Training-free page-sparse decode in llama.cpp: Qwen3-30B-A3B is 1.72× faster at 128k on an RTX 5090, with a 1-point RULER loss
🌐 链接: https://www.reddit.com/r/LocalLLaMA/comments/1xxqdil/trainingfree_pagesparse_decode_in_llamacpp/
🏷️ 分组: Reddit|AI·自托管·VPS
🕒 时间: 2026-11-16 22:13:35
🌐 链接: https://www.reddit.com/r/LocalLLaMA/comments/1xxqdil/trainingfree_pagesparse_decode_in_llamacpp/
🏷️ 分组: Reddit|AI·自托管·VPS
🕒 时间: 2026-11-16 22:13:35