[PD] Fix KV cache corruption on abort by notifying ongoing prefill (#27372)

Signed-off-by: Shangming Cai <csmthu@gmail.com>
This commit is contained in:
Shangming Cai
2026-06-06 00:56:13 +08:00
committed by GitHub
parent 86b9bf5812
commit 57909f731b
2 changed files with 98 additions and 0 deletions
@@ -877,6 +877,7 @@ class CommonKVReceiver(BaseKVReceiver):
self.conclude_state: Optional[KVPoll] = None
self.require_staging: bool = False
self.init_time: Optional[float] = None
self.abort_notified: bool = False
self.kv_mgr.addr_to_rooms_tracker[self.bootstrap_addr].add(self.bootstrap_room)
self.kv_mgr.update_status(self.bootstrap_room, KVPoll.Bootstrapping)
@@ -1062,6 +1063,13 @@ class CommonKVReceiver(BaseKVReceiver):
f"in KVPoll.WaitingForInput",
)
self.kv_mgr.update_status(self.bootstrap_room, KVPoll.Failed)
if (
not self.abort_notified
and hasattr(self, "bootstrap_infos")
and self.bootstrap_infos is not None
):
self._send_abort_notification()
self.abort_notified = True
return KVPoll.Failed
def failure_exception(self):
@@ -1079,6 +1087,36 @@ class CommonKVReceiver(BaseKVReceiver):
)
self.kv_mgr.update_status(self.bootstrap_room, KVPoll.Failed)
self.conclude_state = KVPoll.Failed
if (
not self.abort_notified
and hasattr(self, "bootstrap_infos")
and self.bootstrap_infos is not None
):
self._send_abort_notification()
self.abort_notified = True
def _send_abort_notification(self):
for bootstrap_info in self.bootstrap_infos:
# Best-effort notification to prefill side that this request was aborted.
try:
sock, lock = self._connect_to_bootstrap_server(bootstrap_info)
with lock:
sock.send_multipart(
[
b"ABORT",
str(self.bootstrap_room).encode("ascii"),
self.kv_mgr.local_ip.encode("ascii"),
str(self.kv_mgr.rank_port).encode("ascii"),
]
)
logger.debug(
f"Sent abort notification for room {self.bootstrap_room} "
f"to {bootstrap_info.get('rank_ip', 'unknown')}:{bootstrap_info.get('rank_port', 'unknown')}"
)
except Exception as e:
logger.debug(
f"Failed to send abort notification for room {self.bootstrap_room}: {e}"
)
class CommonKVBootstrapServer(BaseKVBootstrapServer):
@@ -1168,6 +1168,21 @@ class MooncakeKVManager(CommonKVManager):
MooncakeRequestStage.MOONCAKE_WORKER_SEND.level,
)
if (
kv_chunk.room not in self.request_status
or self.check_status(kv_chunk.room) == KVPoll.Failed
):
logger.debug(
f"Skipping chunk for room {kv_chunk.room} because it has already failed or been aborted"
)
if self.enable_trace:
kv_chunk.trace_ctx.trace_slice_end(
MooncakeRequestStage.MOONCAKE_WORKER_SEND.stage_name,
MooncakeRequestStage.MOONCAKE_WORKER_SEND.level,
thread_finish_flag=True,
)
continue
if (
self.enable_staging
and staging_strategy is None
@@ -1386,6 +1401,44 @@ class MooncakeKVManager(CommonKVManager):
handle_staging_rsp(waiting_req_bytes, self.transfer_infos)
continue
# Decode-side abort notification: mark room as failed and ACK
if room == "ABORT":
room_to_be_aborted = int(waiting_req_bytes[1].decode("ascii"))
decode_ip = waiting_req_bytes[2].decode("ascii")
decode_port = int(waiting_req_bytes[3].decode("ascii"))
# No need to abort the room if it has already succeeded
if (
room_to_be_aborted in self.request_status
and self.check_status(room_to_be_aborted) != KVPoll.Success
):
self.update_status(room_to_be_aborted, KVPoll.Failed)
logger.debug(
f"Received abort notification for room {room_to_be_aborted}, "
f"marked as Failed"
)
else:
logger.debug(
f"Received abort notification for room {room_to_be_aborted}, "
f"ignoring (already completed or unknown)"
)
# Send ACK back to decode endpoint
try:
na = NetworkAddress(decode_ip, decode_port)
self._connect(na.to_tcp(), is_ipv6=na.is_ipv6).send_multipart(
[
b"ABORT_ACK",
str(room_to_be_aborted).encode("ascii"),
]
)
logger.debug(
f"Sent ABORT_ACK for room {room_to_be_aborted} to "
f"{decode_ip}:{decode_port}"
)
except Exception as e:
logger.debug(
f"Failed to send ABORT_ACK for room {room_to_be_aborted}: {e}"
)
continue
mooncake_session_id = waiting_req_bytes[3].decode("ascii")
if room == "None":
self.decode_kv_args_table[mooncake_session_id] = (
@@ -1457,6 +1510,13 @@ class MooncakeKVManager(CommonKVManager):
self._handle_staging_req(msg)
continue
# Prefill acknowledges abort notification
if msg[0] == b"ABORT_ACK":
# TODO(shangming): use this info to implement the deferred release mechanism if needed
ack_aborted_room = int(msg[1].decode("ascii"))
logger.debug(f"Received ABORT_ACK for room {ack_aborted_room}")
continue
bootstrap_room, status, prefill_rank = msg
status = int(status.decode("ascii"))
bootstrap_room = int(bootstrap_room.decode("ascii"))