| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Core abstractions for KV cache offloading in vLLM v1.""" |
| 4 | |
| 5 | from abc import ABC, abstractmethod |
| 6 | from collections.abc import Collection, Iterable, Sequence |
| 7 | from dataclasses import dataclass, field |
| 8 | from enum import Enum, auto |
| 9 | from typing import TYPE_CHECKING, Any, ClassVar, NamedTuple, NewType, TypeVar |
| 10 | |
| 11 | import numpy as np |
| 12 | import torch |
| 13 | |
| 14 | if TYPE_CHECKING: |
| 15 | from vllm.distributed.kv_transfer.kv_connector.v1.offloading.metrics import ( |
| 16 | OffloadingConnectorStats, |
| 17 | ) |
| 18 | |
| 19 | from vllm.v1.kv_hints import KvHintsEnvelope |
| 20 | from vllm.v1.kv_offload.config import OffloadingConfig |
| 21 | |
| 22 | # `OffloadKey` identifies an offloaded block. It combines a block hash with |
| 23 | # its KV cache group index, encoded as raw bytes to avoid tuple GC overhead. |
| 24 | # Use the helper functions below to construct / decompose keys. |
| 25 | OffloadKey = NewType("OffloadKey", bytes) |
| 26 | |
| 27 | |
| 28 | def make_offload_key(block_hash: bytes, group_idx: int) -> OffloadKey: |
| 29 | """Pack a block hash and group index into an `OffloadKey`.""" |
| 30 | return OffloadKey(block_hash + group_idx.to_bytes(4, "big", signed=False)) |
| 31 | |
| 32 | |
| 33 | def get_offload_block_hash(key: OffloadKey) -> bytes: |
| 34 | """Extract the block hash from an `OffloadKey`.""" |
| 35 | return key[:-4] |
| 36 | |
| 37 | |
| 38 | def get_offload_group_idx(key: OffloadKey) -> int: |
| 39 | """Extract the group index from an `OffloadKey`.""" |
| 40 | return int.from_bytes(key[-4:], "big", signed=False) |
| 41 | |
| 42 | |
| 43 | _T = TypeVar("_T") |
| 44 | |
| 45 | |
| 46 | class Medium(Enum): |
| 47 | """Storage medium of an offloading tier.""" |
| 48 | |
| 49 | CPU = "CPU" |
| 50 | STORAGE = "STORAGE" |
| 51 | |
| 52 | |
| 53 | class Locality(Enum): |
| 54 | """Locality of a tier's storage relative to the publishing instance.""" |
| 55 | |
| 56 | LOCAL = "LOCAL" |
| 57 | REMOTE = "REMOTE" |
| 58 | |
| 59 | |
| 60 | class TierMatcher(NamedTuple): |
| 61 | medium: Medium | None = None |
| 62 | locality: Locality | None = None |
| 63 | |
| 64 | def matches(self, medium: Medium | None, locality: Locality | None) -> bool: |
| 65 | medium_matches = self.medium is None or medium is None or self.medium == medium |
| 66 | locality_matches = ( |
| 67 | self.locality is None or locality is None or self.locality == locality |
| 68 | ) |
| 69 | return medium_matches and locality_matches |
| 70 | |
| 71 | |
| 72 | @dataclass(frozen=True) |
| 73 | class TierFilter: |
| 74 | """Per-request filter controlling which tiers participate.""" |
| 75 | |
| 76 | matchers: tuple[TierMatcher, ...] = () |
| 77 | |
| 78 | ALL: ClassVar["TierFilter"] |
| 79 | |
| 80 | def allows(self, medium: Medium | None, locality: Locality | None) -> bool: |
| 81 | if self is TierFilter.ALL: |
| 82 | return True |
| 83 | return any(m.matches(medium, locality) for m in self.matchers) |
| 84 | |
| 85 | |
| 86 | TierFilter.ALL = TierFilter(matchers=(TierMatcher(),)) |
| 87 | |
| 88 | |
| 89 | @dataclass |
| 90 | class ReqContext: |
| 91 | req_id: str |
| 92 | kv_transfer_params: dict[str, Any] | None = None |
| 93 | kv_hints: KvHintsEnvelope | None = None |
| 94 | load_tier_filter: TierFilter = TierFilter.ALL |
| 95 | # Per-request scratch space keyed by value type, so a tier can parse |
| 96 | # kv_transfer_params and kv_hints once (in on_new_request) and read the |
| 97 | # result back on later calls for the same request. |
| 98 | _state: dict[type, Any] = field(default_factory=dict, repr=False, init=False) |
| 99 | # End-token position for each key in this request. The scheduler records |
| 100 | # these positions so managers can recover prefix order even when store |
| 101 | # calls arrive out of order (for example, SWA backfills). |
| 102 | _offload_key_positions: dict[OffloadKey, int] = field( |
| 103 | default_factory=dict, repr=False, init=False |
| 104 | ) |
| 105 | |
| 106 | def set_state(self, val: Any) -> None: |
| 107 | self._state[type(val)] = val |
| 108 | |
| 109 | def get_state(self, cls: type[_T]) -> _T | None: |
| 110 | return self._state.get(cls) |
| 111 | |
| 112 | def set_offload_key_position(self, key: OffloadKey, end_token: int) -> None: |
| 113 | self._offload_key_positions[key] = end_token |
| 114 | |
| 115 | def get_offload_key_position(self, key: OffloadKey) -> int | None: |
| 116 | return self._offload_key_positions.get(key) |
| 117 | |
| 118 | |
| 119 | class LookupResult(Enum): |
| 120 | """Result of OffloadingManager.lookup().""" |
| 121 | |
| 122 | MISS = auto() |
| 123 | HIT = auto() |
| 124 | HIT_PENDING = auto() |
| 125 | RETRY = auto() |
| 126 | |
| 127 | |
| 128 | class OffloadPolicy(Enum): |
| 129 | # Offload only newly-computed chunks as they arrive; prefix-hit |
| 130 | # chunks (already offloaded by a prior request) are skipped. |
| 131 | CHUNK_LEVEL = "chunk_level" |
| 132 | # Offload all chunks for the request, including prefix hits. |
| 133 | # Used by tiers that need the complete KV context for a request. |
| 134 | REQUEST_LEVEL = "request_level" |
| 135 | |
| 136 | |
| 137 | @dataclass |
| 138 | class RequestOffloadingContext: |
| 139 | policy: OffloadPolicy = OffloadPolicy.CHUNK_LEVEL |
| 140 | |
| 141 | |
| 142 | class ScheduleEndContext(NamedTuple): |
| 143 | """Per-step scheduling info passed to on_schedule_end().""" |
| 144 | |
| 145 | # Request IDs scheduled for the first time this step. |
| 146 | new_req_ids: Collection[str] |
| 147 | # Request IDs preempted this step. |
| 148 | preempted_req_ids: Collection[str] |
| 149 | |
| 150 | |
| 151 | class LoadStoreSpec: |
| 152 | """Metadata that encapsulates information allowing a worker |
| 153 | to load, and optionally also to store, blocks of KV data. |
| 154 | """ |
| 155 | |
| 156 | |
| 157 | @dataclass |
| 158 | class PrepareStoreOutput: |
| 159 | keys_to_store: list[OffloadKey] |
| 160 | store_spec: LoadStoreSpec |
| 161 | evicted_keys: list[OffloadKey] |
| 162 | |
| 163 | |
| 164 | @dataclass |
| 165 | class OffloadingEvent: |
| 166 | keys: list[OffloadKey] |
| 167 | medium: Medium |
| 168 | # True if blocks are removed, False if stored |
| 169 | removed: bool |
| 170 | locality: Locality | None = None |
| 171 | # Secondary tier identifier that generated the event, or None for primary. |
| 172 | ownership: str | None = None |
| 173 | removal_expected: bool = False |
| 174 | |
| 175 | |
| 176 | """ |
| 177 | OffloadingManager class for managing KV data offloading in vLLM v1 |
| 178 | |
| 179 | This class runs in the scheduler, tracks which blocks are offloaded |
| 180 | and their address. |
| 181 | |
| 182 | The class provides the following primitives: |
| 183 | lookup() - check whether a single block is offloaded and ready. |
