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 2, 06:00 UTC
[Perf] use triton moe backend on hopper by default (#44220)
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Jiangyun Zhu committed
Jun 2, 2026 at 15:52 UTC
dcdfe66bfacf60633b7e18c9c7661a386b88fbad
1 file changed
+6
vllm/model_executor/layers/fused_moe/oracle/unquantized.py
+6
@@ -68,6 +68,12 @@ def _get_priority_backends(moe_config: FusedMoEConfig) -> list[UnquantizedMoeBac
68
UnquantizedMoeBackend.BATCHED_TRITON,
69
]
70
71
+ # On Hopper (SM90), the FlashInfer unquantized MoE kernels are slower
72
+ # than Triton, so prefer Triton by default.
73
+ if current_platform.is_device_capability_family(90):
74
+ _move_to_back(_AVAILABLE_BACKENDS, UnquantizedMoeBackend.FLASHINFER_TRTLLM)
75
+ _move_to_back(_AVAILABLE_BACKENDS, UnquantizedMoeBackend.FLASHINFER_CUTLASS)
76
+
77
# HACK: Qwen3.5 has crash with FLASHINFER_CUTLASS BF16 if DEP.
78
# Updating the oracle querying logic is out of the scope of this
79
# PR. Need to fix the kernel or update structure in follow up.