std::pair<std::unique_ptr<TensorSliceReader::VarToShapeMap>,
            std::unique_ptr<TensorSliceReader::VarToDataTypeMap> >
  BuildV2VarMaps();

  // Invariant: exactly one of "reader_" and "v2_reader_" is non-null.
  std::unique_ptr<TensorSliceReader> reader_;
  std::unique_ptr<BundleReader> v2_reader_;

  std::unique_ptr<TensorSliceReader::VarToShapeMap> var_to_shape_map_;

  TF_Graph* graph = TF_NewGraph();

  // Make two scalar constants.
  TF_Operation* one = ScalarConst(1, graph, s, "one");
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  TF_Operation* two = ScalarConst(2, graph, s, "two");
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  // Add oper.
  TF_Operation* add = Add(one, two, graph, s, "add");
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  size_t finish = block_size_ * ((offset + n) / block_size_);
  if (finish < offset + n) {
    finish += block_size_;
  }
  size_t total_bytes_transferred = 0;
  // Now iterate through the blocks, reading them one at a time.
  for (size_t pos = start; pos < finish; pos += block_size_) {
    Key key = std::make_pair(filename, pos);
    // Look up the block, fetching and inserting it if necessary, and update the

        "lib/gcc/arm-rpi-linux-gnueabihf/**",
        "include/**",
    ]),
)

filegroup(
    name = "aarch64_compiler_pieces",
    srcs = glob([
        "aarch64-none-linux-gnu/**",
        "libexec/**",
        "lib/gcc/aarch64-none-linux-gnu/**",
        "include/**",
    ]),
)

filegroup(
    name = "compiler_components",
    srcs = [
        ":ar",
        ":as",
        ":gcc",

  TFE_DeleteContext(ctx);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  const int num_devices = TF_DeviceListCount(devices);
  EXPECT_GE(num_devices, 1) << "At least one CPU device should exist";
  for (int i = 0; i < num_devices; ++i) {
    EXPECT_NE("", TF_DeviceListName(devices, i, status)) << i;
    EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  }

  }
  return absl::OkStatus();
}

// TODO(agarwal): use an automatic mechanism for handling None arguments to
// gradient functions.
//
// Some gradient functions can accept None arguments for gradients. The
// following maps the operation name to the indices at which the corresponding
// gradient function can accept None values. e.g. FusedBatchNorm outputs 5
// values and hence receives 5 gradient values during backprop. However the

    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.

    if not current_url:
      url_lines['start'] = i
    elif current_url == url:
      url_lines['end'] = i
    else:

# simple way to parameterize this from bazel, hence this helper to
# "normalize" the srcjar layout.
#
# Arguments:
#   src_jar - path to the original srcjar
#   dest_jar - path to the destination
# Returns:
#   None
function cp_normalized_srcjar() {
  src_jar="$1"
  dest_jar="$2"
  tmp_dir=$(mktemp -d)
  cp "${src_jar}" "${tmp_dir}/orig.jar"
  pushd "${tmp_dir}"
  # Extract any src/ files
  jar -xf "${tmp_dir}/orig.jar" src/

**Raspberry Pi 2 and 3**      | [![Status](https://storage.googleapis.com/tensorflow-kokoro-build-badges/rpi23-py3.svg)](https://storage.googleapis.com/tensorflow-kokoro-build-badges/rpi23-py3.html)           | [Py3](https://storage.googleapis.com/tensorflow-nightly/tensorflow-1.10.0-cp34-none-linux_armv7l.whl)

        of built-in training loops such as `tf.keras` `compile` and `fit`, we
        expect reduction value to be `SUM` or `NONE`. Using `AUTO` in that case
        will raise an error.
    *   `NONE`: Weighted losses with one dimension reduced (axis=-1, or axis
        specified by loss function). When this reduction type used with built-in