Issue with C++ inference for model meta-llama/Llama-2-70b-hf

[DDDDDYTS]

I was trying C++ version of spec_infer and incr_decoding in FlexFlow/inference. And I found that when using the model Llama-2-70b-hf, An error occured as below.

 inference/file_loader.cc:252: void load_attention_weights_v2(DT*, int, int, size_t, size_t, std::string, std::string, size_t, int) [with DT = __half; size_t = long unsigned int; std::string = std::__cxx11::basic_string<char>]: Assertion `false && "data size mismatch"' failed.

And after checking issue flexflow/flexflow-train#1154 and the source code of llama.cc and file_loader.cc, I realized that the reason was that the number of attention heads and key value heads are not identical in this model, but the fix was only added to python version in issue flexflow/flexflow-train#1154.

I added a parameter num_key_value_heads to FlexFlow/inference/models/llama.h and passed it to FileDataLoader in FlexFlow/inference/models/llama.cc, and it worked.

FlexFlow/inference/models/llama.h

LLAMAConfig(std::string const &model_config_file_path) {
      std::ifstream config_file(model_config_file_path);
      if (config_file.is_open()) {
        try {
          json model_config;
          config_file >> model_config;
          num_hidden_layers = model_config["num_hidden_layers"];
          vocab_size = model_config["vocab_size"];
          num_attention_heads = model_config["num_attention_heads"];
          hidden_size = model_config["hidden_size"];
          rms_norm_eps = model_config["rms_norm_eps"];
          intermediate_size = model_config["intermediate_size"];
          num_key_value_heads = model_config["num_key_value_heads"]; //modified!!
        } catch (json::exception const &e) {
          std::cerr << "Error parsing LLAMA config from JSON file: " << e.what()
                    << std::endl;
          assert(false);
        }
      } else {
        std::cerr << "Error opening JSON file " << model_config_file_path
                  << std::endl;
        assert(false);
      }
      // max_seq_len = BatchConfig::MAX_SEQ_LENGTH;
      // max_num_tokens = BatchConfig::MAX_NUM_TOKENS;
      max_beam_width = BeamSearchBatchConfig::MAX_BEAM_WIDTH;
      max_beam_depth = BeamSearchBatchConfig::MAX_BEAM_DEPTH;
    }

FlexFlow/inference/models/llama.cc

  FileDataLoader *fileloader = new FileDataLoader(
      "",
      weight_file_path,
      llama_config.num_attention_heads,
      llama_config.num_key_value_heads, //modified!!
      llama_config.hidden_size,
      llama_config.hidden_size / llama_config.num_attention_heads,
      ff.config.tensor_parallelism_degree,
      use_full_precision);