| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """NUMA binding utilities for vLLM worker processes. |
| 4 | |
| 5 | Adapted in part from SGLang's NUMA helper implementation: |
| 6 | https://github.com/sgl-project/sglang/blob/ba6d54d0f08f82f42b8224908ae2459a496b31b3/python/sglang/srt/utils/numa_utils.py |
| 7 | """ |
| 8 | |
| 9 | import ctypes |
| 10 | import logging |
| 11 | import multiprocessing |
| 12 | import os |
| 13 | import subprocess |
| 14 | from contextlib import contextmanager |
| 15 | from functools import cache |
| 16 | from pathlib import Path |
| 17 | from typing import TYPE_CHECKING, NamedTuple |
| 18 | |
| 19 | import psutil |
| 20 | |
| 21 | from vllm import envs |
| 22 | |
| 23 | if TYPE_CHECKING: |
| 24 | from vllm.config import VllmConfig |
| 25 | |
| 26 | logger = logging.getLogger(__name__) |
| 27 | _NUMACTL_ARGS_ENV = "_VLLM_INTERNAL_NUMACTL_ARGS" |
| 28 | _NUMACTL_PYTHON_EXECUTABLE_ENV = "_VLLM_INTERNAL_NUMACTL_PYTHON_EXECUTABLE" |
| 29 | |
| 30 | |
| 31 | @cache |
| 32 | def get_libnuma(): |
| 33 | libnuma = None |
| 34 | for libnuma_so in ["libnuma.so", "libnuma.so.1"]: |
| 35 | try: |
| 36 | libnuma = ctypes.CDLL(libnuma_so) |
| 37 | except OSError: |
| 38 | libnuma = None |
| 39 | if libnuma is not None: |
| 40 | break |
| 41 | return libnuma |
| 42 | |
| 43 | |
| 44 | def _can_set_mempolicy() -> bool: |
| 45 | """Check whether the current process can use NUMA memory policy syscalls.""" |
| 46 | try: |
| 47 | libnuma = get_libnuma() |
| 48 | if libnuma is None or libnuma.numa_available() < 0: |
| 49 | return False |
| 50 | mode = ctypes.c_int() |
| 51 | ret = libnuma.get_mempolicy( |
| 52 | ctypes.byref(mode), None, ctypes.c_ulong(0), None, ctypes.c_ulong(0) |
| 53 | ) |
| 54 | return ret == 0 |
| 55 | except Exception: |
| 56 | return False |
| 57 | |
| 58 | |
| 59 | def _is_auto_numa_available() -> bool: |
| 60 | """Check whether automatic GPU-to-NUMA detection should be attempted.""" |
| 61 | from vllm.platforms import current_platform |
| 62 | |
| 63 | if not current_platform.is_cuda_alike(): |
| 64 | return False |
| 65 | |
| 66 | if not os.path.isdir("/sys/devices/system/node/node1"): |
| 67 | return False |
| 68 | |
| 69 | try: |
| 70 | process = psutil.Process(os.getpid()) |
| 71 | cpu_affinity = process.cpu_affinity() |
| 72 | cpu_count = psutil.cpu_count() |
| 73 | if cpu_count is not None and cpu_affinity != list(range(cpu_count)): |
| 74 | logger.warning( |
| 75 | "CPU affinity is already constrained for this process. " |
| 76 | "Skipping automatic NUMA binding; pass --numa-bind-nodes " |
| 77 | "explicitly to override." |
| 78 | ) |
| 79 | return False |
| 80 | except (AttributeError, NotImplementedError, psutil.Error): |
| 81 | pass |
| 82 | |
| 83 | if not _can_set_mempolicy(): |
| 84 | logger.warning( |
| 85 | "User lacks permission to set NUMA memory policy. " |
| 86 | "Automatic NUMA detection may not work; if you are using Docker, " |
| 87 | "try adding --cap-add SYS_NICE." |
| 88 | ) |
| 89 | return False |
| 90 | |
| 91 | if not hasattr(current_platform, "get_all_device_numa_nodes"): |
| 92 | logger.warning( |
| 93 | "Platform %s does not support automatic NUMA detection", |
| 94 | type(current_platform).__name__, |
| 95 | ) |
| 96 | return False |
| 97 | |
| 98 | return True |
| 99 | |
| 100 | |
| 101 | @cache |
| 102 | def get_auto_numa_nodes() -> list[int] | None: |
| 103 | """Auto-detect NUMA nodes for all visible GPUs.""" |
