LCLMs compress LLM context before decode — 8.8x faster at 16x compression, beating every KV cache method tested. Open-sourced by NYU and Columbia.
It’s well known that the difference in executable size between a compiled binary and one hand-written in optimized assembler ...