[Model][Gemma3] Cast image pixel values already on CPU (#18732)

Signed-off-by: Lukas Geiger <[email protected]>

Author: lgeiger

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:
@@ -543,9 +548,7 @@ 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,