// In the TFv1 convention, TF_Tensor can hold a string serialization of
// DT_RESOURCE. The string serialization is converted back to a
// ResourceHandle during Session run where the TF_Tensor is converted to a
// Tensor.
// TFv2 does not depend on this conversion. There is no matching
// TF_TensorFromTensorV1 because the conversion to string is performed by the
// python side of Session.

  // threadpool, so that we avoid the possibility of running the runner_ in the
  // threadpool of GPU event mgr, as that can trigger more callbacks to be
  // scheduled on that same threadpool, causing a deadlock in cases where the
  // caller of event_mgr->ThenExecute() blocks on the completion of the callback
  // (as in the case of ConstOp kernel creation on GPU, which involves copying a
  // CPU tensor to GPU).

#include "tensorflow/core/platform/logging.h"

namespace tensorflow {

namespace {

// Managing context for the DLManagedTensor, will manage the lifetime of
// DLManagedTensor. When calling DLManagedTensor::deleter, it will notify the
// original framework of destruction, and this context will be deleted also.
struct TfDlManagedTensorCtx {
  TensorReference reference;
  std::vector<int64_t> shape;

  void MarkAsResult(AbstractTensorHandle* gradient) const override;

  void DeleteGradient(AbstractTensorHandle* gradient) const override;

 private:
  // The context where the aggregation op `Add` is to be created.
  AbstractContext* ctx_;
};

// Returns the number of elements in the gradient tensor.
int64_t TapeVSpace::NumElements(AbstractTensorHandle* tensor) const {

  TFE_ContextAsyncWait(ctx, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // Pack 3 variable handles into one TFE_TensorHandle.
  // When remote is false, function device is placed on task0. Handle types are
  // REMOTE, REMOTE, LOCAL on task0. When remote is true, function device is
  // placed on task1, Handle types are LOCAL, REMOTE, LOCAL on task1.
  int num_replicas = 3;

      } else {
        // This node is referenced in inputs. Currently, we place an
        // artificial restriction and require that when num_opers=-1, such
        // nodes must have a single output.
        if (node->num_outputs() != 1) {
          return InvalidArgument(
              "When `num_opers` is set to -1, nodes referenced in `inputs` "
              "must have a single output. Node ",

        PartialTensorShape first_shape;
        TF_RETURN_IF_ERROR(unwrap(tensors_[0].get())->Shape(&first_shape));
        return errors::Unimplemented(absl::StrCat(
            "Computing the shape of a ParallelTensor when the components do "
            "not all have the same rank is not supported. One tensor had "
            "shape ",
            first_shape.DebugString(), " and another had shape ",

                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    /* Given upstream grad U and a matmul op A*B, the gradients are:
     *
     *    dA = U * B.T
     *    dB = A.T * U
     *
     *    where A.T means `transpose(A)`
     */
    AbstractTensorHandle* upstream_grad = grad_outputs[0];

    // Get transpose attrs
    bool t_a;
    TF_RETURN_IF_ERROR(forward_attrs_.Get("transpose_a", &t_a));