diff --git a/vllm/model_executor/models/gemma3_mm.py b/vllm/model_executor/models/gemma3_mm.py
index 743542ec8dfa..2591bbb8ab9d 100644
--- a/vllm/model_executor/models/gemma3_mm.py
+++ b/vllm/model_executor/models/gemma3_mm.py
@@ -263,6 +263,11 @@ def _call_hf_processor(
             mm_data,
             mm_kwargs,
         )
+        if "pixel_values" in processed_outputs:
+            # Cast pixel values to model dtype already here,
+            # so we need to transfer less data to the GPU
+            processed_outputs["pixel_values"] = processed_outputs[
+                "pixel_values"].to(self.info.ctx.model_config.dtype)
 
         # HF processor pops the `num_crops` kwarg, which is needed by vLLM
         if (images := mm_data.get("images")) is not None:
@@ -549,9 +554,7 @@ def _image_pixels_to_features(
         vision_tower: SiglipVisionModel,
         pixel_values: torch.Tensor,
     ) -> torch.Tensor:
-        target_dtype = vision_tower.get_input_embeddings().weight.dtype
-        image_features = vision_tower(pixel_values.to(dtype=target_dtype))
-        return image_features
+        return vision_tower(pixel_values)
 
     def _process_image_input(
         self,