void ExtractPerDeviceValues(
    TFE_Context* context, TFE_TensorHandle* input,
    std::array<TensorHandlePtr, num_replicas>* components, TF_Status* status);

// Helper to pack `num_replicas` TFE_TensorHandles into one parallel handle.
template <std::size_t num_replicas>
TensorHandlePtr CreatePerDeviceValues(
    TFE_Context* context,
    const std::array<TFE_TensorHandle*, num_replicas>& components,
    const char* device, TF_Status* status);

#ifndef TENSORFLOW_C_EXPERIMENTAL_GRADIENTS_ARRAY_GRAD_H_
#define TENSORFLOW_C_EXPERIMENTAL_GRADIENTS_ARRAY_GRAD_H_

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

namespace tensorflow {
namespace gradients {
GradientFunction* IdentityNRegisterer(const ForwardOperation& op);
}  // namespace gradients
}  // namespace tensorflow

typedef jobject jclass;
typedef jobject jthrowable;
typedef jobject jstring;
typedef jobject jarray;
typedef jarray jbooleanArray;
typedef jarray jbyteArray;
typedef jarray jcharArray;
typedef jarray jshortArray;
typedef jarray jintArray;
typedef jarray jlongArray;
typedef jarray jfloatArray;
typedef jarray jdoubleArray;
typedef jarray jobjectArray;

typedef jobject jweak;