vspace, target_tensor_ids, source_tensor_ids, sources_that_are_targets,
      output_gradients, result, /*build_default_zeros_grads*/ false));
  return absl::OkStatus();
}

// Helper functions which delegate to `AbstractOperation`, update
// the state of the ForwardOperation and call the tape as appropriate.
// These APIs are mainly to facilitate testing and are subject to change.
namespace internal {

TFCI_BAZEL_TARGET_SELECTING_CONFIG_PREFIX=linux_arm64
# Note: this is not set to "--cpu", because that changes the package name
# to tensorflow_cpu. These ARM builds are supposed to have the name "tensorflow"
# despite lacking Nvidia CUDA support.
TFCI_BUILD_PIP_PACKAGE_ARGS="--repo_env=WHEEL_NAME=tensorflow"
TFCI_DOCKER_ENABLE=1

# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
# Release jobs are very basic. They don't use any caching or RBE,
# but they do upload logs to resultstore.
# IMPORTANT: trailing slash is required on GCS URIs, as it tells gcloud to
# pretend the path is a directory.
TFCI_ARTIFACT_FINAL_GCS_ENABLE=1

#

# Mac version of Core utilities differ in usage. Since our scripts are written
# with the GNU style, we need to set GNU utilities to be default on Mac.
if [[ -n "$(which grealpath)" ]] &&  [[ -n "$(which gstat)" ]]; then
  alias realpath=grealpath
  alias stat=gstat
  # By default, aliases are only expanded in interactive shells, which means
  # that they are not substituted for their corresponding commands in shell

  // Following are helper functions to assist integrating TFRT with current
  // TF eager runtime.
  // TODO(b/172877902): These helper functions are currently used to support
  // PyFuncOp on TFRT, and might be useful for ops that directly use low
  // level TF APIs. Remove/replace the following functions when TFRT native
  // ops are implemented.

  switch (dtype.code) {
    case DLDataTypeCode::kDLBool:
      if (dtype.bits != 8) {
        return tensorflow::errors::InvalidArgument(
            "Only DLPack bools of bitwidth 8 are supported, got: ", dtype.bits);
      }
      *tf_dtype = TF_DataType::TF_BOOL;
      return absl::OkStatus();

    case DLDataTypeCode::kDLUInt:
      switch (dtype.bits) {
        case 8:

                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    /* Given upstream grad U and a matmul op A*B, the gradients are:
     *
     *    dA = U * B.T
     *    dB = A.T * U
     *
     *    where A.T means `transpose(A)`
     */
    AbstractTensorHandle* upstream_grad = grad_outputs[0];

    // Get transpose attrs

function resultstore_extract_fallback {
  # In case the main script fails somehow.
  cat <<EOF
IMPORTANT: For bazel invocations that uploaded to ResultStore (e.g. RBE), you
can view more detailed results that are probably easier to read than this log.
Try the links below:
EOF
  # Find any "Streaming build results to" line, then print the last word in it,
  # and don't print duplicates

    if (max_age_ == 0) {
      return compute_func(key, value, status);
    }

    // Note: we hold onto mu_ for the rest of this function. In practice, this
    // is okay, as stat requests are typically fast, and concurrent requests are
    // often for the same file. Future work can split this up into one lock per
    // key if this proves to be a significant performance bottleneck.
    absl::MutexLock lock(&mu_);

//
// This has three benefits:
//   * allows having newer plugins than the current core TensorFlow: the
//     additional entries in the plugin's table are just discarded;
//   * allows having older plugins than the current core TensorFlow (though
//     we are still warning users): the entries that core TensorFlow expects
//     but plugins didn't provide will be set to `nullptr` values and core