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 4, 10:00 UTC
Re-enable allreduce rms fusion for DP / PP (#41458)
Signed-off-by: Andy Lo <andy@mistral.ai>
Andy Lo committed
May 2, 2026 at 00:01 UTC
5737770c6c346d918fdfb13e9378f9514f616186
1 file changed
-6
vllm/config/vllm.py
-6
@@ -138,12 +138,6 @@ def enable_allreduce_rms_fusion(cfg: "VllmConfig") -> bool:
138
current_platform.is_device_capability_family(100)
139
or current_platform.is_device_capability(90)
140
)
141
- # tp-dp combination broken:
142
- # https://github.com/vllm-project/vllm/issues/34458
143
- and cfg.parallel_config.data_parallel_size == 1
144
- # tp-pp combination broken:
145
- # https://github.com/vllm-project/vllm/issues/35426
146
- and cfg.parallel_config.pipeline_parallel_size == 1
141
)
142
143