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) {
      return device_.shall_pin_to_this_device(tensorflow::wrap(op), &status);
    }
    return errors::Unimplemented("No custom device pinning implementation.");
  }

 private:
  TFE_Context* context_;
  TFE_CustomDevice device_;
  void* info_;

*   `tf.data`

    * `tf.data` now has an `autotune_options.initial_parallelism` option to
      control the initial parallelism setting used by autotune before the data
      pipeline has started running. The default is 16. A lower value reduces
      initial memory usage, while a higher value improves startup time.

## Keras

*  `keras.layers.experimental.DynamicEmbedding`
    * Added `DynamicEmbedding` Keras layer

  (*config.mutable_device_count())["CPU"] = num_cpu_devices;

  // TODO(b/113217601): This is needed for EagerContext::runner_ to use a
  // threadpool, so that we avoid the possibility of running the runner_ in the
  // threadpool of GPU event mgr, as that can trigger more callbacks to be
  // scheduled on that same threadpool, causing a deadlock in cases where the

    $env:PATH = [Environment]::GetEnvironmentVariable('PATH', 'Machine'); \
    gcloud components install docker-credential-gcr kubectl gsutil;

# Install cygwin and packages
# Running a seperate ps1 file since when running inside a Dockerfile, it does
# not work.
COPY install/install_cygwin.ps1 c:/
RUN c:/install_cygwin.ps1; \
    $old_path = [Environment]::GetEnvironmentVariable(\"PATH\", \"Machine\"); \

  // source tensors. The output gradients are used if not empty and not
  // null. The result is populated with one tensor per target element.
  // When running backward functions, builds zeros-like tensors for
  // incoming grads which are nullptrs, unless `build_default_zeros_grads`
  // is set to false.
  Status ComputeGradient(

EOF
    find . | tr '[A-Z]' '[a-z]' | sort | uniq -d | tee $BATS_FILE_TMPDIR/repeats
    [[ ! -s $BATS_FILE_TMPDIR/repeats ]]
}

# It's unclear why, but running this on //tensorflow/... is faster than running
# only on affected targets, usually. There are targets in //tensorflow/lite that
# don't pass --nobuild, so they're on their own.
#

                                        std::move(mixed_handles), status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK);
  // Can't take the shape of a parallel tensor with varying numbers of axes, but
  // running operations on them is OK.
  TF_ASSERT_OK(unknown_length_vector->Shape(&shape));
  EXPECT_THAT(*shape, ElementsAre(-1));
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> size_op(

  return run_shell([python_bin_path, '-c', 'import sys; print(sys.version[0])'])


def setup_python(environ_cp):
  """Setup python related env variables."""
  # Get PYTHON_BIN_PATH, default is the current running python.
  default_python_bin_path = sys.executable
  ask_python_bin_path = ('Please specify the location of python. [Default is '
                         '{}]: ').format(default_python_bin_path)
  while True:

test:rbe_cross_compile_macos_x86 --config=rbe_cross_compile_base
# Increase the test timeout as tests often take longer on mac.
test:rbe_cross_compile_macos_x86 --test_timeout=300,450,1200,3600
# Limit jobs to 100 to avoid running into "out of memory" issues (b/316266643)
build:rbe_cross_compile_macos_x86 --jobs=100
test:rbe_cross_compile_macos_x86 --jobs=100
# END MACOS CROSS-COMPILE CONFIGS
# END CROSS-COMPILE CONFIGS

  ~DeviceThread();

  // Requests that the worker thread execute the specified operation. Blocks
  // until the previously pending operation (a StartExecute without a Join) has
  // finished, if any.
  //
  // `cancellation_manager` must live until after `Join` finishes and pending
  // `is_async` operations finish. In addition to allowing the caller to cancel
  // the operation, its `StartCancel` method will be called if op execution