A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 7, 15:15 UTC
[Bugfix] disable allreduce_rms_fusion by default when pp size > 1 (#35424)
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Jiangyun Zhu committed
Feb 27, 2026 at 12:18 UTC
487e5c51f72739981d7af3d0d42386ae4fe7824f
1 file changed
+3
vllm/config/vllm.py
+3
@@ -126,6 +126,9 @@ def enable_allreduce_rms_fusion(cfg: "VllmConfig") -> bool:
126
# tp-dp combination broken:
127
# https://github.com/vllm-project/vllm/issues/34458
128
and cfg.parallel_config.data_parallel_size == 1
129
+ # tp-pp combination broken:
130
+ # https://github.com/vllm-project/vllm/issues/35426
131
+ and cfg.parallel_config.pipeline_parallel_size == 1
132
)
133
134