TFE_ContextAddFunctionDef(ctx, function_def.data(), function_def.size(),
                            status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  for (bool clear_cache : {true, false, true}) {
    if (clear_cache) {
      TFE_ContextClearCaches(ctx);
    }
    TFE_TensorHandle* m = TestMatrixTensorHandle(ctx);
    TFE_TensorHandle* retval[1] = {nullptr};
    int num_retvals = 1;

                                                 TF_Buffer* value_buf,
                                                 TF_Status* status);

// Delete configuration key-value. If `key` is a directory, recursively clean up
// all key-values under the path specified by `key`.
TF_CAPI_EXPORT extern void TFE_DeleteConfigKeyValue(TFE_Context* ctx,
                                                    const char* key,

  /// manages the lifetime of the buffer after the call. Thus, all callers of
  /// this method must take ownership of the returned pointer.
  ///
  /// The implementation should clean up paths, including but not limited to,
  /// removing duplicate `/`s, and resolving `..` and `.`.
  ///
  /// Plugins must not return `nullptr`. Returning empty strings is allowed.
  ///

                                      bool reset_context, int keep_alive_secs,
                                      int64_t init_timeout_in_ms, int retries,
                                      bool clear_existing_contexts = false) = 0;

  // Initializes context for the local worker and no contexts will be created
  // for remote workers. Currently this only works for resetting context.

void TF_OutputListSetNumOutputs(TF_OutputList* o, int num_outputs,
                                TF_Status* s) {
  unwrap(o)->expected_num_outputs = num_outputs;
  unwrap(o)->outputs.clear();
  unwrap(o)->outputs.resize(num_outputs);
}
int TF_OutputListNumOutputs(TF_OutputList* o) {
  return unwrap(o)->outputs.size();
}
TF_AbstractTensor* TF_OutputListGet(TF_OutputList* o, int i) {

  DeleteInputValues();
  inputs_.clear();
  for (const auto& p : inputs) {
    inputs_.emplace_back(TF_Output{p.first, 0});
    input_values_.emplace_back(p.second);
  }
}

void CSession::SetOutputs(std::initializer_list<TF_Operation*> outputs) {
  ResetOutputValues();
  outputs_.clear();
  for (TF_Operation* o : outputs) {
    outputs_.emplace_back(TF_Output{o, 0});

Here are the affected tests:
EOF
    while read dep; do
      echo "For dependency $dep:"
      # For every missing dependency, find the tests which directly depend on
      # it, and print that list for debugging. Not really clear if this is
      # helpful since the only examples I've seen are enormous.
      bazel query "rdeps(kind(py_test, $(cat $BATS_TEST_TMPDIR/deps)), $dep, 1)"
    done < $BATS_TEST_TMPDIR/missing_deps
    exit 1
  fi
}

namespace tensorflow {

class ImmediateExecutionContext;
class AbstractOpAttrs;

// Abstract interface to an operation.
class ImmediateExecutionOperation : public AbstractOperation {
 public:
  virtual void Clear() = 0;

  // Returns the inputs of this op.
  virtual absl::Span<ImmediateExecutionTensorHandle* const> GetInputs()
      const = 0;
  virtual Status SetInput(size_t index,

    if (!mutation_warning.empty()) {
      // TODO(b/74949947): turn this back into an error status
      LOG(WARNING) << mutation_warning;
      session->graph->sessions[session].clear();
    }

    const auto num_nodes = graph.num_node_ids();
    if (session->last_num_graph_nodes < num_nodes) {
      // TODO(nolivia): check this on a subset of the graph instead of all of
      // it.

    *   Add tf.keras.layers.AbstractRNNCell as the preferred implementation of
        RNN cell for TF v2. User can use it to implement RNN cell with custom
        behavior.
    *   Adding `clear_losses` API to be able to clear losses at the end of
        forward pass in a custom training loop in eager.
    *   Add support for passing list of lists to the `metrics` param in Keras
        `compile`.