Reenable MNNVL backend for FlashInfer allreduce fusion (#23402)

This commit is contained in:
Shu Wang
2026-06-15 20:19:15 -07:00
committed by GitHub
parent b23477af44
commit 32685874f3
7 changed files with 481 additions and 69 deletions
@@ -2463,7 +2463,7 @@ class ModelRunner(ModelRunnerKVCacheMixin):
(broadcasts, barriers) inside the graph capture context, which can
deadlock with custom_all_reduce.register_graph_buffers.
"""
if not self.server_args.enable_flashinfer_allreduce_fusion:
if self.server_args.flashinfer_allreduce_fusion_backend is None:
return
from sglang.srt.layers.communicator import FUSE_ALLREDUCE_MAX_BATCH_SIZE