diff --git a/vllm/config/vllm.py b/vllm/config/vllm.py index b98c74d459..ba20d75fa1 100644 --- a/vllm/config/vllm.py +++ b/vllm/config/vllm.py @@ -129,13 +129,7 @@ def enable_act_fusion(cfg: "VllmConfig") -> bool: def enable_allreduce_rms_fusion(cfg: "VllmConfig") -> bool: - """Enable if TP > 1, PP == 1, Hopper/Blackwell, and flashinfer installed. - - Gated off for PP > 1: the fused op's GPU-side peer-signal spin-wait - assumes byte-identical kernel launches across TP peers, but concurrent - independent warmup of multiple TP subgroups lets ranks pick divergent - FlashInfer launch configs and deadlock. - """ + """Enable if TP > 1 and Hopper/Blackwell and flashinfer installed.""" from vllm.platforms import current_platform from vllm.utils.flashinfer import has_flashinfer @@ -148,7 +142,6 @@ def enable_allreduce_rms_fusion(cfg: "VllmConfig") -> bool: return ( cfg.parallel_config.tensor_parallel_size > 1 - and cfg.parallel_config.pipeline_parallel_size == 1 and current_platform.is_cuda() and has_flashinfer() and ( diff --git a/vllm/distributed/device_communicators/flashinfer_all_reduce.py b/vllm/distributed/device_communicators/flashinfer_all_reduce.py index 2594c0cf16..38f7bd5ff8 100644 --- a/vllm/distributed/device_communicators/flashinfer_all_reduce.py +++ b/vllm/distributed/device_communicators/flashinfer_all_reduce.py @@ -61,6 +61,7 @@ def _create_workspace( hidden_dim=hidden_dim, dtype=dtype, comm_backend=comm_backend, + group=group, ) except Exception as e: if "multicast" in str(e).lower():