| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import os |
| 5 | from types import SimpleNamespace |
| 6 | |
| 7 | import pytest |
| 8 | |
| 9 | from vllm.config import ParallelConfig |
| 10 | from vllm.utils import numa_utils |
| 11 | |
| 12 | |
| 13 | @pytest.fixture(autouse=True) |
| 14 | def _disable_pct_by_default(monkeypatch): |
| 15 | """Force PCT detection OFF unless a test opts in via ``_patch_pct_gates``. |
| 16 | |
| 17 | The CI / dev machines themselves can be Xeon 6776P with PCT enabled, so a |
| 18 | plain ``cache_clear`` would let the gate auto-detect ``True`` from the |
| 19 | live filesystem and silently re-route "baseline" tests through the PCT |
| 20 | path. Stub ``/proc/cpuinfo`` and ``acpi_cppc/highest_perf`` to a state |
| 21 | that fails the gate; ``_patch_pct_gates`` re-stubs on top when needed. |
| 22 | """ |
| 23 | from io import StringIO |
| 24 | |
| 25 | real_open = open |
| 26 | |
| 27 | def _no_pct_open(path, *args, **kwargs): |
| 28 | if path == numa_utils._PROC_CPUINFO_PATH: |
| 29 | return StringIO("processor\t: 0\nmodel name\t: Generic Test CPU\n") |
| 30 | if path == numa_utils._PCT_HIGHEST_PERF_PATH: |
| 31 | raise OSError("PCT disabled by autouse fixture") |
| 32 | return real_open(path, *args, **kwargs) |
| 33 | |
| 34 | monkeypatch.setattr("builtins.open", _no_pct_open) |
| 35 | numa_utils._pct_sku_config.cache_clear() |
| 36 | yield |
| 37 | numa_utils._pct_sku_config.cache_clear() |
| 38 | |
| 39 | |
| 40 | def _make_config(**parallel_kwargs): |
| 41 | parallel_defaults = dict( |
| 42 | numa_bind=False, |
| 43 | numa_bind_nodes=None, |
| 44 | numa_bind_cpus=None, |
| 45 | distributed_executor_backend="mp", |
| 46 | data_parallel_backend="mp", |
| 47 | nnodes_within_dp=1, |
| 48 | data_parallel_rank_local=0, |
| 49 | data_parallel_index=0, |
| 50 | pipeline_parallel_size=1, |
| 51 | tensor_parallel_size=1, |
| 52 | ) |
| 53 | parallel_defaults.update(parallel_kwargs) |
| 54 | parallel_config = SimpleNamespace(**parallel_defaults) |
| 55 | return SimpleNamespace(parallel_config=parallel_config) |
| 56 | |
| 57 | |
| 58 | def test_get_numactl_args_with_node_binding(): |
| 59 | vllm_config = _make_config(numa_bind=True, numa_bind_nodes=[0, 1]) |
| 60 | assert ( |
| 61 | numa_utils._get_numactl_worker_args(vllm_config.parallel_config, local_rank=1) |
| 62 | == "--cpunodebind=1 --membind=1" |
| 63 | ) |
| 64 | |
| 65 | |
| 66 | def test_get_numactl_args_with_cpu_binding(): |
| 67 | vllm_config = _make_config( |
| 68 | numa_bind=True, |
| 69 | numa_bind_nodes=[0, 1], |
| 70 | numa_bind_cpus=["0-3", "4-7"], |
| 71 | ) |
| 72 | assert ( |
| 73 | numa_utils._get_numactl_worker_args(vllm_config.parallel_config, local_rank=1) |
| 74 | == "--physcpubind=4-7 --membind=1" |
| 75 | ) |
| 76 | |
| 77 | |
| 78 | def _patch_pct_gates( |
| 79 | monkeypatch, |
| 80 | *, |
| 81 | model_match: bool, |
| 82 | highest_perf: int | None, |
| 83 | cpulist: str | None = "0-31,64-95", |
| 84 | cpulist_by_node: dict[int, str | None] | None = None, |
