fix session request with None tokenizer

python/sglang/srt/managers/scheduler.py

@@ -1488,7 +1488,9 @@ def handle_generate_request(
         else:
             # Create a new request from a previous session
             session = self.sessions[recv_req.session_params.id]
-            req = session.create_req(recv_req, self.tokenizer)
+            req = session.create_req(
+                recv_req, self.tokenizer, self.model_config.vocab_size
+            )
             if isinstance(req.finished_reason, FINISH_ABORT):
                 self.init_req_max_new_tokens(req)
                 self._add_request_to_queue(req)

python/sglang/srt/managers/session_controller.py

@@ -65,7 +65,7 @@ def __init__(self, capacity_of_str_len: int, session_id: Optional[str] = None):
         self.capacity_of_str_len = capacity_of_str_len
         self.req_nodes: Dict[str, SessionReqNode] = {}
 
-    def create_req(self, req: TokenizedGenerateReqInput, tokenizer):
+    def create_req(self, req: TokenizedGenerateReqInput, tokenizer, vocab_size: int):
         assert req.session_params is not None
         session_params = req.session_params
 
@@ -144,7 +144,7 @@ def create_req(self, req: TokenizedGenerateReqInput, tokenizer):
             return_logprob=req.return_logprob,
             top_logprobs_num=req.top_logprobs_num,
             token_ids_logprob=req.token_ids_logprob,
-            vocab_size=tokenizer.vocab_size,
+            vocab_size=vocab_size,
         )
         if last_req is not None:
             new_req.multimodal_inputs = last_req.multimodal_inputs