支持直接加载Qwen（一代）的HF模型

TylunasLi · Dec 1, 2024 · 8c7eb27 · 8c7eb27
1 parent e6d4e56
commit 8c7eb27
Show file tree

Hide file tree

Showing 3 changed files with 22 additions and 5 deletions.
diff --git a/docs/models.md b/docs/models.md
@@ -50,10 +50,10 @@
 
 |              模型  | 加载后转换 |  离线转换  |  直接读取  |
 |-------------------: |------------|------------|------------|
-| Qwen/Qwen-7B-Chat   | [✔](#其它模型) | [✔](#qwen模型导出) |  |
-| Qwen/Qwen-14B-Chat  | [✔](#其它模型) | [✔](#qwen模型导出) |  |
-| Qwen/Qwen-72B-Chat  | [✔](#其它模型) | [✔](#qwen模型导出) |  |
-| Qwen/Qwen-1_8B-Chat | [✔](#其它模型) | [✔](#qwen模型导出) |  |
+| Qwen/Qwen-7B-Chat   | [✔](#其它模型) | [✔](#qwen模型导出) | ✔ |
+| Qwen/Qwen-14B-Chat  | [✔](#其它模型) | [✔](#qwen模型导出) | ✔ |
+| Qwen/Qwen-72B-Chat  | [✔](#其它模型) | [✔](#qwen模型导出) | √ |
+| Qwen/Qwen-1_8B-Chat | [✔](#其它模型) | [✔](#qwen模型导出) | ✔ |
 | Qwen/Qwen1.5-0.5B-Chat | [✔](#其它模型) | [✔](#qwen模型导出) | ✔<sup>3</sup> |
 | Qwen/Qwen1.5-1.8B-Chat | [✔](#其它模型) | [✔](#qwen模型导出) | ✔<sup>3</sup> |
 | Qwen/Qwen1.5-4B-Chat   | [✔](#其它模型) | [✔](#qwen模型导出) | ✔<sup>3</sup> |

diff --git a/src/model.cpp b/src/model.cpp
@@ -524,6 +524,18 @@ namespace fastllm {
             model->history_sep = "";
             model->weight.tokenizer.type = Tokenizer::TokenizerType::QWEN;
             model->weight.tokenizer.chatTemplate = "";
+        } else if (tokenizerClass == "QWenTokenizer") {
+            // Qwen用的分词
+            std::vector <std::string> lines, line;
+            SplitString(ReadAllFile(path + "qwen.tiktoken"), {'\n'}, lines);
+            for (int i = 0; i < lines.size(); i++) {
+                SplitString(lines[i], {' '}, line);
+                model->weight.AddTokenizerWord(Base64Decode(line[0]), atoi(line[1].c_str()), 1.0f);
+            }
+            model->weight.tokenizer.type = Tokenizer::TokenizerType::QWEN;
+            model->weight.tokenizer.chatTemplate = "";
+            model->weight.dicts["im_end_id"] = std::to_string(lines.size() + 1);
+            model->weight.dicts["im_start_id"] = std::to_string(lines.size() + 2);
         } else {
             ErrorInFastLLM("Unsupport tokenizer_class: " + tokenizerClass);
         }
@@ -637,7 +649,7 @@ namespace fastllm {
             for (auto &it : generation_config.object_items()) {
                 if ("eos_token_id" == it.first && it.second.type() == json11::Json::ARRAY)
                     continue;
-                model->weight.AddDict(it.first, it.second.dump().c_str());
+                model->weight.AddDict(it.first, it.second.is_string() ? it.second.string_value() : it.second.dump());
             }
             // 更新eos_token_id
             if (generation_config["eos_token_id"].is_array()) {

diff --git a/src/models/qwen.cpp b/src/models/qwen.cpp
@@ -56,6 +56,11 @@ namespace fastllm {
         }
 
         weight.embeddingNames.insert("transformer.wte.weight");
+        weight.linearNames = {
+            "lm_head.weight", "transformer.h.*.ln_1.weight", "transformer.h.*.attn.c_attn.weight",
+            "transformer.h.*.attn.c_proj.weight", "transformer.h.*.ln_2.weight",
+            "transformer.h.*.mlp.w1.weight", "transformer.h.*.mlp.w2.weight", "transformer.h.*.mlp.c_proj.weight"
+        };
     }
 
     int QWenModel::Forward(const Data &inputIds,