diff --git a/tests/kernels/moe/test_deepep_v2_async_finalize.py b/tests/kernels/moe/test_deepep_v2_async_finalize.py
new file mode 100644
index 0000000000..193d8906f3
--- /dev/null
+++ b/tests/kernels/moe/test_deepep_v2_async_finalize.py
@@ -0,0 +1,122 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Contract tests for DeepEPV2PrepareAndFinalize.finalize_async.
+
+Hermetic: fake buffer/event, single process, no GPU collectives — only
+requires deep_ep to be importable.
+
+  T1 finalize_async  -> combine is issued with async_with_compute_stream=True
+                        and the join is deferred to the receiver, which waits
+                        on the combine event (device-side) before the copy.
+  T2 DBO active      -> combine falls back to async_with_compute_stream=False;
+                        the receiver still produces the correct output.
+  T3 finalize (sync) -> combine is issued synchronously and the output is
+                        copied immediately.
+"""
+
+import pytest
+import torch
+
+from vllm.utils.import_utils import has_deep_ep_v2
+
+requires_deep_ep_v2 = pytest.mark.skipif(
+    not has_deep_ep_v2(),
+    reason="Requires DeepEP v2 (ElasticBuffer)",
+)
+
+if has_deep_ep_v2():
+    import vllm.model_executor.layers.fused_moe.prepare_finalize.deepep_v2 as _dv2
+    from vllm.model_executor.layers.fused_moe.prepare_finalize.deepep_v2 import (
+        DeepEPV2PrepareAndFinalize,
+    )
+    from vllm.model_executor.layers.fused_moe.topk_weight_and_reduce import (
+        TopKWeightAndReduceContiguous,
+    )
+
+
+class _FakeEvent:
+    def __init__(self, has_event: bool):
+        self.event = object() if has_event else None
+        self.waited = False
+
+    def current_stream_wait(self):
+        assert self.event is not None
+        self.waited = True
+
+
+class _FakeBuffer:
+    def __init__(self, out: torch.Tensor):
+        self.out = out
+        self.calls: list[dict] = []
+        self.last_event: _FakeEvent | None = None
+
+    def combine(
+        self,
+        x,
+        handle,
+        topk_weights,
+        async_with_compute_stream,
+        allocate_on_comm_stream,
+    ):
+        self.calls.append({"async_with_compute_stream": async_with_compute_stream})
+        # DeepEP returns a completion event only for an async combine.
+        self.last_event = _FakeEvent(has_event=async_with_compute_stream)
+        return self.out, None, self.last_event
+
+
+def _make_pf(out: torch.Tensor):
+    pf = DeepEPV2PrepareAndFinalize(
+        buffer=_FakeBuffer(out),
+        num_dispatchers=1,
+        dp_size=1,
+        rank_expert_offset=0,
+        num_experts=8,
+        num_topk=2,
+    )
+    pf.handles[0] = object()
+    return pf
+
+
+def _run(pf, output: torch.Tensor, do_async: bool):
+    empty = torch.empty(0, dtype=torch.bfloat16)
+    weights = torch.empty(0, 2)
+    ids = torch.empty(0, 2, dtype=torch.int64)
+    fn = pf.finalize_async if do_async else pf.finalize
+    return fn(output, empty, weights, ids, False, TopKWeightAndReduceContiguous())
+
+
+@requires_deep_ep_v2
+def test_finalize_async_defers_join_to_receiver():
+    out_src = torch.full((4, 8), 7.0, dtype=torch.bfloat16)
+    pf = _make_pf(out_src)
+    dst = torch.zeros_like(out_src)
+    recv = _run(pf, dst, do_async=True)
+    assert pf.buffer.calls[0]["async_with_compute_stream"] is True
+    assert not torch.equal(dst, out_src), "join must not happen before receiver"
+    recv()
+    assert pf.buffer.last_event.waited, "receiver must join via event wait"
+    assert torch.equal(dst, out_src)
+
+
+@requires_deep_ep_v2
+def test_finalize_async_under_dbo_uses_sync_combine(monkeypatch):
+    monkeypatch.setattr(_dv2, "dbo_enabled", lambda: True)
+    out_src = torch.full((4, 8), 7.0, dtype=torch.bfloat16)
+    pf = _make_pf(out_src)
+    dst = torch.zeros_like(out_src)
+    recv = _run(pf, dst, do_async=True)
+    assert pf.buffer.calls[0]["async_with_compute_stream"] is False
+    recv()
+    assert not pf.buffer.last_event.waited, "no event to wait on for sync combine"
+    assert torch.equal(dst, out_src)
+
+
+@requires_deep_ep_v2
+def test_finalize_sync():
+    out_src = torch.full((4, 8), 7.0, dtype=torch.bfloat16)
+    pf = _make_pf(out_src)
+    dst = torch.zeros_like(out_src)
+    ret = _run(pf, dst, do_async=False)
+    assert ret is None
+    assert pf.buffer.calls[0]["async_with_compute_stream"] is False
+    assert torch.equal(dst, out_src)
diff --git a/vllm/model_executor/layers/fused_moe/prepare_finalize/deepep_v2.py b/vllm/model_executor/layers/fused_moe/prepare_finalize/deepep_v2.py
index d834c6eff7..b3a146274d 100644
--- a/vllm/model_executor/layers/fused_moe/prepare_finalize/deepep_v2.py
+++ b/vllm/model_executor/layers/fused_moe/prepare_finalize/deepep_v2.py
@@ -21,6 +21,7 @@ from vllm.triton_utils import tl, triton
 from vllm.utils.math_utils import round_up
 from vllm.v1.worker.ubatching import (
     dbo_current_ubatch_id,
+    dbo_enabled,
 )
 
