#gpu

LLMs Use Just 16 of 256 Exponents — So We Compressed the Rest Away

2× compression on Llama-3-8B — and perplexity went down.