TFE_Context* ctx, const char* device_name, TF_DataType, const int64_t* dims,
    int num_dims, void* data, size_t len,
    void (*deallocator)(void* data, size_t len, void* arg),
    void* deallocator_arg, TF_Status* status);

// Retrieves the address space (i.e. job, replia, task) of the local host and
// saves it in the buffer.
TF_CAPI_EXPORT extern void TFE_HostAddressSpace(TFE_Context* ctx,

        TF_NewTensor(
            datatype_enum, /*dims=*/nullptr, /*num_dims=*/0,
            device_value.release(), sizeof(DataType),
            [](void* data, size_t, void* arg) {
              delete reinterpret_cast<DataType*>(data);
            },
            nullptr),
        TF_DeleteTensor);
    // TODO(allenl): Here and when executing regular operations, we could hold

//   are not optimizing for convenience.
// * We make assumption that one session has one graph.  This should be
//   fine since we have the ability to run sub-graphs.
// * We could allow NULL for some arguments (e.g., NULL options arg).
//   However since convenience is not a primary goal, we don't do this.
// * Devices are not in this API.  Instead, they are created/used internally
//   and the API just provides high level controls over the number of

// `tensor_id`, associated with `session`. There must be a graph node named
// "fifo_queue_enqueue_<tensor_id>", to be executed by this API call. It reads
// from a placeholder node "arg_tensor_enqueue_<tensor_id>".
//
// `tensor` is still owned by the caller. This call will be blocked if the queue
// has reached its capacity, and will be unblocked when the queued tensors again

  // Tensor creation functions
  virtual AbstractTensorInterface* CreateTensor(
      DataType dtype, absl::Span<const int64_t> dim_sizes) = 0;

  typedef void (*MemoryReleaser)(void* data, size_t len, void* arg);

  // Create a tensor instance from the given data buffer and description.
  // `memory_releaser` will be called on destruction, and it's responsible for
  // cleaning up the underlying buffer.