Has anyone with 4x sparks tried running 2D parallelism (TPxPP)?
That would require using Tensor Parallelism (TP) over CX-7, and Pipeline Parallelism (PP) over the RJ-45 cable?
Pipeline Parallelism requires much less bandwidth, since we are splitting the model at a boundary, whereas Tensor Parallelism shards per layer.
Has anyone tested this? Curious about the PP bubbles… and for those of us too cheap to buy a fancy network switch…