batteryphil/Primal-Discrete-LLM-Training

ComponentThe "Secret Sauce"MemoryZero-Shadow Training: Training without FP16 master weights.MathPrime-Grid LUT: Better precision-per-bit than standard INT4.StabilityVote Buffering: Making Gradient Accumulation work for discrete weights.

1 best points · 1 comments · 1 Hacker News post · 3 GitHub stars · View on GitHub

Hacker News discussions

• Training LLMs on 1080 Tis without shadow weights (1 points, 1 comments, by batteryphil, 2026-02-12)