From f6895a5fcb4710481bb44726af4fa57f4172330c Mon Sep 17 00:00:00 2001 From: waizuichougou <105572299+waizuichougou@users.noreply.github.com> Date: Mon, 31 Aug 2026 00:19:38 +0800 Subject: [PATCH] [Bugfix][Multimodal] Avoid caching full prompts in fallback (#54439) Signed-off-by: waizuichougou <2082431897@qq.com> Co-authored-by: OpenAI Codex --- vllm/multimodal/processing/processor.py | 9 +++------ 1 file changed, 3 insertions(+), 6 deletions(-) diff --git a/vllm/multimodal/processing/processor.py b/vllm/multimodal/processing/processor.py index 2c3ac6d591..593a48b337 100644 --- a/vllm/multimodal/processing/processor.py +++ b/vllm/multimodal/processing/processor.py @@ -1559,14 +1559,11 @@ class BaseMultiModalProcessor(ABC, Generic[_I]): # let BPE merge tokens across a segment boundary, silently # changing how a non-special-token placeholder is tokenized. new_token_ids = flatten_2d_lists( - [ - cached_encode(tokenizer, text, add_special_tokens=False) - for text in out_texts - ] + [tokenizer.encode(text, add_special_tokens=False) for text in out_texts] ) else: - new_token_ids = cached_encode( - tokenizer, "".join(out_texts), add_special_tokens=False + new_token_ids = tokenizer.encode( + "".join(out_texts), add_special_tokens=False ) return new_token_ids, result