| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """A GPU worker class.""" |
| 4 | |
| 5 | import gc |
| 6 | import os |
| 7 | import time |
| 8 | from collections.abc import Callable |
| 9 | from contextlib import AbstractContextManager, contextmanager, nullcontext |
| 10 | from datetime import timedelta |
| 11 | from fnmatch import filter as fnmatch_filter |
| 12 | from types import NoneType |
| 13 | from typing import TYPE_CHECKING, Any, cast |
| 14 | |
| 15 | import numpy as np |
| 16 | import torch |
| 17 | import torch.nn as nn |
| 18 | |
| 19 | import vllm.envs as envs |
| 20 | from vllm.config import CUDAGraphMode, VllmConfig, set_current_vllm_config |
| 21 | from vllm.config.compilation import CompilationMode |
| 22 | from vllm.device_allocator import get_mem_allocator_instance |
| 23 | from vllm.distributed import ( |
| 24 | ensure_model_parallel_initialized, |
| 25 | init_distributed_environment, |
| 26 | set_custom_all_reduce, |
| 27 | ) |
| 28 | from vllm.distributed.ec_transfer import ( |
| 29 | ensure_ec_transfer_initialized, |
| 30 | ensure_ec_transfer_shutdown, |
| 31 | get_ec_transfer, |
| 32 | has_ec_transfer, |
| 33 | ) |
| 34 | from vllm.distributed.eplb.eplb_utils import override_envs_for_eplb |
| 35 | from vllm.distributed.kv_transfer import ( |
| 36 | ensure_kv_transfer_initialized, |
| 37 | ensure_kv_transfer_shutdown, |
| 38 | get_kv_transfer_group, |
| 39 | has_kv_transfer_group, |
| 40 | ) |
| 41 | from vllm.distributed.kv_transfer.kv_connector.v1.base import ( |
| 42 | KVConnectorHandshakeMetadata, |
| 43 | ) |
| 44 | from vllm.distributed.parallel_state import ( |
| 45 | Handle, |
| 46 | checkpoint_prepare_distributed_state, |
| 47 | checkpoint_restore_distributed_state, |
| 48 | get_pcp_group, |
| 49 | get_pp_group, |
| 50 | get_tp_group, |
| 51 | resume_device_comms, |
| 52 | suspend_device_comms, |
| 53 | ) |
| 54 | from vllm.distributed.weight_transfer import ( |
| 55 | WeightTransferEngine, |
| 56 | WeightTransferEngineFactory, |
| 57 | ) |
| 58 | from vllm.logger import init_logger |
| 59 | from vllm.lora.request import LoRARequest |
| 60 | from vllm.model_executor.model_loader import get_model_loader |
| 61 | from vllm.model_executor.model_loader.weight_checksum import ( |
| 62 | compute_tensor_digests, |
| 63 | zero_weights, |
| 64 | ) |
| 65 | from vllm.model_executor.warmup.kernel_warmup import kernel_warmup |
| 66 | from vllm.multimodal.gpu_ipc_memory import reserve_mm_ipc_gpu_memory |
| 67 | from vllm.platforms import current_platform |
| 68 | from vllm.profiler.wrapper import ( |
| 69 | create_graph_capture_profiler, |
| 70 | create_worker_profiler, |
| 71 | validate_worker_profiler_config, |
| 72 | ) |
| 73 | from vllm.sequence import IntermediateTensors |
| 74 | from vllm.tasks import SupportedTask |
| 75 | from vllm.tracing import instrument |
| 76 | from vllm.utils.gc_utils import freeze_gc_heap, maybe_attach_gc_debug_callback |
| 77 | from vllm.utils.gpu_sync_debug import enable_gpu_sync_check, with_gpu_sync_check |
| 78 | from vllm.utils.mem_constants import GiB_bytes |
| 79 | from vllm.utils.mem_utils import ( |
| 80 | MemoryProfilingResult, |
| 81 | MemorySnapshot, |
| 82 | format_gib, |
| 83 | memory_profiling, |
| 84 | ) |
| 85 | from vllm.utils.torch_utils import set_random_seed, set_torch_threads_for_runtime |
| 86 | from vllm.v1.attention.backends.utils import record_kv_cache_layout |
| 87 | from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput |
| 88 | from vllm.v1.kv_cache_interface import KVCacheConfig, KVCacheSpec |
| 89 | from vllm.v1.outputs import ( |
| 90 | AsyncModelRunnerOutput, |
| 91 | DraftTokenIds, |
| 92 | ModelRunnerOutput, |
| 93 | ) |
| 94 | from vllm.v1.utils import compute_iteration_details, report_usage_stats |
| 95 | from vllm.v1.worker.sentinel.gpu_worker_sentinel import WorkerSentinel |
| 96 | from vllm.v1.worker.startup_plan import ( |
| 97 | maybe_apply_startup_plan, |
| 98 | maybe_save_startup_plan, |
| 99 | ) |
| 100 | from vllm.v1.worker.utils import is_residual_scattered_for_sp |
| 101 | from vllm.v1.worker.worker_base import CompilationTimes, WorkerBase |
| 102 | from vllm.v1.worker.workspace import init_workspace_manager |
| 103 | |
| 104 | from ...model_executor.model_loader import TensorizerLoader |
| 105 | from .gpu.cudagraph_utils import has_compiled_submodule |
| 106 | from .gpu.warmup import warmup_kernels |
| 107 | from .utils import request_memory |
| 108 | |
| 109 | logger = init_logger(__name__) |
| 110 | |
| 111 | |
| 112 | def _num_workspace_lanes(vllm_config: VllmConfig, use_v2_model_runner: bool) -> int: |
| 113 | spec_config = vllm_config.speculative_config |
| 114 | return ( |
| 115 | 2 |
| 116 | if use_v2_model_runner and spec_config is not None and spec_config.use_dspark() |
| 117 | else 1 |
| 118 | ) |
| 119 | |
| 120 | |
| 121 | def maybe_rocm_profiling_fallback(profile_result: MemoryProfilingResult) -> int | None: |
| 122 | """Memory in bytes to size the KV cache from when profiling measured a |
| 123 | release on ROCm, or None to keep what profiling measured. |
| 124 | |
| 125 | The measurement is the drop in free device memory across startup, so it comes |
| 126 | out negative when anything else on the device released memory meanwhile. Torch |
| 127 | reserved is process local, so it stands in as a lower bound instead of refusing |
| 128 | to start. Kept to ROCm, where the AMD CI groups hit this, following #45490. |
| 129 | """ |
| 130 | if profile_result.total_consumed >= 0 or not current_platform.is_rocm(): |
| 131 | return None |
| 132 | |
| 133 | baseline = profile_result.before_create |
| 134 | after_profile = profile_result.after_profile |
| 135 | torch_reserved = max(after_profile.torch_memory - baseline.torch_memory, 0) |
| 136 | logger.warning( |
| 137 | "Free memory grew by %s GiB during profiling (initial %s GiB, current " |
| 138 | "%s GiB). Sizing the KV cache from the %s GiB torch reserved instead of " |
| 139 | "the profiling difference.", |
| 140 | format_gib(-profile_result.total_consumed), |
| 141 | format_gib(baseline.free_memory), |
| 142 | format_gib(after_profile.free_memory), |
| 143 | format_gib(torch_reserved), |
| 144 | ) |
| 145 | return torch_reserved |
| 146 | |
| 147 | |
| 148 | if TYPE_CHECKING: |
| 149 | from vllm.device_allocator.sleep_mode_backend import SleepModeBackend |
| 150 | from vllm.model_executor.model_loader.tensorizer import TensorizerConfig |
| 151 | from vllm.v1.worker.gpu.model_runner import GPUModelRunner as GPUModelRunnerV2 |
| 152 | from vllm.v1.worker.gpu_model_runner import GPUModelRunner |
| 153 | |
| 154 | |
| 155 | class AsyncIntermediateTensors(IntermediateTensors): |
| 156 | """IntermediateTensors with lazy comm synchronization""" |
| 157 | |
| 158 | def __init__( |
| 159 | self, |
| 160 | tensors: dict[str, torch.Tensor], |
| 161 | comm_handles: list[Handle] | None = None, |
| 162 | comm_postprocess: list[Callable[[], None]] | None = None, |
| 163 | ) -> None: |
| 164 | super().__init__(tensors) |
| 165 | self._comm_handles = comm_handles |
| 166 | self._comm_postprocess = comm_postprocess |
| 167 | self._comm_waited = False |
| 168 | |
| 169 | def wait_for_comm(self) -> None: |
| 170 | if self._comm_waited: |
| 171 | return |
| 172 | if self._comm_handles: |
| 173 | for handle in self._comm_handles: |
| 174 | handle.wait() |
| 175 | if self._comm_postprocess: |
| 176 | for fn in self._comm_postprocess: |
| 177 | fn() |
| 178 | self._comm_waited = True |
| 179 | |
| 180 | def __getattribute__(self, name: str): |
| 181 | # ensure `.tensors` is ready before use |
| 182 | if name == "tensors" and not object.__getattribute__(self, "_comm_waited"): |
| 183 | object.__getattribute__(self, "wait_for_comm")() |
| 184 | return object.__getattribute__(self, name) |
| 185 | |
| 186 | |
| 187 | class Worker(WorkerBase): |
| 188 | def __init__( |
| 189 | self, |
| 190 | vllm_config: VllmConfig, |
| 191 | local_rank: int, |
| 192 | rank: int, |
| 193 | distributed_init_method: str, |
| 194 | is_driver_worker: bool = False, |
| 195 | ): |
| 196 | super().__init__( |
| 197 | vllm_config=vllm_config, |
| 198 | local_rank=local_rank, |
| 199 | rank=rank, |
| 200 | distributed_init_method=distributed_init_method, |
| 201 | is_driver_worker=is_driver_worker, |
| 202 | ) |
| 203 | |
| 204 | # configure float32 matmul precision according to vLLM env. |
| 205 | precision = envs.VLLM_FLOAT32_MATMUL_PRECISION |
| 206 | torch.set_float32_matmul_precision(precision) |
| 207 | |
| 208 | from vllm.distributed.elastic_ep.elastic_execute import ElasticEPScalingExecutor |
| 209 | |
| 210 | self.elastic_ep_executor = ElasticEPScalingExecutor(self) |
| 211 | self.worker_sentinel: WorkerSentinel | None = None |
| 212 | if self.parallel_config.enable_fault_tolerance: |
| 213 | self.worker_sentinel = WorkerSentinel(worker=self) |
| 214 | # Buffers saved before sleep |
| 215 | self._sleep_saved_buffers: dict[str, torch.Tensor] = {} |
| 216 | self._sleep_saved_parameters: dict[str, torch.Tensor] = {} |
