Tom Natan
Tom Natan
Hey! The transpose perm is used to reorder the axes in the mesh so they can be used in any order. So in the above example 4 you actually need...
That said, if you already have a mesh with named axes, how hard would it be to go directly to shardy shardings like JAX does? Or is this just for...
This is how I see example 4 in this thread, maybe it’s outdated? 1. PartitionSpec = ('y', None) a. dims = [4, 1, 2] b. transpose_perm = [0, 1] On...
Hey! First of all, there are multiple ways of representing certain shardings in v2, and sometimes you can canonicalize into a more compact version - when you have consecutive dims...