[Model] Move multimodal_cpu_fields definition to field config (#30181)

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

vllm/model_executor/models/hunyuan_vision.py

@@ -563,7 +563,7 @@ def _hunyuan_vl_field_config(hf_inputs: Mapping[str, torch.Tensor]):
     return dict(
         pixel_values=MultiModalFieldConfig.flat_from_sizes("image", image_grid_sizes),
         image_embeds=MultiModalFieldConfig.flat_from_sizes("image", image_grid_sizes),
-        image_grid_thw=MultiModalFieldConfig.batched("image"),
+        image_grid_thw=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
     )
 
 
@@ -786,8 +786,6 @@ class HunYuanVLForConditionalGeneration(
     SupportsQuant,
     SupportsXDRoPE,
 ):
-    multimodal_cpu_fields = {"image_grid_thw"}
-
     # To ensure correct weight loading and mapping.
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={