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

[Fix] Skip cross-node probe in MultimemAllGatherer on single-node runs (fixes mooncake EP segfault) (#30139)

Browse Source
This commit is contained in:
Cheng Wan
2026-07-04 15:36:23 -07:00
committed by GitHub
parent 00f088f6be
commit 754524d8de
1 changed files with 11 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/distributed/device_communicators/triton_symm_mem_ag.py
+11 -2
View File
@@ -466,10 +466,19 @@ class MultimemAllGatherer:
# Lazy import avoids a module-load dependency on the distributed facade. # Lazy import avoids a module-load dependency on the distributed facade.
from sglang.srt.distributed import get_tp_group from sglang.srt.distributed import get_tp_group
from sglang.srt.distributed.parallel_state import in_the_same_node_as from sglang.srt.distributed.parallel_state import in_the_same_node_as
from sglang.srt.runtime_context import get_server_args
tp_group = get_tp_group() tp_group = get_tp_group()
if tp_group.world_size > 1 and not all( # Only probe node topology when the deployment can actually span
in_the_same_node_as(tp_group.cpu_group, source_rank=0) # nodes. Check world_size first so a TP=1 gatherer short-circuits
# before reading server args (which may be unpublished on offline
# paths). On a single node every TP rank is co-located, so skip the
# in_the_same_node_as() all-reduce, which can segfault under some
# EP/mooncake setups, and keep multimem enabled.
if (
tp_group.world_size > 1
and get_server_args().nnodes > 1
and not all(in_the_same_node_as(tp_group.cpu_group, source_rank=0))
): ):
logger.warning( logger.warning(
"multimem all-gather disabled because the TP group spans " "multimem all-gather disabled because the TP group spans "