Add hidden dimension validation for multimodal embedding inputs (#30968)

Signed-off-by: Wenqi Glantz <wglantz@nvidia.com>
This commit is contained in:
Wenqi Glantz
2025-12-19 07:59:36 +00:00
committed by GitHub
parent 096b25c9ed
commit 4924ac582c
4 changed files with 589 additions and 10 deletions
+9 -1
View File
@@ -1330,7 +1330,15 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
of [`MultiModalDataParser`][vllm.multimodal.parse.MultiModalDataParser]
that has additional subparsers.
"""
return MultiModalDataParser()
# Get expected hidden size for embedding validation if mm_embeds enabled
# This validates hidden dimensions to prevent vulnerabilities: embeddings
# with correct ndim but wrong shape could cause crashes at inference time
mm_config = self.info.ctx.model_config.get_multimodal_config()
expected_hidden_size = None
if mm_config.enable_mm_embeds:
expected_hidden_size = self.info.ctx.model_config.get_inputs_embeds_size()
return MultiModalDataParser(expected_hidden_size=expected_hidden_size)
def validate_num_items(
self,