| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import glob |
| 5 | import os |
| 6 | import platform |
| 7 | import subprocess |
| 8 | import sys |
| 9 | from typing import TYPE_CHECKING |
| 10 | |
| 11 | import torch |
| 12 | |
| 13 | from vllm import envs |
| 14 | from vllm.logger import init_logger |
| 15 | from vllm.utils.cpu_resource_utils import ( |
| 16 | DEVICE_CONTROL_ENV_VAR, |
| 17 | get_memory_node_info, |
| 18 | get_visible_memory_node, |
| 19 | ) |
| 20 | from vllm.utils.mem_constants import GiB_bytes |
| 21 | from vllm.v1.attention.backends.registry import AttentionBackendEnum |
| 22 | |
| 23 | from .interface import CpuArchEnum, Platform, PlatformEnum |
| 24 | |
| 25 | logger = init_logger(__name__) |
| 26 | |
| 27 | if TYPE_CHECKING: |
| 28 | from vllm.config import VllmConfig |
| 29 | from vllm.v1.attention.selector import AttentionSelectorConfig |
| 30 | else: |
| 31 | VllmConfig = None |
| 32 | |
| 33 | |
| 34 | def _cpu_mamba_backend(vllm_config: VllmConfig) -> str: |
| 35 | """Return the CPU state backend selected by the model configuration.""" |
| 36 | model_config = vllm_config.model_config |
| 37 | if model_config is None: |
| 38 | return "none" |
| 39 | |
| 40 | hf_config = getattr(model_config, "hf_text_config", None) |
| 41 | model_type = str(getattr(hf_config, "model_type", "")).lower() |
| 42 | architecture = str(getattr(model_config, "architecture", "")).lower() |
| 43 | layer_types = getattr(hf_config, "layer_types", None) |
| 44 | has_linear_attention = isinstance(layer_types, (list, tuple)) and ( |
| 45 | "linear_attention" in layer_types |
| 46 | ) |
| 47 | |
| 48 | try: |
| 49 | has_inner_state = bool(model_config.has_inner_state) |
| 50 | except (AttributeError, RuntimeError): |
| 51 | has_inner_state = False |
| 52 | |
| 53 | if not (has_inner_state or has_linear_attention): |
| 54 | return "none" |
| 55 | |
| 56 | fallback_backend = model_type or architecture or "unknown" |
| 57 | if not has_linear_attention: |
| 58 | return fallback_backend |
| 59 | |
| 60 | try: |
| 61 | model_cls, _ = model_config.registry.resolve_model_cls( |
| 62 | model_config.architecture, |
| 63 | model_config=model_config, |
| 64 | ) |
| 65 | except Exception: |
| 66 | return fallback_backend |
| 67 | |
| 68 | try: |
| 69 | state_dtypes = model_cls.get_mamba_state_dtype_from_config(vllm_config) |
| 70 | except Exception: |
| 71 | return fallback_backend |
| 72 | |
| 73 | if not isinstance(state_dtypes, tuple) or len(state_dtypes) != 2: |
| 74 | return fallback_backend |
| 75 | |
| 76 | if vllm_config.cache_config.mamba_ssm_cache_dtype == "float16": |
| 77 | cache_dtype = torch.float16 |
| 78 | elif vllm_config.cache_config.mamba_ssm_cache_dtype == "bfloat16": |
| 79 | cache_dtype = torch.bfloat16 |
| 80 | else: |
| 81 | return fallback_backend |
| 82 | |
| 83 | if state_dtypes[1] == cache_dtype: |
| 84 | return "gdn" |
| 85 | |
| 86 | return fallback_backend |
| 87 | |
| 88 | |
| 89 | def get_max_threads(pid=0): |
| 90 | if hasattr(os, "sched_getaffinity"): |
| 91 | return len(os.sched_getaffinity(pid)) |
| 92 | elif platform.system() == "Darwin": |
| 93 | return os.cpu_count() |
| 94 | else: |
| 95 | raise NotImplementedError("Unsupported OS") |
| 96 | |
| 97 | |
| 98 | class CpuPlatform(Platform): |
| 99 | _enum = PlatformEnum.CPU |
| 100 | device_name: str = "cpu" |
| 101 | device_type: str = "cpu" |
| 102 | dispatch_key: str = "CPU" |
| 103 | dist_backend: str = "gloo" |
| 104 | device_control_env_var = DEVICE_CONTROL_ENV_VAR |
| 105 | |
| 106 | @property |
| 107 | def supported_dtypes(self) -> list[torch.dtype]: |
| 108 | if self.get_cpu_architecture() == CpuArchEnum.POWERPC: |
| 109 | return [torch.bfloat16, torch.float32, torch.float16] |
| 110 | elif self.get_cpu_architecture() == CpuArchEnum.ARM and sys.platform.startswith( |
| 111 | "darwin" |
| 112 | ): |
