[Bugfix] Share PCP-aware local GPU indexing for NUMA binding (#61237)

This commit is contained in:
Sherif Waly
2026-10-14 10:54:03 +02:00
committed by GitHub
parent 937d6c19df
commit be4ce2e505
6 changed files with 134 additions and 85 deletions
+86 -3
View File
@@ -2,6 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import os
from functools import partial
from types import SimpleNamespace
import pytest
@@ -49,9 +50,18 @@ def _make_config(**parallel_kwargs):
data_parallel_index=0,
pipeline_parallel_size=1,
tensor_parallel_size=1,
prefill_context_parallel_size=1,
)
parallel_defaults.update(parallel_kwargs)
parallel_config = SimpleNamespace(**parallel_defaults)
parallel_config.world_size = (
parallel_config.tensor_parallel_size
* parallel_config.prefill_context_parallel_size
* parallel_config.pipeline_parallel_size
)
parallel_config.get_local_gpu_index = partial(
ParallelConfig.get_local_gpu_index, parallel_config
)
return SimpleNamespace(parallel_config=parallel_config)
@@ -368,13 +378,86 @@ def test_get_numactl_args_user_cpus_override_pct(monkeypatch):
)
def test_get_numactl_args_uses_dp_offset():
@pytest.mark.parametrize("pcp,pp", [(1, 1), (2, 1), (2, 2)])
def test_get_numactl_args_uses_dp_offset(pcp, pp):
workers_per_dp = 2 * pcp * pp
vllm_config = _make_config(
numa_bind=True,
numa_bind_nodes=[0, 0, 1, 1],
numa_bind_nodes=[0] * workers_per_dp + [1] * workers_per_dp,
data_parallel_rank_local=1,
pipeline_parallel_size=1,
pipeline_parallel_size=pp,
tensor_parallel_size=2,
prefill_context_parallel_size=pcp,
)
assert (
numa_utils._get_numactl_worker_args(vllm_config.parallel_config, local_rank=1)
== "--cpunodebind=1 --membind=1"
)
@pytest.mark.parametrize(
"local_dp_rank,global_dp_rank,launch_dp_rank",
[(1, 3, None), (None, 1, None), (0, 0, 1)],
ids=["local-rank", "global-fallback", "launch-override"],
)
def test_pcp_worker_cpu_binding_uses_dp_gpu_index(
local_dp_rank, global_dp_rank, launch_dp_rank
):
"""CPU overrides use the same PCP-aware GPU index as device selection."""
vllm_config = _make_config(
numa_bind=True,
numa_bind_nodes=[0] * 4 + [1] * 4,
numa_bind_cpus=[f"{4 * i}-{4 * i + 3}" for i in range(8)],
tensor_parallel_size=2,
prefill_context_parallel_size=2,
data_parallel_rank_local=local_dp_rank,
data_parallel_index=global_dp_rank,
)
assert (
numa_utils._get_numactl_worker_args(
vllm_config.parallel_config, local_rank=1, dp_local_rank=launch_dp_rank
)
== "--physcpubind=20-23 --membind=1"
)
@pytest.mark.parametrize("dp_rank,expected_nodes", [(0, "0,1"), (1, "2,3")])
def test_pcp_engine_core_covers_all_shard_workers(dp_rank, expected_nodes):
"""EngineCore includes every PCP worker's NUMA node in its DP shard."""
vllm_config = _make_config(
numa_bind=True,
numa_bind_nodes=[0, 0, 1, 1, 2, 2, 3, 3],
tensor_parallel_size=2,
prefill_context_parallel_size=2,
data_parallel_rank_local=dp_rank,
)
assert (
numa_utils._get_numactl_enginecore_args(
vllm_config.parallel_config, local_rank=0
)
== f"--cpunodebind={expected_nodes} --membind={expected_nodes}"
)
@pytest.mark.parametrize(
"parallel_kwargs",
[
{"distributed_executor_backend": "ray"},
{"distributed_executor_backend": "external_launcher"},
{"data_parallel_backend": "ray"},
{"nnodes_within_dp": 2},
],
ids=["ray-executor", "external-launcher", "ray-dp", "multi-node-dp"],
)
def test_pcp_worker_keeps_local_gpu_index_without_dp_offset(parallel_kwargs):
"""Node-local launcher ranks do not acquire a colocated DP offset."""
vllm_config = _make_config(
numa_bind=True,
numa_bind_nodes=[0, 1],
tensor_parallel_size=2,
prefill_context_parallel_size=2,
data_parallel_rank_local=1,
**parallel_kwargs,
)
assert (
numa_utils._get_numactl_worker_args(vllm_config.parallel_config, local_rank=1)