diff --git a/lightllm/server/core/objs/sampling_params.py b/lightllm/server/core/objs/sampling_params.py
index 8e31c5062..25ac7f26f 100644
--- a/lightllm/server/core/objs/sampling_params.py
+++ b/lightllm/server/core/objs/sampling_params.py
@@ -66,31 +66,42 @@ def initialize(self, stop_sequences: Union[str, List[Union[List[int], str]]], to
         elif isinstance(stop_sequences, str):
             stop_sequences = [stop_sequences]
 
-        groups: List[List[int]] = self.stop_sentences_to_token_ids(stop_sequences, tokenizer)
+        # 这里必须使用 (token_ids, 原始条目) 的配对结果：空条目（如 "" 或 []）会被过滤掉，
+        # 若还按原始下标去取 stop_sequences[group_idx]，后面的条目就会和 token id 组错位，
+        # 导致 sequence_str 挂到别的组上（丢失停止字符串，或凭空多出一个停止字符串）。
+        groups: List[Tuple[List[int], Union[List[int], str]]] = self._stop_sentences_to_token_id_pairs(
+            stop_sequences, tokenizer
+        )
         self.size = len(groups)
         assert self.size <= MAX_STOP_SEQUENCES, "Too many stop sequence groups."
 
-        for group_idx in range(self.size):
-            if isinstance(stop_sequences[group_idx], str):
-                self.groups[group_idx].initialize(groups[group_idx], sequence_str=stop_sequences[group_idx])
+        for group_idx, (token_ids, stop_info) in enumerate(groups):
+            if isinstance(stop_info, str):
+                self.groups[group_idx].initialize(token_ids, sequence_str=stop_info)
             else:
-                self.groups[group_idx].initialize(groups[group_idx])
+                self.groups[group_idx].initialize(token_ids)
 
-    def stop_sentences_to_token_ids(self, stop_sequences: List[Union[List[int], str]], tokenizer) -> List[List[int]]:
+    def _stop_sentences_to_token_id_pairs(
+        self, stop_sequences: List[Union[List[int], str]], tokenizer
+    ) -> List[Tuple[List[int], Union[List[int], str]]]:
+        """返回 (token_ids, 原始 stop 条目) 的列表，保证两者一一对应。"""
         new_stop_sequences = []
         for stop_info in stop_sequences:
             if isinstance(stop_info, str):
                 stop_str_ids = self._stop_str_to_token_ids(stop_info, tokenizer)
                 if stop_str_ids is not None and len(stop_str_ids) > 0:
-                    new_stop_sequences.append(stop_str_ids)
+                    new_stop_sequences.append((stop_str_ids, stop_info))
             if isinstance(stop_info, list):
                 if all(isinstance(x, int) for x in stop_info):
                     if len(stop_info) > 0:
-                        new_stop_sequences.append(stop_info)
+                        new_stop_sequences.append((stop_info, stop_info))
                 else:
                     assert False, "stop_sequences item must be type List[int] when it is a list."
         return new_stop_sequences
 
+    def stop_sentences_to_token_ids(self, stop_sequences: List[Union[List[int], str]], tokenizer) -> List[List[int]]:
+        return [token_ids for token_ids, _ in self._stop_sentences_to_token_id_pairs(stop_sequences, tokenizer)]
+
     def _stop_str_to_token_ids(self, stop_str: str, tokenizer) -> List[int]:
         stop_str_ids = tokenizer.encode(stop_str, add_special_tokens=False)
         return stop_str_ids
@@ -202,7 +213,7 @@ class AllowedTokenIds(ctypes.Structure):
     def initialize(self, ids: List[int]):
         self.size = len(ids)
         assert self.size <= ALLOWED_TOKEN_IDS_MAX_LENGTH, "Too many allowed token IDs."
-        assert all(isinstance(e, int) for e in self.ids), "all must be int"
+        assert all(isinstance(e, int) for e in ids), "all must be int"
         self.ids[: self.size] = ids[:]
 
     def to_list(self):
diff --git a/unit_tests/server/core/objs/test_sampling_params.py b/unit_tests/server/core/objs/test_sampling_params.py
index ef3f08d2f..9ba46c152 100644
--- a/unit_tests/server/core/objs/test_sampling_params.py
+++ b/unit_tests/server/core/objs/test_sampling_params.py
@@ -5,10 +5,10 @@
     RegularConstraint,
     AllowedTokenIds,
     ExponentialDecayLengthPenalty,
-    DecodeNode,
     SamplingParams,
     GuidedGrammar,
     GuidedJsonSchema,
+    NodeUUId,
     STOP_SEQUENCE_MAX_LENGTH,
     REGULAR_CONSTRAINT_MAX_LENGTH,
     ALLOWED_TOKEN_IDS_MAX_LENGTH,
@@ -117,22 +117,41 @@ def test_exponential_decay_length_penalty_initialization():
         penalty.initialize((5, 0.5))
 
 
-def test_decode_node_initialization():
-    node = DecodeNode()
-    data = {
-        "node_id": 12345678901234567890,  # 示例 UUID
-        "ip": "192.168.1.1",
-        "rpyc_port": 8080,
-        "max_new_tokens": 10,
-    }
-    node.initialize(data)
-    assert node.exists is True
-    assert node.node_id.node_id_high == (12345678901234567890 >> 64) & 0xFFFFFFFFFFFFFFFF
-    assert node.node_id.node_id_low == 12345678901234567890 & 0xFFFFFFFFFFFFFFFF
-    assert node.ip[0] == 192
-    assert node.ip[1] == 168
-    assert node.ip[2] == 1
-    assert node.ip[3] == 1
+def test_node_uuid_roundtrip():
+    node_id = 12345678901234567890
+    uuid = NodeUUId()
+    uuid.initialize(node_id)
+    assert uuid.node_id_high == (node_id >> 64) & 0xFFFFFFFFFFFFFFFF
+    assert uuid.node_id_low == node_id & 0xFFFFFFFFFFFFFFFF
+    assert uuid.get() == node_id
+
+
+def test_allowed_token_ids_rejects_non_int():
+    allowed_ids = AllowedTokenIds()
+    with pytest.raises(AssertionError):
+        allowed_ids.initialize([1, "2", 3])
+
+
+@pytest.mark.parametrize(
+    "stop_sequences, expected_token_ids, expected_strings",
+    [
+        # 全部有效，无过滤，token id 组与字符串一一对应。
+        (["stop1", "stop2"], [[1, 2], [3, 4]], ["stop1", "stop2"]),
+        # 前置的空字符串会被过滤掉，后面的条目不能因此错位。
+        (["", "stop1"], [[1, 2]], ["stop1"]),
+        # 前置的空 id 列表同理。
+        ([[], "stop2"], [[3, 4]], ["stop2"]),
+        # 被过滤掉的字符串条目不能把自己的字符串挂到后一个条目的 token id 上。
+        (["unknown", "stop2"], [[3, 4]], ["stop2"]),
+        # 纯 id 条目不携带字符串。
+        ([[7, 8], "stop1"], [[7, 8], [1, 2]], ["stop1"]),
+    ],
+)
+def test_stop_sequence_groups_keeps_ids_and_strings_aligned(stop_sequences, expected_token_ids, expected_strings):
+    groups = StopSequenceGroups()
+    groups.initialize(stop_sequences, MockTokenizer())
+    assert groups.to_list() == expected_token_ids
+    assert sorted(groups.to_strings()) == sorted(expected_strings)
 
 
 def test_sampling_params_initialization():