| 104 | from vllm.platforms import current_platform |
| 105 | |
| 106 | if not _is_auto_numa_available(): |
| 107 | return None |
| 108 | |
| 109 | numa_nodes = current_platform.get_all_device_numa_nodes() |
| 110 | if numa_nodes is not None: |
| 111 | logger.info("Auto-detected NUMA nodes for GPUs: %s", numa_nodes) |
| 112 | return numa_nodes |
| 113 | |
| 114 | |
| 115 | # PCT (Priority Core Turbo) auto-detection workaround for Granite Rapids |
| 116 | # Xeon SKUs. |
| 117 | # |
| 118 | # Background: |
| 119 | # * The Linux kernel does not expose PCT priority-core membership via any |
| 120 | # unprivileged sysfs path. The official interface |
| 121 | # (/dev/isst_interface, used by `intel-speed-select`) is root-only, |
| 122 | # which is a non-starter in most production deployments (shared |
| 123 | # clusters, prebuilt containers, managed cloud). |
| 124 | # * Even recent stable kernels (e.g. 6.14, March 2025) do not yet |
| 125 | # preferentially schedule work on PCT priority cores, so vLLM cannot |
| 126 | # just "let the scheduler handle it". |
| 127 | # |
| 128 | # Empirical heuristic (DGX B300 / Xeon 6776P, the SKU we measured): |
| 129 | # * /proc/cpuinfo `model name` contains the SKU number. |
| 130 | # * cpu0 is a PCT priority core on these SKUs, so it reports the |
| 131 | # priority-cohort CPPC `highest_perf` (the value matches the SKU's |
| 132 | # "Max PCT core frequency" in 100 MHz units, e.g. 4.6 GHz -> 46). |
| 133 | # * Priority cores within each NUMA node satisfy `cpu_id % S in (0, 1)` |
| 134 | # intersected with the node's cpulist, where `S` is the SKU's logical |
| 135 | # CPUs per priority "group" (= total threads / 8 priority cores; 16 on |
| 136 | # 64-core SKUs, 18 on 72-core SKUs). |
| 137 | # |
| 138 | # SKU table: |
| 139 | # ``_PCT_CAPABLE_SKUS`` maps each known PCT-capable Granite Rapids part |
| 140 | # to a ``_PctSku(highest_perf, priority_stride)`` config: |
| 141 | # * highest_perf is the expected ``acpi_cppc/highest_perf`` on cpu0, |
| 142 | # derived from Intel ARK's "Max PCT core frequency" * 10 (CPPC max |
| 143 | # ratio reports in 100 MHz units). |
| 144 | # * priority_stride is the SKU's "Total Cores" / 4 (= total HT threads |
| 145 | # / 8 priority cores), used in the ``cpu_id % stride`` filter above. |
| 146 | # Values: |
| 147 | # * 6776P - 4.6 GHz, 64C/128T -> (46, 16) measured on DGX B300 |
| 148 | # * 6774P - 4.6 GHz, 64C/128T -> (46, 16) per Intel ARK, not measured |
| 149 | # * 6962P - 4.4 GHz, 72C/144T -> (44, 18) per Intel ARK, not measured |
| 150 | # The non-measured SKUs are listed best-effort: the gate fails closed |
| 151 | # (no PCT engagement) if a host's actual highest_perf doesn't match the |
| 152 | # table value, so adding entries is safe. If you have access to a 6962P |
| 153 | # or 6774P box and find a different value or cpu-id pattern, update the |
| 154 | # table below. |
| 155 | # |
| 156 | # This whole block is a stop-gap until the kernel exposes PCT membership |
| 157 | # in an unprivileged way; see the tracking issue linked from the PR. |
| 158 | |
| 159 | |
| 160 | class _PctSku(NamedTuple): |
| 161 | """Per-SKU config used by the PCT auto-detection gate.""" |
| 162 | |
| 163 | highest_perf: int |
| 164 | priority_stride: int |
| 165 | |
| 166 | |
| 167 | _PCT_CAPABLE_SKUS: dict[str, _PctSku] = { |
| 168 | "6776P": _PctSku(highest_perf=46, priority_stride=16), |
| 169 | "6774P": _PctSku(highest_perf=46, priority_stride=16), |
| 170 | "6962P": _PctSku(highest_perf=44, priority_stride=18), |
| 171 | } |
| 172 | _PCT_HIGHEST_PERF_PATH = "/sys/devices/system/cpu/cpu0/acpi_cppc/highest_perf" |
| 173 | _PROC_CPUINFO_PATH = "/proc/cpuinfo" |
| 174 | |
| 175 | |
| 176 | def _pct_sku_from_cpuinfo() -> _PctSku | None: |
| 177 | """Return the ``_PctSku`` config for this host's SKU, or None. |
| 178 | |
| 179 | Reads ``/proc/cpuinfo``'s ``model name`` and looks the SKU up in |
| 180 | ``_PCT_CAPABLE_SKUS``. Returns ``None`` when the host is not a known |
| 181 | PCT-capable Granite Rapids Xeon (or when ``/proc/cpuinfo`` is |
| 182 | unreadable). |
| 183 | """ |
| 184 | try: |
| 185 | with open(_PROC_CPUINFO_PATH) as f: |
| 186 | for line in f: |
| 187 | if not line.lstrip().lower().startswith("model name"): |
| 188 | continue |
| 189 | for sku, config in _PCT_CAPABLE_SKUS.items(): |
| 190 | if sku in line: |
| 191 | return config |
| 192 | except OSError: |
| 193 | return None |
| 194 | return None |
| 195 | |
| 196 | |
| 197 | @cache |
| 198 | def _pct_sku_config() -> _PctSku | None: |
| 199 | """Detect a PCT-capable Granite Rapids Xeon with PCT enabled. |
| 200 | |
| 201 | See the comment block above ``_PCT_CAPABLE_SKUS`` for the full context |
| 202 | (why we hard-code SKUs, why we read CPPC ``highest_perf``, etc.). |
| 203 | |
| 204 | Returns the matching ``_PctSku`` config when both gates hold: |
| 205 | * ``/proc/cpuinfo`` ``model name`` contains an SKU listed in |
| 206 | ``_PCT_CAPABLE_SKUS``. |
| 207 | * ``/sys/devices/system/cpu/cpu0/acpi_cppc/highest_perf`` matches |
| 208 | that SKU's expected ``highest_perf``. |
| 209 | Otherwise returns ``None`` and the caller falls back to the default |
| 210 | NUMA-node bind. |
| 211 | """ |
| 212 | sku = _pct_sku_from_cpuinfo() |
| 213 | if sku is None: |
| 214 | return None |
| 215 | |
| 216 | try: |
| 217 | with open(_PCT_HIGHEST_PERF_PATH) as f: |
| 218 | actual = int(f.read().strip()) |
| 219 | except (OSError, ValueError): |
| 220 | return None |
| 221 | if actual != sku.highest_perf: |
| 222 | return None |
| 223 | return sku |
| 224 | |
| 225 | |
| 226 | def _get_gpu_index( |
| 227 | parallel_config, local_rank: int, dp_local_rank: int | None = None |
| 228 | ) -> int: |
| 229 | """Compute the physical GPU index used for NUMA lookup.""" |
| 230 | if ( |
| 231 | parallel_config.distributed_executor_backend not in ("ray", "external_launcher") |
| 232 | and parallel_config.data_parallel_backend != "ray" |
| 233 | and parallel_config.nnodes_within_dp == 1 |
| 234 | ): |
| 235 | if dp_local_rank is None: |
| 236 | dp_local_rank = parallel_config.data_parallel_rank_local |
| 237 | if dp_local_rank is None: |
| 238 | dp_local_rank = parallel_config.data_parallel_index |
| 239 | |
| 240 | tp_pp_world_size = ( |
| 241 | parallel_config.pipeline_parallel_size |
| 242 | * parallel_config.tensor_parallel_size |
| 243 | ) |
| 244 | return local_rank + dp_local_rank * tp_pp_world_size |
| 245 | |
| 246 | return local_rank |
| 247 | |
| 248 | |
| 249 | def _get_numa_node(parallel_config, gpu_index: int) -> int: |
| 250 | numa_nodes = parallel_config.numa_bind_nodes |
| 251 | if numa_nodes is None: |
| 252 | numa_nodes = get_auto_numa_nodes() |
| 253 | if numa_nodes is None: |
| 254 | raise RuntimeError( |
| 255 | "NUMA binding was requested, but vLLM could not detect the " |
| 256 | "GPU-to-NUMA topology automatically. Pass --numa-bind-nodes " |
