diff --git a/python/sglang/srt/layers/moe/token_dispatcher/deepep.py b/python/sglang/srt/layers/moe/token_dispatcher/deepep.py
index f990d7d00172..7fbe8e3088ae 100644
--- a/python/sglang/srt/layers/moe/token_dispatcher/deepep.py
+++ b/python/sglang/srt/layers/moe/token_dispatcher/deepep.py
@@ -51,6 +51,23 @@
 except ImportError:
     use_deepep = False
 
+# DeepEP V2 introduces `ElasticBuffer` alongside the legacy `Buffer`
+# (deepseek-ai/DeepEP#605, merged 2026-04-29). On V2 both classes are
+# exported from `deep_ep.__init__`, so the existing `from deep_ep import
+# Buffer` surface above continues to work unchanged — `ElasticBuffer` is
+# an additional, MoE-shape ctor with auto-QP sizing that callers may
+# opt into. The probe below is orthogonal to `use_deepep` and does not
+# affect the default code path. V2 usage is further gated on
+# `SGLANG_DEEPEP_USE_V2=1`. Mirrors the `HAVE_DEEP_EP_V2` probe shape
+# already used in NVIDIA/Megatron-LM's `fused_a2a.py`.
+try:
+    from deep_ep import ElasticBuffer
+
+    have_deepep_v2 = True
+except ImportError:
+    ElasticBuffer = None
+    have_deepep_v2 = False
+
 from enum import Enum, IntEnum, auto
 
 import torch
@@ -166,6 +183,22 @@ def get_deepep_buffer(
         cls._num_max_dispatch_tokens_per_rank = num_max_dispatch_tokens_per_rank
         cls._num_experts = num_experts
 
+        # Opt-in V2 path: construct `deep_ep.ElasticBuffer` instead of the
+        # legacy `deep_ep.Buffer`. V2 uses a MoE-shape ctor and infers the
+        # dispatch layout internally (no `get_dispatch_config` /
+        # `get_combine_config` / `get_nvl_buffer_size_hint` /
+        # `get_rdma_buffer_size_hint` calls). Gated behind an env var so
+        # the default code path is byte-identical to V1.
+        if have_deepep_v2 and get_bool_env_var("SGLANG_DEEPEP_USE_V2", default="false"):
+            cls._buffer = cls._build_v2_buffer(
+                group,
+                hidden_size,
+                deepep_mode,
+                num_max_dispatch_tokens_per_rank,
+                num_experts,
+            )
+            return cls._buffer
+
         num_nvl_bytes, num_rdma_bytes = 0, 0
         if deepep_mode.enable_normal():
             hidden_bytes = hidden_size * param_bytes
@@ -238,8 +271,72 @@ def get_deepep_buffer(
         )
         return cls._buffer
 
+    @classmethod
+    def _build_v2_buffer(
+        cls,
+        group: dist.ProcessGroup,
+        hidden_size: int,
+        deepep_mode: DeepEPMode,
+        num_max_dispatch_tokens_per_rank: int,
+        num_experts: int,
+    ):
+        """Construct a DeepEP V2 `ElasticBuffer` for opt-in V2 usage.
+
+        DeepEP V2 (deepseek-ai/DeepEP#605) collapses the V1 NVL/RDMA
+        byte-pool ctor into a single MoE-shape ctor and derives the
+        internal buffer size from `num_max_tokens_per_rank`, `hidden`,
+        and `num_topk`. `num_allocated_qps=0` asks V2 to auto-size and
+        auto-cap the Queue-Pair budget — on AWS EFA this clamps to the
+        safe 128-slot GIN ring ceiling (see the EFA fast-path in
+        `deep_ep/buffers/elastic.py`).
+
+        V2 does not expose `get_dispatch_config` / `get_combine_config`
+        / `num_qps_per_rank`, so the matching V1 code above is skipped.
