From 5b0ed8391d497439595a1968d65df93da98265ca Mon Sep 17 00:00:00 2001 From: zhangjf <1061683512@qq.com> Date: Fri, 13 Dec 2024 23:56:19 +0800 Subject: [PATCH] [Bugfix] using len(tokenizer) instead of tokenizer.vocab_size in AllowedTokenIdsLogitsProcessor (#11156) --- vllm/entrypoints/openai/logits_processors.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/vllm/entrypoints/openai/logits_processors.py b/vllm/entrypoints/openai/logits_processors.py index 7913f8720ca73..c8132811de903 100644 --- a/vllm/entrypoints/openai/logits_processors.py +++ b/vllm/entrypoints/openai/logits_processors.py @@ -71,7 +71,7 @@ def get_logits_processors( # Check if token_id is within the vocab size for token_id, bias in clamped_logit_bias.items(): - if token_id < 0 or token_id >= tokenizer.vocab_size: + if token_id < 0 or token_id >= len(tokenizer): raise ValueError(f"token_id {token_id} in logit_bias contains " "out-of-vocab token id") @@ -81,6 +81,6 @@ def get_logits_processors( if allowed_token_ids is not None: logits_processors.append( _get_allowed_token_ids_logits_processor( - frozenset(allowed_token_ids), tokenizer.vocab_size)) + frozenset(allowed_token_ids), len(tokenizer))) return logits_processors