[Bugfix] Fix guided decoding with tokenizer mode mistral (#11046)

2024-12-18 03:34:08 -03:00
parent 866fa4550d
commit 8b79f9e107
7 changed files with 217 additions and 52 deletions
--- a/vllm/transformers_utils/tokenizers/mistral.py
+++ b/vllm/transformers_utils/tokenizers/mistral.py
@@ -314,12 +314,15 @@ class MistralTokenizer:

            if regular_tokens:
                decoded_list.append(
-                    self.decode(regular_tokens))  # type: ignore
+                    self.tokenizer.decode(regular_tokens))  # type: ignore

            decoded = ''.join(decoded_list)

        return decoded

+    # WARN: Outlines logits processors can overwrite this method.
+    # See: guided_decoding/outlines_logits_processors.py::_adapt_tokenizer
+    # for more.
    def decode(self,
               ids: Union[List[int], int],
               skip_special_tokens: bool = True) -> str: