AXN — Native GCN Training in Chapel with PGAS and GPU Acceleration

[arthureleven]

Built a GCN training framework natively in Chapel — forward/backward pass, neighborhood sampling, Adam, mini-batch training on ogbn-products (126M edges), and a PGAS distributed forward pass via coforall over locales.
Repo: https://github.com/arthureleven/axn
Looking for feedback and potentially an arXiv endorser in cs.LG or cs.DC.

[bradcray]

Hi @arthureleven — Thanks for making us aware of this! I wanted to suggest potentially cross-posting this on our community Discourse and/or Discord channels to reach a wider audience—see https://chapel-lang.org/forums/ for links. We've only started using GitHub discussions recently, and primarily just for planning purposes, so it doesn't have a very large audience yet (that I'm aware of, anyway). For example, I think several developers on our team at HPE don't receive notifications about posts here—it took a colleague pointing your post out to me for me to notice it.