Deprecate the parallel getters the context answers, and ratchet them shut (#40342)

test/registered/unit/mem_cache/test_mamba_donated_alloc_ratio.py

@@ -297,7 +297,9 @@ class TestPPMambaPoolSizing(unittest.TestCase):
             server_args=SimpleNamespace(),
             spec_algorithm=SimpleNamespace(is_none=lambda: True),
             layer_info=SimpleNamespace(start_layer=start, end_layer=end),
-            ps=SimpleNamespace(attn_dp_size=1, pp_size=pp_size),
+            # The runner carries its placement as plain attributes.
+            attn_dp_size=1,
+            pp_size=pp_size,
             hybrid_gdn_config=None,
             model_config=SimpleNamespace(
                 hf_config=SimpleNamespace(), num_hidden_layers=cls.TOTAL_LAYERS

test/registered/unit/mem_cache/test_mem_pool_host.py

@@ -304,16 +304,13 @@ class TestHostMemoryBudget(CustomTestCase):
     def test_ranks_per_host_divides_world_size_by_nodes(self):
         # The launcher slices ranks uniformly across nodes, so the co-located
         # rank count is world_size // nnodes — no hostname collective.
-        fake_group = unittest.mock.Mock(world_size=16)
+        # tp_size=16 states the launch width the count divides -- the
+        # published configuration is where ranks_per_host reads it from.
         with (
-            get_context().override_server_args(nnodes=2),
+            get_context().override_server_args(nnodes=2, tp_size=16),
             unittest.mock.patch.object(
                 torch.distributed, "is_initialized", return_value=True
             ),
-            unittest.mock.patch(
-                "sglang.srt.distributed.parallel_state.get_world_group",
-                return_value=fake_group,
-            ),
         ):
             self.assertEqual(base.ranks_per_host(), 8)
 

test/registered/unit/mem_cache/test_unified_mla_views.py

@@ -248,7 +248,7 @@ class TestUnifiedMLATokenToKVPool(unittest.TestCase):
                     ] = float(layer + 1)
 
                 with (
-                    get_parallel().override(dcp_enabled=False),
+                    get_parallel().override(dcp_enabled=False, attn_dcp_rank=0),
                     mock.patch(
                         "sglang.srt.mem_cache.memory_pool.current_platform.synchronize"
                     ),
@@ -411,7 +411,9 @@ class TestMambaAllocatorCpuCopyIsPhysical(unittest.TestCase):
             _FakeKVCache(pool.max_slots("full")),
             _FakeKVCache(pool.max_slots("mamba")),
         )
-        with get_parallel().override(dcp_enabled=False, attn_dcp_size=1):
+        with get_parallel().override(
+            dcp_enabled=False, attn_dcp_size=1, attn_dcp_rank=0
+        ):
             allocator = UnifiedMambaTokenToKVPoolAllocator(
                 unified_buffer=pool, kvcache=kvcache, device=_DEV, page_size=ps
             )