plugin_memory_allocate(TF_READ_ONLY_MEMORY_REGION_OPS_SIZE));
  ops->read_only_memory_region_ops->cleanup =
      tf_read_only_memory_region::Cleanup;
  ops->read_only_memory_region_ops->data = tf_read_only_memory_region::Data;
  ops->read_only_memory_region_ops->length = tf_read_only_memory_region::Length;

  ops->filesystem_ops = static_cast<TF_FilesystemOps*>(
      plugin_memory_allocate(TF_FILESYSTEM_OPS_SIZE));

build:windows --host_copt=/d2ReducedOptimizeHugeFunctions

# Enable the runfiles symlink tree on Windows. This makes it possible to build
# the pip package on Windows without an intermediate data-file archive, as the
# build_pip_package script in its current form (as of Aug 2023) uses the
# runfiles symlink tree to decide what to put into the Python wheel.
startup --windows_enable_symlinks

#ifndef TENSORFLOW_C_EAGER_TAPE_H_
#define TENSORFLOW_C_EAGER_TAPE_H_

// Language-agnostic gradient tape. Does not perform backpropagation, just
// maintains the data structures required to do so.

#include <stack>
#include <unordered_map>
#include <unordered_set>
#include <vector>

#include "tensorflow/core/config/flag_defs.h"
#include "tensorflow/core/config/flags.h"

  GetDims(theta_tensor, theta_dims.data());

  // Initialize auxilary data structures.
  vector<float> thetaPlus_data(num_elems);
  vector<float> thetaMinus_data(num_elems);
  AbstractTensorHandle* f_outputs[1];

  // Numerical Grad Check
  for (int i = 0; i < num_elems; i++) {
    // Get relative epsilon value
    float epsilon = theta_data[i] == 0 ? 1e-4 : std::abs(theta_data[i] * 1e-4);

  CHECK_OK(pjrt_client.status());
  auto c_api_client = down_cast<xla::PjRtCApiClient*>(pjrt_client->get());
  std::vector<int32_t> data(1, 0);
  xla::Shape shape = xla::ShapeUtil::MakeShape(xla::S32, {1});

  auto buffer = c_api_client->pjrt_c_client()->client->BufferFromHostBuffer(
      data.data(), shape.element_type(), shape.dimensions(),
      /*byte_strides=*/std::nullopt,

TFE_TensorHandle* TestMatrixTensorHandle(TFE_Context* ctx);

// Return a tensor handle containing 2D matrix containing given data and
// dimensions
TFE_TensorHandle* TestMatrixTensorHandleWithInput(TFE_Context* ctx,
                                                  float data[], int64_t dims[],
                                                  int num_dims);

// Get a Matrix TensorHandle with given float values and dimensions

void ParallelTensorDeallocator(void* data) {
  delete reinterpret_cast<ParallelTensor*>(data);
}

// Used as an argument to TFE_NewCustomDeviceTensorHandle, for computing the
// number of dimensions of a parallel tensor.
int ParallelTensorNumDims(void* data, TF_Status* status) {
  const std::vector<int64_t>* shape;
  Status s = reinterpret_cast<ParallelTensor*>(data)->Shape(&shape);
  if (!s.ok()) {

  TF_Buffer* result = TF_NewBuffer();
  const std::string& value_str = *value;
  void* data = malloc(value_str.length());
  value_str.copy(static_cast<char*>(data), value_str.length(), 0);
  result->data = data;
  result->length = value_str.length();
  result->data_deallocator = [](void* data, size_t length) { free(data); };
  return result;
}

  TFE_Context* ctx = TFE_NewContext(opts, status);
  EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  TFE_DeleteContextOptions(opts);

  TFE_ContextSetServerDef(ctx, 0, serialized.data(), serialized.size(), status);
  EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);

  const char task0_name[] = "/job:localhost/replica:0/task:0/device:CPU:0";

      EXPECT_EQ(num_values, 1);

      std::vector<TF_Buffer*> in_props_buf(num_values, TF_NewBuffer());

      TF_GetInputPropertiesList(graph_properties, node.name().c_str(),
                                in_props_buf.data(), num_values, status);
      EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

      tensorflow::OpInfo::TensorProperties in_props;
      Status s = tensorflow::BufferToMessage(in_props_buf[0], &in_props);