| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from collections import OrderedDict |
| 4 | from collections.abc import Collection, Iterable |
| 5 | from dataclasses import dataclass, field |
| 6 | |
| 7 | from typing_extensions import override |
| 8 | |
| 9 | from vllm.distributed.kv_transfer.kv_connector.v1.offloading.metrics import ( |
| 10 | OffloadingConnectorStats, |
| 11 | ) |
| 12 | from vllm.v1.kv_offload.base import ( |
| 13 | LoadStoreSpec, |
| 14 | LookupResult, |
| 15 | Medium, |
| 16 | OffloadingEvent, |
| 17 | OffloadingManager, |
| 18 | OffloadKey, |
| 19 | PrepareStoreOutput, |
| 20 | ReqContext, |
| 21 | RequestOffloadingContext, |
| 22 | get_offload_group_idx, |
| 23 | ) |
| 24 | from vllm.v1.kv_offload.cpu.common import ( |
| 25 | CPULoadStoreSpec, |
| 26 | CPUOffloadingMetrics, |
| 27 | ) |
| 28 | from vllm.v1.kv_offload.cpu.policies.base import CachePolicy, ChunkStatus |
| 29 | from vllm.v1.kv_offload.cpu.policies.factory import CachePolicyFactory |
| 30 | |
| 31 | |
| 32 | @dataclass(slots=True) |
| 33 | class _RequestCacheAccess: |
| 34 | """Cache keys observed by one request, grouped in prefix order.""" |
| 35 | |
| 36 | owner: object |
| 37 | cache_generation: int |
| 38 | key_groups: dict[int, list[OffloadKey]] = field(default_factory=dict) |
| 39 | seen_keys: set[OffloadKey] = field(default_factory=set) |
| 40 | inserted_keys: set[OffloadKey] = field(default_factory=set) |
| 41 | reused_keys: set[OffloadKey] = field(default_factory=set) |
| 42 | store_miss_keys: set[OffloadKey] = field(default_factory=set) |
| 43 | finished: bool = False |
| 44 | |
| 45 | |
| 46 | class CPUOffloadingManager(OffloadingManager): |
| 47 | """An OffloadingManager with a pluggable CachePolicy, resolved by name via |
| 48 | CachePolicyFactory (built in: "lru", "arc"; external policies can either |
| 49 | register their own or be loaded out-of-tree via cache_policy_module_path). |
| 50 | |
| 51 | The manager owns all shared logic: ref-counting, event emission, |
| 52 | chunk pool management, and the prepare_store/complete_store skeletons. |
| 53 | Policy-specific chunk organization and eviction decisions are delegated |
| 54 | to the CachePolicy implementation. |
| 55 | """ |
| 56 | |
| 57 | def __init__( |
| 58 | self, |
| 59 | num_chunks: int, |
| 60 | cache_policy: str = "lru", |
| 61 | cache_policy_module_path: str | None = None, |
| 62 | enable_events: bool = False, |
| 63 | store_threshold: int = 1, |
| 64 | max_tracker_size: int = 64_000, |
| 65 | ): |
| 66 | self.medium: Medium = Medium.CPU |
| 67 | self._num_chunks: int = num_chunks |
| 68 | self._num_allocated_chunks: int = 0 |
| 69 | self._free_list: list[int] = [] |
| 70 | self.events: list[OffloadingEvent] | None = [] if enable_events else None |
| 71 | policy_cls = CachePolicyFactory.get_cache_policy_cls( |
| 72 | cache_policy, cache_policy_module_path |
| 73 | ) |
| 74 | self._policy: CachePolicy = policy_cls(cache_capacity=num_chunks) |
| 75 | # Track the number of chunks in the cache that are evictable. i.e. ref_cnt 0. |
| 76 | self._num_evictable_cache_chunks: int = 0 |
| 77 | # Track chunks with an in-flight store (ref_cnt -1, not yet completed). |
| 78 | self._num_write_pending_chunks: int = 0 |
| 79 | |
| 80 | self.store_threshold: int = store_threshold |
| 81 | self.max_tracker_size: int = max_tracker_size |
| 82 | self.stores_skipped_in_current_batch: int = 0 |
| 83 | self.allocation_sizes_in_current_batch: list[int] = [] |
| 84 | self._cache_generation = 0 |
| 85 | |
| 86 | # Number of chunk references. It is ordered so can evict the LRU entry in O(1). |
| 87 | self.counts: OrderedDict[OffloadKey, int] | None = ( |
