if (ic->num_outputs() <= new_src.index) {
    status->status = tensorflow::errors::OutOfRange(
        "Cannot update edge. Output index [", new_src.index,
        "] is greater than the number of total outputs [", ic->num_outputs(),
        "].");
    return;
  }
  tensorflow::shape_inference::ShapeHandle shape = ic->output(new_src.index);

  tensorflow::shape_inference::InferenceContext* ic_dst =

namespace tensorflow {
namespace gradients {

/* Returns numerical grad inside `dtheta_approx` given `forward` model and
 * parameter specified by `input_index`.
 *
 * I.e. if y = <output of the forward model> and w = inputs[input_index],
 * this will calculate dy/dw numerically.
 *
 * `use_function` indicates whether to use graph mode(true) or eager(false).
 *

//
// Must be called only with `index` a valid index in `info->ops`.
static Status ValidateAndRegisterFilesystems(
    const TF_FilesystemPluginInfo* info, int index) {
  TF_RETURN_IF_ERROR(ValidateScheme(info->ops[index].scheme));
  TF_RETURN_IF_ERROR(ValidateABI(&info->ops[index]));
  ValidateAPI(&info->ops[index]);  // we just warn on API number mismatch

# CuDNN: https://docs.nvidia.com/deeplearning/sdk/cudnn-install/index.html#ubuntu-network-installation
libcudnn8=8.9.6.50-1+cuda12.2
libcudnn8-dev=8.9.6.50-1+cuda12.2

# This can be removed once NVIDIA publishes a cuda-12.3.2 Docker image.
# For now it ensures that we install at least version 12.3.107 of PTXAS,
# since 12.3.103 has a bug.

TF_CAPI_EXPORT extern int TFE_TensorDebugInfoOnDeviceNumDims(
    TFE_TensorDebugInfo* debug_info);

// Returns the number of elements in dimension `dim_index`.
// Tensor representation on device can be transposed from its representation
// on host. The data contained in dimension `dim_index` on device
// can correspond to the data contained in another dimension in on-host
// representation. The dimensions are indexed using the standard TensorFlow

TFCI_DOCKER_ENABLE=1
TFCI_DOCKER_IMAGE=gcr.io/tensorflow-sigs/build-arm64:tf-2-16-multi-python
TFCI_DOCKER_PULL_ENABLE=1
TFCI_DOCKER_REBUILD_ARGS="--target=tf ci/official/containers/linux_arm64"
TFCI_INDEX_HTML_ENABLE=1
TFCI_LIB_SUFFIX="-cpu-linux-arm64"
TFCI_OUTPUT_DIR=build_output
TFCI_WHL_AUDIT_ENABLE=1
TFCI_WHL_AUDIT_PLAT=manylinux2014_aarch64
TFCI_WHL_BAZEL_TEST_ENABLE=1
TFCI_WHL_SIZE_LIMIT=240M

namespace {

using ::tensorflow::string;

void ReplaceTaskInServerDef(tensorflow::ServerDef* server_def, int task_index) {
  tensorflow::JobDef* job_def = server_def->mutable_cluster()->mutable_job(0);
  int port = tensorflow::testing::PickUnusedPortOrDie();
  job_def->mutable_tasks()->at(task_index) =
      tensorflow::strings::StrCat("localhost:", port);
}

void CheckTFE_TensorHandleHasFloats(TFE_TensorHandle* handle,

                                        int index) {
  DCHECK(index >= 0 && index < shape_list->num_items);
  TF_ShapeAndType& shape = shape_list->items[index];
  DCHECK(shape.dims == nullptr) << "Shape at " << index << " is already set!";
  shape.num_dims = -1;
  shape.dims = nullptr;
}

void TF_ShapeAndTypeListSetDtype(TF_ShapeAndTypeList* shape_list, int index,
                                 TF_DataType dtype) {

  CHECK(!has_packed_input || func);
  tensorflow::ServerDef server_def = GetServerDef(3);

  // This server def has the task index set to 0.
  string serialized = server_def.SerializeAsString();

  server_def.set_task_index(1);
  std::unique_ptr<tensorflow::GrpcServer> worker_server1;
  ASSERT_TRUE(tensorflow::GrpcServer::Create(

 public:
  virtual void Clear() = 0;

  // Returns the inputs of this op.
  virtual absl::Span<ImmediateExecutionTensorHandle* const> GetInputs()
      const = 0;
  virtual Status SetInput(size_t index,
                          ImmediateExecutionTensorHandle* input) = 0;

  virtual ImmediateExecutionContext* GetContext() const = 0;

  // Following two methods are used to support custom device.