[V1][Structured Output] calculate vocab_size eagerly (vllm-project#14851)

Signed-off-by: Aaron Pham <[email protected]>

Author: aarnphm

vllm/v1/structured_output/__init__.py

@@ -40,7 +40,7 @@ def _delayed_init(self):
         tokenizer_group.ping()
 
         tokenizer = tokenizer_group.get_lora_tokenizer(None)
-        self.vocab_size = tokenizer.max_token_id + 1
+        self.vocab_size = len(tokenizer.get_vocab())
         if isinstance(tokenizer, MistralTokenizer):
             # NOTE: ideally, xgrammar should handle this accordingly.
             # refer to https://github.com/mlc-ai/xgrammar/blob/d77c0a0173ef14779c918e3be7966ba852f7910f/python/xgrammar/tokenizer_info.py#L98