Slice returned routed_experts at routed_experts_start_len

python/sglang/srt/entrypoints/openai/protocol.py

@@ -588,6 +588,10 @@ class ChatCompletionRequest(BaseModel):
     parallel_tool_calls: bool = True
     return_hidden_states: bool = False
     return_routed_experts: bool = False
+    # Skip this many leading token positions in the returned routed_experts
+    # tensor; multi-turn clients that already hold the prefix rows receive
+    # only the rows for newly processed tokens.
+    routed_experts_start_len: int = 0
     return_cached_tokens_details: bool = False
     return_prompt_token_ids: bool = False
     return_completion_token_ids: bool = False

python/sglang/srt/entrypoints/openai/serving_chat.py

@@ -370,6 +370,7 @@ def _convert_to_internal_request(
             disagg_prefill_dp_rank=request.disagg_prefill_dp_rank,
             return_hidden_states=request.return_hidden_states,
             return_routed_experts=request.return_routed_experts,
+            routed_experts_start_len=request.routed_experts_start_len,
             rid=request.rid,
             extra_key=self._compute_extra_key(request),
             require_reasoning=self._get_reasoning_from_request(request),

python/sglang/srt/function_call/multi_format_detector.py

@@ -1030,7 +1030,9 @@ def _deserialize_glm_value(value: str) -> Any:
         except Exception:
             pass
         try:
-            return ast.literal_eval(value)
+            result = ast.literal_eval(value)
+            json.dumps(result)
+            return result
         except Exception:
             pass
         return value

python/sglang/srt/managers/tokenizer_manager.py

@@ -1604,6 +1604,9 @@ def _handle_batch_output(
             if getattr(recv_obj, "routed_experts", None):
                 routed_experts_tensor = recv_obj.routed_experts[i]
                 if routed_experts_tensor is not None:
+                    start_len = getattr(state.obj, "routed_experts_start_len", 0) or 0
+                    if start_len > 0:
+                        routed_experts_tensor = routed_experts_tensor[start_len:]
                     meta_info["routed_experts"] = pybase64.b64encode(
                         routed_experts_tensor.numpy().tobytes()
                     ).decode("utf-8")

test/registered/function_call/test_multi_format_detector.py

@@ -276,6 +276,18 @@ def test_python_tuple_literal_value(self):
         # ast.literal handles tuple syntax; serialized as a JSON array.
         self.assertEqual(args["days"], [1, 2])
 
+    def test_python_set_literal_value_stays_string(self):
+        text = (
+            "<tool_call>get_weather"
+            "<arg_key>days</arg_key><arg_value>{1, 2}</arg_value>"
+            "</tool_call>"
+        )
+        result = self.det.detect_and_parse(text, self.tools)
+        args = json.loads(result.calls[0].parameters)
+        # ast.literal_eval yields a set, which JSON cannot represent; the raw
+        # string is kept so the arguments dict stays serializable.
+        self.assertEqual(args["days"], "{1, 2}")
+
 
 class TestGptOssDialect(unittest.TestCase):
     def setUp(self):