config: make ServerArgs read-only with a single audited mutation entry (#31811)

examples/runtime/token_in_token_out/token_in_token_out_vlm_engine.py

@@ -67,8 +67,8 @@ if __name__ == "__main__":
     ServerArgs.add_cli_args(parser)
     args = [
         "--model-path=Qwen/Qwen2-VL-2B",
+        "--skip-tokenizer-init",
     ]
     args = parser.parse_args(args=args)
     server_args = ServerArgs.from_cli_args(args)
-    server_args.skip_tokenizer_init = True
     token_in_out_example(server_args)

python/sglang/srt/disaggregation/encode_server.py

@@ -3331,8 +3331,9 @@ async def run_dp_worker(
     # 0 when CVD is pinned to one GPU, else the absolute id. rank=0, so
     # MMEncoder runs set_device(base_gpu_id).
     args = copy.deepcopy(server_args)
-    args.base_gpu_id = gpu_id
+    # The copy is already resolved (read-only); route the per-worker
-    args.tp_size = 1
+    # specialization through the audited mutation entry.
+    args.override("encode_server.dp_worker", base_gpu_id=gpu_id, tp_size=1)
     enc = MMEncoder(args, dist_init_method=f"tcp://127.0.0.1:{get_free_port()}", rank=0)
 
     global encoder_metrics_collector

python/sglang/srt/layers/attention/linear/gdn_backend.py

@@ -93,7 +93,12 @@ def maybe_set_default_flashinfer_gdn_prefill(model_runner: ModelRunner) -> None:
     )
 
     if is_flashinfer_gdn_prefill_available():
-        args.linear_attn_prefill_backend = "flashinfer"
+        # server_args is resolved (read-only) by the time backends initialize;
+        # route this load-time default through the audited mutation entry.
+        args.override(
+            "gdn_backend.sm100_flashinfer_default",
+            linear_attn_prefill_backend="flashinfer",
+        )
         rank0_log("Defaulting SM100 GDN prefill backend to FlashInfer.")
 
 

python/sglang/srt/runtime_context.py

@@ -655,6 +655,7 @@ class RuntimeContext:
         "parallel",
         "_server_args",
         "_config_bags",
+        "_overrides_log",
         "flags",
         "resources",
         "forward",
@@ -664,6 +665,7 @@ class RuntimeContext:
         self.parallel = parallel
         self._server_args: ServerArgs | None = None
         self._config_bags: dict | None = None
+        self._overrides_log: list = []
         self.flags = Flags()
         self.resources = Resources()
         self.forward = ForwardFlags()
@@ -725,6 +727,8 @@ class RuntimeContext:
         # truth for config reads). Driven by NS(...) metadata; a mock/partial
         # config with no NS markers yields an empty tree (no bags projected).
         self._config_bags = _build_config_bags(server_args)
+        # Fresh config lifecycle: prior override provenance no longer applies.
+        self._overrides_log = []
 
     def config_bag(self, name: str) -> _ConfigBag:
         """Return the top-level config namespace bag (``device`` / ``model`` /
@@ -736,6 +740,55 @@ class RuntimeContext:
             raise ValueError(f"config namespace {name!r} not published")
         return bags[name]
 
+    def override(self, source: str, **fields) -> None:
+        """The business mutation entry: write resolved config
+        leaves onto the namespace bags — the single source of truth. It does
+        **not** touch ``server_args`` (the pristine startup record) and there is
+        no write-through, so the old "wrote one store, read another" desync class
+        cannot occur.
+
+        Each flat field name is routed to its bag by the ``NS`` metadata (flat
+        names are unique across namespaces). Validation is all-or-nothing: an
+        unknown / unprojected field aborts before any write. ``source`` is
+        recorded for provenance / reproduction.
+        """
+        if not fields:
+            return
+        bags = self._config_bags
+        if bags is None:
+            raise ValueError("config not published; cannot override")
+        from sglang.srt.arg_groups.arg_utils import namespace_of
+
+        nsmap = namespace_of(type(self._server_args))
+        targets = []  # (bag, leaf, value) — resolved before any write
+        for name, value in fields.items():
+            path = nsmap.get(name)
+            if path is None:
+                raise ValueError(
+                    f"override: unknown config field {name!r} (no NS namespace) — "
+                    "not a resolved config leaf"
+                )
+            parts = path.split(".")
+            bag = bags.get(parts[0])
+            if bag is None:
+                raise ValueError(f"override: namespace {parts[0]!r} not published")
+            for seg in parts[1:]:
+                bag = object.__getattribute__(bag, "_subs").get(seg)
+                if bag is None:
+                    raise ValueError(
+                        f"override: subgroup {seg!r} missing under {path!r}"
+                    )
+            if name not in bag:
+                raise ValueError(f"override: field {name!r} not projected on {path!r}")
+            targets.append((bag, name, value))
+        for bag, name, value in targets:
+            bag._set(name, value)
+        self._overrides_log.append((source, dict(fields)))
+
+    def overrides_log(self) -> list:
+        """Provenance of post-publish ``override`` calls: ``[(source, {field: value})]``."""
+        return list(self._overrides_log)
+
     def override_server_args(self, **fields) -> _ServerArgsOverride:
         """Test-only scoped override for the config tier — the sibling of
         ``get_parallel().override()`` and the flag groups' ``override()``:
@@ -927,6 +980,7 @@ def reset_context() -> None:
     """
     _CONTEXT._server_args = None
     _CONTEXT._config_bags = None
+    _CONTEXT._overrides_log = []
     _CONTEXT.flags = Flags()
     _CONTEXT.resources = Resources()
     _CONTEXT.forward = ForwardFlags()

