config: a parallel leaf with no live counterpart is read bare (#36620)

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

test/registered/unit/disaggregation/test_disaggregation_wire.py

@@ -158,16 +158,8 @@ class TestCPReplicatedStateTransfer(unittest.TestCase):
                 manager = object.__new__(CommonKVManager)
                 manager.attn_cp_size = cp_size
                 manager.attn_cp_rank = cp_rank
-                # The policy reads the configured tier, so the stand-in
-                # carries the leaf under `config`, where the bag serves it.
-                parallel = SimpleNamespace(
-                    config=SimpleNamespace(
-                        enable_dsa_cache_layer_split=layer_split,
-                    ),
-                )
-                with patch(
-                    "sglang.srt.disaggregation.common.conn.get_parallel",
-                    return_value=parallel,
+                with get_context().override_server_args(
+                    enable_dsa_cache_layer_split=layer_split,
                 ):
                     self.assertEqual(
                         manager._should_skip_cp_replicated_state_transfer(),
@@ -180,11 +172,8 @@ class TestCPReplicatedStateTransfer(unittest.TestCase):
         manager.attn_cp_rank = 3
         manager.is_hybrid_mla_backend = False
 
-        with patch(
-            "sglang.srt.disaggregation.common.conn.get_parallel",
-            return_value=SimpleNamespace(
-                config=SimpleNamespace(enable_dsa_cache_layer_split=False)
-            ),
+        with get_context().override_server_args(
+            enable_dsa_cache_layer_split=False,
         ):
             self.assertEqual(
                 manager._get_dsa_cache_transfer_skip_flags(None),

test/registered/unit/model_loader/test_presharded_loader.py

@@ -15,6 +15,7 @@ from unittest import mock
 import torch
 
 from sglang.srt.model_loader.loader import PreshardedModelLoader
+from sglang.srt.runtime_context import get_context, get_parallel
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=10, suite="base-a-test-cpu")
@@ -803,22 +804,23 @@ class TestShardConfig(unittest.TestCase):
             "init_expert_location",
             "structural_signature",
         }
-        # Both tiers on one stand-in: bare names are the live groups, `config`
-        # is the published parallel bag.
-        parallel = SimpleNamespace(
+        # The sizes go through both channels: some entries read the published
+        # leaf, others the live property. `get_moe_cp_size` is imported inside
+        # `_collect_shard_config`, so it is patched where it is defined.
+        override = get_context().override_server_args(
             tp_size=8,
-            moe_dp_size=2,
-            moe_ep_size=4,
             pp_size=1,
-            config=SimpleNamespace(
-                moe_dp_size=2,
-                moe_dense_tp_size=1,
-                enable_dp_lm_head=True,
-            ),
+            moe_dp_size=2,
+            moe_dense_tp_size=1,
+            enable_dp_lm_head=True,
         )
-        with mock.patch(
-            "sglang.srt.model_loader.loader.get_parallel",
-            return_value=parallel,
+        override.install()
+        self.addCleanup(override.restore)
+        with get_parallel().override(
+            tp_size=8, pp_size=1, moe_dp_size=2, moe_ep_size=4
+        ), mock.patch(
+            "sglang.srt.layers.dp_attention.get_moe_cp_size",
+            return_value=2,
         ), mock.patch(
             "sglang.srt.model_loader.loader.get_exec",
             return_value=SimpleNamespace(

test/registered/unit/models/test_shared_experts_fusion_gates.py

@@ -332,18 +332,9 @@ class TestBailingMoeV3Gate(_FusionGateCase):
             vocab_size=32000,
             hidden_size=4096,
         )
-        parallel = SimpleNamespace(
-            tp_size=1,
-            moe_ep_size=1,
-            config=SimpleNamespace(enable_dp_lm_head=False),
-        )
+        self._seed(enable_dp_lm_head=False)
         with (
-            unittest.mock.patch.object(
-                bailing_moe_nextn, "get_parallel", return_value=parallel
-            ),
-            unittest.mock.patch.object(
-                bailing_moe_v3, "get_parallel", return_value=parallel
-            ),
+            get_parallel().override(tp_size=1, moe_ep_size=1),
             unittest.mock.patch.object(
                 bailing_moe_v3,
                 "is_shared_experts_fusion_disabled",

test/registered/unit/multimodal/test_tensor_transport_mode.py

@@ -1,10 +1,9 @@
 """Tests for multimodal tensor transport topology detection."""
 
