PR originates with a "more supported" delivery mechanism to cuSignal on aarch64 and involves, I think, 2 main things:

1. Replace conda4aarch64 with miniconda
2. Beef up cusignal packaging to run on Jetson CI/CD and deliver a cusignal aarch64 package. I think the blocker here is via CuPy. @jakirkham

[awthomp]

I was able to get cuSignal up and running using minforge. We can file a PR to change our installation docs and default to miniconda in the future. Nevertheless, the process was the same as delinated here:

To install miniconda on the TX2:

wget https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-Linux-aarch64.sh
bash Miniforge3-Linux-aarch64.sh

I'll submit a PR to update the README install instructions to make use of minforge over conda4aarch64.

[boonedoggle]

Confirming that I have gotten miniforge to work on the AIR-T (Jetson TX2) which is based on L4T 32.2.1.

[jakirkham]

This PR originates with a "more supported" delivery mechanism to cuSignal on aarch64 and involves, I think, 2 main things:

1. Replace conda4aarch64 with miniconda
2. Beef up cusignal packaging to run on Jetson CI/CD and deliver a cusignal aarch64 package. I think the blocker here is via CuPy. @jakirkham

Well we would actually need cudatoolkit for aarch64.[1] Also we would need an NVIDIA CUDA Docker image for aarch64.[2] These are prerequisites before being able to add the other bits needed in conda-forge to do aarch64 builds (as would be needed for cupy).

[1] - AnacondaRecipes/cudatoolkit-feedstock#3
[2] - https://gitlab.com/nvidia/container-images/cuda/-/issues/74

[JasonAtNvidia]

@boonedoggle @jakirkham

Have a look at my github. These feedstocks are about to be worked in to RAPIDS proper soon.
https://github.com/JasonAtNvidia/rapids_l4t_feedstock

Also, I have all the packages you need in my Anaconda channel. https://anaconda.org/rocketsocket/ These are about to be released in a docker container that is aarch64.

[boonedoggle]

@jakirkham curious on why an NVIDIA CUDA Docker image for aarch64 is a prerequisites on an embedded platform.

[awthomp]

This PR originates with a "more supported" delivery mechanism to cuSignal on aarch64 and involves, I think, 2 main things:

1. Replace conda4aarch64 with miniconda
2. Beef up cusignal packaging to run on Jetson CI/CD and deliver a cusignal aarch64 package. I think the blocker here is via CuPy. @jakirkham

Well we would actually need cudatoolkit for aarch64.[1] Also we would need an NVIDIA CUDA Docker image for aarch64.[2] These are prerequisites before being able to add the other bits needed in conda-forge to do aarch64 builds (as would be needed for cupy).

[1] - AnacondaRecipes/cudatoolkit-feedstock#3
[2] - https://gitlab.com/nvidia/container-images/cuda/-/issues/74

This might be due to my ignorance in how conda packages are actually built and released, but isn't there a way for these aarch64 packages to use the system CUDA toolkit installed via JetPack? In our cusignal Jetson conda install instructions, the only dependency that can't be satisfied on aarch64 is cupy, but a simple pip install cupy builds the package for the architecture and then cusignal is good to go.

[mike-wendt]

I'm working on getting Jetsons with JetPack integrated with gpuCI. Will probably be another week with other priorities at the moment. From there we can build bare-metal (so no need for the container), but I will need to create a CTK for aarch64. I have the recipes to do that, but need to get the Jetson AGXs connected to gpuCI first to do the builds.

[jakirkham]

This might be due to my ignorance in how conda packages are actually built and released, but isn't there a way for these aarch64 packages to use the system CUDA toolkit installed via JetPack?

Using the Conda cudatoolkit package was the consensus formed in issue ( conda-forge/conda-forge.github.io#687 ).

[JasonAtNvidia]

@jakirkham AnacondaRecipes/cudatoolkit-feedstock#10

[jakirkham]

Thanks Jason! 😄

[github-actions]

This issue has been marked stale due to no recent activity in the past 30d. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be marked rotten if there is no activity in the next 60d.

[github-actions]

This issue has been marked rotten due to no recent activity in the past 90d. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed.

[awthomp]

Closing issue and tracking #59