config: the lazy imports that buy nothing become eager (#36975)

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

test/registered/unit/server_args/test_page_major_backend_allowlist.py

@@ -30,9 +30,10 @@ under its own default configuration.
 """
 
 import unittest
-from unittest import mock
+from types import SimpleNamespace
 
 from sglang.srt.arg_groups.kv_cache_hook import handle_page_major_kv_layout
+from sglang.srt.configs.model_config import AttentionArch
 from sglang.srt.server_args import ServerArgs
 from sglang.test.ci.ci_register import register_cpu_ci
 
@@ -67,18 +68,18 @@ def _accepts(
         "mamba_backend": "triton",
     }.items():
         object.__setattr__(sa, name, value)
-    # `use_mla_backend` asks the model configuration, which this stand-in has
-    # no room for; the case under test is what the handler does with the answer.
-    # The handler imports it inside the function, so the source module is
-    # where the patch has to go.
-    with mock.patch(
-        "sglang.srt.arg_groups.overrides.use_mla_backend", return_value=use_mla
-    ):
-        try:
-            handle_page_major_kv_layout(sa)
-            return True
-        except AssertionError:
-            return False
+    object.__setattr__(
+        sa,
+        "_model_config",
+        SimpleNamespace(
+            attention_arch=AttentionArch.MLA if use_mla else AttentionArch.MHA
+        ),
+    )
+    try:
+        handle_page_major_kv_layout(sa)
+        return True
+    except AssertionError:
+        return False
 
 
 class TestPageMajorBackendAllowlist(unittest.TestCase):

test/registered/unit/server_args/test_resolution_declarations.py

@@ -934,47 +934,42 @@ class TestDeclaredValuesAreNotEditedLater(CustomTestCase):
         self.addCleanup(restore)
 
     def _resolve_recording_each_entry(self, **supplied):
-        """Resolve, deep-copying every stash entry the moment it is appended."""
-        from sglang.srt.arg_groups import overrides
+        """Resolve, deep-copying every stash entry the moment it is appended.
 
+        The property is about the stash, so the seam is the stash: a list that
+        snapshots on append. Every declaration path -- `declare_resolution`,
+        `declare_late_resolution`, `declare_direct_writes` and the passes --
+        reaches it through `.append`, whatever it was imported as.
+        """
         recorded = []
 
-        def watch(name):
-            original = getattr(overrides, name)
+        class _SnapshotOnAppend(list):
+            def append(self, entry):
+                super().append(entry)
+                recorded.append((len(self) - 1, copy.deepcopy(entry)))
 
-            def wrapper(server_args, *args, **kwargs):
-                result = original(server_args, *args, **kwargs)
-                stash = getattr(server_args, "_resolved_overrides", None) or []
-                while len(recorded) < len(stash):
-                    index = len(recorded)
-                    recorded.append((index, copy.deepcopy(stash[index])))
-                return result
+        class _WatchedArgs(ServerArgs):
+            """Whatever list the pipeline installs, snapshot what lands in it.
 
-            return original, wrapper
+            The pipeline resets the stash at the start of a resolution, so the
+            seam has to survive that assignment rather than precede it.
+            """
 
-        # Every path that appends to the stash.
-        patched = {}
-        for name in (
-            "declare_resolution",
-            "declare_late_resolution",
-            "declare_direct_writes",
-            "run_post_process_pass",
-        ):
-            original, wrapper = watch(name)
-            patched[name] = original
-            setattr(overrides, name, wrapper)
-        try:
-            path = tempfile.mkdtemp(prefix="declared_values_")
-            self.addCleanup(shutil.rmtree, path, ignore_errors=True)
-            with open(os.path.join(path, "config.json"), "w") as handle:
-                json.dump(_MINI_CONFIG, handle)
-            server_args = ServerArgs(
-                model_path=path, device="cuda", random_seed=42, **supplied
-            )
-            server_args.resolve_once()
-        finally:
-            for name, original in patched.items():
-                setattr(overrides, name, original)
+            def __setattr__(self, name, value):
+                if name == "_resolved_overrides" and not isinstance(
+                    value, _SnapshotOnAppend
+                ):
+                    value = _SnapshotOnAppend(value)
+                super().__setattr__(name, value)
+
+        path = tempfile.mkdtemp(prefix="declared_values_")
+        self.addCleanup(shutil.rmtree, path, ignore_errors=True)
+        with open(os.path.join(path, "config.json"), "w") as handle:
+            json.dump(_MINI_CONFIG, handle)
+        server_args = _WatchedArgs(
+            model_path=path, device="cuda", random_seed=42, **supplied
+        )
+        server_args.resolve_once()
         return server_args, recorded
 
