diff --git a/tests/test_config.py b/tests/test_config.py index b7c3740f72..0f01cf4541 100644 --- a/tests/test_config.py +++ b/tests/test_config.py @@ -752,6 +752,10 @@ def test_data_parallel_rpc_port_has_fixed_default(): assert ParallelConfig().data_parallel_rpc_port == 29550 +def test_all2all_backend_has_portable_default(): + assert ParallelConfig().all2all_backend == "allgather_reducescatter" + + @pytest.mark.parametrize("port", [1, 29550, 65535]) def test_data_parallel_rpc_port_accepts_valid_ports(port: int): assert ParallelConfig(data_parallel_rpc_port=port).data_parallel_rpc_port == port diff --git a/vllm/config/parallel.py b/vllm/config/parallel.py index ae1a4c4e32..e7710c59ea 100644 --- a/vllm/config/parallel.py +++ b/vllm/config/parallel.py @@ -192,7 +192,7 @@ class ParallelConfig: with 4 experts and 2 ranks, rank 0 will have experts [0, 2] and rank 1 will have experts [1, 3]. This strategy can help improve load balancing for grouped expert models with no redundant experts.""" - all2all_backend: All2AllBackend = "flashinfer_nvlink_one_sided" + all2all_backend: All2AllBackend = "allgather_reducescatter" """All2All backend for MoE expert parallel communication. Available options: - "allgather_reducescatter": All2all based on allgather and reducescatter