| 217 | self._sleep_saved_draft_buffers: dict[str, torch.Tensor] = {} |
| 218 | |
| 219 | # Weight transfer engine is created in `load_model` once the model |
| 220 | # is available, since the engine needs a reference to the model. |
| 221 | self.weight_transfer_engine: WeightTransferEngine | None = None |
| 222 | self._weight_update_active = False |
| 223 | self._weight_update_is_draft = False |
| 224 | |
| 225 | # Worker profiler. Enabled and configured through profiler_config. |
| 226 | # Profiler wrapper is created lazily in profile() when start is called, |
| 227 | # so we have all the information needed for proper trace naming. |
| 228 | self.profiler: Any | None = None |
| 229 | self.profiler_config = vllm_config.profiler_config |
| 230 | validate_worker_profiler_config(self.profiler_config) |
| 231 | |
| 232 | self.use_v2_model_runner = vllm_config.use_v2_model_runner |
| 233 | |
| 234 | # Device handles of the previous step's PP intermediate-tensor send. |
| 235 | self._pp_send_work: list[Handle] = [] |
| 236 | |
| 237 | # Resolved lazily on first sleep/wake; persists worker-process state. |
| 238 | self._sleep_mode_backend: SleepModeBackend | None = None |
| 239 | |
| 240 | @property |
| 241 | def sleep_mode_backend(self) -> "SleepModeBackend": |
| 242 | if self._sleep_mode_backend is None: |
| 243 | from vllm.device_allocator.sleep_mode_backend import ( |
| 244 | SleepModeBackendFactory, |
| 245 | ) |
| 246 | |
| 247 | self._sleep_mode_backend = SleepModeBackendFactory.create_backend( |
| 248 | self.vllm_config.model_config |
| 249 | ) |
| 250 | return self._sleep_mode_backend |
| 251 | |
| 252 | def _save_sleep_parameters(self, model: nn.Module) -> None: |
| 253 | patterns = self.model_config.sleep_preserve_parameter_names |
| 254 | if not patterns: |
| 255 | self._sleep_saved_parameters = {} |
| 256 | return |
| 257 | parameters = dict(model.named_parameters()) |
| 258 | names: set[str] = set() |
| 259 | for pattern in patterns: |
| 260 | matches = fnmatch_filter(parameters, pattern) |
| 261 | if not matches: |
| 262 | raise ValueError(f"No parameter matches sleep retention: {pattern}") |
| 263 | names.update(matches) |
| 264 | self._sleep_saved_parameters = { |
| 265 | name: param.detach().to("cpu") |
| 266 | for name, param in parameters.items() |
| 267 | if name in names and not param.is_cpu |
| 268 | } |
| 269 | |
| 270 | @torch.no_grad() |
| 271 | def _restore_sleep_parameters(self, model: nn.Module) -> None: |
| 272 | for name, value in self._sleep_saved_parameters.items(): |
| 273 | model.get_parameter(name).copy_(value) |
| 274 | self._sleep_saved_parameters.clear() |
| 275 | |
| 276 | def sleep(self, level: int = 1) -> None: |
| 277 | torch.accelerator.synchronize() |
| 278 | free_bytes_before_sleep = torch.accelerator.get_memory_info()[0] |
| 279 | |
| 280 | # Save the buffers before level 2 sleep |
| 281 | if level == 2: |
| 282 | model = self.model_runner.model |
| 283 | self._save_sleep_parameters(model) |
| 284 | self._sleep_saved_buffers = { |
| 285 | name: buffer.cpu().clone() for name, buffer in model.named_buffers() |
| 286 | } |
| 287 | draft = self.get_draft_model() |
| 288 | if draft is not None: |
| 289 | self._sleep_saved_draft_buffers = { |
| 290 | name: buffer.cpu().clone() for name, buffer in draft.named_buffers() |
| 291 | } |
| 292 | |
| 293 | self.sleep_mode_backend.suspend(level) |
| 294 | if self.vllm_config.model_config.enable_nccl_comm_suspend: |
| 295 | suspend_device_comms() |
| 296 | |
| 297 | torch.accelerator.synchronize() |
| 298 | deadline = time.monotonic() + (5.0 if current_platform.is_rocm() else 0) |
| 299 | while True: |
| 300 | free_bytes_after_sleep, total = torch.accelerator.get_memory_info() |
| 301 | freed_bytes = free_bytes_after_sleep - free_bytes_before_sleep |
| 302 | if freed_bytes >= 0 or time.monotonic() >= deadline: |
| 303 | break |
| 304 | time.sleep(0.1) |
| 305 | |
| 306 | used_bytes = total - free_bytes_after_sleep |
| 307 | assert freed_bytes >= 0, "Memory usage increased after sleeping." |
| 308 | logger.info( |
| 309 | "Sleep mode freed %s GiB memory, %s GiB memory is still in use.", |
| 310 | format_gib(freed_bytes), |
| 311 | format_gib(used_bytes), |
| 312 | ) |
| 313 | |
| 314 | def wake_up(self, tags: list[str] | None = None) -> None: |
| 315 | self.sleep_mode_backend.resume(tags) |
| 316 | if self.vllm_config.model_config.enable_nccl_comm_suspend: |
| 317 | resume_device_comms() |
| 318 | |
| 319 | # Restore the buffers after level 2 sleep |
| 320 | wake_weights = tags is None or "weights" in tags |
| 321 | if wake_weights and self._sleep_saved_parameters: |
| 322 | self._restore_sleep_parameters(self.model_runner.model) |
| 323 | if wake_weights and len(self._sleep_saved_buffers): |
| 324 | model = self.model_runner.model |
| 325 | for name, buffer in model.named_buffers(): |
| 326 | if name in self._sleep_saved_buffers: |
| 327 | buffer.data.copy_(self._sleep_saved_buffers[name].data) |
| 328 | self._sleep_saved_buffers = {} |
| 329 | |
| 330 | if wake_weights and len(self._sleep_saved_draft_buffers): |
| 331 | draft = self.get_draft_model() |
| 332 | if draft is not None: |
| 333 | for name, buffer in draft.named_buffers(): |
| 334 | if name in self._sleep_saved_draft_buffers: |
| 335 | buffer.data.copy_(self._sleep_saved_draft_buffers[name].data) |
| 336 | self._sleep_saved_draft_buffers = {} |
| 337 | |
| 338 | self.synchronize_device() |
| 339 | |
| 340 | def discard(self, tags: tuple[str, ...]) -> None: |
| 341 | self.sleep_mode_backend.discard(tags) |
| 342 | |
| 343 | def checkpoint_prepare(self) -> None: |
| 344 | checkpoint_prepare_distributed_state() |
| 345 | |
| 346 | def checkpoint_restore(self) -> None: |
| 347 | checkpoint_restore_distributed_state() |
| 348 | |
| 349 | def _weight_checksum_prefix(self) -> str: |
| 350 | # data_parallel_rank is 0 in dense workers; only the index is unique. |
| 351 | dp = self.parallel_config.data_parallel_index |
| 352 | pp = get_pp_group().rank_in_group |
| 353 | pcp = get_pcp_group().rank_in_group |
| 354 | tp = get_tp_group().rank_in_group |
| 355 | return f"dp{dp}:pp{pp}:pcp{pcp}:tp{tp}:" |
| 356 | |
| 357 | def compute_weight_checksums(self) -> dict[str, str]: |
| 358 | prefix = self._weight_checksum_prefix() |
| 359 | digests = compute_tensor_digests(self.model_runner.get_model()) |
| 360 | return {prefix + name: digest for name, digest in digests.items()} |
| 361 | |
| 362 | def reset_weights(self) -> None: |
| 363 | zero_weights(self.model_runner.get_model()) |
| 364 | |
| 365 | def _maybe_get_memory_pool_context(self, tag: str) -> AbstractContextManager: |
| 366 | if ( |
| 367 | current_platform.is_cuda_alike() |
| 368 | and not self.vllm_config.model_config.enable_cumem_allocator |
| 369 | ): |
| 370 | return nullcontext() |
| 371 | |
| 372 | if ( |
| 373 | current_platform.is_xpu() |
| 374 | and not self.vllm_config.model_config.enable_sleep_mode |
| 375 | ): |
| 376 | return nullcontext() |
| 377 | |
| 378 | if current_platform.is_cpu(): |
| 379 | return nullcontext() |
| 380 | |
| 381 | allocator = get_mem_allocator_instance() |
| 382 | if tag == "weights": |
| 383 | assert allocator.get_current_usage() == 0, ( |
| 384 | "CuMem allocator can only be used for one instance per process." |
| 385 | ) |
| 386 | return allocator.use_memory_pool(tag=tag) |
| 387 | |
| 388 | @contextmanager |
| 389 | def _scoped_allocator_max_split(self, max_split_size_mb: int): |
| 390 | """Limit CUDA/ROCm allocator splitting, restoring settings on exit.""" |
| 391 | if ( |
| 392 | not current_platform.is_cuda_alike() |
| 393 | or torch.cuda.memory.get_allocator_backend() != "native" |
| 394 | ): |
| 395 | yield |
| 396 | return |
| 397 | |
| 398 | get_settings = getattr(torch._C, "_accelerator_getAllocatorSettings", None) |
| 399 | if get_settings is not None: |
| 400 | original_settings = get_settings() |
| 401 | else: |
| 402 | original_settings = torch.cuda.memory._snapshot()["allocator_settings"][ |
| 403 | "PYTORCH_CUDA_ALLOC_CONF" |
| 404 | ] |
| 405 | |
| 406 | # The setter resets other allocator options unless they are included. |
| 407 | prefix = original_settings.strip().rstrip(",") |
| 408 | settings = f"max_split_size_mb:{max_split_size_mb}" |
| 409 | if prefix: |
| 410 | settings = f"{prefix},{settings}" |
| 411 | try: |
| 412 | torch._C._accelerator_setAllocatorSettings(settings) |
| 413 | yield |
| 414 | finally: |
| 415 | torch._C._accelerator_setAllocatorSettings(original_settings) |
| 416 | |
| 417 | @instrument(span_name="Init device") |
| 418 | def init_device(self): |
| 419 | if self.device_config.device_type == "cuda": |
| 420 | # This env var set by Ray causes exceptions with graph building. |
| 421 | os.environ.pop("NCCL_ASYNC_ERROR_HANDLING", None) |
| 422 | parallel_config = self.parallel_config |
| 423 | if ( |
| 424 | parallel_config.distributed_executor_backend |
| 425 | not in ("ray", "external_launcher") |
| 426 | and parallel_config.data_parallel_backend != "ray" |