+        """
+        # Keep the num_tokens / num_topk source consistent with the V1
+        # low-latency path. For the normal path where
+        # `num_max_dispatch_tokens_per_rank == -1`, fall back to the
+        # SGLang env default that already bounds the V1 path.
+        if num_max_dispatch_tokens_per_rank <= 0:
+            num_max_dispatch_tokens_per_rank = (
+                envs.SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK.get()
+            )
+        # `num_topk` is not known at buffer-construction time in SGLang
+        # (the router choice is per-forward). DeepEP V2 accepts
+        # `num_topk=0` and falls back to a group-size-based conservative
+        # hint, so we pass 0 here and let V2 compute the ceiling.
+        num_topk = 0
+        # Low-latency mode still requires `num_experts % group.size() ==
+        # 0`. The normal mode works with `num_experts == -1`.
+        if deepep_mode.enable_low_latency():
+            assert num_experts != -1 and num_experts % group.size() == 0
+
+        logger.info(
+            "SGLANG_DEEPEP_USE_V2=1: constructing deep_ep.ElasticBuffer "
+            "(num_max_tokens_per_rank=%d, hidden=%d, num_topk=%d).",
+            num_max_dispatch_tokens_per_rank,
+            hidden_size,
+            num_topk,
+        )
+        return ElasticBuffer(
+            group=group,
+            num_max_tokens_per_rank=num_max_dispatch_tokens_per_rank,
+            hidden=hidden_size,
+            num_topk=num_topk,
+            use_fp8_dispatch=False,
+            allow_hybrid_mode=True,
+            # num_allocated_qps=0 -> V2 auto-sizes with an EFA safety cap
+            # (see deep_ep/buffers/elastic.py _is_efa_fabric()).
+            num_allocated_qps=0,
+        )
+
     @classmethod
     def clean_buffer(cls):
+        # DeepEP V2's `ElasticBuffer` does not expose `low_latency_mode`
+        # or `clean_low_latency_buffer` — low-latency cleanup is handled
+        # internally via `EPHandle` lifetime. Fall through for V2.
+        if not hasattr(cls._buffer, "clean_low_latency_buffer"):
+            return
         if not cls._buffer.low_latency_mode:
             return
         cls._buffer.clean_low_latency_buffer(
diff --git a/test/srt/test_deepep_v2_probe.py b/test/srt/test_deepep_v2_probe.py
new file mode 100644
index 000000000000..7eeae8dc59dd
--- /dev/null
+++ b/test/srt/test_deepep_v2_probe.py
@@ -0,0 +1,142 @@
+"""CPU-only unit test for the DeepEP V2 probe plumbing in
+`sglang.srt.layers.moe.token_dispatcher.deepep`.
+
+Exercises the three reachable probe states:
+
+1. V1 `Buffer` available, V2 `ElasticBuffer` missing
+2. V2 `ElasticBuffer` available, V1 `Buffer` also available (normal V2 install)
+3. Neither available
+
+The test does not execute any DeepEP kernels — it imports the module
+under a monkey-patched `deep_ep` stub so it runs on any CPU host without
+CUDA / NCCL / EFA. It is purely a guard against the probe pattern
+silently regressing.
+
+Run:
+    pytest -xvs test/srt/test_deepep_v2_probe.py
+"""
+
+from __future__ import annotations
+
+import importlib
+import sys
+import types
+import unittest
+from contextlib import contextmanager
+
+
+def _make_stub_deep_ep(*, with_buffer: bool, with_elastic_buffer: bool):
+    """Build a fake `deep_ep` module with the attributes we opt to expose."""
+
+    mod = types.ModuleType("deep_ep")
+    if with_buffer:
+
+        class _StubBuffer:  # noqa: D401 - placeholder class
+            num_sms = 20
+
+            @staticmethod
+            def get_dispatch_config(size):  # pragma: no cover - unused in test
+                return None
+
+            @staticmethod
+            def get_combine_config(size):  # pragma: no cover - unused in test
+                return None
+
+        mod.Buffer = _StubBuffer
+        mod.Config = object
+    if with_elastic_buffer:
+
+        class _StubElasticBuffer:  # noqa: D401 - placeholder class
+            pass
+
+        mod.ElasticBuffer = _StubElasticBuffer
+    return mod
+
+
+@contextmanager
+def _patched_deep_ep(*, with_buffer: bool, with_elastic_buffer: bool):
+    """Install a stub `deep_ep` in sys.modules and drop the SGLang
+    deepep module from the cache so it re-runs its imports against the
+    stub. Restores the original state on exit.
+    """
+
+    saved = {
+        k: sys.modules[k]
+        for k in list(sys.modules)
+        if k == "deep_ep" or k.startswith("deep_ep.")
+    }
+    saved_deepep = sys.modules.pop(
+        "sglang.srt.layers.moe.token_dispatcher.deepep", None
+    )
+    # Wipe the stub path
+    sys.modules.pop("deep_ep", None)
+    if with_buffer or with_elastic_buffer:
+        sys.modules["deep_ep"] = _make_stub_deep_ep(
+            with_buffer=with_buffer, with_elastic_buffer=with_elastic_buffer
+        )
+    try:
+        yield
+    finally:
+        # Best-effort restore; we don't try to rewind every transitive
+        # import that the probe triggered.
+        sys.modules.pop("deep_ep", None)
+        for k, v in saved.items():
+            sys.modules[k] = v
+        if saved_deepep is not None:
+            sys.modules["sglang.srt.layers.moe.token_dispatcher.deepep"] = saved_deepep
+
+
+class TestDeepEPV2Probe(unittest.TestCase):
+    """Guard the V2 probe plumbing against silent regression."""
+
+    def _import_probe_flags(self):
+        sys.modules.pop("sglang.srt.layers.moe.token_dispatcher.deepep", None)
+        mod = importlib.import_module("sglang.srt.layers.moe.token_dispatcher.deepep")
+        return (
+            getattr(mod, "use_deepep"),
+            getattr(mod, "have_deepep_v2"),
+        )
+
+    def test_v1_only_installed(self):
+        """Legacy install (V1 `Buffer` only). `use_deepep=True`,
+        `have_deepep_v2=False`. No regression on pre-V2 users."""
+        with _patched_deep_ep(with_buffer=True, with_elastic_buffer=False):
+            try:
+                use_deepep, have_v2 = self._import_probe_flags()
+            except ImportError:
+                # Test env may not have full sglang deps; that's fine,
+                # the probe itself is what we need to prove compiles.
+                self.skipTest("sglang module stack unavailable on test host")
+                return
+        self.assertTrue(use_deepep)
+        self.assertFalse(have_v2)
+
+    def test_v2_installed(self):
+        """V2 install (both `Buffer` legacy + `ElasticBuffer` exported).
+        `use_deepep=True`, `have_deepep_v2=True`. The V2 path is now
+        reachable behind `SGLANG_DEEPEP_USE_V2=1`."""
+        with _patched_deep_ep(with_buffer=True, with_elastic_buffer=True):
+            try:
+                use_deepep, have_v2 = self._import_probe_flags()
+            except ImportError:
+                self.skipTest("sglang module stack unavailable on test host")
+                return
+        self.assertTrue(use_deepep)
+        self.assertTrue(have_v2)
+
+    def test_neither_installed(self):
+        """deep_ep missing entirely. `use_deepep=False`,
+        `have_deepep_v2=False`. Module must still import cleanly so the
+        SGLang package can load on non-MoE targets."""
+        with _patched_deep_ep(with_buffer=False, with_elastic_buffer=False):
+            try:
+                use_deepep, have_v2 = self._import_probe_flags()
+            except ImportError:
+                self.skipTest("sglang module stack unavailable on test host")
+                return
+        self.assertFalse(use_deepep)
+        self.assertFalse(have_v2)
+
+
+if __name__ == "__main__":
+    unittest.main()