返回
RCreddit.com
20
·19小时前·RSS
暂不在当前实时榜单

Warning: llama.cpp --lazy-mode default changed to auto - large tables may stay on disk

查看原文
LlamaQwen

热度趋势

趋势数据积累中

百分比基于当前可用热度信号,而非评论数或独立用户人数。

AI 摘要

The default behavior of llama.cpp's --lazy-mode has changed to "auto" with b10726. This means large tables, such as the 51B-parameter PLE n-gram embedding table of Qwen 3.8 Flash Next, may now remain on disk and be memory-mapped, rather than loaded into RAM. This change can lead to significant performance penalties, including a 50% pp speed penalty and a 15% token generation speed penalty. Users with sufficient RAM are advised to use the flag --lazy-mode off to avoid these performance issues.