| 85 | sku: str = "6776P", |
| 86 | ): |
| 87 | """Force `_pct_sku_config` and node cpulist read to deterministic state. |
| 88 | |
| 89 | ``cpulist`` is the default returned for any node not present in |
| 90 | ``cpulist_by_node``. ``cpulist_by_node`` lets a test return different |
| 91 | cpulists for different NUMA nodes (e.g. node 0 vs node 1). ``sku`` lets |
| 92 | the test pick which Granite Rapids SKU appears in the fake |
| 93 | ``/proc/cpuinfo`` ``model name`` (only used when ``model_match=True``). |
| 94 | """ |
| 95 | import pathlib |
| 96 | from io import StringIO |
| 97 | |
| 98 | import regex as re |
| 99 | |
| 100 | cpuinfo = ( |
| 101 | f"processor\t: 0\nmodel name\t: Intel(R) Xeon(R) Platinum {sku} CPU @ 2.40GHz\n" |
| 102 | if model_match |
| 103 | else "processor\t: 0\nmodel name\t: Intel(R) Xeon(R) Platinum 8480+\n" |
| 104 | ) |
| 105 | |
| 106 | real_open = open |
| 107 | |
| 108 | def fake_open(path, *args, **kwargs): |
| 109 | if path == numa_utils._PROC_CPUINFO_PATH: |
| 110 | return StringIO(cpuinfo) |
| 111 | if path == numa_utils._PCT_HIGHEST_PERF_PATH: |
| 112 | if highest_perf is None: |
| 113 | raise OSError("missing") |
| 114 | return StringIO(f"{highest_perf}\n") |
| 115 | return real_open(path, *args, **kwargs) |
| 116 | |
| 117 | real_read_text = pathlib.Path.read_text |
| 118 | cpulist_by_node = cpulist_by_node or {} |
| 119 | |
| 120 | def fake_read_text(self, *args, **kwargs): |
| 121 | path_str = str(self) |
| 122 | if path_str.endswith("/cpulist") and "/sys/devices/system/node" in path_str: |
| 123 | match = re.search(r"/node(\d+)/cpulist$", path_str) |
| 124 | if match: |
| 125 | node_id = int(match.group(1)) |
| 126 | if node_id in cpulist_by_node: |
| 127 | val = cpulist_by_node[node_id] |
| 128 | if val is None: |
| 129 | raise OSError(f"missing cpulist for node{node_id}") |
| 130 | return val |
| 131 | if cpulist is None: |
| 132 | raise OSError("missing cpulist") |
| 133 | return cpulist |
| 134 | return real_read_text(self, *args, **kwargs) |
| 135 | |
| 136 | monkeypatch.setattr("builtins.open", fake_open) |
| 137 | monkeypatch.setattr("pathlib.Path.read_text", fake_read_text) |
| 138 | numa_utils._pct_sku_config.cache_clear() |
| 139 | |
| 140 | |
| 141 | def test_pct_binding_filters_cpus(monkeypatch): |
| 142 | _patch_pct_gates(monkeypatch, model_match=True, highest_perf=46) |
| 143 | assert numa_utils._maybe_get_pct_cpu_binding([0]) == [0, 1, 16, 17, 64, 65, 80, 81] |
| 144 | |
| 145 | |
| 146 | @pytest.mark.parametrize( |
| 147 | "sku,expected_cpus", |
| 148 | [ |
| 149 | # 64-core SKUs (stride 16): cpus from "0-31,64-95" with cpu_id % 16 |
| 150 | # in (0, 1) -> 0, 1, 16, 17, 64, 65, 80, 81. |
| 151 | ("6776P", [0, 1, 16, 17, 64, 65, 80, 81]), |
| 152 | ("6774P", [0, 1, 16, 17, 64, 65, 80, 81]), |
| 153 | # 72-core SKU (stride 18): cpus from "0-31,64-95" with cpu_id % 18 |
| 154 | # in (0, 1) -> 0, 1, 18, 19, 72, 73, 90, 91. |
| 155 | ("6962P", [0, 1, 18, 19, 72, 73, 90, 91]), |
