Comment by ronsor

No, the problem is that with training, you do care about latency, and you need a crap-ton of bandwidth too! Think of the all_gather; think of the gradients! Inference is actually easier to distribute.