"memory regions");

  return OkStatus();
}

// Validates the operations supplied by the plugin.
//
// Uses the 4 simpler `ValidateHelper(const TF_...*)` to validate each
// individual function table and then checks that the function table for a
// specific file type exists if the plugin offers support for creating that
// type of files.

    // Get transpose attrs
    bool t_a;
    TF_RETURN_IF_ERROR(forward_attrs_.Get("transpose_a", &t_a));

    bool t_b;
    TF_RETURN_IF_ERROR(forward_attrs_.Get("transpose_b", &t_b));

    // Conj each input
    AbstractTensorHandle* conj_output;
    std::string name = "Conj_A_MatMul_Grad";
    TF_RETURN_IF_ERROR(
        SafeConj(ctx, forward_inputs_[0], &conj_output, name.c_str()));

    *   Added a new `rerandomize_each_iteration` argument for the `tf.data.Dataset.random()` operation, which controls whether the sequence of generated random numbers should be re-randomized every epoch or not (the default behavior). If `seed` is set and `rerandomize_each_iteration=True`, the `random()` operation will produce a different (deterministic) sequence of numbers every epoch.

  // (as in the case of ConstOp kernel creation on GPU, which involves copying a
  // CPU tensor to GPU).
  // Setting a larger thread pool does not help with the Swift caller, as we use
  // a different TFE context for each thread of execution (for running graph
  // functions, and their send/recvs corountines).
  config.set_inter_op_parallelism_threads(1);

  TF_Buffer* ret = TF_NewBuffer();
  TF_CHECK_OK(MessageToBuffer(config, ret));

//
// First we filter the tape to just the subset of operations we want to
// differentiate. In the process of doing so we count how many times each Tensor
// is used as an input to an op (so we know when we're done computing gradients
// for that Tensor). We also count, for each tape entry, how many of its output
// Tensors need gradients to be computed (Tensors which are not used do not need
// any gradients to be computed).
//

        'device at: https://developer.nvidia.com/cuda-gpus. Each capability '
        'can be specified as "x.y" or "compute_xy" to include both virtual and'
        ' binary GPU code, or as "sm_xy" to only include the binary '
        'code.\nPlease note that each additional compute capability '
        'significantly increases your build time and binary size, and that '

build:macos --features=archive_param_file

# Settings for MacOS on ARM CPUs.
build:macos_arm64 --cpu=darwin_arm64
build:macos_arm64 --macos_minimum_os=11.0

# iOS configs for each architecture and the fat binary builds.
build:ios --apple_platform_type=ios
build:ios --apple_bitcode=embedded --copt=-fembed-bitcode
build:ios --copt=-Wno-c++11-narrowing
build:ios_armv7 --config=ios

  // A non-blocking version of `Execute`. After each call, `Join` must be called
  // before `StartExecute` is called again. Using `StartExecute` with `Join`
  // allows the caller to schedule computation on multiple ParallelDevices
  // without sequencing those operations (first call `StartExecute` on each
  // parallel device, then call `Join` on each; even if some of the `Join`s

    result_store_line_match = re.search(RESULT_STORE_LINK_RE, line)
    if not result_store_line_match:
      continue

    url = result_store_line_match.group(1)
    url_lines = result_store_links.setdefault(url, {})
    # Each bazel RBE invocation should produce two
    # 'Streaming build results to: ...' lines, one at the start, and one at the
    # end of the invocation.
    # If there's a failure message, it will be found in-between these two.

                       status.get()),
      variable_deleter);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  // Assign an initial value to the variable, mirroring it to each component
  // device.
  {
    TensorHandlePtr initial_value_cpu = FloatTensorHandle(20., status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());