| 427 | and parallel_config.nnodes_within_dp == 1 |
| 428 | ): |
| 429 | # Use local DP rank if available, otherwise use global DP rank. |
| 430 | dp_local_rank = self.parallel_config.data_parallel_rank_local |
| 431 | if dp_local_rank is None: |
| 432 | dp_local_rank = self.parallel_config.data_parallel_index |
| 433 | |
| 434 | tp_pcp_pp_world_size = ( |
| 435 | self.parallel_config.pipeline_parallel_size |
| 436 | * self.parallel_config.prefill_context_parallel_size |
| 437 | * self.parallel_config.tensor_parallel_size |
| 438 | ) |
| 439 | |
| 440 | # DP_LOCAL_RANK * TP_PCP_PP_WORLD_SIZE + TP_LOCAL_RANK |
| 441 | self.local_rank += dp_local_rank * tp_pcp_pp_world_size |
| 442 | |
| 443 | # Publish the logical-to-physical mapping for topology queries |
| 444 | # such as NIC affinity and P2P checks. |
| 445 | assigned_physical_gpu_ids = parallel_config.assigned_physical_gpu_ids |
| 446 | if assigned_physical_gpu_ids is not None: |
| 447 | from vllm.platforms.interface import set_assigned_physical_gpu_ids |
| 448 | |
| 449 | set_assigned_physical_gpu_ids(assigned_physical_gpu_ids) |
| 450 | assert self.local_rank < len(assigned_physical_gpu_ids), ( |
| 451 | f"local_rank {self.local_rank} is out of bounds for " |
| 452 | f"assigned_physical_gpu_ids {assigned_physical_gpu_ids}" |
| 453 | ) |
| 454 | # NOTE(patch pr45026): local_world_size is derived from |
| 455 | # parallel_config.nnodes, which is only set for the "mp" |
| 456 | # multi-node backend. With the "ray"/"external_launcher" |
| 457 | # backends nnodes stays 1, so local_world_size collapses to |
| 458 | # the full world_size and this check wrongly fires on |
| 459 | # cross-node deployments. assigned_physical_gpu_ids is already |
| 460 | # per-node and the local_rank bound above fully validates the |
| 461 | # mapping for these backends, so skip the check for them. |
| 462 | if parallel_config.distributed_executor_backend not in ( |
| 463 | "ray", |
| 464 | "external_launcher", |
| 465 | ): |
| 466 | assert self.parallel_config.local_world_size <= len( |
| 467 | assigned_physical_gpu_ids |
| 468 | ), ( |
| 469 | f"local_world_size ({self.parallel_config.local_world_size})" |
| 470 | " exceeds assigned_physical_gpu_ids count " |
| 471 | f"({len(assigned_physical_gpu_ids)})" |
| 472 | ) |
| 473 | else: |
| 474 | assert self.local_rank < torch.accelerator.device_count(), ( |
| 475 | f"DP adjusted local rank {self.local_rank} is out of " |
| 476 | f"bounds for {torch.accelerator.device_count()} devices." |
| 477 | ) |
| 478 | |
| 479 | visible_device_index = ( |
| 480 | current_platform.logical_device_id_to_visible_device_id(self.local_rank) |
| 481 | ) |
| 482 | self.device = torch.device(f"cuda:{visible_device_index}") |
| 483 | torch.accelerator.set_device_index(self.device) |
| 484 | |
| 485 | current_platform.check_if_supports_dtype(self.model_config.dtype) |
| 486 | |
| 487 | # Initialize the distributed environment BEFORE taking |
| 488 | # memory snapshot |
| 489 | # This ensures NCCL buffers are allocated before we measure |
| 490 | # available memory |
| 491 | init_worker_distributed_environment( |
| 492 | self.vllm_config, |
| 493 | self.rank, |
| 494 | self.distributed_init_method, |
| 495 | self.local_rank, |
| 496 | current_platform.dist_backend, |
| 497 | ) |
| 498 | |
| 499 | if self.use_v2_model_runner: |
| 500 | logger.info_once("Using V2 Model Runner") |
| 501 | |
| 502 | # Set random seed. |
| 503 | set_random_seed(self.model_config.seed) |
| 504 | |
| 505 | # Now take memory snapshot after NCCL is initialized |
| 506 | gc.collect() |
| 507 | torch.accelerator.empty_cache() |
| 508 | |
| 509 | # take current memory snapshot |
| 510 | self.init_snapshot = init_snapshot = MemorySnapshot(device=self.device) |
| 511 | # Weights from external model loader process |
| 512 | external_weight_memory = get_model_loader( |
| 513 | self.load_config |
| 514 | ).get_external_weight_memory(self.vllm_config) |
| 515 | self.requested_memory = request_memory( |
| 516 | init_snapshot, self.cache_config, external_weight_memory |
| 517 | ) |
| 518 | logger.debug("worker init memory snapshot: %r", self.init_snapshot) |
| 519 | logger.debug( |
| 520 | "worker requested memory: %sGiB", format_gib(self.requested_memory) |
| 521 | ) |
| 522 | else: |
| 523 | raise RuntimeError(f"Unsupported device type: {self.device_config.device}") |
| 524 | |
| 525 | self._init_workspace_and_model_runner() |
| 526 | |
| 527 | def _init_workspace_and_model_runner(self) -> None: |
| 528 | """Set up the workspace manager, build the model runner, report usage.""" |
| 529 | # DSpark target and draft CUDA graphs retain workspace views concurrently. |
| 530 | num_ubatches = 2 if self.vllm_config.parallel_config.enable_dbo else 1 |
| 531 | init_workspace_manager( |
| 532 | self.device, |
| 533 | num_ubatches, |
| 534 | _num_workspace_lanes(self.vllm_config, self.use_v2_model_runner), |
| 535 | # Scratch holds no state across steps: discard it on sleep. |
| 536 | alloc_context=lambda: self._maybe_get_memory_pool_context("workspace"), |
| 537 | ) |
| 538 | self.model_runner: GPUModelRunner = self._make_model_runner() |
| 539 | if self.rank == 0: |
| 540 | # If usage stat is enabled, collect relevant info. |
| 541 | report_usage_stats(self.vllm_config) |
| 542 | |
| 543 | def _make_model_runner(self) -> "GPUModelRunner": |
| 544 | """Construct the platform's model runner; subclasses override this.""" |
| 545 | if self.use_v2_model_runner: |
| 546 | if self.vllm_config.is_mm_encoder_only: |
| 547 | from vllm.v1.worker.mm_encoder_model_runner import ( |
| 548 | MMEncoderModelRunner as GPUModelRunnerV2, |
| 549 | ) |
| 550 | else: |
| 551 | from vllm.v1.worker.gpu.model_runner import ( # type: ignore[assignment] |
| 552 | GPUModelRunner as GPUModelRunnerV2, |
| 553 | ) |
| 554 | |
| 555 | # HACK(woosuk): This is a temporary fix to avoid type errors. |
| 556 | return GPUModelRunnerV2(self.vllm_config, self.device) # type: ignore |
| 557 | from vllm.v1.worker.gpu_model_runner import GPUModelRunner as GPUModelRunnerV1 |
| 558 | |
| 559 | return GPUModelRunnerV1(self.vllm_config, self.device) |
| 560 | |
| 561 | def handle_ft_command(self, ft_request): |
| 562 | assert self.worker_sentinel is not None |
| 563 | return self.worker_sentinel.handle_command(ft_request) |
| 564 | |
| 565 | # FIXME(youkaichao & ywang96): Use TorchDispatchMode instead of memory pool |
| 566 | # to hijack tensor allocation. |
| 567 | def load_model(self, *, load_dummy_weights: bool = False) -> None: |
| 568 | with ( |
| 569 | self._maybe_get_memory_pool_context(tag="weights"), |
| 570 | set_current_vllm_config(self.vllm_config), |
| 571 | # 20 MiB is the minimum PyTorch allows for max_split_size_mb. |
| 572 | self._scoped_allocator_max_split(max_split_size_mb=20), |
| 573 | ): |
| 574 | self.model_runner.load_model(load_dummy_weights=load_dummy_weights) |
| 575 | |
| 576 | if has_ec_transfer(): |
| 577 | get_ec_transfer().start_worker_services() |
| 578 | |
| 579 | if self.vllm_config.weight_transfer_config is not None: |
| 580 | self.weight_transfer_engine = WeightTransferEngineFactory.create_engine( |
| 581 | self.vllm_config.weight_transfer_config, |
| 582 | self.vllm_config, |
| 583 | self.device, |
| 584 | self.model_runner.get_model(), |
| 585 | ) |
| 586 | |
| 587 | # Preserve parallel weight loading, then use serving's thread count |
| 588 | # for profiling and compilation so Dynamo's global-state guards remain |
| 589 | # valid when requests arrive. |
| 590 | set_torch_threads_for_runtime() |
| 591 | |
| 592 | def update_config(self, overrides: dict[str, Any]) -> None: |
| 593 | self.model_runner.update_config(overrides) |
| 594 | |
| 595 | def reload_weights(self, *args, **kwargs) -> None: |
| 596 | with set_current_vllm_config(self.vllm_config): |
| 597 | self.model_runner.reload_weights(*args, **kwargs) |
| 598 | |
| 599 | @torch.inference_mode() |
| 600 | def determine_available_memory(self) -> int: |
| 601 | """Profiles the peak memory usage of the model to determine how much |
| 602 | memory can be used for KV cache without OOMs. |
| 603 | |
| 604 | The engine will first conduct a profiling of the existing memory usage. |
| 605 | Then, it calculates the free memory that can be used for KV cache in |
| 606 | bytes. |
| 607 | |
| 608 | Tip: |
| 609 | You may limit the usage of GPU memory |
| 610 | by adjusting the `gpu_memory_utilization` parameter. |
| 611 | |
| 612 | """ |
| 613 | maybe_apply_startup_plan(self) |
| 614 | |
| 615 | if kv_cache_memory_bytes := self.cache_config.kv_cache_memory_bytes: |
| 616 | # still need a profile run which compiles the model for |
| 617 | # max_num_batched_tokens |
| 618 | self.model_runner.profile_run(randomize_inputs=self.randomize_dummy_inputs) |
| 619 | |
| 620 | msg = ( |
| 621 | f"Initial free memory {format_gib(self.init_snapshot.free_memory)} " |
| 622 | f"GiB, reserved {format_gib(kv_cache_memory_bytes)} GiB memory for " |