| 88 | OrderedDict() if store_threshold >= 2 else None |
| 89 | ) |
| 90 | |
| 91 | # --- chunk pool --- |
| 92 | |
| 93 | def _get_num_free_chunks(self) -> int: |
| 94 | return len(self._free_list) + self._num_chunks - self._num_allocated_chunks |
| 95 | |
| 96 | def _allocate_chunks(self, keys: list[OffloadKey]) -> list[ChunkStatus]: |
| 97 | num_fresh = min(len(keys), self._num_chunks - self._num_allocated_chunks) |
| 98 | num_reused = len(keys) - num_fresh |
| 99 | assert len(self._free_list) >= num_reused |
| 100 | |
| 101 | # allocate fresh chunks |
| 102 | chunks: list[ChunkStatus] = [] |
| 103 | for _ in range(num_fresh): |
| 104 | chunks.append(ChunkStatus(self._num_allocated_chunks)) |
| 105 | self._num_allocated_chunks += 1 |
| 106 | |
| 107 | # allocate reused chunks |
| 108 | for _ in range(num_reused): |
| 109 | chunks.append(ChunkStatus(self._free_list.pop())) |
| 110 | return chunks |
| 111 | |
| 112 | def _free_chunk(self, chunk: ChunkStatus) -> None: |
| 113 | self._free_list.append(chunk.chunk_id) |
| 114 | |
| 115 | def _get_load_store_spec( |
| 116 | self, |
| 117 | keys: Iterable[OffloadKey], |
| 118 | chunks: Iterable[ChunkStatus], |
| 119 | ) -> CPULoadStoreSpec: |
| 120 | return CPULoadStoreSpec([chunk.chunk_id for chunk in chunks]) |
| 121 | |
| 122 | def _record_accesses(self, keys: Collection[OffloadKey]) -> None: |
| 123 | """Record an offer without evicting its tracked candidates.""" |
| 124 | assert self.counts is not None |
| 125 | protected: set[OffloadKey] = set() |
| 126 | for key in keys: |
| 127 | if key in self.counts: |
| 128 | self.counts.move_to_end(key) |
| 129 | self.counts[key] += 1 |
| 130 | protected.add(key) |
| 131 | |
| 132 | num_unprotected = len(self.counts) - len(protected) |
| 133 | for key in keys: |
| 134 | if key in self.counts: |
| 135 | continue |
| 136 | if len(self.counts) >= self.max_tracker_size: |
| 137 | if num_unprotected == 0: |
| 138 | continue |
| 139 | self.counts.popitem(last=False) |
| 140 | num_unprotected -= 1 |
| 141 | self.counts[key] = 1 |
| 142 | |
| 143 | def _get_request_cache_access(self, req_context: ReqContext) -> _RequestCacheAccess: |
| 144 | state = req_context.get_state(_RequestCacheAccess) |
| 145 | if ( |
| 146 | state is None |
| 147 | or state.owner is not self |
| 148 | or state.cache_generation != self._cache_generation |
| 149 | ): |
| 150 | state = _RequestCacheAccess( |
| 151 | owner=self, cache_generation=self._cache_generation |
| 152 | ) |
| 153 | req_context.set_state(state) |
| 154 | return state |
| 155 | |
| 156 | def _record_request_cache_access( |
| 157 | self, |
| 158 | keys: Iterable[OffloadKey], |
| 159 | req_context: ReqContext, |
| 160 | inserted_keys: Iterable[OffloadKey] = (), |
| 161 | reused_keys: Iterable[OffloadKey] = (), |
| 162 | ) -> None: |
| 163 | state = self._get_request_cache_access(req_context) |
| 164 | for key in keys: |
| 165 | if key in state.seen_keys: |
| 166 | continue |
| 167 | assert not state.finished, ( |
| 168 | "New cache keys observed after request finalization" |
| 169 | ) |
| 170 | group_idx = get_offload_group_idx(key) |
| 171 | state.key_groups.setdefault(group_idx, []).append(key) |
| 172 | state.seen_keys.add(key) |
| 173 | state.inserted_keys.update(inserted_keys) |
| 174 | # Re-reading a chunk inserted by this request is an internal transfer |
| 175 | # (for example, a tiering cascade), not a second cache access. |
| 176 | state.reused_keys.update( |
| 177 | key for key in reused_keys if key not in state.inserted_keys |
| 178 | ) |
| 179 | |
| 180 | # --- OffloadingManager interface --- |
| 181 | |
| 182 | @override |
| 183 | def on_new_request(self, req_context: ReqContext) -> RequestOffloadingContext: |
| 184 | self._get_request_cache_access(req_context) |
| 185 | return RequestOffloadingContext() |
| 186 | |
| 187 | @override |
| 188 | def lookup(self, key: OffloadKey, req_context: ReqContext) -> LookupResult: |
| 189 | chunk = self._policy.get(key) |
| 190 | if chunk is None: |
| 191 | return LookupResult.MISS |
| 192 | if not chunk.is_ready: |
| 193 | return LookupResult.HIT_PENDING |
| 194 | return LookupResult.HIT |
| 195 | |
| 196 | @override |
| 197 | def prepare_load( |
| 198 | self, |
| 199 | keys: Collection[OffloadKey], |
| 200 | req_context: ReqContext, |
| 201 | ) -> LoadStoreSpec: |
| 202 | return self._prepare_load(keys, req_context, record_access=True) |
| 203 | |
| 204 | def _prepare_load( |
| 205 | self, |
| 206 | keys: Collection[OffloadKey], |
| 207 | req_context: ReqContext, |
| 208 | *, |
| 209 | record_access: bool, |
| 210 | ) -> LoadStoreSpec: |
| 211 | chunks = [] |
| 212 | for key in keys: |
| 213 | chunk = self._policy.get(key) |
| 214 | assert chunk is not None, f"Chunk {key!r} not found in cache" |
| 215 | assert chunk.is_ready, f"Chunk {key!r} is not ready for reading" |
| 216 | if chunk.ref_cnt == 0: |
| 217 | self._policy.mark_non_evictable(key) |
| 218 | self._num_evictable_cache_chunks -= 1 # ref_cnt 0 -> 1 |
| 219 | assert self._num_evictable_cache_chunks >= 0 |
| 220 | chunk.ref_cnt += 1 |
| 221 | chunks.append(chunk) |
| 222 | if record_access: |
| 223 | self._record_request_cache_access(keys, req_context, reused_keys=keys) |
| 224 | return self._get_load_store_spec(keys, chunks) |
| 225 | |
| 226 | @override |
| 227 | def touch(self, keys: Collection[OffloadKey], req_context: ReqContext) -> None: |
| 228 | self._policy.touch(keys, req_context) |
| 229 | |
| 230 | @override |
| 231 | def complete_load( |
| 232 | self, keys: Collection[OffloadKey], req_context: ReqContext |
| 233 | ) -> None: |
| 234 | for key in keys: |
| 235 | chunk = self._policy.get(key) |
| 236 | assert chunk is not None, f"Chunk {key!r} not found" |
| 237 | assert chunk.ref_cnt > 0, f"Chunk {key!r} ref_cnt is already 0" |
| 238 | chunk.ref_cnt -= 1 |
| 239 | if chunk.ref_cnt == 0: |
| 240 | self._num_evictable_cache_chunks += 1 # ref_cnt 1 -> 0 |
| 241 | self._policy.mark_evictable(key) |
| 242 | |
| 243 | @override |
| 244 | def prepare_store( |
| 245 | self, |
| 246 | keys: Collection[OffloadKey], |
| 247 | req_context: ReqContext, |
| 248 | ) -> PrepareStoreOutput | None: |
| 249 | keys = list(keys) |
| 250 | if self.counts is not None: |
| 251 | self._record_accesses(keys) |
| 252 | # Partition the original offer before admission filtering. Ready |
| 253 | # resident chunks are request reuses even if their tracker entry was |
| 254 | # aged out; the threshold applies only to new store candidates. |
| 255 | keys_to_store: list[OffloadKey] = [] |
| 256 | ready_existing_keys: list[OffloadKey] = [] |
| 257 | for key in keys: |
| 258 | chunk = self._policy.get(key) |
| 259 | if chunk is None: |
| 260 | keys_to_store.append(key) |
| 261 | else: |
| 262 | if chunk.is_ready: |
| 263 | ready_existing_keys.append(key) |
| 264 | |
| 265 | if self.counts is not None: |
| 266 | num_store_candidates = len(keys_to_store) |
| 267 | keys_to_store = [ |
| 268 | key |
| 269 | for key in keys_to_store |
| 270 | if self.counts.get(key, 0) >= self.store_threshold |
| 271 | ] |
| 272 | self.stores_skipped_in_current_batch += num_store_candidates - len( |
| 273 | keys_to_store |
| 274 | ) |
| 275 | |
| 276 | state = self._get_request_cache_access(req_context) |
| 277 | new_store_misses = [ |
| 278 | key for key in keys_to_store if key not in state.store_miss_keys |
| 279 | ] |
| 280 | if new_store_misses: |
