diff --git a/python/sglang/srt/disaggregation/base/conn.py b/python/sglang/srt/disaggregation/base/conn.py
index 0cc63380b991..6fbc3b990ac1 100644
--- a/python/sglang/srt/disaggregation/base/conn.py
+++ b/python/sglang/srt/disaggregation/base/conn.py
@@ -132,6 +132,10 @@ def register_to_bootstrap(self):
     # role switch (release transfer resources; the scheduler owns the KV pool).
     supports_role_switch: bool = False
 
+    # Opt-in per backend: set True once the prefill acks a drained abort
+    # (ABORT_ACK carrying the sender rank). Without it a hold only ends on timeout.
+    supports_deferred_decode_kv_release: bool = False
+
     def teardown(self) -> None:
         raise NotImplementedError(
             f"{type(self).__name__} does not support PD role switch teardown"
diff --git a/python/sglang/srt/disaggregation/common/conn.py b/python/sglang/srt/disaggregation/common/conn.py
index 5cf9d060890e..6789c0af8a3a 100644
--- a/python/sglang/srt/disaggregation/common/conn.py
+++ b/python/sglang/srt/disaggregation/common/conn.py
@@ -182,6 +182,7 @@ def __init__(
         self.server_args = server_args
         self.enable_deferred_decode_kv_release = (
             envs.SGLANG_DISAGGREGATION_DEFERRED_DECODE_KV_RELEASE.get()
+            and self.supports_deferred_decode_kv_release
         )
         self._dcp_pack_buffers = None
         self._dcp_pack_max_tokens: Optional[int] = None
diff --git a/python/sglang/srt/disaggregation/mooncake/conn.py b/python/sglang/srt/disaggregation/mooncake/conn.py
index cdbe95e8089d..b54e876aebbd 100644
--- a/python/sglang/srt/disaggregation/mooncake/conn.py
+++ b/python/sglang/srt/disaggregation/mooncake/conn.py
@@ -217,6 +217,8 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
     AUX_DATA_HEADER = b"AUX_DATA"
     # Implements teardown() below, so runtime PD role switching is supported.
     supports_role_switch = True
+    # Bootstrap thread defers the ABORT ack until the transfer worker drains.
+    supports_deferred_decode_kv_release = True
 
     def __init__(
         self,
diff --git a/python/sglang/srt/disaggregation/nixl/conn.py b/python/sglang/srt/disaggregation/nixl/conn.py
index 8d8b7f963652..cc70acf1d3b0 100644
--- a/python/sglang/srt/disaggregation/nixl/conn.py
+++ b/python/sglang/srt/disaggregation/nixl/conn.py
@@ -412,6 +412,8 @@ class NixlKVManager(StagingManagerMixin, CommonKVManager):
     # message is tagged too. It is new to NIXL, hence free to carry the reason.
     kv_status_msg_tag = b"KV_STATUS"
     kv_status_msg_carries_reason = True
+    # ABORT handler defers the ack until the transfer worker drains.
+    supports_deferred_decode_kv_release = True
 
     def __init__(
         self,
diff --git a/python/sglang/srt/environ.py b/python/sglang/srt/environ.py
index 4e502e48b9cd..ac02cc073703 100644
--- a/python/sglang/srt/environ.py
+++ b/python/sglang/srt/environ.py
@@ -698,8 +698,9 @@ class Envs:
     SGLANG_DISAGGREGATION_BOOTSTRAP_ENTRY_CLEANUP_INTERVAL = EnvInt(120)
     # Deferred decode-side KV release: on abort, hold an in-flight request's KV
     # pages/slot until the prefill acks the transfer drained, or the timeout
-    # below fires. Off by default (no behavior/perf impact when disabled).
-    SGLANG_DISAGGREGATION_DEFERRED_DECODE_KV_RELEASE = EnvBool(False)
+    # below fires. Only applies to backends that ack the drain
+    # (supports_deferred_decode_kv_release).
+    SGLANG_DISAGGREGATION_DEFERRED_DECODE_KV_RELEASE = EnvBool(True)
     SGLANG_DISAGGREGATION_DEFERRED_DECODE_KV_RELEASE_TIMEOUT = EnvFloat(30.0)
 
     # ===================================================================
diff --git a/test/registered/unit/disaggregation/test_deferred_decode_kv_release.py b/test/registered/unit/disaggregation/test_deferred_decode_kv_release.py
index 0a8936f219d6..4447f19afe43 100644
--- a/test/registered/unit/disaggregation/test_deferred_decode_kv_release.py
+++ b/test/registered/unit/disaggregation/test_deferred_decode_kv_release.py
@@ -13,8 +13,12 @@
 from unittest.mock import patch
 
 from sglang.srt.disaggregation import decode as decode_mod
+from sglang.srt.disaggregation.base.conn import BaseKVManager
 from sglang.srt.disaggregation.common.conn import CommonKVManager
 from sglang.srt.disaggregation.decode import DecodeTransferQueue
+from sglang.srt.disaggregation.mooncake.conn import MooncakeKVManager
+from sglang.srt.disaggregation.nixl.conn import NixlKVManager
+from sglang.srt.environ import envs
 from sglang.test.ci.ci_register import register_cpu_ci
 from sglang.test.test_utils import CustomTestCase
 
@@ -243,5 +247,24 @@ def test_defer_release_records_deadline_and_idx(self):
         self.assertIsInstance(deadline, float)
 
 
+class TestBackendOptIn(CustomTestCase):
+    """Without a prefill ack, every hold waits out the full release timeout."""
+
+    def test_enabled_by_default(self):
+        self.assertTrue(envs.SGLANG_DISAGGREGATION_DEFERRED_DECODE_KV_RELEASE.get())
+
+    def test_backends_that_ack_opt_in(self):
+        # Ascend inherits Mooncake's threads, so it opts in too.
+        for cls in (MooncakeKVManager, NixlKVManager):
+            with self.subTest(backend=cls.__name__):
+                self.assertTrue(cls.supports_deferred_decode_kv_release)
+
+    def test_backends_without_a_drain_ack_stay_opted_out(self):
+        # Inheriting CommonKVManager is not enough: mori marks the room Failed
+        # without acking.
+        self.assertFalse(BaseKVManager.supports_deferred_decode_kv_release)
+        self.assertFalse(CommonKVManager.supports_deferred_decode_kv_release)
+
+
 if __name__ == "__main__":
     unittest.main()