| 623 | "KV Cache as specified by kv_cache_memory_bytes config and " |
| 624 | "skipped memory profiling. This does not respect the " |
| 625 | "gpu_memory_utilization config. Only use kv_cache_memory_bytes " |
| 626 | "config when you want manual control of KV cache memory " |
| 627 | "size. If OOM'ed, check the difference of initial free " |
| 628 | "memory between the current run and the previous run " |
| 629 | "where kv_cache_memory_bytes is suggested and update it " |
| 630 | "correspondingly." |
| 631 | ) |
| 632 | logger.info(msg) |
| 633 | return reserve_mm_ipc_gpu_memory( |
| 634 | kv_cache_memory_bytes, |
| 635 | self.model_config.multimodal_config, |
| 636 | getattr(self.parallel_config, "_api_process_count", 1), |
| 637 | ) |
| 638 | |
| 639 | # Execute a forward pass with dummy inputs to profile the memory usage |
| 640 | # of the model. |
| 641 | with ( |
| 642 | memory_profiling( |
| 643 | self.init_snapshot, |
| 644 | weights_memory=int(self.model_runner.model_memory_usage), |
| 645 | ) as profile_result, |
| 646 | # Workspaces (e.g. the MoE workspace) grow in steps during this pass, |
| 647 | # freeing each smaller buffer. Without a split limit, a later small |
| 648 | # allocation can be carved out of a freed multi-GiB block and pin the |
| 649 | # whole segment past the empty_cache() in memory_profiling, so it is |
| 650 | # counted as consumed and taken from the KV cache. Blocks above the |
| 651 | # limit are never split, so they stay releasable. Exits before |
| 652 | # memory_profiling measures, restoring the original limit. |
| 653 | self._scoped_allocator_max_split(max_split_size_mb=20), |
| 654 | ): |
| 655 | self.model_runner.profile_run(randomize_inputs=self.randomize_dummy_inputs) |
| 656 | |
| 657 | # Profile CUDA graph memory if graphs will be captured. |
| 658 | # ROCm is included: #44825 moved the profiler to |
| 659 | # torch.accelerator.get_memory_info (reliable on ROCm, as used by |
| 660 | # the AMD-CI mem tests), and graph_pool_handle resolves to the same |
| 661 | # torch.cuda handle the live capture path already uses on ROCm. |
| 662 | cudagraph_memory_estimate = 0 |
| 663 | if ( |
| 664 | current_platform.is_cuda_alike() or current_platform.is_xpu() |
| 665 | ) and self.vllm_config.compilation_config.cudagraph_mode != CUDAGraphMode.NONE: |
| 666 | cudagraph_memory_estimate = self.model_runner.profile_cudagraph_memory() |
| 667 | |
| 668 | # Respect the opt-in flag as originally designed. |
| 669 | cudagraph_memory_estimate_applied = ( |
| 670 | cudagraph_memory_estimate |
| 671 | if envs.VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS |
| 672 | else 0 |
| 673 | ) |
| 674 | |
| 675 | init_free_memory = self.init_snapshot.free_memory |
| 676 | free_gpu_memory = profile_result.after_profile.free_memory |
| 677 | rocm_fallback = maybe_rocm_profiling_fallback(profile_result) |
| 678 | if rocm_fallback is None: |
| 679 | # NOTE(woosuk): Here we assume that the other processes using the same |
| 680 | # GPU did not change their memory usage during the profiling. |
| 681 | assert init_free_memory >= free_gpu_memory, ( |
| 682 | "Error in memory profiling. " |
| 683 | f"Initial free memory {format_gib(init_free_memory)} GiB, " |
| 684 | f"current free memory {format_gib(free_gpu_memory)} GiB. " |
| 685 | "This happens when other processes sharing the same container " |
| 686 | "release GPU memory while vLLM is profiling during initialization. " |
| 687 | "To fix this, ensure consistent GPU memory allocation or " |
| 688 | "isolate vLLM in its own container." |
| 689 | ) |
| 690 | else: |
| 691 | profile_result.total_consumed = rocm_fallback |
| 692 | profile_result.non_kv_cache_memory = ( |
| 693 | profile_result.total_consumed + profile_result.transient_peak_headroom |
| 694 | ) |
| 695 | |
| 696 | self.total_consumed = profile_result.total_consumed |
| 697 | self.peak_activation_memory = ( |
| 698 | profile_result.transient_peak_headroom + cudagraph_memory_estimate_applied |
| 699 | ) |
| 700 | self.cudagraph_memory_estimate = cudagraph_memory_estimate |
| 701 | |
| 702 | self.available_kv_cache_memory_bytes = ( |
| 703 | self.requested_memory |
| 704 | - profile_result.non_kv_cache_memory |
| 705 | - cudagraph_memory_estimate_applied |
| 706 | ) |
| 707 | |
| 708 | unrequested_memory = self.init_snapshot.free_memory - self.requested_memory |
| 709 | logger.debug( |
| 710 | "Initial free memory: %s GiB; Requested memory: %f (util), %s GiB", |
| 711 | format_gib(self.init_snapshot.free_memory), |
| 712 | self.cache_config.gpu_memory_utilization, |
| 713 | format_gib(self.requested_memory), |
| 714 | ) |
| 715 | logger.debug( |
| 716 | "Free memory after profiling: %s GiB (total), %s GiB (within requested)", |
| 717 | format_gib(free_gpu_memory), |
| 718 | format_gib(free_gpu_memory - unrequested_memory), |
| 719 | ) |
| 720 | logger.debug(profile_result) |
| 721 | logger.info_once( |
| 722 | "Available KV cache memory: %s GiB", |
| 723 | format_gib(self.available_kv_cache_memory_bytes), |
| 724 | ) |
| 725 | |
| 726 | if cudagraph_memory_estimate > 0: |
| 727 | total_mem = self.init_snapshot.total_memory |
| 728 | current_util = self.cache_config.gpu_memory_utilization |
| 729 | cg_util_delta = cudagraph_memory_estimate / total_mem |
| 730 | if envs.VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS: |
| 731 | equiv_util = round(current_util - cg_util_delta, 4) |
| 732 | suggested_util = min( |
| 733 | round(current_util + cg_util_delta, 4), |
| 734 | 1.0, |
| 735 | ) |
| 736 | logger.info( |
| 737 | "CUDA graph memory profiling is enabled (default since " |
| 738 | "v0.21.0). The current --gpu-memory-utilization=%.4f is " |
| 739 | "equivalent to --gpu-memory-utilization=%.4f without " |
| 740 | "CUDA graph memory profiling. To maintain the same " |
| 741 | "effective KV cache size as before, increase " |
| 742 | "--gpu-memory-utilization to %.4f. To disable, set " |
| 743 | "VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0.", |
| 744 | current_util, |
| 745 | equiv_util, |
| 746 | suggested_util, |
| 747 | ) |
| 748 | else: |
| 749 | suggested_util = min( |
| 750 | round(current_util + cg_util_delta, 4), |
| 751 | 1.0, |
| 752 | ) |
| 753 | logger.warning( |
| 754 | "CUDA graph memory profiling is disabled " |
| 755 | "(VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0). " |
| 756 | "Without it, CUDA graph memory is not accounted for " |
| 757 | "during KV cache allocation, which may require lowering " |
| 758 | "--gpu-memory-utilization to avoid OOM. Consider " |
| 759 | "re-enabling it (the default as of v0.21.0) and increasing " |
| 760 | "--gpu-memory-utilization from %.4f to %.4f.", |
| 761 | current_util, |
| 762 | suggested_util, |
| 763 | ) |
| 764 | |
| 765 | return reserve_mm_ipc_gpu_memory( |
| 766 | int(self.available_kv_cache_memory_bytes), |
| 767 | self.model_config.multimodal_config, |
| 768 | getattr(self.parallel_config, "_api_process_count", 1), |
| 769 | ) |
| 770 | |
| 771 | def get_kv_connector_handshake_metadata( |
| 772 | self, |
| 773 | ) -> dict[tuple[int, int], KVConnectorHandshakeMetadata] | None: |
| 774 | """Get KV connector metadata from this worker if available. |
| 775 | |
| 776 | Returned dict is keyed by `(pp_rank, tp_rank)`. |
| 777 | """ |
| 778 | if not has_kv_transfer_group(): |
| 779 | return None |
| 780 | |
| 781 | connector = get_kv_transfer_group() |
| 782 | # Return None for connectors that don't need to exchange handshake |
| 783 | # metadata across workers. |
| 784 | if (metadata := connector.get_handshake_metadata()) is None: |
| 785 | return None |
| 786 | |
| 787 | pp_rank = get_pp_group().rank_in_group |
| 788 | tp_rank = get_tp_group().rank_in_group |
| 789 | parallel_config = self.vllm_config.parallel_config |
| 790 | if ( |
| 791 | parallel_config.prefill_context_parallel_size > 1 |
| 792 | and parallel_config.decode_context_parallel_size > 1 |
| 793 | ): |
| 794 | tp_rank += ( |
| 795 | get_pcp_group().rank_in_group * parallel_config.tensor_parallel_size |
| 796 | ) |
| 797 | return {(pp_rank, tp_rank): metadata} |
| 798 | |
| 799 | def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]: |
| 800 | return self.model_runner.get_kv_cache_spec() |
| 801 | |
| 802 | def update_max_model_len(self, max_model_len: int) -> None: |
| 803 | """Update max_model_len after auto-fit to GPU memory. |
| 804 | This is called when max_model_len=-1 is used and the engine |
| 805 | automatically determines the maximum context length that fits |
| 806 | in GPU memory. Workers need to update their cached max_model_len |
| 807 | to match the engine's decision. |
| 808 | """ |
| 809 | self.model_config.max_model_len = max_model_len |
| 810 | if self.model_runner is not None: |
| 811 | self.model_runner.update_max_model_len(max_model_len) |
| 812 | logger.debug("Updated max_model_len to %d", max_model_len) |
| 813 | |
| 814 | @instrument(span_name="Allocate KV cache") |
| 815 | def initialize_from_config(self, kv_cache_config: KVCacheConfig) -> None: |
