[Bugfix] Restore portable all2all backend default (#53952)
Signed-off-by: Kevin Luu <51931015+khluu@users.noreply.github.com> Co-authored-by: Kevin Luu <51931015+khluu@users.noreply.github.com>
vllm-agent committed
Aug 26, 2026 at 16:19 UTC
d1e5e66ee30ba4bc020ac8e14b05e7a8c41b9302
2 files changed
+5
-1
tests/test_config.py
+4
@@ -752,6 +752,10 @@ def test_data_parallel_rpc_port_has_fixed_default():
752
assert ParallelConfig().data_parallel_rpc_port == 29550
753
754
755
+def test_all2all_backend_has_portable_default():
756
+ assert ParallelConfig().all2all_backend == "allgather_reducescatter"
757
+
758
+
759
@pytest.mark.parametrize("port", [1, 29550, 65535])
760
def test_data_parallel_rpc_port_accepts_valid_ports(port: int):
761
assert ParallelConfig(data_parallel_rpc_port=port).data_parallel_rpc_port == port
vllm/config/parallel.py
+1
-1
@@ -192,7 +192,7 @@ class ParallelConfig:
192
with 4 experts and 2 ranks, rank 0 will have experts [0, 2] and rank 1
193
will have experts [1, 3]. This strategy can help improve load balancing
194
for grouped expert models with no redundant experts."""
195
- all2all_backend: All2AllBackend = "flashinfer_nvlink_one_sided"
195
+ all2all_backend: All2AllBackend = "allgather_reducescatter"
196
"""All2All backend for MoE expert parallel communication. Available options:
197
198
- "allgather_reducescatter": All2all based on allgather and reducescatter