move strip_for_generation

Signed-off-by: yechank <[email protected]>

Author: yechank-nvidia

tensorrt_llm/_torch/pyexecutor/model_engine.py

@@ -1262,9 +1262,12 @@ def _prepare_tp_inputs(
             # Multimodal
             multimodal_params = MultimodalParams(
                 multimodal_data=request.py_multimodal_data)
+            multimodal_params.strip_for_generation()
             if multimodal_params.has_content():
-                multimodal_params.strip_for_generation()
-                # re-assign the multimodal_data to the request after strip_for_generation for another generation request
+                multimodal_params.to_device("multimodal_data",
+                                            "cuda",
+                                            pin_memory=True)
+                # re-assign the multimodal_data to the request after strip_for_generation for another generation request,
                 request.py_multimodal_data = multimodal_params.multimodal_data
                 multimodal_params_list.append(multimodal_params)
         extend_requests += extend_dummy_requests