| 816 | """Allocate GPU KV cache with the specified kv_cache_config.""" |
| 817 | # Update local config with adjusted num blocks after profiling, |
| 818 | # so that it's available to the warmup stage. |
| 819 | self.cache_config.num_gpu_blocks = kv_cache_config.num_blocks |
| 820 | |
| 821 | # Adopt the engine core's layout; workers spawned after resolution |
| 822 | # (e.g. elastic EP scale-up) only see it through the config. |
| 823 | if kv_cache_config.kv_cache_layout is not None: |
| 824 | record_kv_cache_layout(self.cache_config, kv_cache_config.kv_cache_layout) |
| 825 | |
| 826 | # Init kv cache connector here, because it requires |
| 827 | # `kv_cache_config`. |
| 828 | # NOTE(Kuntai): This need to be done before `initialize_kv_cache`, |
| 829 | # because `initialize_kv_cache` will inject kv cache groups not |
| 830 | # related to kv cache connector (e.g. kv cache sharing layers). |
| 831 | ensure_kv_transfer_initialized(self.vllm_config, kv_cache_config) |
| 832 | |
| 833 | # If the connector provides a custom memory pool (e.g. Mooncake |
| 834 | # NVLink/BAREX), use it for KV cache allocation; otherwise fall |
| 835 | # back to the standard CuMem pool. |
| 836 | mem_pool_context = ( |
| 837 | get_kv_transfer_group().get_mem_pool_context() |
| 838 | if has_kv_transfer_group() |
| 839 | else None |
| 840 | ) |
| 841 | if mem_pool_context is None: |
| 842 | mem_pool_context = self._maybe_get_memory_pool_context(tag="kv_cache") |
| 843 | |
| 844 | # Offload KV-init state, except on XPU, whose outermost pool would win. |
| 845 | runtime_pool = ( |
| 846 | nullcontext() |
| 847 | if current_platform.is_xpu() |
| 848 | else self._maybe_get_memory_pool_context(tag="runtime") |
| 849 | ) |
| 850 | with runtime_pool: |
| 851 | self.model_runner.initialize_kv_cache( |
| 852 | kv_cache_config, |
| 853 | kv_cache_allocation_context=mem_pool_context, |
| 854 | ) |
| 855 | |
| 856 | # Build KV-zero metadata outside the CuMem pool so the bookkeeping |
| 857 | # GPU tensors (seg_addrs, block-id buffers) use the standard PyTorch |
| 858 | # allocator and are not discarded during sleep/wake cycles. |
| 859 | if kv_cache_config.needs_kv_cache_zeroing and hasattr( |
| 860 | self.model_runner, "_init_kv_zero_meta" |
| 861 | ): |
| 862 | self.model_runner._init_kv_zero_meta() |
| 863 | |
| 864 | @instrument(span_name="Warmup (GPU)") |
| 865 | def compile_or_warm_up_model(self) -> CompilationTimes: |
| 866 | warmup_sizes: list[int] = [] |
| 867 | |
| 868 | if ( |
| 869 | self.vllm_config.compilation_config.mode == CompilationMode.VLLM_COMPILE |
| 870 | and has_compiled_submodule(self.model_runner.get_model()) |
| 871 | ): |
| 872 | # warm up sizes that are not in cudagraph capture sizes, |
| 873 | # but users still want to compile for better performance, |
| 874 | # e.g. for the max-num-batched token size in chunked prefill. |
| 875 | compile_sizes = self.vllm_config.compilation_config.compile_sizes |
| 876 | warmup_sizes = compile_sizes.copy() if compile_sizes is not None else [] # type: ignore[assignment] |
| 877 | cg_capture_sizes: list[int] = [] |
| 878 | |
| 879 | if self.vllm_config.compilation_config.cudagraph_mode != CUDAGraphMode.NONE: |
| 880 | cg_sizes = self.vllm_config.compilation_config.cudagraph_capture_sizes |
| 881 | cg_capture_sizes = [] if cg_sizes is None else cg_sizes |
| 882 | warmup_sizes = [x for x in warmup_sizes if x not in cg_capture_sizes] |
| 883 | |
| 884 | compile_ranges = self.vllm_config.compilation_config.get_compile_ranges() |
| 885 | # For each compile_range, if none of the batch sizes |
| 886 | # in warmup_sizes or cudagraph_capture_sizes are in the range, |
| 887 | # add the end of the range to ensure compilation/warmup. |
| 888 | all_sizes = set(cg_capture_sizes) |
| 889 | all_sizes.update([x for x in warmup_sizes if isinstance(x, int)]) |
| 890 | for compile_range in compile_ranges: |
| 891 | if not any(x in compile_range for x in all_sizes): |
| 892 | warmup_sizes.append(compile_range.end) |
| 893 | |
| 894 | # We skip EPLB here since we don't want to record dummy metrics |
| 895 | for size in sorted(warmup_sizes, reverse=True): |
| 896 | logger.info("Compile and warming up model for size %d", size) |
| 897 | self.model_runner._dummy_run( |
| 898 | size, |
| 899 | skip_eplb=True, |
| 900 | remove_lora=False, |
| 901 | randomize_inputs=self.randomize_dummy_inputs, |
| 902 | ) |
| 903 | self.model_runner.maybe_remove_all_loras(self.model_runner.lora_config) |
| 904 | |
| 905 | # Warmup and tune the kernels used during model execution before |
| 906 | # cuda graph capture. |
| 907 | kernel_warmup(self) |
| 908 | |
| 909 | if self.use_v2_model_runner: # noqa: SIM102 |
| 910 | if self.vllm_config.kernel_config.enable_jit_warmup: |
| 911 | # A workspace resize after capture frees what the graphs point at. |
| 912 | warmup_kernels( |
| 913 | self.model_runner, self.execute_model, self.sample_tokens |
| 914 | ) |
| 915 | |
| 916 | cuda_graph_memory_bytes = 0 |
| 917 | if not self.model_config.enforce_eager: |
| 918 | with self._get_cudagraph_capture_context(): |
| 919 | cuda_graph_memory_bytes = self.model_runner.capture_model() |
| 920 | |
| 921 | # Compare actual vs estimated CUDA graph memory (if we did profiling) |
| 922 | if ( |
| 923 | hasattr(self, "cudagraph_memory_estimate") |
| 924 | and self.cudagraph_memory_estimate > 0 |
| 925 | ): |
| 926 | GiB = lambda b: round(b / GiB_bytes, 2) |
| 927 | diff = abs(cuda_graph_memory_bytes - self.cudagraph_memory_estimate) |
| 928 | logger.info( |
| 929 | "CUDA graph pool memory: %s GiB (actual), %s GiB (estimated), " |
| 930 | "difference: %s GiB (%.1f%%).", |
| 931 | GiB(cuda_graph_memory_bytes), |
| 932 | GiB(self.cudagraph_memory_estimate), |
| 933 | GiB(diff), |
| 934 | 100 * diff / max(cuda_graph_memory_bytes, 1), |
| 935 | ) |
| 936 | |
| 937 | if self.cache_config.kv_cache_memory_bytes is None and hasattr( |
| 938 | self, "peak_activation_memory" |
| 939 | ): |
| 940 | # Suggests optimal kv cache memory size if we rely on |
| 941 | # memory_profiling to guess the kv cache memory size which |
| 942 | # provides peak_activation_memory and a few other memory |
| 943 | # consumption. `memory_profiling` does not consider |
| 944 | # CUDAGraph memory size and may not utilize all gpu memory. |
| 945 | # Users may want fine-grained control to specify kv cache |
| 946 | # memory size. |
| 947 | |
| 948 | # empirically observed that the memory profiling may |
| 949 | # slightly underestimate the memory consumption. |
| 950 | # So leave a small buffer (=150MiB) to avoid OOM. |
| 951 | redundancy_buffer_memory = 150 * (1 << 20) |
| 952 | |
| 953 | non_kv_cache_memory = ( |
| 954 | self.total_consumed |
| 955 | + self.peak_activation_memory |
| 956 | + cuda_graph_memory_bytes |
| 957 | ) |
| 958 | kv_cache_memory_bytes_to_gpu_limit = ( |
| 959 | self.init_snapshot.free_memory |
| 960 | - non_kv_cache_memory |
| 961 | - redundancy_buffer_memory |
| 962 | ) |
| 963 | kv_cache_memory_bytes_to_requested_limit = ( |
| 964 | int(self.requested_memory) |
| 965 | - non_kv_cache_memory |
| 966 | - redundancy_buffer_memory |
| 967 | ) |
| 968 | |
| 969 | msg = ( |
| 970 | f"Free memory on device " |
| 971 | f"({format_gib(self.init_snapshot.free_memory)}/" |
| 972 | f"{format_gib(self.init_snapshot.total_memory)} GiB) on startup. " |
| 973 | f"Desired GPU memory utilization is " |
| 974 | f"({self.cache_config.gpu_memory_utilization}, " |
| 975 | f"{format_gib(self.requested_memory)} GiB). " |
| 976 | f"Actual usage is {format_gib(self.total_consumed)} " |
| 977 | f"GiB for consumed memory (weights + non-torch), " |
| 978 | f"{format_gib(self.peak_activation_memory)} GiB " |
| 979 | f"for peak activation, and {format_gib(cuda_graph_memory_bytes)} " |
| 980 | f"GiB for CUDAGraph memory. Replace gpu_memory_utilization " |
| 981 | f"config with `--kv-cache-memory=" |
| 982 | f"{kv_cache_memory_bytes_to_requested_limit}` " |
| 983 | f"({format_gib(kv_cache_memory_bytes_to_requested_limit)} GiB) to fit " |
| 984 | f"into requested memory, or `--kv-cache-memory=" |
| 985 | f"{kv_cache_memory_bytes_to_gpu_limit}` " |
| 986 | f"({format_gib(kv_cache_memory_bytes_to_gpu_limit)} GiB) to fully " |
| 987 | f"utilize gpu memory. Current kv cache memory in use is " |
| 988 | f"{format_gib(self.available_kv_cache_memory_bytes)} GiB." |
| 989 | ) |
| 990 | |
| 991 | logger.info(msg) |
| 992 | |
| 993 | maybe_save_startup_plan(self, kv_cache_memory_bytes_to_requested_limit) |
| 994 | |
| 995 | if not self.use_v2_model_runner and get_pp_group().is_last_rank: |
| 996 | # V1: Warm up sampler and preallocate memory buffer for logits and other |
| 997 | # sampling related tensors of max possible shape to avoid memory |
| 998 | # fragmentation issue. |
| 999 | # NOTE: This is called after `capture_model` on purpose to prevent |
| 1000 | # memory buffers from being cleared by `torch.accelerator.empty_cache`. |
| 1001 | max_num_reqs = min( |
| 1002 | self.scheduler_config.max_num_seqs, |
