fix bugs (#4031)

Author: Jintao-Huang

swift/llm/argument/base_args/template_args.py

@@ -48,10 +48,15 @@ class TemplateArguments:
     def __post_init__(self):
         if self.template is None and hasattr(self, 'model_meta'):
             self.template = self.model_meta.template
-        if self.system is not None and self.system.endswith('.txt'):
-            assert os.path.isfile(self.system), f'self.system: {self.system}'
-            with open(self.system, 'r') as f:
-                self.system = f.read()
+        if self.system is not None:
+            if self.system.endswith('.txt'):
+                assert os.path.isfile(self.system), f'self.system: {self.system}'
+                with open(self.system, 'r') as f:
+                    self.system = f.read()
+            else:
+                self.system = self.system.replace('\\n', '\n')
+        if self.response_prefix is not None:
+            self.response_prefix = self.response_prefix.replace('\\n', '\n')
         if self.truncation_strategy is None:
             self.truncation_strategy = 'delete'
 

swift/llm/infer/infer_engine/infer_engine.py

@@ -35,6 +35,7 @@ def _post_init(self):
         self.config = self.model_info.config
         if getattr(self, 'default_template', None) is None:
             ckpt_dir = get_ckpt_dir(self.model_dir, getattr(self, 'adapters', None))
+            logger.info('Create the default_template for the infer_engine')
             if ckpt_dir:
                 from swift.llm import BaseArguments
                 args = BaseArguments.from_pretrained(ckpt_dir)

swift/llm/infer/infer_engine/utils.py

@@ -67,7 +67,10 @@ def _align_blank_suffix(self, response: str) -> str:
 
     def _get_response(self, response: str, is_finished: bool, token_len: int) -> str:
         # After the symbol for a new line, we flush the cache.
-        if response.endswith('\n') or is_finished:
+        if self.first_token:
+            printable_text = response
+            self.first_token = False
+        elif response.endswith('\n') or is_finished:
             printable_text = response[self.print_idx:]
             self.cache_idx += token_len
             self.first_num_space = -1
@@ -85,9 +88,10 @@ def _get_response(self, response: str, is_finished: bool, token_len: int) -> str
 
     def get_printable_text(self, raw_tokens: List[int], is_finished: bool) -> str:
         raw_tokens = raw_tokens[self.cache_idx:]
+        if self.first_token:
+            raw_tokens = []
         response = self.template.decode(
             raw_tokens, is_finished=is_finished, tokenizer_kwargs=self.decode_kwargs, first_token=self.first_token)
-        self.first_token = False
         response = self._align_blank_suffix(response)
         return self._get_response(response, is_finished, len(raw_tokens))