} else {
    std::array<Gradient*, 2> to_aggregate;
    to_aggregate[0] = tangent;
    to_aggregate[1] = existing->second;
    // AggregateGradients steals a reference to each of its arguments. We
    // MarkAsResult on `tangent` above so we don't steal a reference to it.
    existing->second = vspace_.AggregateGradients(to_aggregate);
  }
}

  // and not for correctness. The only downside of keeping this 1 seems to be
  // that the gradient accumulation is unbounded and we will never
  // aggressively aggregate accumulated gradients to recover memory.
  // Revisit and fix.
  return 1;
}

// Consumes references to the tensors in the gradient_tensors list and returns
// a tensor with the result.

        `MultiWorkerMirroredStrategy`.
*   `tf.keras`:
    *   Added `experimental_aggregate_gradients` argument to
        `tf.keras.optimizer.Optimizer.apply_gradients`. This allows custom
        gradient aggregation and processing aggregated gradients in custom
        training loop.
    *   Allow `pathlib.Path` paths for loading models via Keras API.
*   `tf.function`/AutoGraph:

        "//tensorflow/compiler/mlir/quantization/stablehlo/python:pywrap_quantization_lib_impl",
        "//tensorflow/compiler/mlir/quantization/tensorflow/calibrator:custom_aggregator_op",
        "//tensorflow/compiler/mlir/quantization/tensorflow/python:quantize_model_cc_impl",
        "//tensorflow/compiler/mlir/quantization/tensorflow:passes",
        "//tensorflow/compiler/mlir/tensorflow",