[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
+35
View File
@@ -1518,6 +1518,41 @@ def test_reconfigure_for_independent_dp_rank_on_multinode_dense_model():
assert parallel_config.world_size == 8
@pytest.mark.parametrize("data_parallel_size", [4, 8])
def test_nnodes_within_dp_when_replicas_outnumber_nodes(data_parallel_size):
"""A replica that fits on one node spans one node, never zero.
External LB pins ``data_parallel_size_local`` to 1, so the ratio rounds
down to 0 as soon as there are more DP replicas than nodes. Anything
dividing by ``nnodes_within_dp`` then raises ZeroDivisionError.
"""
parallel_config = ParallelConfig(
data_parallel_size=data_parallel_size,
data_parallel_size_local=1,
data_parallel_external_lb=True,
distributed_executor_backend="mp",
nnodes=2,
node_rank=1,
)
assert parallel_config.nnodes_within_dp == 1
assert parallel_config.node_rank_within_dp == 0
assert parallel_config.local_world_size == parallel_config.world_size
@pytest.mark.parametrize("nnodes", [4, 9])
def test_nnodes_within_dp_rejects_uneven_internal_lb(nnodes):
parallel_config = ParallelConfig(
data_parallel_size=8,
data_parallel_size_local=1,
distributed_executor_backend="mp",
nnodes=nnodes,
)
with pytest.raises(ValueError, match="Invalid data parallel configuration"):
_ = parallel_config.nnodes_within_dp
def test_draft_model_enables_async_scheduling_by_default():
parallel_config = ParallelConfig(distributed_executor_backend="uni")
model_config = ModelConfig("Qwen/Qwen3-0.6B", max_model_len=2048)