EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[bugfix] bugfix for PD disaggregate (#4319)

Browse Source 
This PR is used to fix mooncake_connector in pcp/dcp case. When
executing function update_done_task_count, it is necessary to ensure
that both pcp/dcp and TP ranks have finished transferring KV cache.

- vLLM version: v0.11.0
- vLLM main:
2918c1b49c

---------

Signed-off-by: wangxiaochao <w00642655@china.huawei.com>
Co-authored-by: wangxiaochao <w00642655@china.huawei.com>
This commit is contained in:
wangxiaochao
2025-11-21 18:08:56 +08:00
committed by GitHub
parent e332e27ec3
commit 3deeea14a0
2 changed files with 28 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
vllm_ascend/distributed/mooncake_connector.py
View File

@@ -150,13 +150,14 @@ class KVCacheTaskTracker:
class KVCacheSendingThread(threading.Thread):
def __init__(self, tp_rank: int, decode_tp_size: int, local_engine_id: str,
side_channel_host: str, side_channel_port: int,
metadata: MooncakeAgentMetadata, ready_event: threading.Event,
kv_caches: dict[str, Any], pcp_rank: int):
def __init__(self, tp_rank: int, prefill_tp_size: int,
local_engine_id: str, side_channel_host: str,
side_channel_port: int, metadata: MooncakeAgentMetadata,
ready_event: threading.Event, kv_caches: dict[str, Any],
pcp_rank: int):
super().__init__(daemon=True, name="KVCacheSendingThread")
self.tp_rank = tp_rank
self.decode_tp_size = decode_tp_size
self.prefill_tp_size = prefill_tp_size
self.local_engine_id = local_engine_id
self.side_channel_host = side_channel_host
self.side_channel_port = side_channel_port
@@ -195,7 +196,7 @@ class KVCacheSendingThread(threading.Thread):
# NOTE(rob): we need each rank to have a unique port. This hack to keeps
# us moving. We will switch when moving to etcd or where we have a
# single ZMQ socket in the scheduler.
handshake_port = self.side_channel_port + self.pcp_rank * self.decode_tp_size \
handshake_port = self.side_channel_port + self.pcp_rank * self.prefill_tp_size \
+ self.tp_rank
path = make_zmq_path("tcp", self.side_channel_host, handshake_port)
logger.info("Starting listening on path: %s", path)
@@ -295,7 +296,7 @@ class KVCacheRecvingThread(threading.Thread):
def add_request(self, request_id: str, local_block_ids: list[int],
remote_block_ids: list[int], remote_engine_id: str,
remote_host: str, remote_handshake_port: int, offset: int,
num_need_pulls: int):
num_need_pulls: int, all_task_done: bool):
"""Add a new request to the queue for processing."""
logger.debug(f"Adding request {request_id} to the queue.")
self.request_queue.put({
@@ -306,7 +307,8 @@ class KVCacheRecvingThread(threading.Thread):
"remote_host": remote_host,
"remote_handshake_port": remote_handshake_port,
"offset": offset,
"num_need_pulls": num_need_pulls
"num_need_pulls": num_need_pulls,
"all_task_done": all_task_done
})
def get_and_clear_finished_requests(self) -> set[str]:
@@ -335,8 +337,7 @@ class KVCacheRecvingThread(threading.Thread):
request_id = req_meta["request_id"]
remote_host = req_meta["remote_host"]
remote_handshake_port = req_meta["remote_handshake_port"]
offset = req_meta["offset"]
num_need_pulls = req_meta["num_need_pulls"]
all_task_done = req_meta["all_task_done"]
try:
logger.debug(
@@ -353,7 +354,7 @@ class KVCacheRecvingThread(threading.Thread):
# remote host.
self._send_done_recv_signal(request_id, remote_host,
remote_handshake_port)
if offset == num_need_pulls - 1:
if all_task_done:
self.task_tracker.update_done_task_count(request_id)
self.request_queue.task_done()
@@ -1091,7 +1092,7 @@ class MooncakeConnectorWorker:
ready_event = threading.Event()
if self.kv_role == 'kv_producer':
self.kv_send_thread = KVCacheSendingThread(
self.tp_rank, self._decode_tp_size, self.engine_id,
self.tp_rank, self._prefill_tp_size, self.engine_id,
self.side_channel_host, self.side_channel_port, metadata,
ready_event, self.kv_caches, self.pcp_rank)
self.kv_send_thread.start()
@@ -1239,7 +1240,10 @@ class MooncakeConnectorWorker:
remote_handshake_port=remote_handshake_port_list[
pcp_dcp_rank][i],
offset=i,
num_need_pulls=self.num_need_pulls)
num_need_pulls=self.num_need_pulls,
all_task_done=(pcp_dcp_rank
== len(remote_handshake_port_list) - 1
and i == self.num_need_pulls - 1))
if self.kv_send_thread is not None:
for req_id, delay_start_time in metadata.requests_to_send.items():