A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 18:30 UTC
[Bugfix][VLM] fix llava processor (#15285)
Signed-off-by: Mengqing Cao <cmq0113@163.com>
Mengqing Cao committed
Mar 21, 2025 at 20:14 UTC
c21b99b91241409c2fdf9f3f8c542e8748b317be
1 file changed
+7
-1
vllm/model_executor/models/llava.py
+7
-1
@@ -233,7 +233,13 @@ class LlavaDummyInputsBuilder(BaseDummyInputsBuilder[_I]):
233
class LlavaProcessingInfo(BaseLlavaProcessingInfo):
234
235
def get_hf_processor(self, **kwargs: object):
236
- return self.ctx.get_hf_processor(LlavaProcessor, **kwargs)
236
+ hf_processor = self.ctx.get_hf_processor(LlavaProcessor, **kwargs)
237
+ # In case patch_size is omitted from `processor_config.json`
238
+ # e.g. for E5-V: https://huggingface.co/royokong/e5-v
239
+ if hf_processor.patch_size is None:
240
+ patch_size = self.get_vision_encoder_info().get_patch_size()
241
+ hf_processor.patch_size = patch_size
242
+ return hf_processor
243
244
245
class BaseLlavaMultiModalProcessor(BaseMultiModalProcessor[_I]):