| 257 | "explicitly or disable --numa-bind." |
| 258 | ) |
| 259 | parallel_config.numa_bind_nodes = numa_nodes |
| 260 | |
| 261 | if gpu_index >= len(numa_nodes): |
| 262 | raise ValueError( |
| 263 | f"GPU index {gpu_index} exceeds numa_bind_nodes size {len(numa_nodes)}. " |
| 264 | "Ensure the binding lists cover every visible GPU." |
| 265 | ) |
| 266 | |
| 267 | return numa_nodes[gpu_index] |
| 268 | |
| 269 | |
| 270 | def _maybe_get_pct_cpu_binding(numa_nodes: list[int]) -> list[int] | None: |
| 271 | """Return the union of PCT priority cores across ``numa_nodes`` (or None). |
| 272 | |
| 273 | PCT (Priority Core Turbo) lets a subset of cores boost above the rest; |
| 274 | we want workers and the EngineCore on those cores. The Linux kernel does |
| 275 | not expose PCT membership without root, so we use the empirical heuristic |
| 276 | documented above ``_PCT_CAPABLE_SKUS``: priority cores within each NUMA |
| 277 | node satisfy ``cpu_id % stride in (0, 1)`` intersected with the node's |
| 278 | ``cpulist``, where ``stride`` is the SKU's logical CPUs per priority |
| 279 | group (16 on 64-core SKUs, 18 on 72-core SKUs). Only triggers on the |
| 280 | SKUs in ``_PCT_CAPABLE_SKUS`` with the expected CPPC ``highest_perf`` |
| 281 | signal; on any other host it returns None and the caller falls back to |
| 282 | the default NUMA-node bind. |
| 283 | |
| 284 | Returns the sorted CPU ids as a ``list[int]``; the caller is expected |
| 285 | to format them for the chosen tool (e.g. comma-joined for |
| 286 | ``numactl --physcpubind``). |
| 287 | """ |
| 288 | sku = _pct_sku_config() |
| 289 | if sku is None: |
| 290 | return None |
| 291 | |
| 292 | from vllm.utils.cpu_resource_utils import parse_id_list |
| 293 | |
| 294 | stride = sku.priority_stride |
| 295 | union_cpus: set[int] = set() |
| 296 | for numa_node in numa_nodes: |
| 297 | cpulist_path = Path(f"/sys/devices/system/node/node{numa_node}/cpulist") |
| 298 | try: |
| 299 | cpulist_raw = cpulist_path.read_text().strip() |
| 300 | except OSError: |
| 301 | continue |
| 302 | if not cpulist_raw: |
| 303 | continue |
| 304 | try: |
| 305 | node_cpus = parse_id_list(cpulist_raw) |
| 306 | except ValueError: |
| 307 | continue |
| 308 | |
| 309 | priority = [cpu for cpu in node_cpus if cpu % stride in (0, 1)] |
| 310 | if not priority: |
| 311 | continue |
| 312 | union_cpus.update(priority) |
| 313 | logger.info( |
| 314 | "Detected PCT-capable Granite Rapids Xeon (stride=%d); " |
| 315 | "NUMA node %d priority cores: %s", |
| 316 | stride, |
| 317 | numa_node, |
| 318 | ",".join(str(c) for c in priority), |
| 319 | ) |
| 320 | |
| 321 | if not union_cpus: |
| 322 | return None |
| 323 | return sorted(union_cpus) |
| 324 | |
| 325 | |
| 326 | def _get_cpu_binding( |
| 327 | parallel_config, gpu_index: int, numa_nodes: list[int] |
| 328 | ) -> str | None: |
| 329 | """Return the CPU list a process should be pinned to (or None).""" |
| 330 | cpu_bindings = parallel_config.numa_bind_cpus |
| 331 | if cpu_bindings is None: |
| 332 | pct_cpus = _maybe_get_pct_cpu_binding(numa_nodes) |
| 333 | if pct_cpus is None: |
| 334 | return None |
| 335 | return ",".join(str(c) for c in pct_cpus) |
| 336 | |
| 337 | if gpu_index >= len(cpu_bindings): |
| 338 | raise ValueError( |
| 339 | f"GPU index {gpu_index} exceeds numa_bind_cpus size " |
| 340 | f"{len(cpu_bindings)}. Ensure the binding lists cover every visible GPU." |
| 341 | ) |
