DeepSeek open-sources inference optimizations with 60-85% faster generation

DeepSeek released a paper detailing inference optimizations that achieve 60-85% faster generation. The optimizations are open-sourced on GitHub.