https://www.tensorflow.org/xla/custom_call for details.
    *   XLA:GPU reductions are deterministic by default (reductions within
        `jit_compile=True` are now deterministic).
    *   XLA:GPU works with Horovod (OSS contribution by Trent Lo from NVidia)
    *   XLA:CPU and XLA:GPU can compile tf.unique and tf.where when shapes are
        provably correct at compile time.
*   `tf.saved_model.save`: