[Bugfix] Fix external LB DP rank handling when replicas share nodes (#53743)

Signed-off-by: Tony Lin <tony.lin@intel.com>
This commit is contained in:
Tony Lin
2026-09-21 17:19:14 +02:00
committed by GitHub
parent e34685dfc0
commit dc479a2d83
5 changed files with 153 additions and 7 deletions
+75
View File
@@ -2,9 +2,11 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from argparse import ArgumentError
from unittest.mock import patch
import pytest
from vllm.config import ModelConfig
from vllm.engine.arg_utils import EngineArgs
from vllm.usage.usage_lib import UsageContext
from vllm.utils.argparse_utils import FlexibleArgumentParser
@@ -121,3 +123,76 @@ def test_data_parallel_start_rank_zero_infers_hybrid_lb():
assert vllm_config.parallel_config.data_parallel_hybrid_lb is True
assert vllm_config.parallel_config.data_parallel_rank == 0
def test_external_lb_preserves_explicit_rank_when_dp_exceeds_nodes():
engine_args = EngineArgs(
model="facebook/opt-125m",
data_parallel_size=4,
data_parallel_rank=3,
data_parallel_external_lb=True,
nnodes=2,
node_rank=1,
)
with patch.object(ModelConfig, "is_moe", new=property(lambda self: True)):
vllm_config = engine_args.create_engine_config(UsageContext.OPENAI_API_SERVER)
assert vllm_config.parallel_config.data_parallel_rank == 3
@pytest.mark.parametrize(
("data_parallel_size", "nnodes", "tensor_parallel_size"),
[(4, 2, 1), (2, 3, 3)],
)
def test_external_lb_requires_explicit_rank_when_nodes_are_not_evenly_partitioned(
data_parallel_size, nnodes, tensor_parallel_size
):
engine_args = EngineArgs(
model="facebook/opt-125m",
data_parallel_size=data_parallel_size,
data_parallel_external_lb=True,
tensor_parallel_size=tensor_parallel_size,
nnodes=nnodes,
node_rank=1,
)
with (
patch.object(ModelConfig, "is_moe", new=property(lambda self: True)),
pytest.raises(
ValueError,
match="Set a unique `--data-parallel-rank`",
),
):
engine_args.create_engine_config(UsageContext.OPENAI_API_SERVER)
def test_external_lb_infers_rank_when_dp_does_not_exceed_nodes():
engine_args = EngineArgs(
model="facebook/opt-125m",
data_parallel_size=2,
data_parallel_external_lb=True,
nnodes=2,
node_rank=1,
)
with patch.object(ModelConfig, "is_moe", new=property(lambda self: True)):
vllm_config = engine_args.create_engine_config(UsageContext.OPENAI_API_SERVER)
assert vllm_config.parallel_config.data_parallel_rank == 1
def test_external_lb_infers_rank_for_multinode_replicas():
engine_args = EngineArgs(
model="facebook/opt-125m",
data_parallel_size=2,
data_parallel_external_lb=True,
tensor_parallel_size=12,
nnodes=4,
node_rank=1,
)
with patch.object(ModelConfig, "is_moe", new=property(lambda self: True)):
vllm_config = engine_args.create_engine_config(UsageContext.OPENAI_API_SERVER)
assert vllm_config.parallel_config.data_parallel_rank == 0