// See TF_AddGradientsWithPrefix, which provides a means to specify a custom
// name prefix for operations added to a graph to compute the gradients.
//
// WARNING: This function does not yet support all the gradients that python
// supports. See
// https://www.tensorflow.org/code/tensorflow/cc/gradients/README.md
// for instructions on how to add C++ more gradients.
TF_CAPI_EXPORT void TF_AddGradients(TF_Graph* g, TF_Output* y, int ny,

class DeviceThread;

// Forwards operations to `devices`, maintaining ParallelTensor with components
// placed on each underlying device.
class ParallelDevice {
 public:
  // Eager async execution is only supported when remote eager is not in use
  // (b/157523095).
  explicit ParallelDevice(const std::vector<std::string>& devices,
                          bool is_async = false, int in_flight_nodes_limit = 0);

  ~ParallelDevice();

// for a GPU tensor this will return a pointer to GPU memory). The pointer is
// only guaranteed to be valid until TFE_DeleteTensorHandle is called on this
// TensorHandle. Only supports POD data types.
TF_CAPI_EXPORT extern void* TFE_TensorHandleDevicePointer(TFE_TensorHandle*,
                                                          TF_Status*);

// Caller must call TF_DeleteTensor() over the returned tensor. If the queue is
// empty, this call is blocked.
//
// Tensors are enqueued via the corresponding TF enqueue op.
// TODO(hongm): Add support for `timeout_ms`.
TF_CAPI_EXPORT extern TF_Tensor* TF_DequeueNamedTensor(TF_Session* session,
                                                       int tensor_id,

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

namespace tensorflow {

class FunctionRecord;

// A traced function: this hides the complexity of converting the serialized
// representation between various supported formats e.g. FunctionDef and Mlir
// function.
class AbstractFunction : public core::RefCounted {
 protected:
  enum AbstractFunctionKind { kGraph, kMlir };

  virtual Status SetInput(size_t index,
                          ImmediateExecutionTensorHandle* input) = 0;

  virtual ImmediateExecutionContext* GetContext() const = 0;

  // Following two methods are used to support custom device.
  // Return true if the inputs contain custom device tensor handle. It means
  // that the argument need to be handled by a custom device.
  virtual bool HasCustomDeviceInput() const = 0;

  size_t struct_size;
  void* ext;  // reserved for future use

  // Graph C API version.
  int32_t major_version;
  int32_t minor_version;
  int32_t patch_version;

  // Backend device type supported by the optimizer.
  const char* device_type;
  TP_OptimizerConfigs* optimizer_configs;  // output, set by plugin
  TP_Optimizer* optimizer;                 // output, set by plugin
} TP_OptimizerRegistrationParams;

/// that plugins provide (some are mandatory, some are optional, with or without
/// a default implementation).
///
/// Each plugin implements the operations that are supported and TensorFlow will
/// properly handle the cases when an operation is not supported (i.e., return
/// the corresponding `Status` value).
///
/// REQUIRED OPERATIONS: All required operations are marked as such, including

  AbstractTensorHandle* GetHandle() const;

 private:
  AbstractTensorHandle* handle_;
};

// A tracing/immediate-execution agnostic tape.
//
// Gradient functions defined for this tape must support handling null incoming
// gradients.
class Tape : protected eager::GradientTape<AbstractTensorHandle,
                                           GradientFunction, TapeTensor> {
 public:

        plugin_memory_free_(std::move(plugin_memory_free)) {}

  ~ModularFileSystem() override { ops_->cleanup(filesystem_.get()); }

  TF_USE_FILESYSTEM_METHODS_WITH_NO_TRANSACTION_SUPPORT;

  Status NewRandomAccessFile(
      const std::string& fname, TransactionToken* token,
      std::unique_ptr<RandomAccessFile>* result) override;
  Status NewWritableFile(const std::string& fname, TransactionToken* token,