change chat template

lvhan028 · Sep 15, 2023 · d1c4e14 · d1c4e14
1 parent 2dec28a
commit d1c4e14
Show file tree

Hide file tree

Showing 2 changed files with 8 additions and 7 deletions.
diff --git a/lmdeploy/model.py b/lmdeploy/model.py
@@ -55,7 +55,7 @@ def get_prompt(self, prompt, sequence_start=True):
 
     @abstractmethod
     def decorate_prompt(self, prompt, sequence_start):
-        pass
+        return prompt
 
     @staticmethod
     def _translate_messages(messages: List):
@@ -176,8 +176,8 @@ class InternLMChat7B(BaseModel):
     def __init__(self,
                  system='',
                  user='<|User|>',
-                 eoh='<eoh>',
-                 eoa='<eoa>',
+                 eoh='',
+                 eoa='',
                  assistant='<|Bot|>',
                  **kwargs):
         super().__init__(**kwargs)
@@ -231,7 +231,7 @@ def messages2prompt(self, messages, sequence_start=True):
     @property
     def stop_words(self):
         """Return the stop-words' token ids."""
-        return [103027, 103028]
+        return [103028]
 
 
 @MODELS.register_module(name='internlm-chat-7b-8k')

diff --git a/lmdeploy/turbomind/tokenizer.py b/lmdeploy/turbomind/tokenizer.py
@@ -112,7 +112,7 @@ class HuggingFaceTokenizer:
 
     def __init__(self, model_dir: str):
         from transformers import (AutoTokenizer, CodeLlamaTokenizerFast,
-                                  LlamaTokenizerFast)
+                                  LlamaTokenizer, LlamaTokenizerFast)
         model_file = osp.join(model_dir, 'tokenizer.model')
         backend_tokenizer_file = osp.join(model_dir, 'tokenizer.json')
         model_file_exists = osp.exists(model_file)
@@ -121,8 +121,9 @@ def __init__(self, model_dir: str):
                   'It may take long time to initialize the tokenizer.')
         self.model = AutoTokenizer.from_pretrained(model_dir,
                                                    trust_remote_code=True)
-        self.need_padding = isinstance(self.model, LlamaTokenizerFast) \
-            or isinstance(self.model, CodeLlamaTokenizerFast)
+        self.need_padding = type(self.model) in [
+            LlamaTokenizer, LlamaTokenizerFast, CodeLlamaTokenizerFast
+        ]
         self._no_prefix_space_tokens = None
         # save tokenizer.json to reuse
         if not osp.exists(backend_tokenizer_file) and model_file_exists: