minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Fix TP deadlock in unified radix cache writing_check / loading_check (#27489)

Browse Source
This commit is contained in:
Ke Bao
2026-06-08 16:08:47 +08:00
committed by GitHub
parent 13dda3b8de
commit d03182cd2d
1 changed files with 5 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/mem_cache/unified_radix_cache.py
+5 -4
View File
@@ -2258,9 +2258,8 @@ class UnifiedRadixCache(KVCacheEventMixin, BasePrefixCache):
assert len(self.ongoing_write_through) == 0 assert len(self.ongoing_write_through) == 0
return return
if len(self.ongoing_write_through) == 0: # Every rank must enter the all_reduce below; ongoing_write_through can
return # diverge across ranks (e.g. write_backup returning 0 on a subset).
finish_count = 0 finish_count = 0
if self.pp_rank == 0: if self.pp_rank == 0:
for _, finish_event, ack_list in cc.ack_write_queue: for _, finish_event, ack_list in cc.ack_write_queue:
@@ -2283,8 +2282,10 @@ class UnifiedRadixCache(KVCacheEventMixin, BasePrefixCache):
def loading_check(self) -> None: def loading_check(self) -> None:
"""Poll load-back completions.""" """Poll load-back completions."""
cc = self.cache_controller cc = self.cache_controller
if cc is None or not self.ongoing_load_back: if cc is None:
return return
# Every rank must enter the all_reduce below; ongoing_load_back can
# diverge across ranks.
finish_count = 0 finish_count = 0
if self.pp_rank == 0: if self.pp_rank == 0:
for _, finish_event, ack_list in cc.ack_load_queue: for _, finish_event, ack_list in cc.ack_load_queue: