You can now train your own LLMs with less memory, on your consumer GPUs
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
An independent AI researcher, Oguz Eldereli, posted on Hacker News a ternary-weight optimizer that cuts per-weight training memory from about 16 bytes (AdamW) to roughly 1 byte or less, reportedly letting a 27B model train on a 24GB RTX 3090.