Support true on-policy (#12058)

test/srt/cpu/test_shared_expert.py

@@ -20,6 +20,7 @@ from utils import (
     torch_w8a8_per_column_moe,
 )
 
+from sglang.srt.server_args import ServerArgs, set_global_server_args_for_scheduler
 from sglang.test.test_utils import CustomTestCase
 
 torch.manual_seed(1234)
@@ -149,6 +150,8 @@ class TestSharedExpert(CustomTestCase):
                 self._int8_shared_expert(*params)
 
     def _fp8_shared_expert(self, M, N, K, routed_scaling_factor):
+        set_global_server_args_for_scheduler(ServerArgs(model_path="dummy"))
+
         dtype = torch.bfloat16
         prepack = True