 
@@ -89,8 +90,10 @@ class DeepEPV2PrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeModular):
         use cudagraphs anyway
       - Provides expert_tokens_meta for efficient batched expert kernels
 
-    Both modes use async_with_compute_stream=False (synchronous from
-    caller's perspective). The ElasticBuffer handles comm internally.
+    Dispatch always uses async_with_compute_stream=False. finalize_async
+    issues the combine with async_with_compute_stream=True (except under
+    DBO) so the modular kernel can overlap the shared-expert FFN with the
+    combine a2a; the returned receiver joins via a device-side event wait.
     """
 
     @staticmethod
@@ -465,15 +468,31 @@ class DeepEPV2PrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeModular):
                 f"got {fused_expert_output.dtype}"
             )
 
+        # DBO drives its own hook/receiver schedule; keep the combine
+        # synchronous there (the receiver then only performs the copy).
+        combine_async = do_async and not dbo_enabled()
         combined_x, _, event = self.buffer.combine(
             x=fused_expert_output,
             handle=handle,
             topk_weights=None,
-            async_with_compute_stream=False,
+            async_with_compute_stream=combine_async,
+            allocate_on_comm_stream=combine_async,
         )
 
-        output.copy_(combined_x, non_blocking=True)
-        return None
+        if do_async:
+            # The combine ran on DeepEP's comm stream; the modular kernel
+            # issues the shared-expert FFN before calling the receiver, which
+            # joins via a device-side cudaStreamWaitEvent (no host sync, so
+            # this is safe inside a captured region).
+            def _receiver():
+                if event.event is not None:
+                    event.current_stream_wait()
+                output.copy_(combined_x, non_blocking=True)
+
+            return _receiver
+        else:
+            output.copy_(combined_x, non_blocking=True)
+            return None
 
     def finalize_async(
         self,
@@ -484,16 +503,17 @@ class DeepEPV2PrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeModular):
         apply_router_weight_on_input: bool,
         weight_and_reduce_impl: mk.TopKWeightAndReduce,
     ) -> Callable:
-        self._finalize(
+        receiver = self._finalize(
             output,
             fused_expert_output,
             topk_weights,
             topk_ids,
             apply_router_weight_on_input,
             weight_and_reduce_impl,
-            False,
+            True,
         )
-        return lambda: None
+        assert receiver is not None
+        return receiver
 
     def finalize(
         self,