| 1003 | self.scheduler_config.max_num_batched_tokens, |
| 1004 | ) |
| 1005 | |
| 1006 | # We skip EPLB here since we don't want to record dummy metrics |
| 1007 | hidden_states, last_hidden_states = self.model_runner._dummy_run( |
| 1008 | num_tokens=max_num_reqs, |
| 1009 | skip_eplb=True, |
| 1010 | cudagraph_runtime_mode=CUDAGraphMode.NONE, |
| 1011 | ) |
| 1012 | if self.model_runner.is_pooling_model: |
| 1013 | self.model_runner._dummy_pooler_run(hidden_states) |
| 1014 | else: |
| 1015 | self.model_runner._dummy_sampler_run(hidden_states=last_hidden_states) |
| 1016 | |
| 1017 | # Reset the seed to ensure that the random state is not affected by |
| 1018 | # the model initialization and profiling. |
| 1019 | set_random_seed(self.model_config.seed) |
| 1020 | |
| 1021 | # Eagerly trigger inductor's once-per-process lazy inits during |
| 1022 | # warmup (rather than on a later compile cache-miss at runtime). |
| 1023 | c_config = self.compilation_config |
| 1024 | if c_config.mode != CompilationMode.NONE and c_config.backend == "inductor": |
| 1025 | from vllm.compilation.compiler_interface import ( |
| 1026 | trigger_inductor_lazy_init, |
| 1027 | ) |
| 1028 | |
| 1029 | trigger_inductor_lazy_init(self.device) |
| 1030 | |
| 1031 | # All warmup is done — start monitoring for unexpected JIT |
| 1032 | # compilations that would cause latency spikes during inference. |
| 1033 | self._maybe_activate_jit_monitor() |
| 1034 | |
| 1035 | # Freeze the worker heap so the GC won't scan static objects |
| 1036 | # (model weights, KV caches, CUDA graphs) during inference. |
| 1037 | freeze_gc_heap() |
| 1038 | maybe_attach_gc_debug_callback() |
| 1039 | |
| 1040 | # Warmup / first-compile is done — activate the `VLLM_GPU_SYNC_CHECK` |
| 1041 | # gate so subsequent `execute_model` / `sample_tokens` calls enforce it. |
| 1042 | enable_gpu_sync_check() |
| 1043 | |
| 1044 | return CompilationTimes( |
| 1045 | language_model=self.compilation_config.compilation_time, |
| 1046 | encoder=self.compilation_config.encoder_compilation_time, |
| 1047 | ) |
| 1048 | |
| 1049 | def _maybe_activate_jit_monitor(self) -> None: |
| 1050 | # When JIT warmup is disabled (e.g. enforce_eager), runtime JIT |
| 1051 | # compilation is expected, so monitoring would only produce noise |
| 1052 | # (or spurious errors in "error" mode). |
| 1053 | if not self.vllm_config.kernel_config.enable_jit_warmup: |
| 1054 | return |
| 1055 | |
| 1056 | from vllm.utils.jit_monitor import activate as activate_jit_monitor |
| 1057 | |
| 1058 | activate_jit_monitor( |
| 1059 | mode=self.observability_config.jit_monitor_mode, |
| 1060 | verbose=self.observability_config.jit_monitor_verbose, |
| 1061 | ) |
| 1062 | |
| 1063 | def _get_cudagraph_capture_context(self) -> AbstractContextManager[None]: |
| 1064 | """Let the configured profiler observe CUDA graph capture.""" |
| 1065 | if not self.use_v2_model_runner: |
| 1066 | return nullcontext() |
| 1067 | if self.profiler is None: |
| 1068 | model_runner = cast("GPUModelRunnerV2", self.model_runner) |
| 1069 | if not model_runner.needs_cudagraph_capture(): |
| 1070 | return nullcontext() |
| 1071 | self.profiler = create_graph_capture_profiler( |
| 1072 | self.profiler_config, global_rank=self.rank |
| 1073 | ) |
| 1074 | if self.profiler is None: |
| 1075 | return nullcontext() |
| 1076 | return self.profiler.capture_cuda_graphs() |
| 1077 | |
| 1078 | def reset_mm_cache(self) -> None: |
| 1079 | self.model_runner.reset_mm_cache() |
| 1080 | |
| 1081 | def reset_encoder_cache(self) -> None: |
| 1082 | self.model_runner.reset_encoder_cache() |
| 1083 | |
| 1084 | def get_model(self) -> nn.Module: |
| 1085 | return self.model_runner.get_model() |
| 1086 | |
| 1087 | def get_draft_model(self) -> nn.Module | None: |
| 1088 | return self.model_runner.get_draft_model() |
| 1089 | |
| 1090 | def supports_draft_weight_updates(self) -> bool: |
| 1091 | engine = self.weight_transfer_engine |
| 1092 | speculative_config = self.speculative_config |
| 1093 | get_draft_model = getattr(self.model_runner, "get_draft_model", None) |
| 1094 | return ( |
| 1095 | engine is not None |
| 1096 | and engine.supports_draft_weight_update |
| 1097 | and callable(get_draft_model) |
| 1098 | and get_draft_model() is not None |
| 1099 | and speculative_config is not None |
| 1100 | and speculative_config.draft_model_config is not None |
| 1101 | ) |
| 1102 | |
| 1103 | def _set_draft_weight_update_target(self) -> None: |
| 1104 | assert self.weight_transfer_engine is not None |
| 1105 | |
| 1106 | draft_model = self.get_draft_model() |
| 1107 | if draft_model is None: |
| 1108 | raise RuntimeError( |
| 1109 | "Draft model weight update requested, but no draft model is configured." |
| 1110 | ) |
| 1111 | |
| 1112 | speculative_config = self.speculative_config |
| 1113 | if speculative_config is None or speculative_config.draft_model_config is None: |
| 1114 | raise RuntimeError( |
| 1115 | "Draft model weight update requested, but no draft model " |
| 1116 | "config is configured." |
| 1117 | ) |
| 1118 | |
| 1119 | self.weight_transfer_engine.set_weight_update_target( |
| 1120 | draft_model, speculative_config.draft_model_config |
| 1121 | ) |
| 1122 | |
| 1123 | def get_supported_tasks(self) -> tuple[SupportedTask, ...]: |
| 1124 | return self.model_runner.get_supported_tasks() |
| 1125 | |
| 1126 | def get_compilation_match_table(self) -> dict[str, int]: |
| 1127 | from vllm.compilation.passes.vllm_inductor_pass import get_match_table |
| 1128 | |
| 1129 | return get_match_table() |
| 1130 | |
| 1131 | def get_encoder_timing_stats(self) -> dict[str, dict[str, float | int]]: |
| 1132 | """Get encoder timing stats from model runner.""" |
| 1133 | return self.model_runner.get_encoder_timing_stats() |
| 1134 | |
| 1135 | def annotate_profile(self, scheduler_output): |
| 1136 | # add trace annotation so that we can easily distinguish |
| 1137 | # context/generation request numbers in each iteration. |
| 1138 | # A context request is a request that has not yet generated any tokens |
| 1139 | if not self.profiler: |
| 1140 | return nullcontext() |
| 1141 | |
| 1142 | self.profiler.step() |
| 1143 | if not self.profiler.should_annotate: |
| 1144 | return nullcontext() |
| 1145 | |
| 1146 | iteration_details = compute_iteration_details(scheduler_output) |
| 1147 | |
| 1148 | if self.vllm_config.profiler_config.detailed_trace_annotation: |
| 1149 | # Compute roofline-model metrics per request, split by phase |
| 1150 | # (context vs generation). These help estimate compute and |
| 1151 | # memory intensity from the trace. |
| 1152 | # |
| 1153 | # Per-request quantities: |
| 1154 | # query_len = number of scheduled (new) tokens for this request |
| 1155 | # seq_len = total sequence length (computed + scheduled tokens) |
| 1156 | # |
| 1157 | # Aggregated across requests in each phase |
| 1158 | # (ctx_=context, gen_=generation): |
| 1159 | # seq_len_sum = sum of seq_len (total KV length) |
| 1160 | # qq_compute = sum of query_len*query_len |
| 1161 | # (proxy for QK^T compute cost) |
| 1162 | # qk_compute = sum of query_len*seq_len |
| 1163 | # (proxy for QK^T compute cost for decode and |
| 1164 | # chunked prefill) |
| 1165 | # total_scheduled_tokens = scheduled tokens across all requests |
| 1166 | ctx_seq_len_sum = 0 |
| 1167 | ctx_qq_compute = 0 |
| 1168 | ctx_qk_compute = 0 |
| 1169 | gen_seq_len_sum = 0 |
| 1170 | gen_qq_compute = 0 |
| 1171 | gen_qk_compute = 0 |
| 1172 | total_scheduled_tokens = 0 |
| 1173 | |
| 1174 | # Build a map of req_id -> num_computed_tokens for all requests |
| 1175 | new_req_ids = { |
| 1176 | new_req.req_id for new_req in scheduler_output.scheduled_new_reqs |
| 1177 | } |
| 1178 | num_computed_tokens_ids = { |
| 1179 | new_req.req_id: new_req.num_computed_tokens |
| 1180 | for new_req in scheduler_output.scheduled_new_reqs |
| 1181 | } |
| 1182 | for req_id, num_computed_tokens in zip( |
| 1183 | scheduler_output.scheduled_cached_reqs.req_ids, |
| 1184 | scheduler_output.scheduled_cached_reqs.num_computed_tokens, |
| 1185 | ): |
| 1186 | num_computed_tokens_ids[req_id] = num_computed_tokens |
| 1187 | |
| 1188 | # Accumulate per-phase metrics |
| 1189 | for req_id, num_tokens in scheduler_output.num_scheduled_tokens.items(): |
| 1190 | query_len = num_tokens |
| 1191 | total_scheduled_tokens += query_len |
| 1192 | seq_len = num_computed_tokens_ids.get(req_id, 0) + query_len |
| 1193 | if ( |
| 1194 | scheduler_output.scheduled_cached_reqs.is_context_phase(req_id) |
| 1195 | or req_id in new_req_ids |
| 1196 | ): |
| 1197 | ctx_seq_len_sum += seq_len |
| 1198 | ctx_qq_compute += query_len * query_len |
| 1199 | ctx_qk_compute += query_len * seq_len |
| 1200 | else: |
| 1201 | gen_seq_len_sum += seq_len |
| 1202 | gen_qq_compute += query_len * query_len |
| 1203 | gen_qk_compute += query_len * seq_len |
| 1204 | annotation = "".join( |
| 1205 | [ |
| 1206 | "execute_", |
| 1207 | str(total_scheduled_tokens), |
| 1208 | "_context_", |
