num_elements *= shape[i];
  }
  std::vector<float> data(num_elements);
  for (size_t j = 0; j < num_elements; ++j) {
    data[j] = j;
  }
  DLManagedTensor dlm_in = {};
  DLTensor* dltensor_in = &dlm_in.dl_tensor;
  dltensor_in->data = data.data();
  dltensor_in->device = {kDLCPU, 0};
  dltensor_in->ndim = static_cast<int32_t>(shape.size());
  dltensor_in->dtype = {kDLFloat, 32, 1};

#ifndef TENSORFLOW_C_EAGER_DLPACK_H_
#define TENSORFLOW_C_EAGER_DLPACK_H_

#include "tensorflow/c/eager/c_api.h"

namespace tensorflow {

// PyCapsule name for DLPack Tensor
const char* const kDlTensorCapsuleName = "dltensor";

// Converts eager tensor handle to DLPack (DLManagedTensor*), and return the
// void* for further PyCapsule construction.
TF_CAPI_EXPORT extern void* TFE_HandleToDLPack(TFE_TensorHandle* h,

  DLManagedTensor* dlm_tensor = &tf_dlm_tensor_ctx->tensor;
  dlm_tensor->manager_ctx = tf_dlm_tensor_ctx;
  dlm_tensor->deleter = &DLManagedTensorDeleter;
  dlm_tensor->dl_tensor.device = tf_dlm_context;
  int ndim = tensor->dims();
  dlm_tensor->dl_tensor.ndim = ndim;
  dlm_tensor->dl_tensor.data = tf_dlm_data;
  dlm_tensor->dl_tensor.dtype = tf_dlm_type;

*   API changes under `tf.experimental.dtensor`:
    *   New API for initialization of CPU/GPU/TPU in dtensor.
        `dtensor.initialize_accelerator_system` and
        `dtensor.shutdown_accelerator_system`.
    *   The following existing API will be removed:
        `dtensor.initialize_multi_client`, `dtensor.initialize_tpu_system`, and
        `dtensor.shutdown_tpu_system`.

## Deprecations

        "TFE_TensorHandleDeviceMemorySize may not be called on a ",
        handle->TypeString(), " tensor handle.");
    return 0;
  }
  const tensorflow::Tensor* tensor;
  status->status = handle->Tensor(&tensor);
  if (!status->status.ok()) {
    return 0;
  }
  return tensor->TotalBytes();
}

TFE_Op* TFE_NewOp(TFE_Context* ctx, const char* op_or_function_name,

// TODO(srbs): Figure out if we can avoid declaring this in the public header.
// Wrapper for a tensor output of an operation executing under a tape.
//
// `GetID` returns a unique id for the wrapped tensor which is used to maintain
// a map (`tensorflow::eager::TensorTape`) from the wrapped tensor to the id of
// the op that produced it (or -1 if this tensor was watched using
// `GradientTape::Watch`.) The op_id is simply a unique index assigned to each

                                                       TF_Status* status) const;

  // Construct a parallel tensor consisting of the scalar values from `values`.
  template <typename DataType>
  std::unique_ptr<ParallelTensor> ScalarsFromSequence(
      absl::Span<const DataType> values, TFE_Context* context,
      TF_Status* status) const;

    TFE_TensorHandle* h, TF_Status* status) {
  tensorflow::TensorHandle* handle =
      TensorHandleFromInterface(tensorflow::unwrap(h));
  const tensorflow::Tensor* tensor;
  status->status = handle->Tensor(&tensor);
  if (!status->status.ok()) {
    return nullptr;
  }

  std::vector<int64_t> dev_dims = TensorShapeAsVector(*handle, &status->status);
  if (!status->status.ok()) {
    return nullptr;

  // Returns a tensor for the handle. If tensor is remote, it will be copied.
  virtual AbstractTensorInterface* Resolve(absl::Status* status) = 0;

  std::string DebugString() const override;

  // Returns a Boolean hint indicating whether callers should prefer
  // `SummarizeValue` to resolving this handle and formatting the tensor.
  //
  // For example some tensor handles may represent distributed values, in which

#include "tensorflow/core/protobuf/tensorflow_server.pb.h"

// Return a tensor handle containing a float scalar
TFE_TensorHandle* TestScalarTensorHandle(TFE_Context* ctx, float value);

// Return a tensor handle containing a int scalar
TFE_TensorHandle* TestScalarTensorHandle(TFE_Context* ctx, int value);

// Return a tensor handle containing a bool scalar
TFE_TensorHandle* TestScalarTensorHandle(TFE_Context* ctx, bool value);