python/sglang/srt/server_args.py

@@ -7877,20 +7877,21 @@ class ServerArgs:
             object.__setattr__(self, "_in_override", False)
 
     def __setattr__(self, name, value):
-        # After materialization the fields are the resolved configuration:
+        # after materialization the fields are the resolved startup
-        # under the strict test harness, a bare assignment outside
+        # configuration -- the pristine, READ-ONLY record. A bare assignment
-        # ServerArgs.override() (and the resolution pipeline itself) raises.
+        # outside ServerArgs.override() (and the resolution pipeline, which runs
+        # before materialization) always raises; resolved config is mutated on
+        # the context bags via get_context().override(...), not here. (Formerly
+        # gated on SGLANG_STRICT_CONFIG_MUTATION; now unconditional.)
         if (
             not name.startswith("_")
             and getattr(self, "_declarations_materialized", False)
             and not getattr(self, "_in_override", False)
         ):
-            from sglang.srt.environ import envs
-
-            if envs.SGLANG_STRICT_CONFIG_MUTATION.get():
             raise AttributeError(
-                    f"server_args.{name} assigned after resolution; use "
+                f"server_args.{name} assigned after resolution; server_args is "
-                    "server_args.override(source, ...) instead."
+                "read-only -- use get_context().override(source, ...) to change "
+                "resolved config."
             )
         object.__setattr__(self, name, value)
 

test/registered/unit/layers/attention/test_gdn_prefill_backend_policy.py

@@ -35,6 +35,14 @@ def make_runner(
     for name, value in arg_overrides.items():
         setattr(args, name, value)
 
+    # The policy routes its load-time default through the audited mutation entry
+    # (server_args.override); mirror that on the stub so the write lands.
+    def _override(source, **fields):
+        for _field, _value in fields.items():
+            setattr(args, _field, _value)
+
+    args.override = _override
+
     return SimpleNamespace(
         server_args=args,
         model_config=SimpleNamespace(),

test/registered/unit/test_runtime_context_override.py

@@ -0,0 +1,88 @@
+"""Context-first mutation.
+
+``get_context().override(source, **fields)`` is the business mutation entry: it
+writes the resolved config bags (the single source of truth) and never touches
+``server_args`` (the pristine startup record). Routing is by NS metadata; a bad
+field aborts before any write; provenance is recorded.
+"""
+
+import unittest
+
+from sglang.srt import runtime_context as rc
+from sglang.srt.server_args import ServerArgs
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_cpu_ci(est_time=5, suite="base-a-test-cpu")
+
+
+class TestContextOverride(CustomTestCase):
+    def setUp(self):
+        rc.reset_context()
+
+    def tearDown(self):
+        rc.reset_context()
+
+    def _publish(self):
+        sa = ServerArgs(model_path="dummy")
+        rc.get_context().set_server_args(sa)
+        return sa
+
+    def test_override_writes_bag_not_server_args(self):
+        sa = self._publish()
+        before = sa.hicache_ratio
+        rc.get_context().override("test", hicache_ratio=before + 1.0)
+        self.assertEqual(rc.get_memory().hicache_ratio, before + 1.0)
+        # server_args stays the pristine startup record.
+        self.assertEqual(sa.hicache_ratio, before)
+
+    def test_override_routes_across_namespaces(self):
+        self._publish()
+        rc.get_context().override(
+            "test",
+            moe_runner_backend="triton",
+            page_size=64,
+            disaggregation_mode="decode",
+        )
+        self.assertEqual(rc.get_exec().moe.moe_runner_backend, "triton")
+        self.assertEqual(rc.get_schedule().page_size, 64)
+        self.assertEqual(rc.get_disagg().disaggregation_mode, "decode")
+
+    def test_override_unknown_field_raises_and_is_atomic(self):
+        self._publish()
+        before = rc.get_memory().hicache_ratio
+        with self.assertRaises(ValueError):
+            rc.get_context().override(
+                "test", hicache_ratio=before + 5.0, not_a_real_field=1
+            )
+        # No partial write: the valid field was not applied.
+        self.assertEqual(rc.get_memory().hicache_ratio, before)
+
+    def test_override_before_publish_raises(self):
+        with self.assertRaises(ValueError):
+            rc.get_context().override("test", page_size=32)
+
+    def test_override_provenance_recorded(self):
+        self._publish()
+        rc.get_context().override("srcA", page_size=16)
+        log = rc.get_context().overrides_log()
+        self.assertEqual(log[-1], ("srcA", {"page_size": 16}))
+
+    def test_republish_resets_provenance(self):
+        self._publish()
+        rc.get_context().override("srcA", page_size=16)
+        self.assertTrue(rc.get_context().overrides_log())
+        self._publish()
+        self.assertEqual(rc.get_context().overrides_log(), [])
+
+    def test_bare_server_args_write_raises_after_resolution(self):
+        # server_args is read-only after resolution regardless of the
+        # SGLANG_STRICT_CONFIG_MUTATION env; write via override instead.
+        sa = ServerArgs(model_path="dummy")
+        object.__setattr__(sa, "_declarations_materialized", True)
+        with self.assertRaises(AttributeError):
+            sa.page_size = 999
+
+
+if __name__ == "__main__":
+    unittest.main()