Add DCP to runtime parallel context (#30478)

test/registered/unit/test_parallel_adoption_ratchet.py

@@ -7,10 +7,10 @@ which gives one import, one naming scheme, and the scoped ``override()``
 test primitive. Direct calls to the ``parallel_state`` size/rank getters in
 these directories are regressions against that sweep.
 
-Exemptions, pinned by path: ``layers/dp_attention.py`` is delegation
-substrate (the wrapper's attn-DP dims delegate TO it), and ``layers/dcp/``
-is the DCP subsystem's own plumbing, booked for a follow-up sweep. Sweeping
-an exempt path must remove it from the pin.
+Exemptions, pinned by path: ``runtime_context.py`` and
+``layers/dp_attention.py`` are delegation substrate, while
+``layers/dcp/comm.py`` retains deprecated DCP compatibility shims for
+out-of-tree callers. Sweeping an exempt path must remove it from the pin.
 """
 
 from sglang.test.ci.ci_register import register_cpu_ci
@@ -27,7 +27,7 @@ from sglang.test.test_utils import CustomTestCase
 _SRT_ROOT = Path(next(iter(sglang.srt.__path__)))
 
 _BANNED_CALLS = re.compile(
-    r"\bget_(?:"
+    r"\b(?:dcp_enabled|get_(?:"
     r"tensor_model_parallel_(?:world_size|rank)"
     r"|pipeline_model_parallel_(?:world_size|rank)"
     r"|moe_expert_parallel_(?:world_size|rank)"
@@ -36,8 +36,10 @@ _BANNED_CALLS = re.compile(
     r"|attn_tensor_model_parallel_(?:world_size|rank)"
     r"|attn_context_model_parallel_(?:world_size|rank)"
     r"|dcp_(?:world_size|rank)"
+    r"|dcp_group(?:_no_assert)?"
+    r"|attention_dcp_(?:world_size|rank)"
     r"|attention_(?:tp|cp)_(?:group|rank|size)"
-    r")\(\)"
+    r"))\(\)"
 )
 
 # The whole package is swept; the exemptions are the substrate itself.
@@ -45,7 +47,9 @@ _SWEPT_DIRS = ("",)
 
 _EXEMPT = (
     "distributed/",  # parallel_state: defines the canonical getters
+    "runtime_context.py",  # delegates DCP reads to canonical getters
     "layers/dp_attention.py",  # delegation substrate for the attn-DP dims
+    "layers/dcp/comm.py",  # deprecated out-of-tree DCP compatibility shims
     # The dumper's megatron plugin calls third-party getters that share the
     # parallel_state names (self._mpu.get_tensor_model_parallel_rank()).
     "debug_utils/dumper.py",

test/registered/unit/test_runtime_context.py

@@ -32,6 +32,8 @@ SIZE_RANK_DELEGATIONS = [
     ("world_rank", f"{_PS}.get_world_rank"),
     ("tp_size", f"{_PS}.get_tensor_model_parallel_world_size"),
     ("tp_rank", f"{_PS}.get_tensor_model_parallel_rank"),
+    ("dcp_size", f"{_PS}.get_dcp_world_size"),
+    ("dcp_rank", f"{_PS}.get_dcp_rank"),
     ("pp_size", f"{_PS}.get_pipeline_model_parallel_world_size"),
     ("pp_rank", f"{_PS}.get_pipeline_model_parallel_rank"),
     ("moe_ep_size", f"{_PS}.get_moe_expert_parallel_world_size"),
@@ -51,6 +53,7 @@ SIZE_RANK_DELEGATIONS = [
 GROUP_DELEGATIONS = [
     ("world_group", f"{_PS}.get_world_group"),
     ("tp_group", f"{_PS}.get_tp_group"),
+    ("dcp_group", f"{_PS}.get_dcp_group"),
     ("pp_group", f"{_PS}.get_pp_group"),
     ("moe_ep_group", f"{_PS}.get_moe_ep_group"),
     ("moe_dp_group", f"{_PS}.get_moe_dp_group"),
@@ -151,6 +154,40 @@ class TestParallelOverride(_IsolatedOverrides):
         self.assertEqual(p._overrides, {})
 
 
+class TestParallelDCP(_IsolatedOverrides):
+    def test_attn_dcp_defaults_when_group_is_uninitialized(self):
+        with (
+            patch(f"{_PS}.get_dcp_group_no_assert", return_value=None),
+            patch(f"{_PS}.get_dcp_world_size", side_effect=AssertionError),
+            patch(f"{_PS}.get_dcp_rank", side_effect=AssertionError),
+        ):
+            self.assertFalse(get_parallel().dcp_enabled)
+            self.assertEqual(get_parallel().attn_dcp_size, 1)
+            self.assertEqual(get_parallel().attn_dcp_rank, 0)
+
+    def test_attn_dcp_delegates_when_enabled(self):
+        with (
+            patch(f"{_PS}.get_dcp_group_no_assert", return_value=object()),
+            patch(f"{_PS}.get_dcp_world_size", return_value=8),
+            patch(f"{_PS}.get_dcp_rank", return_value=3),
+        ):
+            self.assertTrue(get_parallel().dcp_enabled)
+            self.assertEqual(get_parallel().attn_dcp_size, 8)
+            self.assertEqual(get_parallel().attn_dcp_rank, 3)
+
+    def test_dcp_enablement_is_platform_agnostic(self):
+        with (
+            patch(f"{_PS}.get_dcp_group_no_assert", return_value=object()),
+            patch("sglang.srt.utils.is_cuda", return_value=False) as is_cuda,
+            patch(f"{_PS}.get_dcp_world_size", return_value=8),
+            patch(f"{_PS}.get_dcp_rank", return_value=3),
+        ):
+            self.assertTrue(get_parallel().dcp_enabled)
+            self.assertEqual(get_parallel().attn_dcp_size, 8)
+            self.assertEqual(get_parallel().attn_dcp_rank, 3)
+            is_cuda.assert_not_called()
+
+
 class _IsolatedServerArgs(CustomTestCase):
     """Save/restore the published ServerArgs around each test (the slot is
     process-global; another test file sharing the process may have published)."""