| 281 | # ARC learns from B1/B2 before insert() removes the ghost entry. |
| 282 | # Deduplication makes this one policy observation per request. |
| 283 | self._policy.on_store_miss(new_store_misses, req_context) |
| 284 | state.store_miss_keys.update(new_store_misses) |
| 285 | |
| 286 | recorded_keys = set(keys_to_store) |
| 287 | recorded_keys.update(ready_existing_keys) |
| 288 | # Record once before any allocation-related early return. Store |
| 289 | # candidates are classified as insertions only after allocation succeeds. |
| 290 | self._record_request_cache_access( |
| 291 | (key for key in keys if key in recorded_keys), |
| 292 | req_context, |
| 293 | reused_keys=ready_existing_keys, |
| 294 | ) |
| 295 | |
| 296 | if not keys_to_store: |
| 297 | return PrepareStoreOutput( |
| 298 | keys_to_store=[], |
| 299 | store_spec=self._get_load_store_spec([], []), |
| 300 | evicted_keys=[], |
| 301 | ) |
| 302 | |
| 303 | self.allocation_sizes_in_current_batch.append(len(keys_to_store)) |
| 304 | num_chunks_to_evict = len(keys_to_store) - self._get_num_free_chunks() |
| 305 | |
| 306 | to_evict: list[OffloadKey] = [] |
| 307 | if num_chunks_to_evict > 0: |
| 308 | if num_chunks_to_evict > self._num_evictable_cache_chunks: |
| 309 | # Eviction will fail. |
| 310 | return None |
| 311 | # There is a still a chance for eviction failure as some of the |
| 312 | # idle chunks might be in the protected list. |
| 313 | |
| 314 | # Chunks from the original input are excluded from eviction candidates: |
| 315 | # a chunk that was already stored must remain in the cache after this call. |
| 316 | protected = set(keys) |
| 317 | evicted = self._policy.evict(num_chunks_to_evict, protected) |
| 318 | if evicted is None: |
| 319 | return None |
| 320 | |
| 321 | # cache-policy removes only idle chunks. |
| 322 | self._num_evictable_cache_chunks -= len(evicted) |
| 323 | assert self._num_evictable_cache_chunks >= 0 |
| 324 | |
| 325 | for key, chunk in evicted: |
| 326 | self._free_chunk(chunk) |
| 327 | to_evict.append(key) |
| 328 | |
| 329 | if to_evict and self.events is not None: |
| 330 | self.events.append( |
| 331 | OffloadingEvent( |
| 332 | keys=to_evict, |
| 333 | medium=self.medium, |
| 334 | removed=True, |
| 335 | ) |
| 336 | ) |
| 337 | |
| 338 | chunks = self._allocate_chunks(keys_to_store) |
| 339 | assert len(chunks) == len(keys_to_store), ( |
| 340 | "Chunk pool did not allocate the expected number of chunks" |
| 341 | ) |
| 342 | |
| 343 | for key, chunk in zip(keys_to_store, chunks): |
| 344 | self._policy.insert(key, chunk) |
| 345 | self._num_write_pending_chunks += len(keys_to_store) |
| 346 | state.inserted_keys.update(keys_to_store) |
| 347 | |
| 348 | # build store specs for allocated chunks |
| 349 | store_spec = self._get_load_store_spec(keys_to_store, chunks) |
| 350 | |
| 351 | return PrepareStoreOutput( |
| 352 | keys_to_store=keys_to_store, |
| 353 | store_spec=store_spec, |
| 354 | evicted_keys=to_evict, |
| 355 | ) |
| 356 | |
| 357 | @override |
| 358 | def complete_store( |
| 359 | self, |
| 360 | keys: Collection[OffloadKey], |
| 361 | req_context: ReqContext, |
| 362 | success: bool = True, |
| 363 | ) -> None: |
| 364 | stored_keys: list[OffloadKey] = [] |
| 365 | |
| 366 | if success: |
| 367 | for key in keys: |
| 368 | chunk = self._policy.get(key) |
| 369 | if chunk is not None and not chunk.is_ready: |
| 370 | chunk.ref_cnt = 0 |
| 371 | self._num_write_pending_chunks -= 1 |
| 372 | self._num_evictable_cache_chunks += 1 |
| 373 | self._policy.mark_evictable(key) |
| 374 | stored_keys.append(key) |
| 375 | else: |
| 376 | for key in keys: |
| 377 | chunk = self._policy.get(key) |
| 378 | if chunk is not None and not chunk.is_ready: |
| 379 | self._num_write_pending_chunks -= 1 |
| 380 | self._policy.remove(key) |
| 381 | self._free_chunk(chunk) |
| 382 | |
| 383 | if stored_keys and self.events is not None: |
| 384 | self.events.append( |
| 385 | OffloadingEvent( |
| 386 | keys=stored_keys, |
| 387 | medium=self.medium, |
| 388 | removed=False, |
| 389 | ) |
| 390 | ) |
| 391 | |
| 392 | @override |
| 393 | def on_request_finished(self, req_context: ReqContext) -> None: |
| 394 | state = req_context.get_state(_RequestCacheAccess) |
| 395 | if ( |
| 396 | state is None |
| 397 | or state.owner is not self |
| 398 | or state.cache_generation != self._cache_generation |
| 399 | or state.finished |
| 400 | ): |
| 401 | return |
| 402 | state.finished = True |
| 403 | key_groups = [] |
| 404 | for group_idx in sorted(state.key_groups): |
| 405 | keys = state.key_groups[group_idx] |
| 406 | positions = { |
| 407 | key: position |
| 408 | for key in keys |
| 409 | if (position := req_context.get_offload_key_position(key)) is not None |
| 410 | } |
| 411 | if len(positions) == len(keys): |
| 412 | keys = sorted(keys, key=positions.__getitem__) |
| 413 | key_groups.append(tuple(keys)) |
| 414 | self._policy.on_request_finished( |
| 415 | tuple(key_groups), |
| 416 | state.inserted_keys - state.reused_keys, |
| 417 | state.reused_keys, |
| 418 | req_context, |
| 419 | ) |
| 420 | |
| 421 | @override |
| 422 | def reset_cache(self) -> None: |
| 423 | # Clear ALL chunks unconditionally. The scheduler's _stale_job_threshold |
| 424 | # guarantees that complete_load / complete_store are never called for |
| 425 | # pre-reset jobs, so no lazy cleanup is needed. The scheduler also |
| 426 | # flushes in-flight load job IDs to the workers before any new stores |
| 427 | # can begin, preventing a cross-direction data race on reused offload chunk IDs. |
| 428 | self._policy.clear() |
| 429 | self._cache_generation += 1 |
| 430 | self._num_evictable_cache_chunks = 0 |
| 431 | self._num_write_pending_chunks = 0 |
| 432 | |
| 433 | self._free_list.clear() |
| 434 | self._num_allocated_chunks = 0 |
| 435 | |
| 436 | @override |
| 437 | def take_events(self) -> Iterable[OffloadingEvent]: |
| 438 | if self.events is not None: |
| 439 | yield from self.events |
| 440 | self.events.clear() |
| 441 | |
| 442 | def get_stats(self) -> OffloadingConnectorStats | None: |
| 443 | stats = OffloadingConnectorStats() |
| 444 | |
| 445 | # Compute cache usage. |
| 446 | num_used = ( |
| 447 | self._num_allocated_chunks |
| 448 | - len(self._free_list) |
| 449 | - self._num_evictable_cache_chunks |
| 450 | ) |
| 451 | usage = num_used / self._num_chunks if self._num_chunks > 0 else 0.0 |
| 452 | stats.set_gauge(CPUOffloadingMetrics.CPU_CACHE_USAGE_PERC, usage) |
| 453 | |
| 454 | for allocation_size in self.allocation_sizes_in_current_batch: |
| 455 | stats.observe_histogram( |
| 456 | CPUOffloadingMetrics.CPU_ALLOCATION_SIZE, allocation_size |
| 457 | ) |
| 458 | self.allocation_sizes_in_current_batch.clear() |
| 459 | |
| 460 | write_usage = ( |
| 461 | self._num_write_pending_chunks / self._num_chunks |
| 462 | if self._num_chunks > 0 |
| 463 | else 0.0 |
| 464 | ) |
| 465 | read_usage = max(usage - write_usage, 0.0) |
| 466 | stats.set_gauge(CPUOffloadingMetrics.CPU_CACHE_WRITE_USAGE_PERC, write_usage) |
| 467 | stats.set_gauge(CPUOffloadingMetrics.CPU_CACHE_READ_USAGE_PERC, read_usage) |
| 468 | |
| 469 | if self.store_threshold >= 2: |
| 470 | stats.increase_counter( |
| 471 | CPUOffloadingMetrics.STORES_SKIPPED, |
| 472 | self.stores_skipped_in_current_batch, |
| 473 | ) |
| 474 | self.stores_skipped_in_current_batch = 0 |
| 475 | |
| 476 | return stats |