| 113 | # sysctl exits non-zero when the OID is absent, so check_output would |
| 114 | # raise instead of letting the fp16/fp32 fallback below run. |
| 115 | bf16 = ( |
| 116 | subprocess.run( |
| 117 | ["sysctl", "-n", "hw.optional.arm.FEAT_BF16"], capture_output=True |
| 118 | ).stdout.strip() |
| 119 | == b"1" |
| 120 | ) |
| 121 | if bf16: |
| 122 | return [torch.bfloat16, torch.float16, torch.float32] |
| 123 | return [torch.float16, torch.float32] |
| 124 | elif self.get_cpu_architecture() == CpuArchEnum.RISCV: |
| 125 | return [torch.bfloat16, torch.float16, torch.float32] |
| 126 | # x86/aarch64 CPU has supported both bf16 and fp16 natively. |
| 127 | return [torch.bfloat16, torch.float16, torch.float32] |
| 128 | |
| 129 | @classmethod |
| 130 | def check_runner_kv_caches_multi_layer(cls) -> None: |
| 131 | pass |
| 132 | |
| 133 | @classmethod |
| 134 | def get_device_name(cls, device_id: int = 0) -> str: |
| 135 | return "cpu" |
| 136 | |
| 137 | @classmethod |
| 138 | def get_attn_backend_cls( |
| 139 | cls, |
| 140 | selected_backend: "AttentionBackendEnum", |
| 141 | attn_selector_config: "AttentionSelectorConfig", |
| 142 | num_heads: int | None = None, |
| 143 | ) -> str: |
| 144 | if attn_selector_config.use_sparse: |
| 145 | raise NotImplementedError("Sparse Attention is not supported on CPU.") |
| 146 | if attn_selector_config.use_mla: |
| 147 | amx_available = ( |
| 148 | cls.get_cpu_architecture() == CpuArchEnum.X86 |
| 149 | and torch.cpu._is_amx_tile_supported() |
| 150 | ) |
| 151 | if amx_available and selected_backend != AttentionBackendEnum.CPU_MLA: |
| 152 | # Prefer AMX when available, unless CPU_MLA was explicitly requested. |
| 153 | if ( |
| 154 | selected_backend |
| 155 | and selected_backend != AttentionBackendEnum.AMX_MLA |
| 156 | ): |
| 157 | logger.info("Cannot use %s backend on CPU.", selected_backend) |
| 158 | logger.info_once("Using %s backend.", AttentionBackendEnum.AMX_MLA.name) |
| 159 | return AttentionBackendEnum.AMX_MLA.get_path() |
| 160 | # Reference MLA implementation on CPU. Performance is not the |
| 161 | # goal here; the backend simply wires the CPU decode kernel |
| 162 | # (`mla_decode_kvcache`) and an SDPA-based prefill together with |
| 163 | # the shared MLA scaffolding so that DeepSeek-style models can |
| 164 | # execute on CPU. |
| 165 | if selected_backend and selected_backend not in ( |
| 166 | AttentionBackendEnum.CPU_MLA, |
| 167 | AttentionBackendEnum.AMX_MLA, |
| 168 | ): |
| 169 | logger.info("Cannot use %s backend on CPU.", selected_backend) |
| 170 | logger.info_once("Using %s backend.", AttentionBackendEnum.CPU_MLA.name) |
| 171 | return AttentionBackendEnum.CPU_MLA.get_path() |
| 172 | if selected_backend and selected_backend != AttentionBackendEnum.CPU_ATTN: |
| 173 | logger.info("Cannot use %s backend on CPU.", selected_backend) |
| 174 | return AttentionBackendEnum.CPU_ATTN.get_path() |
| 175 | |
| 176 | @classmethod |
| 177 | def get_device_total_memory(cls, device_id: int = 0) -> int: |
| 178 | meminfo = get_memory_node_info(device_id) |
| 179 | |
| 180 | return meminfo.total_memory |
| 181 | |
| 182 | @classmethod |
| 183 | def set_device(cls, device: torch.device) -> None: |
| 184 | """Set the device for the current platform.""" |
| 185 | torch.cpu.set_device(device) |
| 186 | |
| 187 | @classmethod |
| 188 | def manual_seed_all(cls, seed: int) -> None: |
| 189 | pass |
| 190 | |
| 191 | @classmethod |
| 192 | def inference_mode(cls): |
| 193 | return torch.no_grad() |
| 194 | |
| 195 | @classmethod |
| 196 | def check_and_update_config(cls, vllm_config: VllmConfig) -> None: |
| 197 | model_config = vllm_config.model_config |
| 198 | |
| 199 | if model_config is not None: |
| 200 | model_config.disable_cascade_attn = True |
| 201 | |
| 202 | # Import lazily: vllm.triton_utils imports vllm.platforms.current_platform, |
| 203 | # which is still being resolved while this platform class is loading. |
