[PD] Gate deferred decode KV release on backend capability (#37454)

Co-authored-by: yhzhuang <yhzhuang@fb.com>

python/sglang/srt/disaggregation/base/conn.py

@@ -101,6 +101,8 @@ class KVPoll:
 class BaseKVManager(ABC):
     """Base class for managing transfer states"""
 
+    enable_deferred_decode_kv_release: bool = False
+
     @abstractmethod
     def __init__(
         self,

python/sglang/srt/disaggregation/decode.py

@@ -2326,6 +2326,7 @@ class DecodeTransferQueue(DecodeHiCacheTransferMixin):
                     self.scheduler.hisparse_coordinator.request_finished(decode_req.req)
                 if (
                     self.enable_deferred_kv_release
+                    and decode_req.kv_receiver.kv_mgr.enable_deferred_decode_kv_release
                     and decode_req.kv_receiver.abort_notified
                 ):
                     # Decode-initiated abort: a prefill write may still target

python/sglang/srt/disaggregation/fake/conn.py

@@ -99,6 +99,8 @@ class FakeKVReceiver(BaseKVReceiver):
         bootstrap_addr: str,
         bootstrap_room: Optional[int] = None,
     ):
+        self.kv_mgr = mgr
+        self.abort_notified: bool = False
         self.bootstrap_done = False
         self.has_sent_metadata = False
         self.require_staging: bool = False