註冊並分享邀請連結,可獲得影片播放與邀請獎勵。

Crusoe
@CrusoeAI
The AI factory company. We are on a mission to accelerate the abundance of energy and intelligence.
加入 January 2019
373 正在關注    19.9K 粉絲
LoRA freezes the base weights. So why store them in bf16? We keep them in mxfp4 in HBM and dequantize inside the kernel. The optimizer only touches the LoRA adapters. Up to 75% less base-weight memory, with no accuracy loss in our internal evals. ➕ The math:
顯示更多