星颖 发表于 2026-3-16 20:40:10

月之暗面发布Attention Residuals技术:48B模型训练效率飙升1.25倍,算力消耗直降20%

https://cdn3.ldstatic.com/optimized/4X/c/a/c/cac6687eea945c48d3200ef4e64ec7e991613419_2_409x500.jpeg
github.com (https://github.com/MoonshotAI/Attention-Residuals/blob/master/Attention_Residuals.pdf)
Attention_Residuals.pdf
页: [1]
查看完整版本: 月之暗面发布Attention Residuals技术:48B模型训练效率飙升1.25倍,算力消耗直降20%