| 204 | from vllm.triton_utils import HAS_TRITON |
| 205 | |
| 206 | if cls.get_cpu_architecture() == CpuArchEnum.X86 and not HAS_TRITON: |
| 207 | logger.warning_once( |
| 208 | "Triton is not installed. triton-cpu is expected on x86 CPUs." |
| 209 | ) |
| 210 | |
| 211 | cache_config = vllm_config.cache_config |
| 212 | |
| 213 | is_deepseek_v4 = ( |
| 214 | model_config is not None |
| 215 | and getattr(model_config.hf_config, "model_type", None) == "deepseek_v4" |
| 216 | ) |
| 217 | |
| 218 | # The CPU MLA decode kernel only compiles with block_size=16 today |
| 219 | # (see csrc/cpu/mla_decode.cpp). If the model uses MLA we override |
| 220 | # the default block size regardless of user preference to avoid a |
| 221 | # runtime kernel dispatch failure. AMX MLA has no such constraint |
| 222 | # (same AMX-available condition as get_attn_backend_cls), so it's |
| 223 | # excluded from this override. DeepSeek-V4 is also excluded here and |
| 224 | # handled in its own branch below: its sparse-MLA and indexer |
| 225 | # backends declare block_size=256 as their only supported kernel |
| 226 | # block size (DeepseekV4SparseMLABackend/DeepseekV4IndexerBackend), |
| 227 | # so it needs the same override-regardless-of-preference treatment |
| 228 | # as CPU MLA, just with a different value. |
| 229 | cpu_mla_enabled = ( |
| 230 | not is_deepseek_v4 |
| 231 | and model_config is not None |
| 232 | and getattr(model_config, "use_mla", False) |
| 233 | ) |
| 234 | amx_mla_enabled = ( |
| 235 | cpu_mla_enabled |
| 236 | and cls.get_cpu_architecture() == CpuArchEnum.X86 |
| 237 | and torch.cpu._is_amx_tile_supported() |
| 238 | and vllm_config.attention_config.backend != AttentionBackendEnum.CPU_MLA |
| 239 | ) |
| 240 | reference_cpu_mla_enabled = cpu_mla_enabled and not amx_mla_enabled |
| 241 | # DeepSeek-V4's CPU attention/indexer kernels |
| 242 | # (csrc/cpu/sgl-kernels/{flash_mla,store_cache,compressor, |
| 243 | # paged_mqa_logits,topk}.cpp) are AMX-kernel-backed and built on the |
| 244 | # same paged/position-indexed conventions as the GPU/XPU backends |
| 245 | # (block_table/slot_mapping addressing throughout, chunk metadata |
| 246 | # that already carries per-token causal offsets for partial/extend |
| 247 | # continuation) -- so they support chunked prefill and prefix |
| 248 | # caching the same way. `amx_mla_enabled` above deliberately |
| 249 | # excludes DeepSeek-V4 (it has its own block-size requirements, |
| 250 | # unrelated to chunked-prefill support), so it can't be reused here. |
| 251 | amx_mla_or_dsv4_enabled = amx_mla_enabled or ( |
| 252 | is_deepseek_v4 |
| 253 | and cls.get_cpu_architecture() == CpuArchEnum.X86 |
| 254 | and torch.cpu._is_amx_tile_supported() |
| 255 | ) |
| 256 | if reference_cpu_mla_enabled: |
| 257 | if cache_config.user_specified_block_size and cache_config.block_size != 16: |
| 258 | logger.warning( |
| 259 | "CPU MLA backend requires block_size=16, overriding " |
| 260 | "user-specified block_size=%s.", |
| 261 | cache_config.block_size, |
| 262 | ) |
| 263 | cache_config.block_size = 16 |
| 264 | elif is_deepseek_v4: |
| 265 | if ( |
| 266 | cache_config.user_specified_block_size |
| 267 | and cache_config.block_size != 256 |
| 268 | ): |
| 269 | logger.warning( |
| 270 | "DeepSeek-V4 CPU backend requires block_size=256, " |
| 271 | "overriding user-specified block_size=%s.", |
| 272 | cache_config.block_size, |
| 273 | ) |
| 274 | cache_config.block_size = 256 |
| 275 | elif not cache_config.user_specified_block_size: |
| 276 | cache_config.block_size = 128 |
| 277 | |
| 278 | if not reference_cpu_mla_enabled and cache_config.block_size % 32 != 0: |
| 279 | logger.warning( |
| 280 | "CPU backend prefers block_size is multiples of 32, " |
| 281 | "otherwise the performance is not optimized." |
| 282 | ) |
| 283 | |
| 284 | # Accelerated GDN uses AMX tiles or AVX-512BF16 VDPBF16PS. |
| 285 | if ( |
| 286 | torch.cpu._is_avx512_bf16_supported() |
| 287 | and cache_config.mamba_ssm_cache_dtype != "float32" |
| 288 | ): |
| 289 | mamba_backend = _cpu_mamba_backend(vllm_config) |
| 290 | if ( |
| 291 | cache_config.mamba_ssm_cache_dtype in ("float16", "bfloat16") |
| 292 | and mamba_backend == "gdn" |
| 293 | ): |
| 294 | logger.info( |
| 295 | "Using %s SSM state storage for the CPU accelerated GDN backend.", |
| 296 | cache_config.mamba_ssm_cache_dtype, |
| 297 | ) |
| 298 | else: |
| 299 | cache_config.mamba_ssm_cache_dtype = "float32" |
| 300 | logger.warning( |
| 301 | "Reset SSM cache type to float32 for accelerated GDN mamba " |
| 302 | "backend '%s'.", |
| 303 | mamba_backend, |
| 304 | ) |
| 305 | |
| 306 | # Lagecy setting |
| 307 | env_key = "VLLM_CPU_KVCACHE_SPACE" |
| 308 | if env_key in os.environ and os.environ[env_key] != "": |
| 309 | kv_cache_space = int(os.environ[env_key]) |
| 310 | cache_config.kv_cache_memory_bytes = kv_cache_space * GiB_bytes |
| 311 | |
| 312 | scheduler_config = vllm_config.scheduler_config |
| 313 | # async scheduling is not required on CPU |
| 314 | scheduler_config.async_scheduling = False |
| 315 | |
| 316 | parallel_config = vllm_config.parallel_config |
| 317 | if ( |
| 318 | os.environ.get("VLLM_ENABLE_V1_MULTIPROCESSING", "1") == "1" |
| 319 | and parallel_config.distributed_executor_backend == "uni" |
| 320 | ): |
| 321 | # OMP requires the MP executor to function correctly, UniProc |
| 322 | # is not supported as it is not possible to set the OMP |
| 323 | # environment correctly |
| 324 | parallel_config.distributed_executor_backend = "mp" |
| 325 | |
| 326 | if parallel_config.worker_cls == "auto": |
| 327 | parallel_config.worker_cls = "vllm.v1.worker.cpu_worker.CPUWorker" |
| 328 | # Disable DBO |
| 329 | if parallel_config.enable_dbo: |
| 330 | logger.warning_once("Dual-Batch Overlap is not supported on CPU, disabled.") |
| 331 | parallel_config.enable_dbo = False |
| 332 | |
| 333 | # Note: workaround for v1 gpu_model_runner |
| 334 | from vllm.config import CompilationMode |
| 335 | |
| 336 | vllm_config.compilation_config.cudagraph_capture_sizes = [] |
| 337 | |
| 338 | compilation_config = vllm_config.compilation_config |
| 339 | if vllm_config.compilation_config.mode == CompilationMode.VLLM_COMPILE: |
| 340 | # Note: vLLM V1 is using PIECEWISE level compilation, which will |
| 341 | # take time to compile kernels just-in-time with the inductor |
| 342 | # backend. For CPU CI tests, most of them are executed fast and |
| 343 | # compilations consume too much time, even with torch compile |
| 344 | # cache. So use VLLM_CPU_CI_ENV to indicate the CI environment, |
| 345 | # and just execute model with dynamo + eager mode to save time. |
| 346 | # VLLM_CPU_CI_ENV is only used as an internal variable. |
| 347 | backend = "eager" if envs.VLLM_CPU_CI_ENV else "inductor" |
| 348 | |
| 349 | compilation_config.mode = CompilationMode.DYNAMO_TRACE_ONCE |
| 350 | compilation_config.backend = backend |
| 351 | compilation_config.inductor_compile_config.update( |
| 352 | { |
| 353 | "dce": True, |
| 354 | "size_asserts": False, |
| 355 | "nan_asserts": False, |
| 356 | "epilogue_fusion": True, |
| 357 | "cpp.dynamic_threads": True, |
| 358 | } |
| 359 | ) |
| 360 | compilation_config.ir_enable_torch_wrap = False |
| 361 | |
| 362 | if vllm_config.lora_config is not None: |
| 363 | compilation_config.mode = CompilationMode.NONE |
| 364 | |
| 365 | if ( |
| 366 | cls.get_cpu_architecture() == CpuArchEnum.ARM |
| 367 | and "+gelu" not in compilation_config.custom_ops |
| 368 | and "-gelu" not in compilation_config.custom_ops |
| 369 | ): |
| 370 | compilation_config.custom_ops.append("+gelu") |
| 371 | if ( |
| 372 | cls.get_cpu_architecture() == CpuArchEnum.ARM |
| 373 | and "+gelu_tanh" not in compilation_config.custom_ops |
