From d0cd736caadafea1ec1721737af432d8b0a7e919 Mon Sep 17 00:00:00 2001 From: hallerite Date: Mon, 9 Mar 2026 22:30:51 -0700 Subject: [PATCH] [Bugfix] Fix `RuntimeError: Already borrowed` that degrades VLM serving throughput under concurrent load. (#36557) Signed-off-by: hallerite Signed-off-by: hallerite Co-authored-by: Cyrus Leung --- vllm/renderers/base.py | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-) diff --git a/vllm/renderers/base.py b/vllm/renderers/base.py index a82646688f..853a48945e 100644 --- a/vllm/renderers/base.py +++ b/vllm/renderers/base.py @@ -1,6 +1,7 @@ # SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project import asyncio +import copy import time from abc import ABC, abstractmethod from collections.abc import Mapping, Sequence @@ -90,10 +91,17 @@ class BaseRenderer(ABC, Generic[_T]): mm_processor_cache = mm_registry.processor_cache_from_config(config) + # Deep-copy the tokenizer so the multimodal processor gets its + # own Rust tokenizer backend. Without this, concurrent access + # from AsyncMicrobatchTokenizer and call_hf_processor causes + # "RuntimeError: Already borrowed" from the Rust RefCell. + # See: https://github.com/huggingface/tokenizers/issues/537 + mm_tokenizer = copy.deepcopy(tokenizer) + with set_default_torch_num_threads(): self.mm_processor = mm_registry.create_processor( config.model_config, - tokenizer=tokenizer, + tokenizer=mm_tokenizer, cache=mm_processor_cache, )