undo formatting

Author: carlesonielfa

vllm_bart_plugin/bart.py

@@ -1043,7 +1043,7 @@ def _call_hf_processor(
 
         # 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"]
@@ -1155,7 +1155,8 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             config.vocab_size, config.d_model, embed_scale=embed_scale
         )
         # Bias added to logits after lm_head, matching HuggingFace approach
-        self.register_buffer("final_logits_bias", torch.zeros((1, config.vocab_size)))
+        self.register_buffer("final_logits_bias",
+                             torch.zeros((1, config.vocab_size)))
         self.logits_processor = LogitsProcessor(
             self.unpadded_vocab_size, config.vocab_size
         )
@@ -1343,9 +1344,7 @@ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
                         "Shared weight embedding already loaded with name "
                         "%s, skipping. This is expected on facebook/bart-large"
                         " like models, where the same shared embedding is "
-                        "present multiple times.",
-                        name,
-                    )
+                        "present multiple times.", name)
                     continue
 
         loader = AutoWeightsLoader(
@@ -1383,4 +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