[Bugfix] Fix auto dtype casting for BatchFeature

vllm/inputs/registry.py

@@ -168,10 +168,12 @@ def maybe_cast_dtype(x):
         try:
             output = hf_processor(**data, **merged_kwargs, return_tensors="pt")
             # this emulates output.to(dtype=self.model_config.dtype)
-            cast_output = json_map_leaves(maybe_cast_dtype, output)
             if isinstance(output, BatchFeature):
+                cast_output = json_map_leaves(maybe_cast_dtype, output.data)
                 return BatchFeature(cast_output)
 
+            cast_output = json_map_leaves(maybe_cast_dtype, output)
+
             logger.warning_once(
                 f"{type(hf_processor).__name__} did not return `BatchFeature`. "
                 "Make sure to match the behaviour of `ProcessorMixin` when "