     def test_no_entry_changes_after_it_is_recorded(self):

test/registered/unit/server_args/test_server_args.py

@@ -979,10 +979,7 @@ class TestFa4PageSizeAutoForce(CustomTestCase):
         return args
 
     @patch("sglang.srt.arg_groups.overrides.is_sm100_supported", return_value=True)
-    @patch("sglang.srt.arg_groups.overrides.use_mla_backend", return_value=False)
-    def test_combined_attention_backend_fa4_forces_page_size_128(
-        self, _mock_mla, _mock_sm100
-    ):
+    def test_combined_attention_backend_fa4_forces_page_size_128(self, _mock_sm100):
         # `--attention-backend fa4` (combined): prefill/decode fields stay None.
         args = self._make_args(attention_backend="fa4")
 
@@ -994,8 +991,7 @@ class TestFa4PageSizeAutoForce(CustomTestCase):
         self.assertEqual(resolved_view(args).page_size, 128)
 
     @patch("sglang.srt.arg_groups.overrides.is_sm100_supported", return_value=True)
-    @patch("sglang.srt.arg_groups.overrides.use_mla_backend", return_value=False)
-    def test_explicit_prefill_fa4_forces_page_size_128(self, _mock_mla, _mock_sm100):
+    def test_explicit_prefill_fa4_forces_page_size_128(self, _mock_sm100):
         # `--prefill-attention-backend fa4`: the previously-covered path.
         args = self._make_args(attention_backend=None, prefill="fa4", page_size=1)
 
@@ -1869,12 +1865,7 @@ class TestCudaGraphDisaggregationRoles(CustomTestCase):
             is_multimodal=False,
             is_multimodal_piecewise_cuda_graph_supported=False,
         )
-        with (
-            patch("sglang.srt.utils.is_cuda", return_value=True),
-            patch(
-                "sglang.srt.arg_groups.overrides.use_mla_backend", return_value=False
-            ),
-        ):
+        with patch("sglang.srt.utils.is_cuda", return_value=True):
             handle_cuda_graph_config(args)
         return args
 
@@ -1944,12 +1935,7 @@ class TestPrefillCudaGraphLoRACompatibility(CustomTestCase):
             is_multimodal=False,
             is_multimodal_piecewise_cuda_graph_supported=False,
         )
-        with (
-            patch("sglang.srt.utils.is_cuda", return_value=True),
-            patch(
-                "sglang.srt.arg_groups.overrides.use_mla_backend", return_value=False
-            ),
-        ):
+        with patch("sglang.srt.utils.is_cuda", return_value=True):
             handle_cuda_graph_config(args)
         return args
 
@@ -2010,12 +1996,7 @@ class TestBreakableCudaGraphMultimodalAllowlist(CustomTestCase):
             is_multimodal_piecewise_cuda_graph_supported=False,
             is_multimodal_breakable_cuda_graph_supported=allowlisted,
         )
-        with (
-            patch("sglang.srt.utils.is_cuda", return_value=True),
-            patch(
-                "sglang.srt.arg_groups.overrides.use_mla_backend", return_value=False
-            ),
-        ):
+        with patch("sglang.srt.utils.is_cuda", return_value=True):
             handle_cuda_graph_config(args)
         return args