| 342 | return cpu_bindings[gpu_index] |
| 343 | |
| 344 | |
| 345 | def _get_numactl_worker_args( |
| 346 | parallel_config, local_rank: int, dp_local_rank: int | None = None |
| 347 | ) -> str: |
| 348 | """Compute the numactl args for a single TP/PP worker subprocess.""" |
| 349 | gpu_index = _get_gpu_index(parallel_config, local_rank, dp_local_rank) |
| 350 | numa_node = _get_numa_node(parallel_config, gpu_index) |
| 351 | cpu_binding = _get_cpu_binding(parallel_config, gpu_index, [numa_node]) |
| 352 | |
| 353 | if cpu_binding is not None: |
| 354 | logger.info( |
| 355 | "Binding worker subprocess (local_rank=%s, gpu_index=%s) to CPUs %s and NUMA node %s", # noqa: E501 |
| 356 | local_rank, |
| 357 | gpu_index, |
| 358 | cpu_binding, |
| 359 | numa_node, |
| 360 | ) |
| 361 | return f"--physcpubind={cpu_binding} --membind={numa_node}" |
| 362 | |
| 363 | logger.info( |
| 364 | "Binding worker subprocess (local_rank=%s, gpu_index=%s) to NUMA node %s", |
| 365 | local_rank, |
| 366 | gpu_index, |
| 367 | numa_node, |
| 368 | ) |
| 369 | return f"--cpunodebind={numa_node} --membind={numa_node}" |
| 370 | |
| 371 | |
| 372 | def _get_enginecore_numa_nodes( |
| 373 | parallel_config, dp_local_rank: int | None = None |
| 374 | ) -> list[int]: |
| 375 | """Return the sorted, unique NUMA nodes of the EngineCore's DP shard.""" |
| 376 | numa_nodes = parallel_config.numa_bind_nodes |
| 377 | if numa_nodes is None: |
| 378 | # Trigger auto-detection (it caches into parallel_config). |
| 379 | _get_numa_node(parallel_config, 0) |
| 380 | numa_nodes = parallel_config.numa_bind_nodes |
| 381 | |
| 382 | if ( |
| 383 | parallel_config.distributed_executor_backend not in ("ray", "external_launcher") |
| 384 | and parallel_config.data_parallel_backend != "ray" |
| 385 | and parallel_config.nnodes_within_dp == 1 |
| 386 | ): |
| 387 | if dp_local_rank is None: |
| 388 | dp_local_rank = parallel_config.data_parallel_rank_local |
| 389 | if dp_local_rank is None: |
| 390 | dp_local_rank = parallel_config.data_parallel_index |
| 391 | |
| 392 | tp_pp_world_size = ( |
| 393 | parallel_config.pipeline_parallel_size |
| 394 | * parallel_config.tensor_parallel_size |
| 395 | ) |
| 396 | shard_start = dp_local_rank * tp_pp_world_size |
| 397 | shard_end = min(shard_start + tp_pp_world_size, len(numa_nodes)) |
| 398 | shard_indices: range | tuple[int, ...] = range(shard_start, shard_end) |
| 399 | else: |
| 400 | shard_indices = range(len(numa_nodes)) |
| 401 | |
| 402 | if not shard_indices: |
| 403 | return [numa_nodes[0]] |
| 404 | return sorted({numa_nodes[i] for i in shard_indices}) |
| 405 | |
| 406 | |
| 407 | def _get_numactl_enginecore_args( |
| 408 | parallel_config, local_rank: int, dp_local_rank: int | None = None |
| 409 | ) -> str: |
| 410 | """Compute the numactl args for an EngineCore subprocess. |
| 411 | |
| 412 | ``--numa-bind-cpus`` is deliberately ignored here: the user provides a |
| 413 | per-worker CPU list, and binding EngineCore to any of those entries |
| 414 | would shrink its ``cpus_allowed`` below the strict-superset that the |
| 415 | workers' ``--physcpubind`` spawns require. We fall back to |
| 416 | ``--cpunodebind=<shard nodes>`` instead, which is always a safe |
| 417 | superset. PCT auto-detection still applies when the user did not pass |
| 418 | ``--numa-bind-cpus`` (its priority-core union across the shard nodes |
| 419 | is also a safe superset by construction). |
