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

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

@@ -93,6 +93,9 @@ def _make_prefill_aware_swa_runner(
         page_size=1,
         attn_cp_size=1,
         tp_size=1,
+        # The backend still reads the runner's frozen record for these two;
+        # same single-rank placement, stated where it looks for it.
+        ps=SimpleNamespace(attn_cp_size=1, tp_size=1),
         is_draft_worker=False,
         server_args=server_args,
         attention_chunk_size=None,

test/registered/unit/layers/moe/test_mega_moe_deepgemm_api.py

@@ -14,6 +14,7 @@ from sglang.test.test_utils import CustomTestCase, maybe_stub_sgl_kernel
 maybe_stub_sgl_kernel()
 
 from sglang.srt.layers.moe import mega_moe
+from sglang.srt.runtime_context import get_parallel
 
 register_cpu_ci(est_time=8, suite="base-a-test-cpu")
 
@@ -271,11 +272,8 @@ class TestDeepGemmMegaMoeApi(CustomTestCase):
                 "init_new",
                 return_value=object(),
             ),
-            patch(
-                "sglang.srt.runtime_context.get_parallel",
-                return_value=SimpleNamespace(
-                    moe_ep_group=SimpleNamespace(device_group=object())
-                ),
+            get_parallel().override(
+                moe_ep_group=SimpleNamespace(device_group=object())
             ),
         ):
             mega_moe._run_mega_routed(