@embedded_jdk//
^//$
@ruy//
EOF

  license_query "attr('licenses', 'notice', deps($BUILD_TARGET))" > $BATS_TEST_TMPDIR/expected_licenses
  license_query "deps($LICENSES_TARGET)" > $BATS_TEST_TMPDIR/actual_licenses

  # Column 1 is left only, Column 2 is right only, Column 3 is shared lines
  # Select lines unique to actual_licenses, i.e. extra licenses.

# downloaded TF as a zip archive).
export TFCI_GIT_DIR=$(cd $(dirname "$0"); realpath ../../)
cd "$TFCI_GIT_DIR"

# "TFCI" may optionally be set to the name of an env-type file with TFCI
# variables in it, OR may be left empty if the user has already exported the
# relevant variables in their environment. Because of 'set -o allexport' above
# (which is equivalent to "set -a"), every variable in the file is exported
# for other files to use.
#

  int64_t* p = storage;
  int storage_left = storage_size;
  for (int i = 0; i < len; ++i) {
    // shapes[i].dims() == -1 for shapes with an unknown rank.
    int64_t n = shapes[i].dims();
    num_dims[i] = n;
    dims[i] = p;
    if (n < 0) {
      continue;
    }
    if (storage_left < n) {
      status->status = InvalidArgument(

// "vspace" because a subset of these are related to a vector space, such as
// adding gradients, getting zeroes, etc. Currently cannot be implemented
// without using tensorflow python code, hence left unspecified here.
//
// Gradient is the type returned by gradient functions. In Python TF it's either
// Tensor or IndexedSlices or None, which here we map to nullptr. Gradients need

// through custom device implementations).
//
// TODO(allenl): Currently these are black boxes, but we should have some way to
// inspect values. This would let people e.g. copy over most attributes and then
// modify some based on their values.

// A reference to an op's name -> attribute mapping
typedef struct TFE_OpAttrs TFE_OpAttrs;

        tensorflow::unwrap(handles[i]);
    if (tensorflow::CustomDeviceTensorHandle::classof(unwrapped_handle)) {
      // One of the inputs we're trying to pack is on a custom device. We'll let
      // the first custom device we see handle all of the packing.
      auto* custom_device_handle =
          tensorflow::down_cast<tensorflow::CustomDeviceTensorHandle*>(
              unwrapped_handle);

    *   Various random number generation improvements:
    *   Add a fast path for default `random_uniform`
    *   `random_seed` documentation improvement.
    *   `RandomBinomial` broadcasts and appends the sample shape to the left
        rather than the right.
    *   Added `tf.random.stateless_binomial`, `tf.random.stateless_gamma`,
        `tf.random.stateless_poisson`
    *   `tf.random.stateless_uniform` now supports unbounded sampling of `int`

    return status.status;
  }

  absl::StatusOr<bool> ShallPinToThisDevice(
      const ImmediateExecutionOperation* op) override {
    TF_Status status;
    // Let this custom device choose the device to pin this op on if it
    // implements the pinning function.
    if (device_.shall_pin_to_this_device != nullptr) {

    # https://github.com/tensorflow/tensorflow/issues/10521
    # Overriding eigen strong inline speeds up the compiling of
    # conv_grad_ops_3d.cc and conv_ops_3d.cc by 20 minutes,
    # but this also hurts the performance. Let users decide what they want.
    write_to_bazelrc('build --define=override_eigen_strong_inline=true')


def config_info_line(name, help_text):