| 420 | """ |
| 421 | shard_nodes = _get_enginecore_numa_nodes(parallel_config, dp_local_rank) |
| 422 | membind_arg = ",".join(str(n) for n in shard_nodes) |
| 423 | |
| 424 | pct_cpus = ( |
| 425 | None |
| 426 | if parallel_config.numa_bind_cpus is not None |
| 427 | else _maybe_get_pct_cpu_binding(shard_nodes) |
| 428 | ) |
| 429 | |
| 430 | if pct_cpus is not None: |
| 431 | cpu_binding = ",".join(str(c) for c in pct_cpus) |
| 432 | logger.info( |
| 433 | "Binding EngineCore subprocess (local_rank=%s) to CPUs %s " |
| 434 | "and NUMA nodes %s", |
| 435 | local_rank, |
| 436 | cpu_binding, |
| 437 | membind_arg, |
| 438 | ) |
| 439 | return f"--physcpubind={cpu_binding} --membind={membind_arg}" |
| 440 | |
| 441 | logger.info( |
| 442 | "Binding EngineCore subprocess (local_rank=%s) to NUMA nodes %s", |
| 443 | local_rank, |
| 444 | membind_arg, |
| 445 | ) |
| 446 | return f"--cpunodebind={membind_arg} --membind={membind_arg}" |
| 447 | |
| 448 | |
| 449 | def _log_numactl_show(label: str) -> bool: |
| 450 | try: |
| 451 | result = subprocess.run( |
| 452 | ["numactl", "--show"], |
| 453 | check=True, |
| 454 | capture_output=True, |
| 455 | text=True, |
| 456 | ) |
| 457 | except (FileNotFoundError, subprocess.CalledProcessError) as e: |
| 458 | logger.warning("Failed to run `numactl --show` for %s: %s", label, e) |
| 459 | return False |
| 460 | |
| 461 | output = result.stdout.strip() |
| 462 | if not output: |
| 463 | logger.warning("`numactl --show` returned no output for %s", label) |
| 464 | return False |
| 465 | |
| 466 | summary = ", ".join(line.strip() for line in output.splitlines() if line.strip()) |
| 467 | logger.debug("%s affinity: %s", label, summary) |
| 468 | return True |
| 469 | |
| 470 | |
| 471 | def log_current_affinity_state(label: str) -> None: |
| 472 | """Log the process's effective NUMA affinity state.""" |
| 473 | _log_numactl_show(label) |
| 474 | |
| 475 | |
| 476 | def _probe_numactl_args(numactl_args: str) -> bool: |
| 477 | """Whether ``numactl <args> true`` succeeds in this (parent) environment.""" |
| 478 | try: |
| 479 | result = subprocess.run( |
| 480 | ["numactl", *numactl_args.split(), "true"], |
| 481 | capture_output=True, |
| 482 | timeout=10, |
| 483 | ) |
| 484 | except (OSError, subprocess.SubprocessError): |
| 485 | return False |
| 486 | return result.returncode == 0 |
| 487 | |
| 488 | |
| 489 | def _resolve_numactl_args(numactl_args: str) -> str: |
| 490 | """Drop ``--membind`` if the container rejects it, keeping CPU binding.""" |
| 491 | cpu_only = " ".join( |
| 492 | t for t in numactl_args.split() if not t.startswith("--membind=") |
| 493 | ) |
| 494 | for candidate in (numactl_args, cpu_only, ""): |
| 495 | if _probe_numactl_args(candidate): |
| 496 | if candidate != numactl_args: |
| 497 | logger.warning( |
| 498 | "numactl args %r rejected; falling back to %r. Add " |
| 499 | "--cap-add SYS_NICE for full NUMA binding.", |
| 500 | numactl_args, |
| 501 | candidate or "no binding", |
| 502 | ) |
| 503 | return candidate |
| 504 | return "" |
| 505 | |
| 506 | |
| 507 | @contextmanager |
| 508 | def configure_subprocess( |
| 509 | vllm_config: "VllmConfig", |
| 510 | local_rank: int, |
| 511 | dp_local_rank: int | None = None, |
| 512 | process_kind: str = "worker", |
| 513 | ): |
| 514 | """Temporarily replace the multiprocessing executable with a numactl wrapper.""" |
| 515 | parallel_config = vllm_config.parallel_config |
