read_only_memory_region_ops_;
  std::function<void*(size_t)> plugin_memory_allocate_;
  std::function<void(void*)> plugin_memory_free_;
  ModularFileSystem(const ModularFileSystem&) = delete;
  void operator=(const ModularFileSystem&) = delete;
};

class ModularRandomAccessFile final : public RandomAccessFile {
 public:
  ModularRandomAccessFile(const std::string& filename,

    if (!metadata) {
      TF_SetStatusFromGCSStatus(metadata.status(), status);
      return;
    }
    // We have to delete the temporary object after composing.
    auto delete_status = gcs_client->DeleteObject(bucket, temporary_object);
    if (!delete_status.ok()) {
      TF_SetStatusFromGCSStatus(delete_status, status);
      return;
    }
    // We truncate the data that are already uploaded.
    if (!outfile->truncate()) {

fi

# Repair wheels with auditwheel and delete the old one.
if [[ "$TFCI_WHL_AUDIT_ENABLE" == "1" ]]; then
  python3 -m auditwheel repair --plat "$TFCI_WHL_AUDIT_PLAT" --wheel-dir . *.whl
  # if the wheel is already named correctly, auditwheel won't rename it. so we
  # list all .whl files by their modification time (ls -t) and delete anything
  # other than the most recently-modified one (the new one).

    strategy:
      matrix:
        python-version: [python3.9, python3.10, python3.11, python3.12]
    permissions:
      contents: read
      pull-requests: write
    steps:
      - name: Delete unnecessary tools folder
        run: |
          df -h
          rm -rf /opt/hostedtoolcache
          df -h
      -
        name: Checkout

// GraphOperation wraps and populates a TF_OperationDescription.
class GraphOperation : public TracingOperation {
 public:
  explicit GraphOperation(TF_Graph* g) : TracingOperation(kGraph), g_(g) {}
  void Release() override { delete this; }
  Status Reset(const char* op, const char* raw_device_name) override {
    if (op_) {
      return errors::FailedPrecondition("Reset called on already built op.");
    }
    if (raw_device_name) {

  memcpy(&dnumerical[0], TF_TensorData(numerical_tensor),
         TF_TensorByteSize(numerical_tensor));

  for (int j = 0; j < num_grad; j++) {
    ASSERT_NEAR(dnumerical[j], expected_grad[j], abs_error);
  }
  delete[] dnumerical;
  TF_DeleteTensor(numerical_tensor);
}

Status MatMulModel(AbstractContext* ctx,
                   absl::Span<AbstractTensorHandle* const> inputs,

  ///
  /// DEFAULT IMPLEMENTATION: Does a BFS traversal of tree rooted at `path`,
  /// deleting entries as needed. Needs `path_exists`, `get_children`,
  /// `is_directory`, `delete_file`, and `delete_dir`.
  void (*delete_recursively)(const TF_Filesystem* filesystem, const char* path,
                             uint64_t* undeleted_files,
                             uint64_t* undeleted_dirs, TF_Status* status);

  ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);
  TF_GraphCopyFunction(host_graph_, func_, nullptr, s_);
  ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);

  // Delete the gradient func.
  // It is safe to delete after adding a copy to host graph.
  TF_DeleteFunction(grad_func);

  // Check that GraphDef did not change
  GraphDef gdef2;
  GetGraphDef(host_graph_, &gdef2);

    TFE_ContextOptions* options, TFE_ContextDevicePlacementPolicy policy) {
  options->device_placement_policy = policy;
}

void TFE_DeleteContextOptions(TFE_ContextOptions* options) { delete options; }

TFE_Context* TFE_NewContext(const TFE_ContextOptions* opts, TF_Status* status) {
  if (opts->use_tfrt) {
    status->status = tensorflow::errors::Unimplemented("TFRT is not supported");
    return nullptr;

        *   Compared to `tf.lookup.experimental.DenseHashTable` this offers
            lower overall memory usage, and a cleaner API. It does not require
            specifying a `delete_key` and `empty_key` that cannot be inserted
            into the table.
    *   Added support for specifying number of subdivisions in all reduce host