fix: don't skip first special token. (#1497)

vllm/transformers_utils/tokenizer.py

@@ -120,6 +120,10 @@ def detokenize_incrementally(
         # tokenizers (bigger = more conservative).
         # Subtract 1 extra to account for the generated token.
         prefix_offset = max(len(output_tokens) - 6, 0)
+        # If the first new token is a special token, we can't skip 1 extra token
+        if skip_special_tokens and new_token_id in tokenizer.all_special_ids:
+            read_offset = max(len(output_tokens), 0)
+        else:
             read_offset = max(len(output_tokens) - 1, 0)
     else:
         # Put new_token_id in a list so skip_special_tokens is respected