[Bugfix] fix KeyError on top logprobs are special tokens (vllm-project#17637)

Signed-off-by: chaunceyjiang <[email protected]>

Author: chaunceyjiang

vllm/entrypoints/openai/serving_chat.py

@@ -1111,7 +1111,8 @@ def _create_chat_logprobs(
             return_as_token_id is not None else self.return_tokens_as_token_ids
         for i, token_id in enumerate(token_ids):
             step_top_logprobs = top_logprobs[i]
-            if step_top_logprobs is None:
+            if step_top_logprobs is None or step_top_logprobs.get(
+                    token_id) is None:
                 token = tokenizer.decode(token_id)
                 if should_return_as_token_id:
                     token = f"token_id:{token_id}"