Data Parallel MNIST with DTensor and TensorFlow Core

You’ll train a simple MLP on MNIST using TensorFlow Core plus DTensor in a data-parallel setup: create a one-dimensional mesh (“batch”), keep model weights replicated (DVariables), shard the global batch across devices via pack/repack, and run a standard loop with tf.GradientTape, custom Adam, and accuracy/loss metrics. The code shows how mesh/layout choices propagate through ops, how to write DTensor-aware layers, and how to evaluate/plot results. Saving is limited today—DTensor models must be fully replicated to export, and saved models lose DTensor annotations.