diff --git a/python/sglang/srt/managers/router/model_rpc.py b/python/sglang/srt/managers/router/model_rpc.py
index f59c3f0a1..40ae6dd98 100644
--- a/python/sglang/srt/managers/router/model_rpc.py
+++ b/python/sglang/srt/managers/router/model_rpc.py
@@ -432,9 +432,14 @@ class ModelRpcServer(rpyc.Service):
                 req.logprob = logprobs[pt : pt + req.extend_input_len - 1]
                 req.normalized_logprob = normalized_logprobs[i]
 
-                token_ids = req.input_ids + [next_token_ids[i]]
-                token_logprobs = [None] + req.logprob + [last_logprobs[i]]
+                # If logprob_start_len > 0, then first logprob_start_len prompt tokens
+                # will be ignored.
+                prompt_token_len = len(req.logprob)
+                token_ids = req.input_ids[-prompt_token_len :] + [next_token_ids[i]]
+                token_logprobs = req.logprob + [last_logprobs[i]]
                 req.token_logprob = list(zip(token_ids, token_logprobs))
+                if req.logprob_start_len == 0:
+                    req.token_logprob = [(req.input_ids[0], None)] + req.token_logprob
                 pt += req.extend_input_len
 
         self.handle_finished_requests(batch)