| 516 | if not parallel_config.numa_bind: |
| 517 | yield |
| 518 | return |
| 519 | |
| 520 | if process_kind == "EngineCore": |
| 521 | numactl_args = _get_numactl_enginecore_args( |
| 522 | parallel_config, local_rank, dp_local_rank |
| 523 | ) |
| 524 | elif process_kind == "worker": |
| 525 | numactl_args = _get_numactl_worker_args( |
| 526 | parallel_config, local_rank, dp_local_rank |
| 527 | ) |
| 528 | else: |
| 529 | raise ValueError( |
| 530 | f"Unknown process_kind {process_kind!r}; expected 'worker' or 'EngineCore'." |
| 531 | ) |
| 532 | |
| 533 | executable, debug_str = _get_numactl_executable() |
| 534 | numactl_args = _resolve_numactl_args(numactl_args) |
| 535 | if not numactl_args: |
| 536 | # No NUMA binding possible here; launch without the wrapper. |
| 537 | yield |
| 538 | return |
| 539 | python_executable = os.fsdecode(multiprocessing.spawn.get_executable()) |
| 540 | with ( |
| 541 | _set_numa_wrapper_env(numactl_args, python_executable), |
| 542 | _mp_set_executable(executable, debug_str), |
| 543 | ): |
| 544 | yield |
| 545 | |
| 546 | |
| 547 | def _get_numactl_executable() -> tuple[str, str]: |
| 548 | """Return the fixed wrapper executable used to launch numactl.""" |
| 549 | from shutil import which |
| 550 | |
| 551 | if which("numactl") is None: |
| 552 | raise RuntimeError( |
| 553 | "numactl is required for NUMA binding but is not installed or " |
| 554 | "not available on PATH." |
| 555 | ) |
| 556 | |
| 557 | script_path = Path(__file__).with_name("numa_wrapper.sh") |
| 558 | return str(script_path), f"{script_path} via {_NUMACTL_ARGS_ENV}" |
| 559 | |
| 560 | |
| 561 | @contextmanager |
| 562 | def _set_numa_wrapper_env(numactl_args: str, python_executable: str): |
| 563 | old_numactl_args = os.environ.get(_NUMACTL_ARGS_ENV) |
| 564 | old_python_executable = os.environ.get(_NUMACTL_PYTHON_EXECUTABLE_ENV) |
| 565 | os.environ[_NUMACTL_ARGS_ENV] = numactl_args |
| 566 | os.environ[_NUMACTL_PYTHON_EXECUTABLE_ENV] = python_executable |
| 567 | try: |
| 568 | yield |
| 569 | finally: |
| 570 | if old_numactl_args is None: |
| 571 | os.environ.pop(_NUMACTL_ARGS_ENV, None) |
| 572 | else: |
| 573 | os.environ[_NUMACTL_ARGS_ENV] = old_numactl_args |
| 574 | |
| 575 | if old_python_executable is None: |
| 576 | os.environ.pop(_NUMACTL_PYTHON_EXECUTABLE_ENV, None) |
| 577 | else: |
| 578 | os.environ[_NUMACTL_PYTHON_EXECUTABLE_ENV] = old_python_executable |
| 579 | |
| 580 | |
| 581 | @contextmanager |
| 582 | def _mp_set_executable(executable: str, debug_str: str): |
| 583 | start_method = envs.VLLM_WORKER_MULTIPROC_METHOD |
| 584 | if start_method != "spawn": |
| 585 | logger.warning( |
| 586 | "NUMA binding requires spawn method but got '%s'. " |
| 587 | "NUMA binding will be ineffective. " |
| 588 | "Set VLLM_WORKER_MULTIPROC_METHOD=spawn to enable NUMA binding.", |
| 589 | start_method, |
| 590 | ) |
| 591 | yield |
| 592 | return |
| 593 | |
| 594 | old_executable = os.fsdecode(multiprocessing.spawn.get_executable()) |
| 595 | multiprocessing.spawn.set_executable(executable) |
| 596 | try: |
| 597 | yield |
| 598 | finally: |
| 599 | assert os.fsdecode(multiprocessing.spawn.get_executable()) == executable, ( |
| 600 | "Executable was changed during NUMA binding context: " |
| 601 | f"expected {executable}, got {multiprocessing.spawn.get_executable()}" |
| 602 | ) |
| 603 | multiprocessing.spawn.set_executable(old_executable) |