xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-03-25 00:01:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix: handle None tokenizer in multimodal processor initialization

Browse Source 
When skip_tokenizer_init=True is set, the tokenizer is None. Previously,
this None value was unconditionally passed to the processor, which
overrode the processor's ability to load its own tokenizer from the
model path. This caused crashes in multimodal models like gemma-3 that
require a tokenizer during processor initialization.

The fix is to only pass the tokenizer kwarg when it's not None, allowing
the processor to load its own tokenizer when skip_tokenizer_init=True.

Fixes #31123

Signed-off-by: yurekami <69337011+yurekami@users.noreply.github.com>

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Signed-off-by: yurekami <yurekami@users.noreply.github.com>
This commit is contained in:
yurekami 2025-12-24 23:23:59 +09:00
parent 7cd288a4b3
commit 2843784c1c
1 changed files with 10 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
vllm/multimodal/processing.py
View File

@ -1046,10 +1046,19 @@ class InputProcessingContext:
typ = ProcessorMixin
# Only pass tokenizer if not None to allow the processor to
# load its own tokenizer from the model path when skip_tokenizer_init
# is True. Passing tokenizer=None would override the processor's
# tokenizer loading and cause crashes in multimodal models that
# require a tokenizer during processor initialization.
tokenizer_kwargs = {}
if self.tokenizer is not None:
tokenizer_kwargs["tokenizer"] = self.tokenizer
return cached_processor_from_config(
self.model_config,
processor_cls=typ,
tokenizer=self.tokenizer,
**tokenizer_kwargs,
**kwargs,
)