[Core] Automatically cast multi-modal input dtype (#18756)

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

vllm/model_executor/models/gemma3_mm.py

@@ -263,11 +263,6 @@ class Gemma3MultiModalProcessor(BaseMultiModalProcessor[Gemma3ProcessingInfo]):
             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: