| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import inspect |
| 5 | import os |
| 6 | from collections.abc import Callable, Hashable, Iterator |
| 7 | from contextlib import AbstractContextManager, contextmanager, nullcontext |
| 8 | from contextvars import ContextVar |
| 9 | from itertools import accumulate |
| 10 | from math import prod |
| 11 | from typing import TypeVar, cast |
| 12 | |
| 13 | import torch |
| 14 | |
| 15 | import vllm.envs as envs |
| 16 | from vllm.logger import init_logger |
| 17 | from vllm.utils.math_utils import round_up |
| 18 | from vllm.v1.worker.ubatching import dbo_current_ubatch_id |
| 19 | |
| 20 | logger = init_logger(__name__) |
| 21 | T = TypeVar("T") |
| 22 | |
| 23 | |
| 24 | def _compute_bytes(shape: tuple[int, ...], dtype: torch.dtype) -> int: |
| 25 | return prod(shape) * dtype.itemsize |
| 26 | |
| 27 | |
| 28 | # Constants |
| 29 | _MB = 1024**2 |
| 30 | _GiB = 1024**3 |
| 31 | |
| 32 | # Global workspace manager instance |
| 33 | _manager: "WorkspaceManager | None" = None |
| 34 | _workspace_lane: ContextVar[int] = ContextVar("vllm_workspace_lane", default=0) |
| 35 | |
| 36 | |
| 37 | @contextmanager |
| 38 | def use_workspace_lane(lane: int) -> Iterator[None]: |
| 39 | """Select an independent workspace owner for this execution context.""" |
| 40 | if lane < 0: |
| 41 | raise ValueError(f"Workspace lane must be non-negative, got {lane}.") |
| 42 | token = _workspace_lane.set(lane) |
| 43 | try: |
| 44 | yield |
| 45 | finally: |
| 46 | _workspace_lane.reset(token) |
| 47 | |
| 48 | |
| 49 | class WorkspaceManager: |
| 50 | """Manager for workspace allocation. |
| 51 | |
| 52 | Manages shared scratch and keyed persistent resources per ``(ubatch, lane)``. |
| 53 | Can be locked to prevent growth or new resources during execution. |
| 54 | """ |
| 55 | |
| 56 | def __init__( |
| 57 | self, |
| 58 | device: torch.device, |
| 59 | num_ubatches: int | None = None, |
| 60 | num_lanes: int = 1, |
| 61 | alloc_context: Callable[[], AbstractContextManager] = nullcontext, |
| 62 | ): |
| 63 | self._device = device |
| 64 | self._alloc_context = alloc_context |
| 65 | # Cache num ubatches at init based on configuration (default to 1) |
| 66 | self._num_ubatches = num_ubatches if num_ubatches is not None else 1 |
| 67 | if num_lanes < 1: |
| 68 | raise ValueError(f"num_lanes must be at least one, got {num_lanes}.") |
| 69 | self._num_lanes = num_lanes |
| 70 | self._current_workspaces: list[torch.Tensor | None] = [None] * ( |
| 71 | self._num_ubatches * self._num_lanes |
| 72 | ) |
| 73 | self._persistent_resources: list[dict[Hashable, object]] = [ |
| 74 | {} for _ in self._current_workspaces |
| 75 | ] |
| 76 | self._locked: bool = False |
| 77 | |
| 78 | @staticmethod |
| 79 | def _workspace_size_bytes(workspace: torch.Tensor | None) -> int: |
| 80 | """Get size of workspace in bytes.""" |
| 81 | if workspace is None: |
| 82 | return 0 |
| 83 | return workspace.numel() * workspace.element_size() |
| 84 | |
| 85 | def lock(self) -> None: |
| 86 | """Lock the workspace to prevent growth or new persistent resources. |
| 87 | |
| 88 | Larger scratch requests and new persistent resources raise an assertion |
| 89 | error. Existing resources retain their storage throughout execution. |
| 90 | """ |
| 91 | self._locked = True |
| 92 | if envs.VLLM_DEBUG_WORKSPACE: |
| 93 | logger.info( |
| 94 | "[WORKSPACE DEBUG] Workspace locked. Current sizes: %s", |
| 95 | [ |
| 96 | self._workspace_size_bytes(ws) / _MB |
| 97 | for ws in self._current_workspaces |
| 98 | if ws is not None |
| 99 | ], |
| 100 | ) |
| 101 | |
| 102 | def unlock(self) -> None: |
| 103 | """Unlock the workspace to allow growth. |
| 104 | |