| 374 | and "-gelu_tanh" not in compilation_config.custom_ops |
| 375 | ): |
| 376 | compilation_config.custom_ops.append("+gelu_tanh") |
| 377 | if ( |
| 378 | cls.get_cpu_architecture() == CpuArchEnum.ARM |
| 379 | and "+gelu_and_mul" not in compilation_config.custom_ops |
| 380 | and "-gelu_and_mul" not in compilation_config.custom_ops |
| 381 | ): |
| 382 | compilation_config.custom_ops.append("+gelu_and_mul") |
| 383 | |
| 384 | vllm_config.profiler_config.torch_profiler_dump_cuda_time_total = False |
| 385 | |
| 386 | assert vllm_config.device_config.device_type == "cpu" |
| 387 | |
| 388 | # |
| 389 | # Environment variables for CPU executor |
| 390 | # |
| 391 | |
| 392 | os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn" |
| 393 | |
| 394 | # Note: to avoid the error 'nthreads cannot be larger than environment |
| 395 | # variable "NUMEXPR_MAX_THREADS" (64)'. |
| 396 | os.environ["NUMEXPR_MAX_THREADS"] = str(get_max_threads()) |
| 397 | |
| 398 | # Disable torch async compiling which won't work with daemonic processes |
| 399 | os.environ["TORCHINDUCTOR_COMPILE_THREADS"] = "1" |
| 400 | |
| 401 | # Disable multi-stream for shared experts as no Stream on CPU |
| 402 | os.environ["VLLM_DISABLE_SHARED_EXPERTS_STREAM"] = "1" |
| 403 | |
| 404 | # Avoid inductor generates num_thread() and breaks the thread binding |
| 405 | os.environ["TORCHINDUCTOR_CPP_DYNAMIC_THREADS"] = "1" |
| 406 | |
| 407 | # NIXL's Mamba descriptors require DS conv state storage. Select it |
| 408 | # before cache shapes are created, while preserving an explicit layout. |
| 409 | conv_state_layout_env = "VLLM_SSM_CONV_STATE_LAYOUT" |
| 410 | if conv_state_layout_env not in os.environ: |
| 411 | kv_transfer_config = vllm_config.kv_transfer_config |
| 412 | uses_nixl = kv_transfer_config is not None and any( |
| 413 | kv_transfer_config.has_connector(name) |
| 414 | for name in ( |
| 415 | "NixlConnector", |
| 416 | "NixlPullConnector", |
| 417 | "NixlPushConnector", |
| 418 | ) |
| 419 | ) |
| 420 | if uses_nixl: |
| 421 | os.environ[conv_state_layout_env] = "DS" |
| 422 | elif torch.cpu._is_avx512_bf16_supported(): |
| 423 | os.environ[conv_state_layout_env] = "SD" |
| 424 | |
| 425 | ld_preload_str = os.getenv("LD_PRELOAD", "") |
| 426 | cpu_architecture = Platform.get_cpu_architecture() |
| 427 | |
| 428 | if ( |
| 429 | platform.system() == "Linux" |
| 430 | and cpu_architecture |
| 431 | in ( |
| 432 | CpuArchEnum.ARM, |
| 433 | CpuArchEnum.POWERPC, |
| 434 | CpuArchEnum.X86, |
| 435 | CpuArchEnum.S390X, |
| 436 | ) |
| 437 | and not ( |
| 438 | "libomp" in ld_preload_str |
| 439 | or "libgomp" in ld_preload_str |
| 440 | or "libiomp" in ld_preload_str |
| 441 | ) |
| 442 | ): |
| 443 | # We need to LD_PRELOAD PyTorch's libgomp, otherwise only |
| 444 | # one core will be properly utilized when we thread-bind |
| 445 | # See: https://github.com/vllm-project/vllm/issues/27369 |
| 446 | # TODO: Remove once: |
| 447 | # https://github.com/pytorch/pytorch/issues/166087 is fixed |
| 448 | |
| 449 | # We need to find the location of PyTorch's libgomp |
| 450 | torch_pkg = os.path.dirname(torch.__file__) |
| 451 | site_root = os.path.dirname(torch_pkg) |
| 452 | # Search both torch.libs and torch/lib - See: |
| 453 | # https://github.com/vllm-project/vllm/issues/30470 |
| 454 | torch_libs_paths = [ |
| 455 | os.path.join(site_root, "torch.libs"), |
| 456 | os.path.join(torch_pkg, "lib"), |
| 457 | ] |
| 458 | pytorch_libgomp_so_candidates = [] |
| 459 | for torch_libs in torch_libs_paths: |
| 460 | pytorch_libgomp_so_candidates.extend( |
| 461 | glob.glob(os.path.join(torch_libs, "libgomp*.so*")) |
| 462 | ) |
| 463 | if pytorch_libgomp_so_candidates: |
| 464 | pytorch_libgomp_so = pytorch_libgomp_so_candidates[0] |
| 465 | if ld_preload_str: |
| 466 | ld_preload_str += ":" |
| 467 | ld_preload_str += pytorch_libgomp_so |
| 468 | os.environ["LD_PRELOAD"] = ld_preload_str |
