@yimlin I’ll be breaking this PR up in to a set of fixes:
I’ve been mainly working on fixing regressions and adding new features for easier debugging.
My main goal will be to add a new llamacpp backend to make LLMs faster.
I’ll also be updating the cuda versions.
Overall, dl4j as a framework going forward will be simpler. There’s been a lot of feature creep over the years and other frameworks dominate the industry.
Focusing more on being an easy to use framework that can run models people expect at a performance they expect will be the baseline going forward.