| 105 | This is used during elastic EP scaling when the workspace size |
| 106 | needs to grow due to changes in the number of experts. |
| 107 | """ |
| 108 | self._locked = False |
| 109 | if envs.VLLM_DEBUG_WORKSPACE: |
| 110 | logger.info( |
| 111 | "[WORKSPACE DEBUG] Workspace unlocked. Current sizes: %s", |
| 112 | [ |
| 113 | self._workspace_size_bytes(ws) / _MB |
| 114 | for ws in self._current_workspaces |
| 115 | if ws is not None |
| 116 | ], |
| 117 | ) |
| 118 | |
| 119 | def is_locked(self) -> bool: |
| 120 | """Check if workspace is locked.""" |
| 121 | return self._locked |
| 122 | |
| 123 | def _get_workspace_id(self) -> int: |
| 124 | lane = _workspace_lane.get() |
| 125 | if lane >= self._num_lanes: |
| 126 | raise RuntimeError( |
| 127 | f"Workspace lane {lane} is not configured; manager has " |
| 128 | f"{self._num_lanes} lane(s)." |
| 129 | ) |
| 130 | return dbo_current_ubatch_id() * self._num_lanes + lane |
| 131 | |
| 132 | def get_persistent_resource(self, key: Hashable, factory: Callable[[], T]) -> T: |
| 133 | """Return the resource cached for ``key`` in the current ubatch and lane. |
| 134 | |
| 135 | The first request calls ``factory`` and stores its result for the lifetime |
| 136 | of this manager. Later requests return the same object without calling |
| 137 | the factory again. Each ubatch/lane pair has a separate cache. Once the |
| 138 | manager is locked, requesting a key missing from that cache raises an |
| 139 | AssertionError. |
| 140 | |
| 141 | Keys must identify the resource type and configuration. Callers must not |
| 142 | resize or replace tensor storage held by a cached resource. Uses of the |
| 143 | same resource must not overlap; include the CUDA stream in the key when |
| 144 | streams can execute concurrently. |
| 145 | |
| 146 | With torch.compile, call this inside a runtime custom op so resource |
| 147 | selection happens during execution. |
| 148 | """ |
| 149 | resources = self._persistent_resources[self._get_workspace_id()] |
| 150 | if key not in resources: |
| 151 | if self._locked: |
| 152 | raise AssertionError( |
| 153 | f"Workspace is locked but {key!r} was not allocated during " |
| 154 | "warmup for the current ubatch and lane." |
| 155 | ) |
| 156 | resources[key] = factory() |
| 157 | return cast(T, resources[key]) |
| 158 | |
| 159 | def get_persistent( |
| 160 | self, |
| 161 | key: Hashable, |
| 162 | shape: tuple[int, ...], |
| 163 | dtype: torch.dtype, |
| 164 | *, |
| 165 | zero_init: bool = False, |
| 166 | ) -> torch.Tensor: |
| 167 | """Get a fixed-shape tensor, optionally zeroed on first allocation only.""" |
| 168 | factory = torch.zeros if zero_init else torch.empty |
| 169 | workspace = self.get_persistent_resource( |
| 170 | (torch.Tensor, key), |
| 171 | lambda: factory(shape, dtype=dtype, device=self._device), |
| 172 | ) |
| 173 | if workspace.shape != shape or workspace.dtype != dtype: |
| 174 | raise ValueError( |
| 175 | f"Persistent workspace {key!r} has shape={workspace.shape}, " |
| 176 | f"dtype={workspace.dtype}; requested shape={shape}, dtype={dtype}." |
| 177 | ) |
| 178 | return workspace |
| 179 | |
| 180 | def get_simultaneous( |
| 181 | self, *shapes_and_dtypes: tuple[tuple[int, ...], torch.dtype] |
| 182 | ) -> list[torch.Tensor]: |
| 183 | """Get multiple workspace tensors simultaneously from a single allocation. |
| 184 | |
| 185 | Args: |
| 186 | *shapes_and_dtypes: One or more (shape, dtype) tuples. |
| 187 | |
| 188 | Returns: |
| 189 | List of tensor views into the workspace buffer, one per shape/dtype pair. |
| 190 | |
| 191 | """ |
| 192 | actual_bytes = [_compute_bytes(s, d) for s, d in shapes_and_dtypes] |
| 193 | aligned_bytes = [round_up(actual, 256) for actual in actual_bytes] |
| 194 | total_bytes = sum(aligned_bytes) |
| 195 | |
| 196 | # Calculate cumulative offsets using itertools.accumulate |
| 197 | offsets = list(accumulate([0] + aligned_bytes[:-1])) |
| 198 | |
| 199 | current_workspace = self._ensure_workspace_size(total_bytes) |
| 200 | |
| 201 | return [ |
| 202 | current_workspace[offsets[i] : offsets[i] + actual_bytes[i]] |
| 203 | .view(shapes_and_dtypes[i][1]) |
| 204 | .reshape(shapes_and_dtypes[i][0]) |
| 205 | for i in range(len(shapes_and_dtypes)) |
| 206 | ] |
| 207 | |
| 208 | def _ensure_workspace_size(self, required_bytes: int) -> torch.Tensor: |
| 209 | """Ensure workspace is allocated and large enough, return current workspace. |
| 210 | |
| 211 | Args: |
| 212 | required_bytes: The number of bytes required. |
| 213 | |
| 214 | Returns: |
| 215 | The current workspace tensor. |
| 216 | |
| 217 | """ |
| 218 | workspace_id = self._get_workspace_id() |
| 219 | current_workspace = self._current_workspaces[workspace_id] |
| 220 | current_size = self._workspace_size_bytes(current_workspace) |
| 221 | |
| 222 | if current_size < required_bytes: |
| 223 | |
| 224 | def get_caller_info() -> str: |
| 225 | """Find first frame outside WorkspaceManager.""" |
| 226 | curr_frame = inspect.currentframe() |
| 227 | if curr_frame is None: |
| 228 | return "unknown" |
| 229 | # Walk up the stack skipping WorkspaceManager frames |
| 230 | curr_frame = curr_frame.f_back |
| 231 | while curr_frame is not None: |
| 232 | # TODO: This only catches instance methods (self), missing |
| 233 | # classmethods and staticmethods. Once Python 3.11+ is the |
| 234 | # minimum supported version, use co_qualname instead: |
| 235 | # qualname = curr_frame.f_code.co_qualname |
| 236 | # if qualname.startswith("WorkspaceManager."): |
| 237 | if isinstance(curr_frame.f_locals.get("self"), WorkspaceManager): |
| 238 | curr_frame = curr_frame.f_back |
| 239 | continue |
| 240 | filename = os.path.basename(curr_frame.f_code.co_filename) |
| 241 | return ( |
| 242 | f"{filename}:{curr_frame.f_lineno}:{curr_frame.f_code.co_name}" |
| 243 | ) |
| 244 | return "unknown" |
| 245 | |
| 246 | if self._locked: |
| 247 | raise AssertionError( |
| 248 | f"Workspace is locked but allocation from '{get_caller_info()}' " |
| 249 | f"requires {required_bytes / _MB:.2f} MB, current size is " |
| 250 | f"{current_size / _MB:.2f} MB. " |
| 251 | "Workspace growth is not allowed after locking." |
| 252 | ) |
| 253 | |
| 254 | # Only resize the requesting ubatch/lane workspace. Other slots |
| 255 | # resize lazily on their next get_simultaneous call. |
| 256 | # Resizing all ubatches here would orphan the other ubatch's |
| 257 | # old tensor when it still holds views into it (DBO leak). |
| 258 | self._current_workspaces[workspace_id] = None |
| 259 | del current_workspace |
| 260 | # Release the freed segment back to CUDA so the caching |
| 261 | # allocator can reuse the GPU memory for the larger |
| 262 | # allocation below. Without this, each resize may leave a |
| 263 | # dead segment in reserved memory which can cause higher peak |
| 264 | # memory usage. |
| 265 | torch.accelerator.empty_cache() |
| 266 | # In sleep mode, entering the pool also trims outgrown buffers. |
| 267 | with self._alloc_context(): |
| 268 | self._current_workspaces[workspace_id] = torch.empty( |
| 269 | (required_bytes,), dtype=torch.uint8, device=self._device |
| 270 | ) |
| 271 | current_workspace = self._current_workspaces[workspace_id] |
| 272 | |
| 273 | if envs.VLLM_DEBUG_WORKSPACE: |
| 274 | logger.info( |
| 275 | "[WORKSPACE DEBUG] Resized workspace from '%s': %.2f MB -> " |
| 276 | "%.2f MB (ubatch %d, lane %d)", |
| 277 | get_caller_info(), |
