status->status = InvalidArgument("Invalid GraphDef");
    return;
  }
  status->status = s->session->Extend(g);
}

}  // end extern "C"

// Reset helper for converting character arrays to string vectors.
static void TF_Reset_Helper(const TF_SessionOptions* opt,
                            const char** containers, int ncontainers,
                            TF_Status* status) {

#include "absl/strings/str_cat.h"
#include "tensorflow/c/eager/c_api_unified_experimental_internal.h"
#include "tensorflow/c/tf_datatype.h"
#include "tensorflow/c/tf_status.h"
#include "tensorflow/c/tf_status_helper.h"
#include "tensorflow/core/framework/types.pb.h"
#include "tensorflow/core/lib/llvm_rtti/llvm_rtti.h"
#include "tensorflow/core/platform/errors.h"
#include "tensorflow/core/platform/types.h"

using tensorflow::string;

typedef struct TFE_TensorHandle TFE_TensorHandle;

TF_CAPI_EXPORT extern TFE_TensorHandle* TFE_NewTensorHandle(const TF_Tensor* t,
                                                            TF_Status* status);
// Indicates that the caller will not be using `h` any more.
TF_CAPI_EXPORT extern void TFE_DeleteTensorHandle(TFE_TensorHandle* h);
TF_CAPI_EXPORT extern TF_DataType TFE_TensorHandleDataType(TFE_TensorHandle* h);

  // before `StartExecute` is called again. Using `StartExecute` with `Join`
  // allows the caller to schedule computation on multiple ParallelDevices
  // without sequencing those operations (first call `StartExecute` on each
  // parallel device, then call `Join` on each; even if some of the `Join`s
  // return a bad status the caller must run all of the `Join`s or any future
  // `StartExecute`s will deadlock).
  //

// allocated to size `ninputs`. The caller should build `cond_graph` and
// `body_graph` starting from the inputs, and store the final outputs in
// `cond_output` and `body_outputs`.
//
// If `status` is OK, the caller must call either TF_FinishWhile or
// TF_AbortWhile on the returned TF_WhileParams. If `status` isn't OK, the
// returned TF_WhileParams is not valid, and the caller should not call

#include "tensorflow/c/experimental/ops/math_ops.h"
#include "tensorflow/c/tf_tensor.h"

namespace tensorflow {
namespace gradients {

using namespace std;

// ================== Helper functions =================

// Fills data with values [start,end) with given step size.
void Range(vector<int32_t>* data, int32_t start, int32_t end,
           int32_t step = 1) {

mkdir -p "$TFCI_OUTPUT_DIR"

# In addition to dumping all script output to the terminal, place it into
# $TFCI_OUTPUT_DIR/script.log
exec > >(tee "$TFCI_OUTPUT_DIR/script.log") 2>&1

# Setup tfrun, a helper function for executing steps that can either be run
# locally or run under Docker. setup_docker.sh, below, redefines it as "docker
# exec".
# Important: "tfrun foo | bar" is "( tfrun foo ) | bar", not "tfrun (foo | bar)".

#include "tensorflow/c/experimental/gradients/tape/tape_context.h"
#include "tensorflow/c/experimental/ops/array_ops.h"
#include "tensorflow/c/experimental/ops/math_ops.h"
#include "tensorflow/c/tf_status_helper.h"
#include "tensorflow/c/tf_tensor.h"
#include "tensorflow/core/lib/llvm_rtti/llvm_rtti.h"
#include "tensorflow/core/platform/errors.h"
#include "tensorflow/core/platform/test.h"

namespace tensorflow {

                   &FloatDeallocator, nullptr),
      TF_DeleteTensor);
  return TensorHandlePtr(TFE_NewTensorHandle(tensor.get(), status));
}

// Helper to un-pack `num_replicas` TFE_TensorHandles from one parallel handle.
template <std::size_t num_replicas>
void ExtractPerDeviceValues(
    TFE_Context* context, TFE_TensorHandle* input,

TF_Buffer* ProcessGetKeyValueResult(absl::StatusOr<std::string> value,
                                    TF_Status* status) {
  status->status = value.status();
  if (!value.ok()) {
    return nullptr;
  }
  // Caller is responsible to call `TF_DeleteBuffer` to release the buffer.
  TF_Buffer* result = TF_NewBuffer();
  const std::string& value_str = *value;
  void* data = malloc(value_str.length());