absl::MutexLock l(&block->mu);
  if (block->state != FetchState::FINISHED) {
    return true;  // No need to check for staleness.
  }
  if (max_staleness_ == 0) return true;  // Not enforcing staleness.
  return timer_seconds_() - block->timestamp <= max_staleness_;
}

std::shared_ptr<RamFileBlockCache::Block> RamFileBlockCache::Lookup(
    const Key& key) {
  absl::MutexLock lock(&mu_);
  auto entry = block_map_.find(key);

    const ForwardOperation& op,
    std::unique_ptr<GradientFunction>* gradient_function) const {
  auto iter = registry_.find(op.op_name);
  if (iter == registry_.end()) {
    const string error_msg = "No gradient defined for op: " + op.op_name + ".";
    return errors::NotFound(error_msg);
  }
  gradient_function->reset(iter->second(op));
  return absl::OkStatus();
}

  if (!status->status.ok()) return -1;
  auto iter = name_ranges.find(arg_name);
  if (iter == name_ranges.end()) {
    status->status = InvalidArgument("Output arg '", arg_name, "' not found");
    return -1;
  }
  return iter->second.second - iter->second.first;
}

int TF_OperationNumInputs(TF_Operation* oper) {
  return oper->node.num_inputs();
}

    return;
  }
  std::vector<tensorflow::CoordinatedTask> task_vec(tasks.length);
  auto* task_iter = static_cast<const tensorflow::CoordinatedTask*>(tasks.data);
  for (size_t i = 0; i < tasks.length; ++i) {
    task_vec[i].set_job_name(task_iter->job_name());
    task_vec[i].set_task_id(task_iter->task_id());
    ++task_iter;
  }
  auto results = coord_agent->GetTaskState(task_vec);
  if (!results.ok()) {

    for (const OpDef::ArgDef& arg : fdef.signature().output_arg()) {
      const auto& iter = fdef.ret().find(arg.name());
      if (iter != fdef.ret().end()) {
        const auto& v = a_edges.insert({iter->second, arg.name()});
        ASSERT_TRUE(v.second) << "Duplicate edge " << iter->second << " -> "
                              << arg.name() << ". fdef: " << fdef.DebugString();
      } else {

/// automatically determined, so ignored for the rest of this comment).
///
/// Since filesystem plugins are outside of TensorFlow's code tree, they are not
/// tied with TensorFlow releases and should have their own versioning metadata
/// in addition with the data discussed in this section. Each plugin author can
/// use a custom scheme, but it should only relate to changes in plugin code.

        for dynamically-sized TensorArray inside `tf.function`.
*   `tf.math`:
    *   Narrow down `argmin`/`argmax` contract to always return the smallest
        index for ties.
    *   `tf.math.reduce_variance` and `tf.math.reduce_std` return correct
        computation for complex types and no longer support integer types.
    *   Add Bessel functions of order 0,1 to `tf.math.special`.

  auto iter = builder->attr_names.insert(attr_name).first;
  builder->Set(*iter, static_cast<tensorflow::DataType>(value));
}

void TF_AttrBuilderSetTypeList(TF_AttrBuilder* builder, const char* attr_name,
                               const TF_DataType* values, int num_values) {
  auto iter = builder->attr_names.insert(attr_name).first;

                                    " into function '", fn_name, "'");

    input_tensors->emplace_back(node, idx);

    const auto& iter = input_nodes->find(node);
    if (iter == input_nodes->end()) {
      input_nodes->insert({node, {idx}});
    } else {
      auto& indices = iter->second;
      if (std::find(indices.begin(), indices.end(), idx) != indices.end()) {
        return InvalidArgument("TF_Output ", node->name(), ":", idx,

  // lifetime through ref counting. Thus clients MUST call Release() in order to
  // destroy an instance of this class.
  virtual void Release() = 0;

  // Creates an operation builder and ties it to this context.
  // The returned object can be used for setting operation's attributes,
  // adding inputs and finally executing (immediately or lazily as in tracing)
  // it in this context.