[Model] Use merge_by_field_config for MM models (Llava family) (#26280)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

vllm/model_executor/models/pixtral.py

@@ -64,7 +64,7 @@ from vllm.transformers_utils.tokenizer import (
 from vllm.utils.tensor_schema import TensorSchema, TensorShape
 
 from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
-from .utils import flatten_bn, init_vllm_registered_model, maybe_prefix
+from .utils import init_vllm_registered_model, maybe_prefix
 from .vision import (
     VisionEncoderInfo,
     VisionFeatureSelectStrategy,
@@ -365,6 +365,8 @@ class PixtralMultiModalProcessor(BaseMultiModalProcessor[PixtralProcessingInfo])
     dummy_inputs=PixtralDummyInputsBuilder,
 )
 class PixtralForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
+    merge_by_field_config = True
+
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> Optional[str]:
         if modality.startswith("image"):
@@ -424,7 +426,7 @@ class PixtralForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP)
 
         return PixtralImagePixelInputs(
             type="pixel_values",
-            images=flatten_bn(images),
+            images=images,
         )
 
     def _process_image_input(