| 469 | |
| 470 | # LD_PRELOAD libtcmalloc, bundled under vllm/libs to reduce |
| 471 | # memory allocation overhead |
| 472 | if ( |
| 473 | platform.system() == "Linux" |
| 474 | and cpu_architecture |
| 475 | in (CpuArchEnum.ARM, CpuArchEnum.X86, CpuArchEnum.S390X) |
| 476 | and "libtcmalloc" not in ld_preload_str |
| 477 | ): |
| 478 | vllm_pkg = os.path.dirname(os.path.dirname(__file__)) |
| 479 | tcmalloc_so = None |
| 480 | for pattern in ("libtcmalloc_minimal*.so*", "libtcmalloc.so*"): |
| 481 | tcmalloc_so_candidates = glob.glob( |
| 482 | os.path.join(vllm_pkg, "libs", pattern) |
| 483 | ) |
| 484 | if tcmalloc_so_candidates: |
| 485 | tcmalloc_so = tcmalloc_so_candidates[0] |
| 486 | break |
| 487 | |
| 488 | if tcmalloc_so is not None: |
| 489 | if ld_preload_str: |
| 490 | ld_preload_str = f"{tcmalloc_so}:{ld_preload_str}" |
| 491 | else: |
| 492 | ld_preload_str = tcmalloc_so |
| 493 | os.environ["LD_PRELOAD"] = ld_preload_str |
| 494 | |
| 495 | os.environ["LOCAL_WORLD_SIZE"] = str( |
| 496 | vllm_config.parallel_config.tensor_parallel_size |
| 497 | ) |
| 498 | |
| 499 | if ( |
| 500 | model_config is not None |
| 501 | and model_config.use_mla |
| 502 | and not amx_mla_or_dsv4_enabled |
| 503 | ): |
| 504 | logger.info_once( |
| 505 | "MLA is enabled on a non-GPU platform; forcing chunked " |
| 506 | "prefill and prefix caching to be disabled." |
| 507 | ) |
| 508 | vllm_config.scheduler_config.enable_chunked_prefill = False |
| 509 | vllm_config.cache_config.enable_prefix_caching = False |
| 510 | vllm_config.scheduler_config.max_num_batched_tokens = max( |
| 511 | vllm_config.model_config.max_model_len, |
| 512 | vllm_config.scheduler_config.DEFAULT_MAX_NUM_BATCHED_TOKENS, |
| 513 | ) |
| 514 | |
| 515 | @classmethod |
| 516 | def update_block_size_for_backend(cls, vllm_config: "VllmConfig") -> None: |
| 517 | model_config = vllm_config.model_config |
| 518 | if model_config is None or not model_config.is_hybrid: |
| 519 | return |
| 520 | |
| 521 | # reconcile attention and mamba page sizes |
| 522 | backend_classes = cls._find_non_ssm_backends(vllm_config) |
| 523 | if not backend_classes: |
| 524 | return |
| 525 | |
| 526 | cls._align_hybrid_block_size(vllm_config, backend_classes[0]) |
| 527 | |
| 528 | @classmethod |
| 529 | def discover_numa_topology(cls) -> list[list[int]]: |
| 530 | """Discover NUMA topology and keep the last physical core of each numa |
| 531 | into one core group list for nixl start_kv_load() |
| 532 | """ |
| 533 | SYS_NODE = "/sys/devices/system/node" |
| 534 | SYS_CPU = "/sys/devices/system/cpu" |
| 535 | |
| 536 | if not (os.path.exists(SYS_NODE) and os.path.exists(SYS_CPU)): |
| 537 | return [] |
| 538 | |
| 539 | use_highest_sibling = cls.get_cpu_architecture() == CpuArchEnum.X86 |
| 540 | core_rsv_for_kv = [] |
| 541 | for node in os.listdir(SYS_NODE): |
| 542 | if not node.startswith("node") or not node[4:].isdigit(): |
| 543 | continue |
| 544 | node_path = f"{SYS_NODE}/{node}" |
| 545 | |
| 546 | seen_phys = set() |
| 547 | for cpu in os.listdir(node_path): |
| 548 | if not cpu.startswith("cpu") or not cpu[3:].isdigit(): |
| 549 | continue |
| 550 | |
| 551 | cpu_id = int(cpu[3:]) |
| 552 | # thread_siblings based on cpu_id |
| 553 | path = f"{SYS_CPU}/cpu{cpu_id}/topology/thread_siblings_list" |
| 554 | |
| 555 | if os.path.exists(path): |
| 556 | try: |
| 557 | with open(path) as f: |
| 558 | s = f.read() |
| 559 | cpus: list[int] = [] |
| 560 | for part in s.strip().split(","): |
| 561 | if "-" in part: |
| 562 | a, b = map(int, part.split("-")) |
| 563 | cpus.extend(range(a, b + 1)) |
| 564 | else: |
| 565 | cpus.append(int(part)) |
| 566 | siblings = cpus if cpus else [cpu_id] |
| 567 | except (OSError, ValueError): |
| 568 | siblings = [cpu_id] |
| 569 | else: |
| 570 | siblings = [cpu_id] |
| 571 | |
| 572 | phys = max(siblings) if use_highest_sibling else min(siblings) |