| 278 | current_size / _MB, |
| 279 | required_bytes / _MB, |
| 280 | workspace_id // self._num_lanes, |
| 281 | workspace_id % self._num_lanes, |
| 282 | ) |
| 283 | |
| 284 | return current_workspace |
| 285 | |
| 286 | |
| 287 | def is_workspace_manager_initialized() -> bool: |
| 288 | """Check if workspace manager has been initialized. |
| 289 | |
| 290 | Returns: |
| 291 | True if workspace manager is initialized, False otherwise. |
| 292 | |
| 293 | """ |
| 294 | return _manager is not None |
| 295 | |
| 296 | |
| 297 | def current_workspace_manager() -> "WorkspaceManager": |
| 298 | """Get the current workspace manager instance. |
| 299 | |
| 300 | Raises: |
| 301 | AssertionError: If workspace manager has not been initialized. |
| 302 | |
| 303 | """ |
| 304 | assert _manager is not None, ( |
| 305 | "WorkspaceManager not initialized. Call init_workspace_manager() " |
| 306 | "with a device before using workspace functions." |
| 307 | ) |
| 308 | return _manager |
| 309 | |
| 310 | |
| 311 | def init_workspace_manager( |
| 312 | device: torch.device, |
| 313 | num_ubatches: int | None = None, |
| 314 | num_lanes: int = 1, |
| 315 | alloc_context: Callable[[], AbstractContextManager] = nullcontext, |
| 316 | ) -> None: |
| 317 | """Initialize the workspace manager with a device. |
| 318 | |
| 319 | Must be called before using any workspace functions. Typically called |
| 320 | from GPUModelRunner.__init__. |
| 321 | |
| 322 | Args: |
| 323 | device: The device to allocate workspace on. |
| 324 | num_ubatches: Number of workspace ubatch slots. Defaults to 1. |
| 325 | num_lanes: Number of independent execution lanes per ubatch. Defaults to 1. |
| 326 | alloc_context: Factory for the context scratch is allocated in. |
| 327 | |
| 328 | """ |
| 329 | global _manager |
| 330 | if _manager is not None: |
| 331 | logger.warning( |
| 332 | "WorkspaceManager already initialized on device %s, " |
| 333 | "reinitializing on device %s", |
| 334 | _manager._device, |
| 335 | device, |
| 336 | ) |
| 337 | _manager = WorkspaceManager(device, num_ubatches, num_lanes, alloc_context) |
| 338 | |
| 339 | |
| 340 | def lock_workspace() -> None: |
| 341 | """Lock the workspace to prevent growth or new persistent resources. |
| 342 | |
| 343 | Larger scratch requests and new persistent resources raise an AssertionError. |
| 344 | All required resources must be initialized in each execution slot during |
| 345 | warmup, so their storage remains fixed during execution. |
| 346 | |
| 347 | Example: |
| 348 | # During initialization |
| 349 | init_workspace_manager(device) |
| 350 | reserve_workspace(shape1, dtype1) |
| 351 | reserve_workspace(shape2, dtype2) |
| 352 | |
| 353 | # Lock after warmup/profiling |
| 354 | lock_workspace() |
| 355 | |
| 356 | # Now all get_workspace calls must fit in pre-allocated size |
| 357 | |
| 358 | """ |
| 359 | current_workspace_manager().lock() |
| 360 | |
| 361 | |
| 362 | def unlock_workspace() -> None: |
| 363 | """Unlock the workspace to allow growth. |
| 364 | |
| 365 | This is used during elastic EP scaling when the workspace size |
| 366 | needs to grow due to changes in the number of experts. |
| 367 | After scaling operations complete, lock_workspace() should be |
| 368 | called again to prevent unexpected allocations. |
| 369 | """ |
| 370 | current_workspace_manager().unlock() |
| 371 | |
| 372 | |
| 373 | def reset_workspace_manager() -> None: |
| 374 | """Reset the workspace manager to uninitialized state. |
| 375 | |
| 376 | This is primarily intended for testing purposes to allow tests |
| 377 | to reinitialize the workspace manager cleanly. |
| 378 | """ |
| 379 | global _manager |
| 380 | _manager = None |