| 1209 | str(iteration_details.num_ctx_requests), |
| 1210 | "(sq", |
| 1211 | str(iteration_details.num_ctx_tokens), |
| 1212 | "sk", |
| 1213 | str(ctx_seq_len_sum), |
| 1214 | "sqsq", |
| 1215 | str(ctx_qq_compute), |
| 1216 | "sqsk", |
| 1217 | str(ctx_qk_compute), |
| 1218 | ")_generation_", |
| 1219 | str(iteration_details.num_generation_requests), |
| 1220 | "(sq", |
| 1221 | str(iteration_details.num_generation_tokens), |
| 1222 | "sk", |
| 1223 | str(gen_seq_len_sum), |
| 1224 | "sqsq", |
| 1225 | str(gen_qq_compute), |
| 1226 | "sqsk", |
| 1227 | str(gen_qk_compute), |
| 1228 | ")", |
| 1229 | ] |
| 1230 | ) |
| 1231 | else: |
| 1232 | annotation = "".join( |
| 1233 | [ |
| 1234 | "execute_context_", |
| 1235 | str(iteration_details.num_ctx_requests), |
| 1236 | "(", |
| 1237 | str(iteration_details.num_ctx_tokens), |
| 1238 | ")", |
| 1239 | "_generation_", |
| 1240 | str(iteration_details.num_generation_requests), |
| 1241 | "(", |
| 1242 | str(iteration_details.num_generation_tokens), |
| 1243 | ")", |
| 1244 | ] |
| 1245 | ) |
| 1246 | return self.profiler.annotate_context_manager(annotation) |
| 1247 | |
| 1248 | @torch.inference_mode() |
| 1249 | @with_gpu_sync_check |
| 1250 | def sample_tokens( |
| 1251 | self, grammar_output: "GrammarOutput | None" |
| 1252 | ) -> ModelRunnerOutput | AsyncModelRunnerOutput: |
| 1253 | return self.model_runner.sample_tokens(grammar_output) |
| 1254 | |
| 1255 | @torch.inference_mode() |
| 1256 | @with_gpu_sync_check |
| 1257 | def execute_model( |
| 1258 | self, scheduler_output: "SchedulerOutput" |
| 1259 | ) -> ModelRunnerOutput | AsyncModelRunnerOutput | None: |
| 1260 | # Wait for the previous step's sends so this forward pass cannot |
| 1261 | # overwrite buffers they are still reading. |
| 1262 | if self._pp_send_work: |
| 1263 | for handle in self._pp_send_work: |
| 1264 | handle.wait() |
| 1265 | self._pp_send_work = [] |
| 1266 | |
| 1267 | intermediate_tensors = None |
| 1268 | forward_pass = scheduler_output.total_num_scheduled_tokens > 0 |
| 1269 | num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens |
| 1270 | all_gather_tensors = {} |
| 1271 | compilation_config = self.vllm_config.compilation_config |
| 1272 | parallel_config = self.vllm_config.parallel_config |
| 1273 | |
| 1274 | if ( |
| 1275 | parallel_config.pipeline_parallel_size > 1 |
| 1276 | and compilation_config.pass_config.enable_sp |
| 1277 | and forward_pass |
| 1278 | ): |
| 1279 | # currently only supported by V1 GPUModelRunner |
| 1280 | assert not self.use_v2_model_runner |
| 1281 | num_scheduled_tokens_np = np.array( |
| 1282 | list(scheduler_output.num_scheduled_tokens.values()), |
| 1283 | dtype=np.int32, |
| 1284 | ) |
| 1285 | # TODO(lucas): This is pretty gross; ideally we should only ever call |
| 1286 | # `_determine_batch_execution_and_padding` once (will get called again |
| 1287 | # in `execute_model`) but this requires a larger refactor of PP. |
| 1288 | _, batch_desc, _, _, _ = ( |
| 1289 | self.model_runner._determine_batch_execution_and_padding( |
| 1290 | num_tokens=num_scheduled_tokens, |
| 1291 | num_reqs=len(num_scheduled_tokens_np), |
| 1292 | num_scheduled_tokens_np=num_scheduled_tokens_np, |
| 1293 | max_num_scheduled_tokens=num_scheduled_tokens_np.max(), |
| 1294 | use_cascade_attn=False, # TODO(lucas): Handle cascade attention |
| 1295 | ) |
| 1296 | ) |
| 1297 | all_gather_tensors = { |
| 1298 | "residual": not is_residual_scattered_for_sp( |
| 1299 | self.vllm_config, batch_desc.num_tokens |
| 1300 | ) |
| 1301 | } |
| 1302 | |
| 1303 | if forward_pass and not get_pp_group().is_first_rank: |
| 1304 | tensor_dict, comm_handles, comm_postprocess = ( |
| 1305 | get_pp_group().irecv_tensor_dict( |
| 1306 | all_gather_group=get_tp_group(), |
| 1307 | all_gather_tensors=all_gather_tensors, |
| 1308 | ) |
| 1309 | ) |
| 1310 | assert tensor_dict is not None |
| 1311 | intermediate_tensors = AsyncIntermediateTensors( |
| 1312 | tensor_dict, |
| 1313 | comm_handles=comm_handles, |
| 1314 | comm_postprocess=comm_postprocess, |
| 1315 | ) |
| 1316 | |
| 1317 | with self.annotate_profile(scheduler_output): |
| 1318 | output = self.model_runner.execute_model( |
| 1319 | scheduler_output, intermediate_tensors |
| 1320 | ) |
| 1321 | if ( |
| 1322 | self.use_v2_model_runner |
| 1323 | and self.model_runner.is_pooling_model |
| 1324 | and output is None |
| 1325 | ): |
| 1326 | output = self.model_runner.pool() # type: ignore |
| 1327 | if isinstance( |
| 1328 | output, ModelRunnerOutput | AsyncModelRunnerOutput | NoneType |
| 1329 | ): |
| 1330 | return output |
| 1331 | |
| 1332 | assert isinstance(output, IntermediateTensors) |
| 1333 | parallel_config = self.vllm_config.parallel_config |
| 1334 | assert ( |
| 1335 | parallel_config.distributed_executor_backend != "external_launcher" |
| 1336 | and not get_pp_group().is_last_rank |
| 1337 | ) |
| 1338 | |
| 1339 | # Non-blocking send of the intermediate tensors. The metadata handle |
| 1340 | # is reaped lazily by the GroupCoordinator; the device handles are |
| 1341 | # waited at the top of the next step. |
| 1342 | handles = get_pp_group().isend_tensor_dict( |
| 1343 | output.tensors, |
| 1344 | all_gather_group=get_tp_group(), |
| 1345 | all_gather_tensors=all_gather_tensors, |
| 1346 | ) |
| 1347 | self._pp_send_work = handles[1:] |
| 1348 | |
| 1349 | if self.use_v2_model_runner and self.model_runner.is_pooling_model: |
| 1350 | return self.model_runner.pool() # type: ignore |
| 1351 | return None |
| 1352 | |
| 1353 | def take_draft_token_ids(self) -> DraftTokenIds | None: |
| 1354 | return self.model_runner.take_draft_token_ids() |
| 1355 | |
| 1356 | def profile(self, is_start: bool = True, profile_prefix: str | None = None): |
| 1357 | # Check if profiling is enabled |
| 1358 | if self.profiler_config is None or self.profiler_config.profiler is None: |
| 1359 | raise RuntimeError( |
| 1360 | "Profiling is not enabled. Please set --profiler-config to enable " |
| 1361 | "profiling. Example: " |
| 1362 | "'--profiler-config.profiler=torch --profiler-config.torch_profiler_dir" |
| 1363 | "=YOUR_DIR_PATH_TO_DUMP_TRACE'" |
| 1364 | ) |
| 1365 | |
| 1366 | if is_start: |
| 1367 | # Generate the trace name by combining prefix with comprehensive rank suffix |
| 1368 | from vllm.distributed.utils import get_worker_rank_suffix |
| 1369 | |
| 1370 | rank_suffix = get_worker_rank_suffix(global_rank=self.rank) |
| 1371 | |
| 1372 | # Build the full trace name |
| 1373 | if profile_prefix: |
| 1374 | trace_name = f"{profile_prefix}_{rank_suffix}" |
| 1375 | else: |
| 1376 | trace_name = rank_suffix |
| 1377 | |
| 1378 | if self.profiler_config.profiler == "proton" and self.profiler is not None: |
| 1379 | self.profiler.set_output_name(trace_name) |
| 1380 | |
| 1381 | # Create the profiler wrapper only on the first start call |
| 1382 | if self.profiler is None: |
| 1383 | self.profiler = create_worker_profiler( |
| 1384 | self.profiler_config, |
| 1385 | worker_name=trace_name, |
| 1386 | local_rank=self.local_rank, |
| 1387 | ) |
| 1388 | |
| 1389 | self.profiler.start() |
| 1390 | else: |
| 1391 | if self.profiler is None: |
| 1392 | logger.warning("Profiler was not started, nothing to stop.") |
| 1393 | return |
| 1394 | try: |
| 1395 | self.profiler.stop() |
| 1396 | finally: |
| 1397 | if self.profiler_config.profiler == "proton" and not ( |
| 1398 | self.profiler.has_cuda_graph_session |
| 1399 | ): |
| 1400 | # Proton output names are fixed when the wrapper is constructed. |
| 1401 | # Recreate it so the next profile_prefix is honored. |
| 1402 | self.profiler = None |
| 1403 | |
| 1404 | @property |
| 1405 | def randomize_dummy_inputs(self) -> bool: |
| 1406 | # Not cached: elastic EP rewrites parallel_config in place on reconfigure. |
| 1407 | return ( |
| 1408 | envs.VLLM_RANDOMIZE_DP_DUMMY_INPUTS |
| 1409 | and self.parallel_config.data_parallel_size > 1 |
| 1410 | ) |
| 1411 | |
| 1412 | def execute_dummy_batch(self) -> None: |
| 1413 | num_tokens = getattr(self.model_runner, "uniform_decode_query_len", 1) |
| 1414 | self.model_runner._dummy_run( |
| 1415 | num_tokens, |
| 1416 | uniform_decode=True, |
| 1417 | randomize_inputs=self.randomize_dummy_inputs, |
| 1418 | ) |
| 1419 | |
| 1420 | def add_lora(self, lora_request: LoRARequest) -> bool: |
| 1421 | return self.model_runner.add_lora(lora_request) |
| 1422 | |
| 1423 | def remove_lora(self, lora_id: int) -> bool: |
| 1424 | return self.model_runner.remove_lora(lora_id) |
| 1425 | |
| 1426 | def list_loras(self) -> set[int]: |
| 1427 | return self.model_runner.list_loras() |
| 1428 | |
| 1429 | def pin_lora(self, lora_id: int) -> bool: |
| 1430 | return self.model_runner.pin_lora(lora_id) |
| 1431 | |
| 1432 | def check_health(self) -> None: |
| 1433 | # worker will always be healthy as long as it's running. |
| 1434 | return |
| 1435 | |
| 1436 | def save_sharded_state( |
| 1437 | self, |
| 1438 | path: str, |
| 1439 | pattern: str | None = None, |
