Created by: NHZlX

PR types

Bug fixes

PR changes

Others

Describe

The passes for inference do a lot of fuse operation during graph analysis. During this period, The delete ops、vars, insert new ops or new vars operation will be carried out during this period . When adding new variables, the variable name is determined by https://github.com/PaddlePaddle/Paddle/blob/a7944904d3b50aeffcaae61f00ecd5d6920805b3/paddle/fluid/framework/ir/graph_pattern_detector.h#L357 .

static std::string PDNodeName(const std::string& name_scope,
                              const std::string& repr) {
  return string::Sprintf("%s/%s/%d", name_scope, repr,
                         KeyCounter::Instance().IncCounter(repr));
}

We can see, the name is generated by KeyCounter::Instance().IncCounter, which is a singleton instance.

struct KeyCounter {
  static KeyCounter& Instance() {
    static KeyCounter x;
    return x;
  }

  int IncCounter(const std::string& key) { return dic_[key]++; }

 private:
  std::unordered_map<std::string, size_t> dic_;
};

So，two identical models produce variables with different names during pass optimization.

We should avoid using this method to generate internal variable names