GPU运算如何优化环境达到最高速率？

同样的环境 3090 size:66 0.59 steps/s
3060 size:16 0.65 steps/s
实在是搞不懂 而且CUDA11.6时 运算速率会随着模型增大而逐步变慢，3060一开始时1.1 steps/s 运行二十几个小时后逐步掉到0.65 steps/s
requirements.txt里有好几个支持的版本是在30系显卡有冲突或者不兼容的（也有可能是和win11不匹配），作者能否重新配置一下版本信息？

Start by reviewing requirements.txt alongside the reported CUDA 11.6, Windows 11, and GPU configurations, then try to reproduce the steps-per-second decline. Done means documenting compatible dependency versions and determining whether the reported slowdown or 30-series incompatibility is reproducible.

Written by the indexing model from the issue text.