xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-18 01:06:59 +08:00
Code Issues Packages Projects Releases Wiki Activity

[BugFix]Reclaim resources to prevent memory leaks when use LMCacheMPConnector (#30745)

Browse Source 
Signed-off-by: wz1qqx <ziqi.wang@novita.ai>
Co-authored-by: wz1qqx <ziqi.wang@novita.ai>
This commit is contained in:
wz1qqx 2025-12-19 03:09:51 +08:00 committed by GitHub
parent 058926d48c
commit 889f8bb250
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 24 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/distributed/kv_transfer/kv_connector/v1/lmcache_integration/multi_process_adapter.py
View File

@ -147,6 +147,14 @@ class LMCacheMPSchedulerAdapter:
""" """
return self.blocks_in_chunk return self.blocks_in_chunk
def _cleanup_lookup_result(self, request_id: str) -> None:
"""
Clean up lookup future for a finished request to prevent memory leak.
Args:
request_id: The ID of the finished request.
"""
self.lookup_futures.pop(request_id, None)
# Helper functions # Helper functions
def _create_key(self, block_hash: bytes) -> IPCCacheEngineKey: def _create_key(self, block_hash: bytes) -> IPCCacheEngineKey:
"""Convert a block hash to an IPC cache engine key""" """Convert a block hash to an IPC cache engine key"""

16
vllm/distributed/kv_transfer/kv_connector/v1/lmcache_mp_connector.py
View File

@ -701,6 +701,8 @@ class LMCacheMPConnector(KVConnectorBase_V1):
if condition if condition
else LMCacheMPRequestState.READY else LMCacheMPRequestState.READY
) )
# Clean up lookup future in scheduler adapter
self.scheduler_adapter._cleanup_lookup_result(request.request_id)
def build_connector_meta( def build_connector_meta(
self, scheduler_output: SchedulerOutput self, scheduler_output: SchedulerOutput
@ -754,6 +756,8 @@ class LMCacheMPConnector(KVConnectorBase_V1):
Optional KVTransferParams to be included in the request outputs Optional KVTransferParams to be included in the request outputs
returned by the engine. returned by the engine.
""" """
# Clean up request tracker to prevent memory leak
self._cleanup_request_tracker(request.request_id)
return True, None return True, None
def take_events(self) -> Iterable["KVCacheEvent"]: def take_events(self) -> Iterable["KVCacheEvent"]:
@ -915,3 +919,15 @@ class LMCacheMPConnector(KVConnectorBase_V1):
new_tracker = LMCacheMPRequestTracker(request) new_tracker = LMCacheMPRequestTracker(request)
self.request_trackers[request_id] = new_tracker self.request_trackers[request_id] = new_tracker
return self.request_trackers[request_id] return self.request_trackers[request_id]
def _cleanup_request_tracker(self, request_id: str) -> None:
"""
Clean up request tracker and associated lookup future for a request.
This should be called when a request is finished to prevent memory leak.
"""
# Clean up request tracker
if self.request_trackers.pop(request_id, None):
logger.debug(
"[KVConnector] Cleaned up request_tracker for request %s",
request_id,
)