Prevent add_special_tokens overwrite for BART

Signed-off-by: Carles Onielfa <carlesonielfa@gmail.com>

Author: carlesonielfa

vllm_bart_plugin/bart.py

@@ -1041,6 +1041,9 @@ def _call_hf_processor(
         has_encoder_data = mm_data is not None and "texts" in mm_data
         result = {}
 
+        # vLLM may pass add_special_tokens in tok_kwargs; we set it ourselves
+        tok_kwargs = {k: v for k, v in tok_kwargs.items() if k != "add_special_tokens"}
+        
         if has_encoder_data:
             # Tokenize the encoder text from mm_data
             encoder_texts = mm_data["texts"]
@@ -1379,4 +1382,4 @@ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         for key in self.keys_to_ignore_on_load_missing:
             loaded_params.add(key)
 
-        return loaded_params
+        return loaded_params