DTensor has famously terrible eager mode performance; for example, this paper measured a 35-60% slowdown in end-to-end training performance with and without DTensor (with DTensor operations taking at least 7x longer than actually running the computation for real).
