// allows code like "tf.constant(1.)" or "tf.reduce_sum(..., axis=1)"
        // (where the value starts on the host), without allowing other implicit
        // copies/broadcasts. Other implicit copies may be supported eventually,
        // but need special handling for gradients (gradient of copy-on is not
        // just copy-off but includes a sum) and consideration of performance.
        //

// LINT.IfChange
// Note: Keep in sync with internal copy of enum in eager/context.h.
typedef enum TFE_ContextDevicePlacementPolicy {
  // Running operations with input tensors on the wrong device will fail.
  TFE_DEVICE_PLACEMENT_EXPLICIT = 0,
  // Copy the tensor to the right device but log a warning.
  TFE_DEVICE_PLACEMENT_WARN = 1,
  // Silently copy the tensor, which has a performance cost since the operation

struct TfDlManagedTensorCtx {
  TensorReference reference;
  std::vector<int64_t> shape;
  std::vector<int64_t> strides;
  DLManagedTensor tensor;

  explicit TfDlManagedTensorCtx(const TensorReference& ref) : reference(ref) {}
};

// Gets tensor from eager tensor handle.
const Tensor* GetTensorFromHandle(TFE_TensorHandle* h, TF_Status* status) {
  if (h == nullptr) {

AbstractTensorHandle* TapeTensor::ZerosLike() const { return nullptr; }

class TapeVSpace
    : public eager::VSpace<AbstractTensorHandle, GradientFunction, TapeTensor> {
 public:
  explicit TapeVSpace(AbstractContext* ctx) : ctx_(ctx) {}
  ~TapeVSpace() override {}

  // Returns the number of elements in the gradient tensor.
  int64_t NumElements(AbstractTensorHandle* tensor) const override;

  TFE_Execute(matmul.get(), &retval, &num_retvals, status.get());
  EXPECT_TRUE(executed);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TFE_DeleteTensorHandle(retval);

  // Explicit placement still forces the op onto the requested device
  matmul.reset(MatMulOp(context.get(), hcustom0.get(), hcpu.get()));
  TFE_OpSetDevice(matmul.get(), "/job:localhost/replica:0/task:0/device:CPU:0",

    grad_inputs[0]->Ref();
    grad_inputs[1]->Ref();
    return absl::OkStatus();
  }
  ~AddGradientFunction() override {}
};

class ExpGradientFunction : public GradientFunction {
 public:
  explicit ExpGradientFunction(AbstractTensorHandle* exp) : exp_(exp) {
    exp->Ref();
  }
  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,

  // otherwise be the name of this custom device. Ops are placed onto a custom
  // device if any of their inputs are on that custom device, but custom devices
  // are free to set a bad status in order to require explicit placement.
  void (*execute)(const TFE_Op* op, int* num_outputs,
                  TFE_TensorHandle** outputs, TF_Status* s, void* device_info);

  // Method to delete a device.

// GraphTensor wraps a `TF_Output`, i.e. a pointer to TF_Operation and the index
// into the list of outputs for the operation.
class GraphTensor : public TracingTensorHandle {
 public:
  explicit GraphTensor(TF_Output output, TF_Graph* graph)
      : TracingTensorHandle(kGraph), output_(output), graph_(graph) {}

  tensorflow::DataType DataType() const override {

// LINT.IfChange
// Note: Keep in sync with exported copy of enum in eager/c_api.h.
enum ContextDevicePlacementPolicy {
  // Running operations with input tensors on the wrong device will fail.
  DEVICE_PLACEMENT_EXPLICIT = 0,
  // Copy the tensor to the right device but log a warning.
  DEVICE_PLACEMENT_WARN = 1,
  // Silently copy the tensor, which has a performance cost since the operation

  // functions (and hence the tensors they keep alive). Instead, everything
  // is deleted in ~GradientTape. Persistent GradientTapes are useful when
  // users want to compute multiple gradients over the same tape.
  explicit GradientTape(bool persistent) : persistent_(persistent) {}
  ~GradientTape() {
    for (const auto& pair : op_tape_) {
      pair.second.backward_function_deleter(pair.second.backward_function);
    }
  }