return ptr->getKind() == kGraph || ptr->getKind() == kMlir;
  }
};

// An abstract operation describes an operation by its type, name, and
// attributes. It can be "executed" by the context with some input tensors.
// It is allowed to reusing the same abstract operation for multiple execution
// on a given context, with the same or different input tensors.
class TracingOperation : public AbstractOperation {

==============================================================================*/
#ifndef TENSORFLOW_C_EAGER_TRACING_UTILS_H_
#define TENSORFLOW_C_EAGER_TRACING_UTILS_H_

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

namespace tensorflow {
namespace tracing {
Status MaybeSetOpName(AbstractOperation*, const char* op_name);
}  // namespace tracing
}  // namespace tensorflow

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
  // will be blocked till the copy completes. This is the default policy.

                         absl::Span<AbstractTensorHandle* const> grad_outputs,
                         absl::Span<AbstractTensorHandle*> grad_inputs) = 0;
  virtual ~GradientFunction() {}
};

// Metadata from the forward operation that is made available to the
// gradient registerer to instantiate a GradientFunction.
struct ForwardOperation {
 public:
  string op_name;
  std::vector<AbstractTensorHandle*> inputs;

   *                 v   v
   */
  // Define
  TF_Operation* feed1 = Placeholder(func_graph_, s_, "feed1");
  TF_Operation* feed2 = Placeholder(func_graph_, s_, "feed2");
  Define(-1, {}, {feed1, feed2}, {feed2, feed1}, {});

  // Use, run, and verify
  TF_Operation* two = ScalarConst(2, host_graph_, s_);
  TF_Operation* func_feed = Placeholder(host_graph_, s_);
  TF_Operation* func_op = Use({two, func_feed});

limitations under the License.
==============================================================================*/
#ifndef TENSORFLOW_C_EXPERIMENTAL_GRADIENTS_TAPE_TAPE_OPERATION_H_
#define TENSORFLOW_C_EXPERIMENTAL_GRADIENTS_TAPE_TAPE_OPERATION_H_

#include "tensorflow/c/eager/abstract_operation.h"
#include "tensorflow/c/eager/gradients.h"

namespace tensorflow {
namespace gradients {
class TapeOperation : public AbstractOperation {
 public:

        operation is a scalar.
    *   Fix `NNAPI` delegate failure when Axis input for reduce operation is a
        scalar.
    *   Expose option to limit the number of partitions that will be delegated
        to `NNAPI`.
    *   If a target accelerator is specified, use its feature level to determine
        operations to delegate instead of SDK version.
*   `tf.random`:

*   [Java license example](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/java/src/main/java/org/tensorflow/Graph.java#L1)
*   [Go license example](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/go/operation.go#L1)
*   [Bash license example](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/ci_build/ci_build.sh#L2)

TF_Operation* Min(TF_Operation* l, TF_Operation* r, TF_Graph* graph,
                  TF_Status* s, const char* name = "min");

TF_Operation* Mul(TF_Operation* l, TF_Operation* r, TF_Graph* graph,
                  TF_Status* s, const char* name = "mul");

// If `op_device` is non-empty, set the created op on that device.
TF_Operation* MinWithDevice(TF_Operation* l, TF_Operation* r, TF_Graph* graph,

    TracingOperationPtr operation(CreateOperation());
    TF_RETURN_IF_ERROR(operation->Reset("Placeholder", nullptr));
    TF_RETURN_IF_ERROR(
        operation->SetOpName(absl::StrCat("_input_", inputs_.size()).c_str()));
    TF_RETURN_IF_ERROR(operation->SetAttrType("dtype", dtype));
    if (!shape.unknown_rank()) {
      TF_RETURN_IF_ERROR(operation->SetAttrShape(