If you have a 3090, or other 30xx for local LLMs, I have something for you

I have a custom fork of llama.cpp designed around the ampere architecture specifically (though many of the upgrades also translate to faster performance of blackwell + lovelace). The recommended config supports 90+ TPS (for agentic/coding, at temp 1; greedy will of course be faster) through 100K tok

中文全文翻译正在补齐，可先查看下方英文原文。

I have a custom fork of llama.cpp designed around the ampere architecture specifically (though many of the upgrades also translate to faster performance of blackwell + lovelace). The recommended config supports 90+ TPS (for agentic/coding, at temp 1; greedy will of course be faster) through 100K tok