const char* filename;
  int fd;
} PosixFile;

static void Cleanup(TF_RandomAccessFile* file) {
  auto posix_file = static_cast<PosixFile*>(file->plugin_file);
  close(posix_file->fd);
  // This would be safe to free using `free` directly as it is only opaque.
  // However, it is better to be consistent everywhere.
  plugin_memory_free(const_cast<char*>(posix_file->filename));
  delete posix_file;
}

    // These XLA flags are needed to trigger XLA properly from C (more generally
    // non-Python) clients. If this API is called again with `enable` set to
    // false, it is safe to keep these flag values as is.
    tensorflow::MarkForCompilationPassFlags* flags =
        tensorflow::GetMarkForCompilationPassFlags();
    flags->tf_xla_cpu_global_jit = true;
    flags->tf_xla_min_cluster_size = 1;

}  // namespace

// Allows a single op at a time to be launched without blocking.
//
// DeviceThread itself is thread-safe, in that StartExecute will block if there
// is a pending execution. Since StartExecute is equivalent to grabbing a lock,
// multiple DeviceThreads should always be accessed in the same order to avoid
// deadlocks.
class DeviceThread {
 public:
  // Starts a background thread waiting for `StartExecute`.

      gtl::ArraySlice<AbstractTensorHandle*> output_gradients,
      absl::Span<AbstractTensorHandle*> result) const override;

  // Builds a tensor filled with ones with the same shape and dtype as `t`.
  Status BuildOnesLike(const TapeTensor& t,
                       AbstractTensorHandle** result) const override;

  // Looks up the ID of a Gradient.

  TFE_Execute(matmul.get(), &retval, &num_retvals, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TFE_DeleteTensorHandle(retval);

  // Custom device: inputs in same custom device works.
  matmul.reset(MatMulOp(context.get(), hcustom0.get(), hcustom0.get()));
  num_retvals = 1;
  executed = false;
  TFE_Execute(matmul.get(), &retval, &num_retvals, status.get());

             : absl::OkStatus();
}

// Converts `ninputs` and `inputs` into `inputs_tensors` and `input_nodes` and
// does various checks while doing so. `input_nodes` will contain the same
// information as input_tensors just in a different structure to make
// following processing easier. TODO(iga): Simplify this nested structure.
Status ProcessInputs(
    const TF_Graph* fn_body, const char* fn_name, int ninputs,

      context.get(), components, device_name, status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  // Run a collective sum, so each component should now be the same.
  TensorHandlePtr reduced(
      CollectiveSum(context.get(), parallel_value.get(), 2, status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

}

// Note: this function looks up a thread local policy. So it should be called in
// the appropriate client thread. In particular, in async mode, it may not be
// safe to call this function from the async EagerExecutor threads.
extern TFE_ContextDevicePlacementPolicy TFE_ContextGetDevicePlacementPolicy(
    TFE_Context* ctx) {
  return static_cast<TFE_ContextDevicePlacementPolicy>(

// default as the multiple API calls required add a risk of stranding temporary
// objects.
constexpr char kComposeAppend[] = "compose";

// We can cast `google::cloud::StatusCode` to `TF_Code` because they have the
// same integer values. See
// https://github.com/googleapis/google-cloud-cpp/blob/6c09cbfa0160bc046e5509b4dd2ab4b872648b4a/google/cloud/status.h#L32-L52
static inline void TF_SetStatusFromGCSStatus(