| 156 | ], |
| 157 | ) |
| 158 | def test_pct_binding_fires_on_every_capable_sku(monkeypatch, sku, expected_cpus): |
| 159 | """Each SKU in ``_PCT_CAPABLE_SKUS`` engages the gate at its own |
| 160 | expected ``highest_perf`` and uses its own priority-core stride.""" |
| 161 | sku_config = numa_utils._PCT_CAPABLE_SKUS[sku] |
| 162 | _patch_pct_gates( |
| 163 | monkeypatch, |
| 164 | model_match=True, |
| 165 | highest_perf=sku_config.highest_perf, |
| 166 | sku=sku, |
| 167 | ) |
| 168 | assert numa_utils._maybe_get_pct_cpu_binding([0]) == expected_cpus |
| 169 | |
| 170 | |
| 171 | def test_pct_binding_fails_closed_when_sku_perf_mismatch(monkeypatch): |
| 172 | """6962P with 6776P's highest_perf (46 vs expected 44) must fail closed.""" |
| 173 | _patch_pct_gates(monkeypatch, model_match=True, highest_perf=46, sku="6962P") |
| 174 | assert numa_utils._maybe_get_pct_cpu_binding([0]) is None |
| 175 | |
| 176 | |
| 177 | def test_pct_binding_disabled_when_cpu_model_mismatch(monkeypatch): |
| 178 | _patch_pct_gates(monkeypatch, model_match=False, highest_perf=46) |
| 179 | assert numa_utils._maybe_get_pct_cpu_binding([0]) is None |
| 180 | |
| 181 | |
| 182 | def test_pct_binding_disabled_when_highest_perf_does_not_match(monkeypatch): |
| 183 | _patch_pct_gates(monkeypatch, model_match=True, highest_perf=42) |
| 184 | assert numa_utils._maybe_get_pct_cpu_binding([0]) is None |
| 185 | |
| 186 | |
| 187 | def test_pct_binding_disabled_when_files_missing(monkeypatch): |
| 188 | _patch_pct_gates(monkeypatch, model_match=True, highest_perf=None) |
| 189 | assert numa_utils._maybe_get_pct_cpu_binding([0]) is None |
| 190 | |
| 191 | |
| 192 | def test_pct_binding_returns_none_when_node_cpulist_filter_empty(monkeypatch): |
| 193 | _patch_pct_gates( |
| 194 | monkeypatch, |
| 195 | model_match=True, |
| 196 | highest_perf=46, |
| 197 | cpulist="2-15,18-31", |
| 198 | ) |
| 199 | assert numa_utils._maybe_get_pct_cpu_binding([0]) is None |
| 200 | |
| 201 | |
| 202 | def test_pct_binding_returns_none_when_node_cpulist_missing(monkeypatch): |
| 203 | _patch_pct_gates(monkeypatch, model_match=True, highest_perf=46, cpulist=None) |
| 204 | assert numa_utils._maybe_get_pct_cpu_binding([0]) is None |
| 205 | |
| 206 | |
| 207 | def test_get_numactl_args_uses_pct_when_user_did_not_specify_cpus(monkeypatch): |
| 208 | _patch_pct_gates(monkeypatch, model_match=True, highest_perf=46) |
| 209 | vllm_config = _make_config(numa_bind=True, numa_bind_nodes=[0, 1]) |
| 210 | assert ( |
| 211 | numa_utils._get_numactl_worker_args(vllm_config.parallel_config, local_rank=1) |
| 212 | == "--physcpubind=0,1,16,17,64,65,80,81 --membind=1" |
| 213 | ) |
| 214 | |
| 215 | |
| 216 | def test_get_numactl_args_engine_core_baseline_single_node_shard(): |
| 217 | """Baseline (no PCT): single-NUMA shard -> single-node bind.""" |
| 218 | vllm_config = _make_config(numa_bind=True, numa_bind_nodes=[0, 1]) |
| 219 | assert ( |
| 220 | numa_utils._get_numactl_enginecore_args( |
