Merge branch 'lizhg1-v0.8.5.post1-dev-patch-30255' into 'v0.8.5.post1-dev'

Update sequence.py fix assert error。

See merge request dcutoolkit/deeplearing/vllm!121

vllm/zero_overhead/sequence.py

 
-
 from typing import Union
 from vllm.sequence import Sequence
 from typing import Sequence as GenericSequence
@@ -12,12 +11,12 @@ class ZeroOverheadSequence(Sequence):
 
     def fix_last_token_id(self, token_id: int) -> None:
         effect_offset = self.effective_output_len - len(self.data.output_token_ids)
-        assert effect_offset < 0
-        self.data._output_token_ids[effect_offset] = token_id
-        if len(self.data._new_appended_tokens) >= effect_offset * -1:
-            self.data._new_appended_tokens[effect_offset] = token_id
-        self.data._cached_all_token_ids[effect_offset] = token_id
-        self.effective_output_len += 1
+        if effect_offset < 0:
+            self.data._output_token_ids[effect_offset] = token_id
+            if len(self.data._new_appended_tokens) >= effect_offset * -1:
+                self.data._new_appended_tokens[effect_offset] = token_id
+            self.data._cached_all_token_ids[effect_offset] = token_id
+            self.effective_output_len += 1
     
     def remove_last_place_holder(self, count):
         self.data._output_token_ids = self.data._output_token_ids[:-1 * count]
@@ -62,4 +61,4 @@ class ZeroOverheadSequence(Sequence):
             return []
 
         effect_offset = self.effective_output_len - len(self.data.output_token_ids)
-        return self.data._cached_all_token_ids[-num_new_tokens : effect_offset]
\ No newline at end of file
+        return self.data._cached_all_token_ids[-num_new_tokens : effect_offset]