 import unittest
-from types import SimpleNamespace
-from unittest.mock import patch
 
 from sglang.srt.multimodal.transport import determine_tensor_transport_mode
+from sglang.srt.runtime_context import get_context
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
@@ -22,17 +21,9 @@ class TestTensorTransportMode(CustomTestCase):
 
         for nnodes, dist_init_addr, expected in cases:
             with self.subTest(nnodes=nnodes, dist_init_addr=dist_init_addr):
-                # `nnodes` is a config-only leaf, so the stand-in carries it
-                # under `config`, where the published bag serves it.
-                parallel = SimpleNamespace(
-                    config=SimpleNamespace(
-                        nnodes=nnodes,
-                        dist_init_addr=dist_init_addr,
-                    ),
-                )
-                with patch(
-                    "sglang.srt.multimodal.transport.get_parallel",
-                    return_value=parallel,
+                with get_context().override_server_args(
+                    nnodes=nnodes,
+                    dist_init_addr=dist_init_addr,
                 ):
                     self.assertEqual(determine_tensor_transport_mode(), expected)
 

test/registered/unit/server_args/test_resolution_is_reproducible.py

@@ -863,7 +863,7 @@ class TestACopyStaysResolved(_RestoresProcessState, CustomTestCase):
         self.addCleanup(reset_context)
         reset_context()
         publish(copy_, role="scheduler")
-        self.assertEqual(get_parallel().config.dist_init_addr, "1.2.3.4:5000")
+        self.assertEqual(get_parallel().dist_init_addr, "1.2.3.4:5000")
         self.assertEqual(
             get_schedule().chunked_prefill_size,
             resolution_result(parent, "chunked_prefill_size"),

test/registered/unit/test_launch_path_reads_configured_sizes.py

@@ -304,6 +304,7 @@ class TestLaunchPathsReadConfiguredSizes(CustomTestCase):
         from unittest.mock import patch
 
         from sglang.srt.runtime_context import (
+            ParallelContext,
             get_parallel,
             publish,
             reset_context,
@@ -383,15 +384,20 @@ class TestLaunchPathsReadConfiguredSizes(CustomTestCase):
                         f"get_parallel().config.{name} followed the live topology "
                         "instead of the published configuration",
                     )
-        # A bare read of a leaf with no live property is not a config read any
-        # more, and the error says where it went. Spelled through `getattr` so a
-        # mechanical `.config` sweep cannot "fix" the very read under test.
-        with self.assertRaisesRegex(
-            AttributeError, r"read it as get_parallel\(\)\.config\.nccl_port"
-        ):
-            getattr(get_parallel(), "nccl_port")
+        from sglang.srt.arg_groups.overrides import resolution_result
+
+        self.assertEqual(
+            resolution_result(server_args, "nccl_port"),
+            getattr(get_parallel(), "nccl_port"),
+            "a config-only leaf read bare disagreed with what resolution decided",
+        )
         reset_context()
 
+        with self.assertRaisesRegex(ValueError, r"'parallel' not published"):
+            getattr(ParallelContext(), "nccl_port")
+        with self.assertRaisesRegex(AttributeError, r"has no 'not_a_leaf'"):
+            getattr(ParallelContext(), "not_a_leaf")
+
     def test_no_live_topology_read_before_distributed_init(self):
         offenders = []
         for rel, tree in _launch_paths():

test/registered/unit/test_runtime_context.py

@@ -911,11 +911,11 @@ class TestForwardFlags(_IsolatedServerArgs):
             @torch.compile(fullgraph=True, backend="eager", dynamic=False)
             def probe(x):
                 par = get_parallel()
-                if par.config.enable_prefill_context_parallel:
+                if par.enable_prefill_context_parallel:
                     x = x + 1
-                if par.config.moe_dense_tp_size == 1:
+                if par.moe_dense_tp_size == 1:
                     x = x + 2
-                if par.config.dwdp_size > 1:
+                if par.dwdp_size > 1:
                     x = x + 4
                 return x