well over the past 3 days i have been setting up tiny messy public build infrastructure for tensorflow for old pre-avx cpus i was doing it using github actions to get the business-energy-support and so it could run on its own i got pretty far! check out https://github.com/karl3wm/tensorflow-builder/actions?page=3 today i ran out of gratuity minutes for the month on github actions. it says i have to pay to continue, i visit billing it says they would charge $200 for this if i wasn't a classy open-source public developer it will keep automatically trying where i left it to run every sunday, but i had only 2 builds enabled for testing. that might move those two builds forward by the end of next month. i didn't let any of the builds get so far as to actually produce downloadable wheels so that part of the builds has not been tested at all. i left off making macos work.
if i continue this this month i'd either move to a different ci/forge or pay github money. but it was an intense run for the past 3 days! i have tried to make tensorflow work on a low-end system so many times over the years and never made it do so; i imagine many many others have had a similar experience. it was pleasant to have a remote ci infrastructure holding part of the challenge of keeping it running for a long time and having a reasonably powerful system to perform the build. i just wanted to try birdnet, a tiny 50MB model that reliably identifies nearby birds. but it's written in tensorflow where the public builds assume avx. and my cpu doesn't seem to have avx. so i figured it was more useful to try to make public tensorflow builds for everyone without avx than to continue with birdnet. didn't quite get there though. good to spend my time doing soemthing else though.
