feat(huixiangdou/service/feature_store.py): support language option

.gitignore

@@ -2,6 +2,7 @@ models/
 repodir/
 workdir/
 write_toml.py
+modeling_internlm2.py
 config.ini
 config-template.ini
 logs/

huixiangdou/service/feature_store.py

@@ -25,10 +25,12 @@ class FeatureStore:
 
     def __init__(self,
                  device: str = 'cuda',
-                 config_path: str = 'config.ini') -> None:
+                 config_path: str = 'config.ini',
+                 language: str = 'zh') -> None:
         """Init with model device type and config."""
         self.config_path = config_path
         self.reject_throttle = -1
+        self.language = language
         with open(config_path, encoding='utf8') as f:
             config = pytoml.load(f)['feature_store']
             model_path = config['model_path']
@@ -153,7 +155,8 @@ def ingress_response(self, markdown_dir: str, work_dir: str):
                 full_text = str(p).rsplit('/_',
                                           maxsplit=1)[-1] + '\n' + f.read()
                 if '.md' in str(p):
-                    if not self.is_chinese_doc(full_text):
+                    if self.language == 'zh' and not self.is_chinese_doc(
+                            full_text):  # noqa E501
                         continue
 
                 full_texts.append(full_text)

huixiangdou/service/worker.py

@@ -41,7 +41,7 @@ def __init__(self, work_dir: str, config_path: str, language: str = 'zh'):
             language (str, optional): Specifies the language to be used. Defaults to 'zh' (Chinese).  # noqa E501
         """
         self.llm = ChatClient(config_path=config_path)
-        self.fs = FeatureStore(config_path=config_path)
+        self.fs = FeatureStore(config_path=config_path, language=language)
         self.fs.load_feature(work_dir=work_dir)
         self.config_path = config_path
         self.config = None

tests/test_internlm2.py

@@ -10,8 +10,6 @@
                                              device_map='auto',
                                              torch_dtype='auto').eval()
 
-# 不能像某些 LLM 一样 AutoModelForCausalLM.from_pretrained(.. fp16=True) 这样写，会 Internlm2Config.__init__() 报错
-
 queries = ['how to install mmdeploy ?']
 for query in queries:
     pdb.set_trace()