| 1440 | max_size: int | None = None, |
| 1441 | ) -> None: |
| 1442 | from vllm.model_executor.model_loader import ShardedStateLoader |
| 1443 | |
| 1444 | ShardedStateLoader.save_model( |
| 1445 | self.model_runner.model, |
| 1446 | path, |
| 1447 | pattern=pattern, |
| 1448 | max_size=max_size, |
| 1449 | ) |
| 1450 | |
| 1451 | def save_tensorized_model(self, tensorizer_config: "TensorizerConfig") -> None: |
| 1452 | TensorizerLoader.save_model( |
| 1453 | self.get_model(), |
| 1454 | tensorizer_config=tensorizer_config, |
| 1455 | model_config=self.model_config, |
| 1456 | ) |
| 1457 | |
| 1458 | def _check_weight_transfer_engine(self) -> None: |
| 1459 | if self.weight_transfer_engine is None: |
| 1460 | raise RuntimeError( |
| 1461 | "Weight transfer not configured. " |
| 1462 | "Please set weight_transfer_config to enable weight transfer." |
| 1463 | ) |
| 1464 | |
| 1465 | def init_weight_transfer_engine(self, init_info: dict) -> None: |
| 1466 | """Initialize weight transfer mechanism. |
| 1467 | For NCCL backend, this creates a process group with the trainer. |
| 1468 | |
| 1469 | Args: |
| 1470 | init_info: Dictionary containing backend-specific initialization info |
| 1471 | |
| 1472 | """ |
| 1473 | self._check_weight_transfer_engine() |
| 1474 | assert self.weight_transfer_engine is not None |
| 1475 | # Parse dict into backend-specific typed dataclass |
| 1476 | typed_init_info = self.weight_transfer_engine.parse_init_info(init_info) |
| 1477 | self.weight_transfer_engine.init_transfer_engine(typed_init_info) |
| 1478 | |
| 1479 | def start_weight_update(self) -> None: |
| 1480 | """Start a new weight update session. |
| 1481 | |
| 1482 | Delegates engine-specific preparation (e.g. layerwise reload setup) to |
| 1483 | the configured weight transfer engine. The worker only tracks that a |
| 1484 | session is active. |
| 1485 | """ |
| 1486 | with set_current_vllm_config(self.vllm_config): |
| 1487 | self._start_weight_update() |
| 1488 | |
| 1489 | def start_draft_weight_update(self) -> None: |
| 1490 | """Like start_weight_update, but retargets the engine at the speculative |
| 1491 | draft model for this session. |
| 1492 | """ |
| 1493 | with set_current_vllm_config(self.vllm_config): |
| 1494 | self._start_weight_update(is_draft=True) |
| 1495 | |
| 1496 | def _start_weight_update(self, is_draft: bool = False) -> None: |
| 1497 | self._check_weight_transfer_engine() |
| 1498 | assert self.weight_transfer_engine is not None |
| 1499 | |
| 1500 | if is_draft and not self.weight_transfer_engine.supports_draft_weight_update: |
| 1501 | raise RuntimeError( |
| 1502 | f"{type(self.weight_transfer_engine).__name__} does not support " |
| 1503 | "draft model weight updates." |
| 1504 | ) |
| 1505 | |
| 1506 | if self._weight_update_active: |
| 1507 | raise RuntimeError( |
| 1508 | "start_weight_update called while a weight update is already " |
| 1509 | "active. Call finish_weight_update first." |
| 1510 | ) |
| 1511 | |
| 1512 | try: |
| 1513 | if is_draft: |
| 1514 | self._set_draft_weight_update_target() |
| 1515 | self.weight_transfer_engine.start_weight_update() |
| 1516 | except BaseException: |
| 1517 | self.weight_transfer_engine.reset_weight_update_target() |
| 1518 | raise |
| 1519 | self._weight_update_active = True |
| 1520 | self._weight_update_is_draft = is_draft |
| 1521 | |
| 1522 | def update_weights(self, update_info: dict | list[dict]) -> None: |
| 1523 | """Receive one weight update chunk from the trainer. |
| 1524 | |
| 1525 | start_weight_update must be called before update_weights and |
| 1526 | finish_weight_update must be called after all chunks have been sent. |
| 1527 | Every chunk loads into whichever model the session's start_weight_update |
| 1528 | / start_draft_weight_update call selected. |
| 1529 | |
| 1530 | Args: |
| 1531 | update_info: Backend-specific update info, or a list indexed by |
| 1532 | global worker rank across data parallel replicas. |
| 1533 | |
| 1534 | """ |
| 1535 | self._check_weight_transfer_engine() |
| 1536 | assert self.weight_transfer_engine is not None |
| 1537 | |
| 1538 | if not self._weight_update_active: |
| 1539 | raise RuntimeError( |
| 1540 | "start_weight_update must be called before update_weights." |
| 1541 | ) |
| 1542 | |
| 1543 | with set_current_vllm_config(self.vllm_config): |
| 1544 | try: |
| 1545 | if isinstance(update_info, list): |
| 1546 | parallel_config = self.vllm_config.parallel_config |
| 1547 | local_update_info = update_info[ |
| 1548 | parallel_config.data_parallel_index * parallel_config.world_size |
| 1549 | + self.rank |
| 1550 | ] |
| 1551 | else: |
| 1552 | local_update_info = update_info |
| 1553 | self.weight_transfer_engine.update_weights(local_update_info) |
| 1554 | except BaseException: |
| 1555 | self._weight_update_active = False |
| 1556 | self.weight_transfer_engine.reset_weight_update_target() |
| 1557 | raise |
| 1558 | |
| 1559 | def finish_weight_update(self) -> None: |
| 1560 | """Finish the current weight update session.""" |
| 1561 | self._check_weight_transfer_engine() |
| 1562 | assert self.weight_transfer_engine is not None |
| 1563 | |
| 1564 | if not self._weight_update_active: |
| 1565 | raise RuntimeError( |
| 1566 | "finish_weight_update called without a matching start_weight_update." |
| 1567 | ) |
| 1568 | |
| 1569 | with set_current_vllm_config(self.vllm_config): |
| 1570 | self.weight_transfer_engine.finish_weight_update() |
| 1571 | self.weight_transfer_engine.reset_weight_update_target() |
| 1572 | self._weight_update_active = False |
| 1573 | |
| 1574 | # Weight transfer bypasses GPUModelRunner.reload_weights(). |
| 1575 | if not self._weight_update_is_draft: |
| 1576 | self.model_runner.reset_lora_state() |
| 1577 | |
| 1578 | def shutdown(self) -> None: |
| 1579 | gc.unfreeze() |
| 1580 | |
| 1581 | # has_kv_transfer_group can be None during interpreter shutdown. |
| 1582 | if ensure_kv_transfer_shutdown is not None: |
| 1583 | ensure_kv_transfer_shutdown() |
| 1584 | if ensure_ec_transfer_shutdown is not None: |
| 1585 | ensure_ec_transfer_shutdown() |
| 1586 | if self.profiler is not None: |
| 1587 | self.profiler.shutdown() |
| 1588 | |
| 1589 | if weight_transfer_engine := getattr(self, "weight_transfer_engine", None): |
| 1590 | weight_transfer_engine.shutdown() |
| 1591 | |
| 1592 | self.elastic_ep_executor.shutdown() |
| 1593 | |
| 1594 | # Release GPU resources held by the model runner so that memory |
| 1595 | # can be reclaimed when running in-process |
| 1596 | if model_runner := getattr(self, "model_runner", None): |
| 1597 | model_runner.shutdown() |
| 1598 | |
| 1599 | # Release kept-alive cumem pools while the pluggable allocator wrappers |
| 1600 | # and callbacks are still alive, so MemPool teardown is not deferred to |
| 1601 | # interpreter finalization (pytorch/pytorch#145168). |
| 1602 | if current_platform.is_cuda_alike(): |
| 1603 | from vllm.device_allocator.cumem import CuMemAllocator |
| 1604 | |
| 1605 | if CuMemAllocator.instance is not None: |
| 1606 | CuMemAllocator.instance.release_pools() |
| 1607 | |
| 1608 | def elastic_ep_execute(self, execute_method: str, *args, **kwargs): |
| 1609 | return self.elastic_ep_executor.execute(execute_method, *args, **kwargs) |
| 1610 | |
| 1611 | |
| 1612 | def init_worker_distributed_environment( |
| 1613 | vllm_config: VllmConfig, |
| 1614 | rank: int, |
| 1615 | distributed_init_method: str | None = None, |
| 1616 | local_rank: int = -1, |
| 1617 | backend: str = "nccl", |
| 1618 | ) -> None: |
| 1619 | """Initialize the distributed environment.""" |
| 1620 | parallel_config = vllm_config.parallel_config |
| 1621 | from vllm.model_executor.determinism.batch_invariant import init_batch_invariance |
| 1622 | |
| 1623 | init_batch_invariance() |
| 1624 | override_envs_for_eplb( |
| 1625 | parallel_config, |
| 1626 | moe_backend=getattr(vllm_config.kernel_config, "moe_backend", None), |
| 1627 | ) |
| 1628 | set_custom_all_reduce(not parallel_config.disable_custom_all_reduce) |
| 1629 | |
| 1630 | init_method = distributed_init_method or "env://" |
| 1631 | |
| 1632 | timeout = None |
| 1633 | if parallel_config.distributed_timeout_seconds is not None: |
| 1634 | timeout = timedelta(seconds=parallel_config.distributed_timeout_seconds) |
| 1635 | |
| 1636 | init_distributed_environment( |
| 1637 | parallel_config.world_size, |
| 1638 | rank, |
| 1639 | init_method, |
| 1640 | local_rank, |
| 1641 | backend, |
| 1642 | timeout, |
| 1643 | ) |
| 1644 | |
| 1645 | ensure_model_parallel_initialized( |
| 1646 | parallel_config.tensor_parallel_size, |
| 1647 | parallel_config.pipeline_parallel_size, |
| 1648 | parallel_config.prefill_context_parallel_size, |
| 1649 | parallel_config.decode_context_parallel_size, |
| 1650 | ) |
| 1651 | |
| 1652 | # Init ec connector here before KV caches init |
| 1653 | # NOTE: We do not init KV caches for Encoder-only instance in EPD disagg mode |
| 1654 | ensure_ec_transfer_initialized(vllm_config) |