#include "tensorflow/c/eager/c_api_unified_experimental_internal.h"
#include "tensorflow/c/eager/graph_function.h"
#include "tensorflow/c/tf_datatype.h"
#include "tensorflow/c/tf_status.h"
#include "tensorflow/c/tf_status_helper.h"
#include "xla/tsl/c/tsl_status_internal.h"
#include "tensorflow/core/framework/full_type.pb.h"
#include "tensorflow/core/framework/shape_inference.h"
#include "tensorflow/core/framework/tensor_shape.h"

    // GCS will return an empty header (e.g no `content-length` header). In this
    // case, we will set read to `0` and continue.
    read = 0;
  } else if (!absl::SimpleAtoi(content_length->second, &read)) {
    TF_SetStatus(status, TF_UNKNOWN, "Could not get content-length header");
    return -1;
  }
  // `TF_OUT_OF_RANGE` isn't considered as an error. So we clear it here.

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 {

#include "tensorflow/c/eager/tfe_context_internal.h"
#include "tensorflow/c/eager/tfe_op_internal.h"
#include "tensorflow/c/eager/tfe_tensorhandle_internal.h"
#include "tensorflow/c/tf_status.h"
#include "tensorflow/c/tf_status_helper.h"
#include "xla/tsl/c/tsl_status_internal.h"
#include "xla/tsl/distributed_runtime/coordination/coordination_service_agent.h"
#include "tensorflow/core/common_runtime/composite_device.h"

  // (b/157523095).
  explicit ParallelDevice(const std::vector<std::string>& devices,
                          bool is_async = false, int in_flight_nodes_limit = 0);

  ~ParallelDevice();

  // Helper to copy a tensor handle from another device once for each component
  // of the ParallelDevice.
  //
  // Sets a bad status and returns a nullptr if `tensor` is already on the

                   &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,

# BEGIN TF CACHE HELPER OPTIONS
# Options when using remote execution
# WARNING: THESE OPTIONS WONT WORK IF YOU DO NOT HAVE PROPER AUTHENTICATION AND PERMISSIONS

# Use --config=tf_public_cache to try and use the TensorFlow public build cache

        "//tensorflow/c:c_api_internal",
        "//tensorflow/c:c_api_no_xla",
        "//tensorflow/c:checkpoint_reader",
        "//tensorflow/c:tensor_interface",
        "//tensorflow/c:tf_status_helper",
        "//tensorflow/c:tf_tensor_internal",
        "//tensorflow/cc/saved_model:loader",
        "//tensorflow/compiler/mlir/lite/metrics:error_collector",

    return UseT(ToOutput(inputs));
  }

  TF_Operation* UseT(const std::vector<TF_Output>& inputs) {
    TF_Operation* op;
    UseHelper(inputs, &op);
    return op;
  }

  // All the *Helper methods are used as a workaround for the restrictions that
  // one cannot call ASSERT_* methods in non-void-returning functions (when
  // exceptions are disabled during compilation)