After a long time waiting the release will be coming out. You can get a preview of what to expect here with the links below:
Some upcoming things to expect:
Dl4j as it stands will NOT be what it was. The focus is much narrower now. No automl, no reinforcement learning, save RLHF and newer techniques, no spark. Dl4j as an api is in maintenance mode.
This is a substantial rewrite and should be treated as a preview.
You can expect test coverage to be good. It’s been a long time coming. I wanted to put out a framework that was at least usable in the age of LLMs. This means import for pytorch, ggml, multi gpu support, huggingface support,good performance and most of all: a proper deep learning compiler framework, a proper finetuning framework, (LORA,distillation and co).
Things are still being finished up yet. There’s a lot of polish to do. I’ve been working on publishing releases to snapshots for now. More soon. I will be doing polish work but at least feel like there’s enough docs out there that people can reasonably browse if interested.
@AllenWGX yep you can! Most of what you’ll have is import GGML → finetune since that’s what most people want but what you’d expect will be there (qwen, kimi,multi modal models,..) basically whatever shows up on hugging face should work.
