DocumentQuestionAnsweringPipeline only for fast ⚡ tokenizers (#22745)

* fix

---------

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

src/transformers/pipelines/document_question_answering.py

@@ -131,6 +131,11 @@ class DocumentQuestionAnsweringPipeline(ChunkPipeline):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
+        if self.tokenizer is not None and not self.tokenizer.__class__.__name__.endswith("Fast"):
+            raise ValueError(
+                "`DocumentQuestionAnsweringPipeline` requires a fast tokenizer, but a slow tokenizer "
+                f"(`{self.tokenizer.__class__.__name__}`) is provided."
+            )
 
         if self.model.config.__class__.__name__ == "VisionEncoderDecoderConfig":
             self.model_type = ModelType.VisionEncoderDecoder