// If heavy_load_on_streaming_rpc is true, send some rpc requests before the one
// which creates a remote input, to simulate a scenario that the remote input
// is not ready when we start running an op or a function.
void TestRemoteExecuteSilentCopies(bool async, bool remote, bool func,
                                   bool heavy_load_on_streaming_rpc,

// Some TF ops need a step container to be set to limit the lifetime of some
// resources (mostly TensorArray and Stack, used in while loop gradients in
// graph mode). Calling this on a context tells it to start a step.
TF_CAPI_EXPORT extern void TFE_ContextStartStep(TFE_Context* ctx);

// Ends a step. When there is no active step (that is, every started step has

// for that Tensor). We also count, for each tape entry, how many of its output
// Tensors need gradients to be computed (Tensors which are not used do not need
// any gradients to be computed).
//
// Finally, we start a backprop stack with a set of tape entries for which we
// have all gradients available. This set usually is a subset of the set of
// targets (not all since targets which have outputs in the tape will not have

//           can be null. The names used for function inputs are normalized
//           names of the operations (usually placeholders) pointed to by
//           `inputs`. These operation names should start with a letter.
//           Normalization will convert all letters to lowercase and
//           non-alphanumeric characters to '_' to make resulting names match

  /// be set to the reason why one entry couldn't be removed and the proper
  /// count must be updated. If the deletion is unsuccessful because the
  /// traversal couldn't start, `*undeleted_files` must be set to 0 and
  /// `*undeleted_dirs` must be set to 1.
  ///
  /// TODO(b/139060984): After all filesystems are converted, consider
  /// invariant about `*undeleted_files` and `*undeleted_dirs`.