config: the runner and scheduler read resolved config from the bags (#34095)

test/registered/unit/layers/quantization/test_fp4_kv_cache_quant_method.py

@@ -72,10 +72,16 @@ class TestKVCacheQuantRegistry(CustomTestCase):
         from types import SimpleNamespace
 
         from sglang.srt.model_executor.model_runner import ModelRunner
+        from sglang.srt.runtime_context import get_context
 
         runner = object.__new__(ModelRunner)
-        runner.server_args = SimpleNamespace(kv_cache_dtype="fp4_e2m1")
+        runner.server_args = SimpleNamespace()
         runner.draft_attention_backend = None
+        # The runner reads the requested dtype off the model bag, so the double
+        # publishes it rather than carrying it on a stand-in config.
+        override = get_context().override_server_args(kv_cache_dtype="fp4_e2m1")
+        override.install()
+        self.addCleanup(override.restore)
         with self.assertRaisesRegex(ValueError, "fp4_mx_block16"):
             runner.configure_kv_cache_dtype()
 

test/registered/unit/managers/test_scheduler_init_req_max_new_tokens.py

@@ -4,6 +4,7 @@ from types import SimpleNamespace
 
 from sglang.srt.environ import envs
 from sglang.srt.managers.scheduler import Scheduler
+from sglang.srt.runtime_context import get_parallel
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=1, suite="base-a-test-cpu")
@@ -36,6 +37,14 @@ class TestSchedulerInitReqMaxNewTokens(unittest.TestCase):
     def tearDownClass(cls):
         cls._scheduler_logger.setLevel(cls._old_level)
 
+    def setUp(self):
+        # The scheduler scales the budget by the live DCP size
+        # (`get_parallel().attn_dcp_size`), so the double states a topology
+        # rather than publishing a config it does not otherwise need.
+        cm = get_parallel().override(attn_dcp_size=1)
+        cm.__enter__()
+        self.addCleanup(cm.__exit__, None, None, None)
+
     def _new_scheduler(
         self,
         max_req_len: int = 128,