EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Forbid DeepEP racing condition when too many tokens (#9567)

Browse Source
This commit is contained in:
fzyzcjy
2025-09-05 19:47:05 +08:00
committed by GitHub
parent 13705dae06
commit adf73175d6
1 changed files with 3 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
python/sglang/srt/layers/moe/token_dispatcher/deepep.py
View File

@@ -272,6 +272,9 @@ class _DeepEPDispatcherImplBase:
self.num_max_dispatch_tokens_per_rank = get_int_env_var( self.num_max_dispatch_tokens_per_rank = get_int_env_var(
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK", 128 "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK", 128
) )
# DeepEP internode_ll dispatch uses FINISHED_SUM_TAG=1024
# and the logic requires num-tokens-sent-from-one-rank-to-another-rank less than it
assert self.num_max_dispatch_tokens_per_rank <= 1024
self.handle = None self.handle = None