| 221 | vllm_config.parallel_config, local_rank=0 |
| 222 | ) |
| 223 | == "--cpunodebind=0 --membind=0" |
| 224 | ) |
| 225 | |
| 226 | |
| 227 | def test_get_numactl_args_engine_core_baseline_spans_shard_numa_nodes(): |
| 228 | """Baseline (no PCT): a TP=4 shard spanning both NUMA nodes -> bind to both.""" |
| 229 | vllm_config = _make_config( |
| 230 | numa_bind=True, |
| 231 | numa_bind_nodes=[0, 0, 1, 1], |
| 232 | tensor_parallel_size=4, |
| 233 | ) |
| 234 | assert ( |
| 235 | numa_utils._get_numactl_enginecore_args( |
| 236 | vllm_config.parallel_config, local_rank=0 |
| 237 | ) |
| 238 | == "--cpunodebind=0,1 --membind=0,1" |
| 239 | ) |
| 240 | |
| 241 | |
| 242 | def test_get_numactl_args_engine_core_pct_spans_shard_numa_nodes(monkeypatch): |
| 243 | """PCT: EngineCore for a multi-NUMA shard binds to the union of priority |
| 244 | cores across all shard nodes, so worker `--physcpubind` is always a |
| 245 | subset of EngineCore's `cpus_allowed`.""" |
| 246 | _patch_pct_gates( |
| 247 | monkeypatch, |
| 248 | model_match=True, |
| 249 | highest_perf=46, |
| 250 | cpulist_by_node={0: "0-31,128-159", 1: "64-95,192-223"}, |
| 251 | ) |
| 252 | vllm_config = _make_config( |
| 253 | numa_bind=True, |
| 254 | numa_bind_nodes=[0, 0, 1, 1], |
| 255 | tensor_parallel_size=4, |
| 256 | ) |
| 257 | assert numa_utils._get_numactl_enginecore_args( |
| 258 | vllm_config.parallel_config, local_rank=0 |
| 259 | ) == ( |
| 260 | "--physcpubind=" |
| 261 | "0,1,16,17,64,65,80,81,128,129,144,145,192,193,208,209" |
| 262 | " --membind=0,1" |
| 263 | ) |
| 264 | |
| 265 | |
| 266 | def test_get_numactl_args_engine_core_pct_dp_shard_picks_local_nodes(monkeypatch): |
| 267 | """With DP=2, each shard's EngineCore binds only to its own NUMA nodes.""" |
| 268 | _patch_pct_gates( |
| 269 | monkeypatch, |
| 270 | model_match=True, |
| 271 | highest_perf=46, |
| 272 | cpulist_by_node={0: "0-31,128-159", 1: "64-95,192-223"}, |
| 273 | ) |
| 274 | vllm_config = _make_config( |
| 275 | numa_bind=True, |
| 276 | numa_bind_nodes=[0, 0, 1, 1], |
| 277 | tensor_parallel_size=2, |
| 278 | data_parallel_rank_local=1, |
| 279 | ) |
| 280 | # Shard 1 owns gpu_indices 2 and 3 -> nodes [1, 1] -> {1}. |
| 281 | assert ( |
| 282 | numa_utils._get_numactl_enginecore_args( |
| 283 | vllm_config.parallel_config, local_rank=0 |
| 284 | ) |
| 285 | == "--physcpubind=64,65,80,81,192,193,208,209 --membind=1" |
| 286 | ) |
| 287 | |
| 288 | |
| 289 | def test_get_numactl_args_engine_core_pct_external_launcher_spans_local_nodes( |
| 290 | monkeypatch, |
| 291 | ): |
| 292 | """external_launcher (or multi-node-within-DP, or Ray) hits the |
| 293 | fallback branch. EngineCore must still span every local NUMA node |
| 294 | so it can mp-spawn its local workers without ``--physcpubind`` |
| 295 | strict-validation failures.""" |
| 296 | _patch_pct_gates( |
| 297 | monkeypatch, |
| 298 | model_match=True, |
| 299 | highest_perf=46, |
| 300 | cpulist_by_node={0: "0-31,128-159", 1: "64-95,192-223"}, |
| 301 | ) |
| 302 | vllm_config = _make_config( |
| 303 | numa_bind=True, |
| 304 | numa_bind_nodes=[0, 0, 0, 0, 1, 1, 1, 1], |
| 305 | distributed_executor_backend="external_launcher", |
| 306 | tensor_parallel_size=8, |
| 307 | ) |
| 308 | assert numa_utils._get_numactl_enginecore_args( |
| 309 | vllm_config.parallel_config, local_rank=0 |
| 310 | ) == ( |
| 311 | "--physcpubind=" |
| 312 | "0,1,16,17,64,65,80,81,128,129,144,145,192,193,208,209" |
| 313 | " --membind=0,1" |
| 314 | ) |
| 315 | |
| 316 | |
| 317 | def test_get_numactl_args_engine_core_baseline_multi_node_within_dp_spans_locals(): |
| 318 | """Multi-node-within-DP fallback: bind EngineCore to all local NUMA |
| 319 | nodes that the visible ``numa_bind_nodes`` reference.""" |
| 320 | vllm_config = _make_config( |
| 321 | numa_bind=True, |
| 322 | numa_bind_nodes=[0, 0, 0, 0, 1, 1, 1, 1], |
| 323 | nnodes_within_dp=2, |
| 324 | tensor_parallel_size=8, |
| 325 | ) |
| 326 | assert ( |
| 327 | numa_utils._get_numactl_enginecore_args( |
| 328 | vllm_config.parallel_config, local_rank=0 |
| 329 | ) |
| 330 | == "--cpunodebind=0,1 --membind=0,1" |
| 331 | ) |
| 332 | |
| 333 | |
| 334 | def test_get_numactl_args_engine_core_skips_user_cpu_list(monkeypatch): |
| 335 | """EngineCore ignores ``--numa-bind-cpus``. |
| 336 | |
| 337 | Those are per-worker lists; binding EngineCore to any of them would |
| 338 | shrink its ``cpus_allowed`` below the strict-superset workers' |
| 339 | ``--physcpubind`` spawns need. We fall back to ``--cpunodebind`` over |
| 340 | the shard's NUMA nodes instead. PCT auto-detect is also bypassed when |
| 341 | the user is explicit (its priority-core union may not be a superset |
| 342 | of the user's per-worker cores).""" |
| 343 | _patch_pct_gates(monkeypatch, model_match=True, highest_perf=46) |
| 344 | vllm_config = _make_config( |
| 345 | numa_bind=True, |
| 346 | numa_bind_nodes=[0, 0, 1, 1], |
| 347 | numa_bind_cpus=["0-3", "4-7", "64-67", "68-71"], |
| 348 | tensor_parallel_size=4, |
| 349 | ) |
| 350 | assert ( |
| 351 | numa_utils._get_numactl_enginecore_args( |
| 352 | vllm_config.parallel_config, local_rank=0 |
| 353 | ) |
| 354 | == "--cpunodebind=0,1 --membind=0,1" |
| 355 | ) |
| 356 | |
| 357 | |
| 358 | def test_get_numactl_args_user_cpus_override_pct(monkeypatch): |
| 359 | _patch_pct_gates(monkeypatch, model_match=True, highest_perf=46) |
| 360 | vllm_config = _make_config( |
| 361 | numa_bind=True, |
| 362 | numa_bind_nodes=[0, 1], |
| 363 | numa_bind_cpus=["0-3", "4-7"], |
| 364 | ) |
| 365 | assert ( |
| 366 | numa_utils._get_numactl_worker_args(vllm_config.parallel_config, local_rank=1) |
| 367 | == "--physcpubind=4-7 --membind=1" |
| 368 | ) |
| 369 | |
| 370 | |
| 371 | def test_get_numactl_args_uses_dp_offset(): |
| 372 | vllm_config = _make_config( |
| 373 | numa_bind=True, |
| 374 | numa_bind_nodes=[0, 0, 1, 1], |
| 375 | data_parallel_rank_local=1, |
| 376 | pipeline_parallel_size=1, |
| 377 | tensor_parallel_size=2, |
| 378 | ) |
| 379 | assert ( |
| 380 | numa_utils._get_numactl_worker_args(vllm_config.parallel_config, local_rank=1) |
| 381 | == "--cpunodebind=1 --membind=1" |
| 382 | ) |
| 383 | |
| 384 | |
| 385 | def test_get_numactl_args_requires_detectable_nodes(monkeypatch): |
| 386 | vllm_config = _make_config(numa_bind=True) |
| 387 | monkeypatch.setattr(numa_utils, "get_auto_numa_nodes", lambda: None) |
| 388 | with pytest.raises(RuntimeError): |
| 389 | numa_utils._get_numactl_worker_args(vllm_config.parallel_config, local_rank=0) |
| 390 | |
| 391 | |
| 392 | def test_configure_subprocess_rejects_unknown_process_kind(): |
| 393 | """configure_subprocess only knows 'worker' and 'EngineCore'; anything |
| 394 | else must raise ValueError instead of silently routing to the worker |
| 395 | path.""" |
| 396 | vllm_config = _make_config(numa_bind=True, numa_bind_nodes=[0]) |
| 397 | with ( |
| 398 | pytest.raises(ValueError, match="process_kind"), |
| 399 | numa_utils.configure_subprocess( |
| 400 | vllm_config, local_rank=0, process_kind="bogus" |
| 401 | ), |
| 402 | ): |
| 403 | pass |
| 404 | |
| 405 | |
| 406 | def test_log_numactl_show(monkeypatch): |
| 407 | log_lines = [] |
| 408 | |
| 409 | def fake_debug(msg, *args): |
| 410 | log_lines.append(msg % args) |
| 411 | |
| 412 | monkeypatch.setattr(numa_utils.logger, "debug", fake_debug) |
| 413 | monkeypatch.setattr( |
| 414 | numa_utils.subprocess, |
| 415 | "run", |
| 416 | lambda *args, **kwargs: SimpleNamespace( |
| 417 | stdout="policy: bind\nphyscpubind: 0 1 2 3\n", returncode=0 |
| 418 | ), |
| 419 | ) |
| 420 | |
| 421 | assert numa_utils._log_numactl_show("Worker_0") is True |
| 422 | assert log_lines == [ |
| 423 | "Worker_0 affinity: policy: bind, physcpubind: 0 1 2 3", |
| 424 | ] |
| 425 | |
| 426 | |
| 427 | def test_get_numactl_executable_points_to_fixed_wrapper(monkeypatch): |
| 428 | monkeypatch.setattr("shutil.which", lambda name: "/usr/bin/numactl") |
| 429 | executable, debug_str = numa_utils._get_numactl_executable() |
| 430 | assert executable.endswith("/vllm/utils/numa_wrapper.sh") |
| 431 | assert "_VLLM_INTERNAL_NUMACTL_ARGS" in debug_str |
| 432 | |
| 433 | |
| 434 | def test_set_numa_wrapper_env_restores_previous_values(): |
| 435 | os.environ[numa_utils._NUMACTL_ARGS_ENV] = "old-args" |
| 436 | os.environ[numa_utils._NUMACTL_PYTHON_EXECUTABLE_ENV] = "old-python" |
| 437 | |
| 438 | with numa_utils._set_numa_wrapper_env("new-args", "new-python"): |
| 439 | assert os.environ[numa_utils._NUMACTL_ARGS_ENV] == "new-args" |
| 440 | assert os.environ[numa_utils._NUMACTL_PYTHON_EXECUTABLE_ENV] == "new-python" |
| 441 | |
| 442 | assert os.environ[numa_utils._NUMACTL_ARGS_ENV] == "old-args" |
| 443 | assert os.environ[numa_utils._NUMACTL_PYTHON_EXECUTABLE_ENV] == "old-python" |
| 444 | |
| 445 | |
| 446 | def test_set_numa_wrapper_env_clears_values_when_unset(): |
| 447 | os.environ.pop(numa_utils._NUMACTL_ARGS_ENV, None) |