| 573 | |
| 574 | if phys not in seen_phys: |
| 575 | seen_phys.add(phys) |
| 576 | |
| 577 | if len(seen_phys) > 0: |
| 578 | core_rsv_for_kv.append(list(seen_phys)) |
| 579 | |
| 580 | return core_rsv_for_kv |
| 581 | |
| 582 | @classmethod |
| 583 | def is_pin_memory_available(cls) -> bool: |
| 584 | return False |
| 585 | |
| 586 | @classmethod |
| 587 | def get_punica_wrapper(cls) -> str: |
| 588 | return "vllm.lora.punica_wrapper.punica_cpu.PunicaWrapperCPU" |
| 589 | |
| 590 | @classmethod |
| 591 | def get_device_communicator_cls(cls) -> str: |
| 592 | """Get device specific communicator class for distributed communication.""" |
| 593 | return "vllm.distributed.device_communicators.cpu_communicator.CpuCommunicator" # noqa |
| 594 | |
| 595 | @classmethod |
| 596 | def supports_structured_output(cls) -> bool: |
| 597 | return True |
| 598 | |
| 599 | @classmethod |
| 600 | def opaque_attention_op(cls) -> bool: |
| 601 | return True |
| 602 | |
| 603 | @classmethod |
| 604 | def support_hybrid_kv_cache(cls) -> bool: |
| 605 | return True |
| 606 | |
| 607 | @classmethod |
| 608 | def num_compute_units(cls, device_id: int = 0) -> int: |
| 609 | return torch.get_num_threads() |
| 610 | |
| 611 | @classmethod |
| 612 | def import_kernels(cls) -> None: |
| 613 | if Platform.get_cpu_architecture() in (CpuArchEnum.X86,): |
| 614 | # Note: The lib name is _C_AVX2/AVX512, but the module name is _C. |
| 615 | # This will cause a exception "dynamic module does define |
| 616 | # module export function". But the library is imported |
| 617 | # successfully. So ignore the exception for now, until we find |
| 618 | # a solution. |
| 619 | ignored_msg = "dynamic module does not define module export function" |
| 620 | if torch.cpu._is_avx512_supported(): |
| 621 | if torch.cpu._is_avx512_bf16_supported(): |
| 622 | try: |
| 623 | import vllm._C # noqa: F401 |
| 624 | except ImportError as e: |
| 625 | logger.warning_once( |
| 626 | "Failed to import from vllm._C: %s", repr(e) |
| 627 | ) |
| 628 | else: |
| 629 | try: |
| 630 | import vllm._C_AVX512 # noqa: F401 |
| 631 | except ImportError as e: |
| 632 | if ignored_msg not in e.msg: |
| 633 | logger.warning_once( |
| 634 | "Failed to import from vllm._C_AVX512: %s", repr(e) |
| 635 | ) |
| 636 | else: |
| 637 | try: |
| 638 | import vllm._C_AVX2 # noqa: F401 |
| 639 | except ImportError as e: |
| 640 | if ignored_msg not in e.msg: |
| 641 | logger.warning_once( |
| 642 | "Failed to import from vllm._C_AVX2: %s", repr(e) |
| 643 | ) |
| 644 | else: |
| 645 | try: |
| 646 | import vllm._C # noqa: F401 |
| 647 | except ImportError as e: |
| 648 | logger.warning_once("Failed to import from vllm._C: %s", repr(e)) |
| 649 | |
| 650 | @classmethod |
| 651 | def register_triton_kernel_overrides(cls) -> None: |
| 652 | """Fallback C++ implementations for core Triton kernels. |
| 653 | |
| 654 | Used when the Triton-CPU backend is unavailable. When Triton-CPU is |
| 655 | present, the Triton kernels can run natively and the overrides |
| 656 | would only shadow them, so registration is skipped. |
| 657 | """ |
| 658 | from vllm.triton_utils import HAS_TRITON |
| 659 | |
| 660 | if HAS_TRITON: |
| 661 | return |
| 662 | |
| 663 | from vllm.triton_utils.dispatcher import register_kernels |
| 664 | from vllm.utils import cpu_triton_utils as cpu_tl |
| 665 | |
| 666 | register_kernels( |
| 667 | { |
| 668 | "vllm.v1.worker.block_table.ComputeSlotMappingKernel.kernel": ( |
| 669 | cpu_tl._compute_slot_mapping_kernel_impl |
| 670 | ), |
| 671 | "vllm.v1.spec_decode.utils.eagle_step_slot_mapping_metadata_kernel": ( |
| 672 | cpu_tl._eagle_step_slot_mapping_metadata_kernel_impl |
| 673 | ), |
| 674 | "vllm.v1.spec_decode.utils.eagle_prepare_inputs_padded_kernel": ( |
| 675 | cpu_tl._eagle_prepare_inputs_padded_kernel_impl |
| 676 | ), |
