[Feature Request] Prebuilt Docker images with CUDA runtime

[stevapple]

Is your feature request related to a problem? Please describe

OpenSearch provides built-in ML support through ml-commons plugin. While the plugin provides CUDA integration out of the box, the official Docker images doesn’t have CUDA built-in, so one need to build their own Docker image if they want to use GPU acceleration for machine learning.

Describe the solution you'd like

Add a -cuda variant for OpenSearch’s Docker distribution with CUDA built-in (optionally a -neuron one), and serve it via DockerHub.

Related component

Other

Describe alternatives you've considered

No response

Additional context

No response

[prudhvigodithi]

[Triage]
Thanks @stevapple can you please add more details on this and some helpful links.
Thank you

[rishabh6788]

In case you are building your own image using OS image as base image, can you please add the instructions on how you are building cuda enabled image?

[rishabh6788]

@stevapple any update on this?

[peterzhuamazon]

I think having a GPU enabled image such as opensearchproject/opensearch:2.x.x-cuda12 is doable, but this means a few more things to be cleared up before proceeding:

1. ML Team would need to be the main drive and supporter of this image as it is related to ML plugin (cc: @vamshin ).
2. A new pipeline stage would needs to be established to handle the creation of this. Plus, new tests specifically related to GPU features needs to be created by ML Team, and those hosts should be available for Infra to run the tests during release process.
3. What is the demand of such image from official repo vs. custom images from users writing their own dockerfile on top of the official image.

Happy to discuss,

Thanks.