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 6, 05:45 UTC
[Core] Enable allreduce fusion by default for SM 10.3 (B300/GB300) (#37755)
Signed-off-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>
Mohammad Miadh Angkad committed
Mar 22, 2026 at 03:40 UTC
88f1b374f53e1d079dda26966dfd4422cf84a978
1 file changed
+1
-1
vllm/config/vllm.py
+1
-1
@@ -120,7 +120,7 @@ def enable_allreduce_rms_fusion(cfg: "VllmConfig") -> bool:
120
and current_platform.is_cuda()
121
and has_flashinfer()
122
and (
123
- current_platform.is_device_capability(100)
123
+ current_platform.is_device_capability_family(100)
124
or current_platform.is_device_capability(90)
125
)
126
# tp-dp combination broken: