撸
这个梯度压缩的效果比我预期好太多了!!带宽直接降一半 🚀🚀
这个梯度压缩的效果比我预期好太多了!!带宽直接降一半 🚀🚀
压缩比多少的时候开始掉精度?我这边一直没敢往上调
我测到8:1还行 再往上就看得出来了 😅
Worth noting the degradation is usually not uniform across layers. In my experience the early layers tolerate aggressive compression much better than the final ones, so a per-layer ratio beats a global one if you have the patience to tune it.