[Bugfix][Model] Mistral3: fix image placeholder grid for processor size overrides (#52874)
Signed-off-by: Anmol Gupta <14880251+anmolgupt@users.noreply.github.com>
This commit is contained in:
@@ -38,7 +38,18 @@ from vllm.multimodal.processing import (
|
||||
_I = TypeVar("_I", bound=Mistral3ProcessingInfo)
|
||||
|
||||
|
||||
class LightOnOCRMultiModalProcessor(BaseMultiModalProcessor[Mistral3ProcessingInfo]):
|
||||
class LightOnOCRProcessingInfo(Mistral3ProcessingInfo):
|
||||
def get_vision_encoder_info(
|
||||
self,
|
||||
mm_processor_kwargs: Mapping[str, object] | None = None,
|
||||
) -> PixtralHFEncoderInfo:
|
||||
# LightOnOCR prompt updates use vision_config.image_size directly;
|
||||
# preserve that behavior rather than inheriting Mistral3's processor
|
||||
# longest_edge override.
|
||||
return PixtralHFEncoderInfo(self.get_hf_config())
|
||||
|
||||
|
||||
class LightOnOCRMultiModalProcessor(BaseMultiModalProcessor[LightOnOCRProcessingInfo]):
|
||||
def _get_hf_processor_text(self, mm_counts: Mapping[str, int]) -> str:
|
||||
return self.dummy_inputs.get_dummy_text(mm_counts)
|
||||
|
||||
@@ -124,7 +135,7 @@ class LightOnOCRMultiModalProcessor(BaseMultiModalProcessor[Mistral3ProcessingIn
|
||||
|
||||
@MULTIMODAL_REGISTRY.register_processor(
|
||||
LightOnOCRMultiModalProcessor,
|
||||
info=Mistral3ProcessingInfo,
|
||||
info=LightOnOCRProcessingInfo,
|
||||
dummy_inputs=Mistral3DummyInputsBuilder,
|
||||
)
|
||||
class LightOnOCRForConditionalGeneration(Mistral3ForConditionalGeneration):
|
||||
|
||||
Reference in New Issue
Block a user