| 448 | os.environ.pop(numa_utils._NUMACTL_PYTHON_EXECUTABLE_ENV, None) |
| 449 | |
| 450 | with numa_utils._set_numa_wrapper_env("new-args", "new-python"): |
| 451 | assert os.environ[numa_utils._NUMACTL_ARGS_ENV] == "new-args" |
| 452 | assert os.environ[numa_utils._NUMACTL_PYTHON_EXECUTABLE_ENV] == "new-python" |
| 453 | |
| 454 | assert numa_utils._NUMACTL_ARGS_ENV not in os.environ |
| 455 | assert numa_utils._NUMACTL_PYTHON_EXECUTABLE_ENV not in os.environ |
| 456 | |
| 457 | |
| 458 | def test_parallel_config_validates_numa_bind_nodes(): |
| 459 | with pytest.raises(ValueError, match="non-negative"): |
| 460 | ParallelConfig(numa_bind_nodes=[0, -1]) |
| 461 | |
| 462 | |
| 463 | @pytest.mark.parametrize("cpuset", ["", "abc", "1-", "4-1", "1,,2", "1:2"]) |
| 464 | def test_parallel_config_rejects_invalid_numa_bind_cpus(cpuset): |
| 465 | with pytest.raises(ValueError, match="numa_bind_cpus"): |
| 466 | ParallelConfig(numa_bind_cpus=[cpuset]) |
| 467 | |
| 468 | |
| 469 | def _fake_numactl_run(rejected_args): |
| 470 | """Fake ``numactl`` that fails when any of ``rejected_args`` is present.""" |
| 471 | |
| 472 | def run(cmd, *args, **kwargs): |
| 473 | arg_str = " ".join(cmd[1:-1]) |
| 474 | ok = not any(bad in arg_str for bad in rejected_args) |
| 475 | return SimpleNamespace(returncode=0 if ok else 1) |
| 476 | |
| 477 | return run |
| 478 | |
| 479 | |
| 480 | def test_configure_subprocess_numa_fallback(monkeypatch): |
| 481 | import multiprocessing |
| 482 | |
| 483 | monkeypatch.setattr("shutil.which", lambda name: "/usr/bin/numactl") |
| 484 | monkeypatch.setattr(numa_utils.envs, "VLLM_WORKER_MULTIPROC_METHOD", "spawn") |
| 485 | node_config = _make_config(numa_bind=True, numa_bind_nodes=[0]) |
| 486 | |
| 487 | monkeypatch.setattr(numa_utils.subprocess, "run", _fake_numactl_run([])) |
| 488 | with numa_utils.configure_subprocess(node_config, local_rank=0): |
| 489 | assert os.environ[numa_utils._NUMACTL_ARGS_ENV] == "--cpunodebind=0 --membind=0" |
| 490 | |
| 491 | membind_fails = _fake_numactl_run(["--membind="]) |
| 492 | monkeypatch.setattr(numa_utils.subprocess, "run", membind_fails) |
| 493 | with numa_utils.configure_subprocess(node_config, local_rank=0): |
| 494 | assert os.environ[numa_utils._NUMACTL_ARGS_ENV] == "--cpunodebind=0" |
| 495 | |
| 496 | cpu_config = _make_config( |
| 497 | numa_bind=True, |
| 498 | numa_bind_nodes=[0], |
| 499 | numa_bind_cpus=["0-3"], |
| 500 | ) |
| 501 | with numa_utils.configure_subprocess(cpu_config, local_rank=0): |
| 502 | assert os.environ[numa_utils._NUMACTL_ARGS_ENV] == "--physcpubind=0-3" |
| 503 | |
| 504 | before = multiprocessing.spawn.get_executable() |
| 505 | monkeypatch.setattr( |
| 506 | numa_utils.subprocess, |
| 507 | "run", |
| 508 | _fake_numactl_run(["--cpunodebind=", "--membind="]), |
| 509 | ) |
| 510 | with numa_utils.configure_subprocess(node_config, local_rank=0): |
| 511 | assert multiprocessing.spawn.get_executable() == before |
| 512 | assert numa_utils._NUMACTL_ARGS_ENV not in os.environ |