// which returns a dict of string key -> EITHER:
// 1. python callable (taking a checkpoint key) returning SaveableObject OR
// 2. variable (partitioned/resource/reference or otherwise)
// https://github.com/tensorflow/tensorflow/blob/ddc1bbad3dfd4a089eb96014f26cc16664b1b2f8/tensorflow/python/training/saving/saveable_object.py#L58.

  }

  /**
   * Guides the allowed concurrency among update operations. Used as a hint for internal sizing. The
   * table is internally partitioned to try to permit the indicated number of concurrent updates
   * without contention. Because assignment of entries to these partitions is not necessarily
   * uniform, the actual concurrency observed may vary. Ideally, you should choose a value to

  }

  /**
   * Guides the allowed concurrency among update operations. Used as a hint for internal sizing. The
   * table is internally partitioned to try to permit the indicated number of concurrent updates
   * without contention. Because assignment of entries to these partitions is not necessarily
   * uniform, the actual concurrency observed may vary. Ideally, you should choose a value to

def TPUResourceReadsWritesPartitioningPass : Pass<"tf-tpu-resource-partition", "mlir::func::FuncOp"> {
  let summary = "Partitions unpartitioned resource read/write to partitioned resource variables.";

  let description = [{
    This pass creates individual resource reads/writes from the unpartitioned
    resource variable (from `tf.TPUPartitionedInput`) to individual partitioned

// RUN: tf-opt %s -split-input-file -verify-diagnostics -tf-tpu-reorder-replicate-partitioned-inputs | FileCheck %s

// CHECK-LABEL:func @simple
// CHECK-SAME: ([[ARG0:%.*]]: tensor<!tf_type.resource<tensor<10x3xf32>>>, [[ARG1:%.*]]: tensor<!tf_type.resource<tensor<10x3xf32>>>, [[ARG2:%.*]]: tensor<!tf_type.resource<tensor<10x3xf32>>>, [[ARG3:%.*]]: tensor<!tf_type.resource<tensor<10x3xf32>>>)

        from = partitionPoint + 1;
      }
    }
  }

  /**
   * Performs a partition operation on the slice of {@code array} with elements in the range [{@code
   * from}, {@code to}]. Uses the median of {@code from}, {@code to}, and the midpoint between them
   * as a pivot. Returns the index which the slice is partitioned around, i.e. if it returns {@code

        from = partitionPoint + 1;
      }
    }
  }

  /**
   * Performs a partition operation on the slice of {@code array} with elements in the range [{@code
   * from}, {@code to}]. Uses the median of {@code from}, {@code to}, and the midpoint between them
   * as a pivot. Returns the index which the slice is partitioned around, i.e. if it returns {@code

emulates renames through copy and delete APIs. This interaction pattern causes significant loss of performance because of the write amplification. _Netflix_, for example, developed two new staging committers - the Directory staging committer and the Partitioned staging committer - to take full advantage of native object storage operations. These committers do not require rename operation. The two staging committers were evaluated, along with another new addition called the Magic committer for benchmarking....

   * <p><b>Note:</b> if {@code iterable} is a {@link List}, use {@link Lists#partition(List, int)}
   * instead.
   *
   * @param iterable the iterable to return a partitioned view of
   * @param size the desired size of each partition (the last may be smaller)
   * @return an iterable of unmodifiable lists containing the elements of {@code iterable} divided
   *     into partitions
   * @throws IllegalArgumentException if {@code size} is nonpositive

   * <p><b>Note:</b> if {@code iterable} is a {@link List}, use {@link Lists#partition(List, int)}
   * instead.
   *
   * @param iterable the iterable to return a partitioned view of
   * @param size the desired size of each partition (the last may be smaller)
   * @return an iterable of unmodifiable lists containing the elements of {@code iterable} divided
   *     into partitions
   * @throws IllegalArgumentException if {@code size} is nonpositive