[HotFix] Fix final output truncation with stop string + streaming (#8468)

2024-09-13 19:31:12 +01:00
parent f57092c00b
commit 18e9e1f7b3
2 changed files with 24 additions and 6 deletions
--- a/vllm/sequence.py
+++ b/vllm/sequence.py
@@ -477,7 +477,9 @@ class Sequence:
        if not delta:
            return self.output_text[:-buffer_length] if truncate else (
                self.output_text)
-        length = len(self.output_text) - buffer_length
+        length = len(self.output_text)
+        if truncate:
+            length -= buffer_length
        last_offset = self._last_output_text_offset
        if last_offset < length:
            self._last_output_text_offset = length