fix

src/runtime/file_loader.cc

@@ -219,10 +219,10 @@ void load_attention_weights_v2(DT *ptr,
                                int tensor_parallelism_degree) {
   // layers_0_attention_wq_weight
   // layers_0_self_attn_q_proj_weight
-  std::string q_file = layer_name + "_wq_weight";
-  std::string k_file = layer_name + "_wk_weight";
-  std::string v_file = layer_name + "_wv_weight";
-  std::string o_file = layer_name + "_wo_weight";
+  std::string q_file = layer_name + ".q_proj.weight";
+  std::string k_file = layer_name + ".k_proj.weight";
+  std::string v_file = layer_name + ".v_proj.weight";
+  std::string o_file = layer_name + ".o_proj.weight";
   std::vector<std::string> weight_filenames = {q_file, k_file, v_file};
   int file_index = 0;
 
@@ -409,10 +409,10 @@ void load_attention_weights_quantized(char *ptr,
                                       bool use_full_precision) {
   // layers_0_attention_wq_weight
   // layers_0_self_attn_q_proj_weight
-  std::string q_file = layer_name + "_wq_weight";
-  std::string k_file = layer_name + "_wk_weight";
-  std::string v_file = layer_name + "_wv_weight";
-  std::string o_file = layer_name + "_wo_weight";
+  std::string q_file = layer_name + ".q_proj.weight";
+  std::string k_file = layer_name + ".k_proj.weight";
+  std::string v_file = layer_name + ".v_proj.weight";
+  std::string o_file = layer_name + ".o_proj.weight";
   std::vector<std::string> weight_filenames = {q_file, k_file, v_file, o_file};
 
   int file_index = 0;
@@ -690,7 +690,7 @@ void FileDataLoader::load_quantization_weight(FFModel *ff,
     if (weight_idx > 0) {
       assert(weight_idx == 0 || weight_idx == 1);
       if (weight_filename != "embed_tokens_weight_lm_head") {
-        weight_filename += weight_idx == 0 ? "_weight" : "_bias";
+        weight_filename += weight_idx == 0 ? ".weight" : ".bias";
       }
     }
     load_from_quantized_file(data,
@@ -728,12 +728,9 @@ void FileDataLoader::load_single_weight_tensor(FFModel *ff,
   if (l->op_type == OP_INC_MULTIHEAD_SELF_ATTENTION ||
       l->op_type == OP_SPEC_INC_MULTIHEAD_SELF_ATTENTION ||
       l->op_type == OP_TREE_INC_MULTIHEAD_SELF_ATTENTION) {
-    if (weight_filename.find("self_attention") != std::string::npos) {
-      load_attention_weights_multi_query(
-          data, weight_filename, weights_folder, hidden_dim, num_heads);
-    } else if (weight_filename.find("attention") != std::string::npos &&
-               weight_filename.rfind("attention") ==
-                   weight_filename.length() - strlen("attention")) {
+    if (weight_filename.find("attention") != std::string::npos &&
+        weight_filename.rfind("attention") ==
+            weight_filename.length() - strlen("attention")) {
       if (weight_idx == 0) {
         load_attention_weights_v2(data,
                                   num_heads,
@@ -765,7 +762,7 @@ void FileDataLoader::load_single_weight_tensor(FFModel *ff,
     assert(weight_idx >= 0 || weight_idx <= 2);
     weight_filename += (weight_idx == 0)
                            ? "_attn_bias"
-                           : ((weight_idx == 1) ? "_weight" : "_bias");
+                           : ((weight_idx == 1) ? ".weight" : ".bias");
     std::cout << "Loading weight file " << weight_filename << std::endl;
     std::string weight_filepath = join_path({weights_folder, weight_filename});
     load_from_file(data, volume, weight_filepath);
@@ -774,7 +771,7 @@ void FileDataLoader::load_single_weight_tensor(FFModel *ff,
     assert(weight_idx == 0 || weight_idx == 1);
     // handle exception
     if (weight_filename != "embed_tokens_weight_lm_head") {
-      weight_filename += weight_idx == 0 ? "_weight" : "_bias";
+      weight_filename += weight_idx == 0 ? ".weight" : ".bias";
     }
     std::cout << "Loading weight file " << weight_filename << std::endl;
     std::string weight_filepath = join_path({weights_folder, weight_filename});