fix: Update generation arguments in constant.py and server.py to include eos_token_id and pad_token_id, and improve handling of bad_words_ids in GlobalModelManager.

Author: lxchtan

utils/constant.py

@@ -71,6 +71,8 @@
     'max_new_tokens': 512,
     'repetition_penalty': 1.2,
     'length_penalty': 1.0,
+    'eos_token_id': 151645,
+    'pad_token_id': 151643,
 }
 
 MAX_HISTORY_TURNS = 8  # Keep only the latest 8 rounds of conversation (16 messages: 8 user + 8 assistant)

web_demo/server/server.py

@@ -182,15 +182,10 @@ def initialize(self, model_path: str, target_sample_rate: int = 16000):
         ).to(self.device)
 
         # set gen args
-        self.gen_kwargs = DEFAULT_S2M_GEN_KWARGS
-        if self.gen_kwargs['eos_token_id'] is None:
-            self.gen_kwargs['eos_token_id'] = [self.processor.tokenizer.eos_token_id,
-                                        self.processor.tokenizer.convert_tokens_to_ids("<|im_end|>")]
-        if self.gen_kwargs['pad_token_id'] is None:
-            self.gen_kwargs['pad_token_id'] = self.processor.tokenizer.pad_token_id
-        if self.gen_kwargs['bad_words_ids'] is None:
+        self.gen_kwargs = DEFAULT_S2M_GEN_KWARGS.copy()
+        if 'bad_words_ids' not in self.gen_kwargs or self.gen_kwargs['bad_words_ids'] is None:
             self.gen_kwargs['bad_words_ids'] = [[self.processor.tokenizer.convert_tokens_to_ids('<|audio_bos|>'),
-                                            self.processor.tokenizer.convert_tokens_to_ids('<|sil|>')]]
+                                                self.processor.tokenizer.convert_tokens_to_ids('<|sil|>')]]
 
         self.model.sp_gen_kwargs = DEFAULT_SP_GEN_KWARGS.copy()