fprintf(stderr, "TF_GetLocalTempDirectories returned no results\n");
    return 1;
  }

  char file_name[100];
  time_t t = time(NULL);
  snprintf(file_name, sizeof(file_name), "test-%d-%ld.txt", getpid(), t);

  size_t length = 2 + strlen(path) + strlen(file_name);
  char* full_path = malloc(length);
  snprintf(full_path, length, "%s/%s", path, file_name);

  TF_WritableFileHandle* h;

#ifndef TF_Bool
#define TF_Bool unsigned char
#endif  // TF_Bool

// Macro used to calculate struct size for maintaining ABI stability across
// different struct implementations.
#ifndef TF_OFFSET_OF_END
#define TF_OFFSET_OF_END(TYPE, MEMBER) \
  (offsetof(TYPE, MEMBER) + sizeof(((TYPE *)0)->MEMBER))
#endif  // TF_OFFSET_OF_END

void TF_DeleteExecutionContext(TF_ExecutionContext*);

// Represents a (partially-defined) shape.
typedef struct TF_Shape {
  int num_dims;  // Must be >= -1; -1 represents unknown rank.
  int64_t* dim_sizes;
} TF_Shape;

// Add a new parameter to a TensorFlow Function.
TF_AbstractTensor* TF_AddFunctionParameter(TF_ExecutionContext* func,
                                           TF_DataType dtype, TF_Shape shape,

  uint64_t block_size;  // Reads smaller than block_size will trigger a read
                        // of block_size.
  std::unique_ptr<ExpiringLRUCache<GcsFileStat>> stat_cache;
  GCSFile(google::cloud::storage::Client&& gcs_client);
  // This constructor is used for testing purpose only.
  GCSFile(google::cloud::storage::Client&& gcs_client, bool compose,
          uint64_t block_size, size_t max_bytes, uint64_t max_staleness,

// 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,
    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,

typedef struct TF_Thread TF_Thread;

typedef struct TF_ThreadOptions {
  // Thread stack size to use (in bytes), zero implies that the system default
  // will be used.
  size_t stack_size;

  // Guard area size to use near thread stacks to use (in bytes), zero implies
  // that the system default will be used.
  size_t guard_size;

  // The NUMA node to use, -1 implies that there should be no NUMA affinity for
  // this thread.

#include <sys/stat.h>

namespace tf_posix_filesystem {

// Transfers up to `size` bytes from `dst_fd` to `src_fd`.
//
// This method uses `sendfile` if available (i.e., linux 2.6.33 or later) or an
// intermediate buffer if not.
//
// Returns number of bytes transferred or -1 on failure.
int CopyFileContents(int dst_fd, int src_fd, off_t size);

}  // namespace tf_posix_filesystem

// Return a shape op fetching the shape of `a`.
TFE_Op* ShapeOp(TFE_Context* ctx, TFE_TensorHandle* a);

// Return an allreduce op adding up input tensor `in` from `group_size` workers.
TFE_Op* AllReduceOp(TFE_Context* ctx, TFE_TensorHandle* in, int group_size);

// Return a SendOp op `op_name` with send input tensor `in` and attributes
// `send_device`, `recv_device`, and `send_device_incarnation` set.

  return SetAttrShape(attr_name, shape.dim_sizes().data(), shape.dims());
}

inline Status AbstractOperation::SetAttrStringList(
    const char* attr_name, absl::Span<string const> values) {
  std::vector<const char*> raw_strs;
  std::vector<size_t> lengths;
  raw_strs.reserve(values.size());
  lengths.reserve(values.size());
  for (const auto& s : values) {
    raw_strs.emplace_back(s.data());

 public:
  // Returns number of dimensions.
  virtual Status NumDims(int* num_dims) const = 0;
  // Returns number of elements across all dimensions.
  virtual Status NumElements(int64_t* num_elements) const = 0;
  // Returns size of specified dimension
  //
  // -1 indicates an unknown axis length; this is unreachable for most standard
  // ImmediateExecutionTensorHandles, but comes up for example when computing