From 6432739ef1b35be382733c6c081dab76696b1f96 Mon Sep 17 00:00:00 2001 From: Nick Hill Date: Fri, 5 Sep 2025 22:30:22 -0700 Subject: [PATCH] [Bugfix] Catch and log invalid token ids in detokenizer (#24351) Signed-off-by: Nick Hill --- vllm/v1/engine/detokenizer.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/vllm/v1/engine/detokenizer.py b/vllm/v1/engine/detokenizer.py index 0ccbe65493499..38f435f5166e0 100644 --- a/vllm/v1/engine/detokenizer.py +++ b/vllm/v1/engine/detokenizer.py @@ -233,6 +233,11 @@ class FastIncrementalDetokenizer(BaseIncrementalDetokenizer): def _protected_step(self, next_token_id: int) -> Optional[str]: try: token = self.stream.step(self.tokenizer, next_token_id) + except OverflowError: + # Handle rare observed overflow, still to be diagnosed. + # See https://github.com/vllm-project/vllm/issues/21951. + logger.exception("Encountered invalid token id: %d", next_token_id) + token = None except Exception as e: if not str(e).startswith(INVALID_PREFIX_ERR_MSG): raise e