diff --git a/python/sglang/srt/models/utils.py b/python/sglang/srt/models/utils.py
index 96bec941e9bc..16df5d01b60a 100644
--- a/python/sglang/srt/models/utils.py
+++ b/python/sglang/srt/models/utils.py
@@ -41,7 +41,7 @@
 )
 from sglang.srt.model_executor.runner import get_is_capture_mode
 from sglang.srt.model_loader.weight_utils import default_weight_loader
-from sglang.srt.runtime_context import get_exec
+from sglang.srt.runtime_context import get_exec, get_parallel
 from sglang.srt.utils import get_current_device_stream_fast, is_cpu, is_cuda, is_hip
 from sglang.srt.utils.custom_op import register_custom_op
 
@@ -296,6 +296,9 @@ def enable_fused_set_kv_buffer(forward_batch: ForwardBatch):
     `.view(-> 4D NHD)` reshape, and let the rotary forward pass
     `flash_layout=False`. See `create_fused_set_kv_buffer_arg` below.
     """
+    # The fused RoPE writer cannot localize DCP slots or filter token owners.
+    if get_parallel().dcp_enabled:
+        return False
     pool = get_token_to_kv_pool()
     return (
         _is_cuda
diff --git a/test/registered/unit/models/test_fused_set_kv_buffer_gate.py b/test/registered/unit/models/test_fused_set_kv_buffer_gate.py
index c71fafa2f31f..a56aa1251767 100644
--- a/test/registered/unit/models/test_fused_set_kv_buffer_gate.py
+++ b/test/registered/unit/models/test_fused_set_kv_buffer_gate.py
@@ -95,6 +95,11 @@ def _hip_gate(pool, dcp_kv_mask=None) -> bool:
     with (
         patch.object(model_utils, "_is_hip", True),
         patch.object(model_utils, "_is_cuda", False),
+        patch.object(
+            model_utils,
+            "get_parallel",
+            return_value=SimpleNamespace(dcp_enabled=False),
+        ),
         patch.object(model_utils, "get_token_to_kv_pool", return_value=pool),
     ):
         return model_utils.enable_fused_set_kv_buffer(
@@ -156,5 +161,35 @@ def test_decode_context_parallel_mask_declines(self):
         self.assertFalse(_hip_gate(_static_window_pool(), dcp_kv_mask=object()))
 
 
+class TestDcpFusedStoreGate(CustomTestCase):
+    def test_dcp_without_a_batch_mask_uses_the_backend_store(self):
+        pool = SimpleNamespace(
+            dtype=torch.bfloat16,
+            layer_router=SimpleNamespace(sub_pool=lambda name: None),
+        )
+        for cuda in (True, False):
+            for dcp_size in (1, 2, 4):
+                with (
+                    self.subTest(cuda=cuda, dcp_size=dcp_size),
+                    patch.object(model_utils, "_is_cuda", cuda),
+                    patch.object(model_utils, "_is_hip", not cuda),
+                    patch.object(model_utils, "is_cp_active", return_value=False),
+                    patch.object(
+                        model_utils, "get_token_to_kv_pool", return_value=pool
+                    ),
+                    patch.object(
+                        model_utils,
+                        "get_parallel",
+                        return_value=SimpleNamespace(dcp_enabled=dcp_size > 1),
+                    ),
+                ):
+                    self.assertEqual(
+                        model_utils.enable_fused_set_kv_buffer(
+                            SimpleNamespace(dcp_kv_mask=None)
+                        ),
+                        dcp_size == 1,
+                    )
+
+
 if __name__ == "__main__":
     unittest.main()