| 677 | "vllm.v1.spec_decode.utils.eagle_prepare_next_token_padded_kernel": ( |
| 678 | cpu_tl._eagle_prepare_next_token_padded_kernel_impl |
| 679 | ), |
| 680 | "vllm.v1.spec_decode.utils.copy_and_expand_eagle_inputs_kernel": ( |
| 681 | cpu_tl._copy_and_expand_eagle_inputs_kernel_impl |
| 682 | ), |
| 683 | "vllm.v1.spec_decode.utils.copy_and_expand_dflash_inputs_kernel": ( |
| 684 | cpu_tl._copy_and_expand_dflash_inputs_kernel_impl |
| 685 | ), |
| 686 | "vllm.v1.sample.rejection_sampler.rejection_greedy_sample_kernel": ( |
| 687 | cpu_tl._rejection_greedy_sample_kernel_impl |
| 688 | ), |
| 689 | "vllm.v1.sample.rejection_sampler.rejection_random_sample_kernel": ( |
| 690 | cpu_tl._rejection_random_sample_kernel_impl |
| 691 | ), |
| 692 | "vllm.v1.sample.rejection_sampler.expand_kernel": ( |
| 693 | cpu_tl._expand_kernel_impl |
| 694 | ), |
| 695 | "vllm.v1.sample.rejection_sampler.sample_recovered_tokens_kernel": ( |
| 696 | cpu_tl._sample_recovered_tokens_kernel_impl |
| 697 | ), |
| 698 | "vllm.v1.worker.mamba_utils.batch_memcpy_kernel": ( |
| 699 | cpu_tl._batch_memcpy_impl |
| 700 | ), |
| 701 | } |
| 702 | ) |
| 703 | |
| 704 | @classmethod |
| 705 | def pack_kv_cache( |
| 706 | cls, |
| 707 | kv_cache: torch.Tensor, |
| 708 | indices: torch.Tensor, |
| 709 | ) -> None: |
| 710 | """Rewrite the kv cache shape for the current platform.""" |
| 711 | # Import lazily: cpu_attn pulls in _custom_ops, which needs a fully |
| 712 | # initialized vllm.platforms (avoid circular import while CpuPlatform loads). |
| 713 | from vllm._custom_ops import cpu_attn_reshape_and_cache |
| 714 | from vllm.v1.attention.backends.cpu_attn import _get_attn_isa |
| 715 | |
| 716 | # MLA uses a single latent cache of shape [N, block_size, head_size], |
| 717 | # so the classic key/value split does not apply. The MLA backend |
| 718 | # writes the cache itself via `concat_and_cache_mla` inside |
| 719 | # `do_kv_cache_update`, so there is nothing to pack here. |
| 720 | if kv_cache.dim() == 3: |
| 721 | return |
| 722 | |
| 723 | num_blocks, num_kv_heads, block_size, fused_head_size = kv_cache.shape |
| 724 | head_size = fused_head_size // 2 |
| 725 | |
| 726 | # Fused path used by heterogeneous NIXL CPU_ATTN post-processing. |
| 727 | blocks_to_update = kv_cache.index_select(0, indices) |
| 728 | key = blocks_to_update[..., :head_size] |
| 729 | value = blocks_to_update[..., head_size:] |
| 730 | |
| 731 | key_cache, value_cache = kv_cache.view( |
| 732 | num_blocks, num_kv_heads, block_size * 2, head_size |
| 733 | ).chunk(2, dim=2) |
| 734 | |
| 735 | dtype = key.dtype |
| 736 | # For CPU_ATTN, the shape is [N, num_kv_heads, block_size, head_size] |
| 737 | key = key.permute(0, 2, 1, 3).flatten(0, 1) |
| 738 | value = value.permute(0, 2, 1, 3).flatten(0, 1) |
| 739 | |
| 740 | isa = _get_attn_isa(dtype, block_size, head_size) |
| 741 | block_offsets = torch.arange(block_size, device="cpu", dtype=torch.long) |
| 742 | num_blocks = indices.numel() |
| 743 | slot_mapping = ( |
| 744 | block_offsets.reshape(1, block_size) |
| 745 | + indices.reshape(num_blocks, 1) * block_size |
| 746 | ).flatten() |
| 747 | if key_cache.dtype == torch.uint8: |
| 748 | raise NotImplementedError( |
| 749 | "FP8 KV cache is not yet supported with KV transfer on CPU" |
| 750 | ) |
| 751 | cpu_attn_reshape_and_cache( |
| 752 | key, |
| 753 | value, |
| 754 | key_cache, |
| 755 | value_cache, |
| 756 | slot_mapping, |
| 757 | isa, |
| 758 | ) |
| 759 | |
| 760 | @classmethod |
| 761 | def get_current_memory_usage( |
| 762 | cls, device: torch.types.Device | None = None |
| 763 | ) -> float: |
| 764 | allowed_mem_node_list = get_visible_memory_node() |
| 765 | mem_status_list = [get_memory_node_info(i) for i in allowed_mem_node_list] |
| 766 | memory_usage = 0 |
| 767 | for s in mem_status_list: |
| 768 | memory_usage += s.total_memory - s.available_memory |
| 769 | |
| 770 | return memory_usage |