dltensor_out->shape[i + 1] * dltensor_out->strides[i + 1]);
      }
    }
  }
  const float* data_in = static_cast<const float*>(dltensor_in->data);
  const float* data_out = static_cast<const float*>(dltensor_out->data);
  for (size_t j = 0; j < num_elements; ++j) {
    EXPECT_EQ(data_out[j], data_in[j]);
  }

  TFE_CallDLManagedTensorDeleter(dlm_out);
  TFE_DeleteTensorHandle(handle);
}

     */

    AbstractTensorHandle* upstream_grad = grad_outputs[0];
    DCHECK(upstream_grad);

    // Recover data format from forward pass for gradient.
    std::string data_format;
    TF_RETURN_IF_ERROR(forward_attrs_.Get("data_format", &data_format));

    // Grad for A
    grad_inputs[0] = upstream_grad;
    grad_inputs[0]->Ref();

    // Grad for bias

      auto* out_shape_and_type = handle_data.add_shape_and_type();
      ic->ShapeHandleToProto(p.shape, out_shape_and_type->mutable_shape());
      out_shape_and_type->set_dtype(p.dtype);
      *out_shape_and_type->mutable_type() = p.type;
    }
  }
  string str_data;
  handle_data.SerializeToString(&str_data);

  TF_Buffer* result = TF_NewBufferFromString(str_data.c_str(), str_data.size());
  return result;
}

                  server_def, tensorflow::Env::Default(), &worker_server2)
                  .ok());
  ASSERT_TRUE(worker_server2->Start().ok());

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

  BasicTestsForTwoDevices(context.get(),

  tsl::Set_TF_Status_from_Status(
      s, unwrap(op)->Execute(
             absl::MakeSpan(reinterpret_cast<AbstractTensorHandle**>(
                                unwrap(o)->outputs.data()),
                            unwrap(o)->outputs.size()),
             &num_outputs));
}

void TF_DeleteAbstractFunction(TF_AbstractFunction* func) {
  unwrap(func)->Unref();
}

  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,

};

int64_t LoggedTensorDim(void* data, int dim_index, TF_Status* status) {
  return TFE_TensorHandleDim(reinterpret_cast<LoggedTensor*>(data)->tensor,
                             dim_index, status);
}

int LoggedTensorNumDims(void* data, TF_Status* status) {
  return TFE_TensorHandleNumDims(reinterpret_cast<LoggedTensor*>(data)->tensor,
                                 status);
}

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;
}