Lossless LLM compression for efficient GPU inference via dynamic-length float

Lossless LLM compression for efficient GPU inference via dynamic-length float. 117 comments, 411 points on Hacker News.