| 184 | prepare_load() - prepare given blocks to be read. |
| 185 | The given blocks will be protected from eviction. |
| 186 | This function returns a LoadSpec which encapsulates |
| 187 | information required for performing the load. |
| 188 | touch() - marks the give blocks as recently used. Can be used |
| 189 | to track block's LRU. This function is separated from the |
| 190 | prepare_load function to allow setting block recency even |
| 191 | for blocks which do not need reading from the cache, such as |
| 192 | blocks that are cached by the GPU prefix cache. |
| 193 | complete_load() - mark blocks which were previously prepared to be |
| 194 | loaded as done loading. This is to re-allow their eviction. |
| 195 | prepare_store() - prepare the given blocks to be written. |
| 196 | Returns a StoreSpec encapsulating offloading information, |
| 197 | as well as a list of blocks that were evicted as a result. |
| 198 | complete_store() - marks a previous store as completed. |
| 199 | Following this call, the given blocks will become loadable. |
| 200 | """ |
| 201 | |
| 202 | |
| 203 | @dataclass(frozen=True) |
| 204 | class OffloadingMetricMetadata: |
| 205 | documentation: str |
| 206 | labelnames: tuple[str, ...] = () |
| 207 | |
| 208 | |
| 209 | @dataclass(frozen=True) |
| 210 | class OffloadingCounterMetadata(OffloadingMetricMetadata): |
| 211 | pass |
| 212 | |
| 213 | |
| 214 | @dataclass(frozen=True) |
| 215 | class OffloadingGaugeMetadata(OffloadingMetricMetadata): |
| 216 | pass |
| 217 | |
| 218 | |
| 219 | @dataclass(frozen=True) |
| 220 | class OffloadingHistogramMetadata(OffloadingMetricMetadata): |
| 221 | buckets: tuple[float, ...] | None = None |
| 222 | |
| 223 | |
| 224 | @dataclass(frozen=True) |
| 225 | class OffloadingKVEventsConfig: |
| 226 | # Global vLLM KV event publishing flag. When false, connector-specific |
| 227 | # event capture must stay inert because take_events() is not drained. |
| 228 | enable_kv_cache_events: bool |
| 229 | # OffloadingConnector opt-in for self-describing BlockStored payloads. |
| 230 | # Effective only when enable_kv_cache_events is true. |
| 231 | self_describing_kv_events: bool |
| 232 | |
| 233 | |
| 234 | class OffloadingManager(ABC): |
| 235 | @abstractmethod |
| 236 | def lookup(self, key: OffloadKey, req_context: ReqContext) -> LookupResult: |
| 237 | """Checks whether a single block is offloaded and ready to be read. |
| 238 | |
| 239 | Args: |
| 240 | key: the key identifying the block to lookup. |
| 241 | req_context: per-request context (e.g. kv_transfer_params). |
| 242 | |
| 243 | Returns: |
| 244 | HIT if the block is offloaded and ready, MISS if not found, |
| 245 | HIT_PENDING if found but not yet readable, or RETRY if the |
| 246 | lookup should be retried later. |
| 247 | |
| 248 | """ |
| 249 | pass |
| 250 | |
| 251 | @abstractmethod |
| 252 | def prepare_load( |
| 253 | self, |
| 254 | keys: Collection[OffloadKey], |
| 255 | req_context: ReqContext, |
| 256 | ) -> LoadStoreSpec: |
| 257 | """Prepare the given blocks to be read. |
| 258 | The given blocks will be protected from eviction until |
| 259 | complete_load is called. |
| 260 | It assumes all given blocks are offloaded. |
| 261 | |
| 262 | Args: |
| 263 | keys: the keys identifying the blocks. |
| 264 | req_context: per-request context (e.g. kv_transfer_params). |
| 265 | |
| 266 | Returns: |
| 267 | A LoadStoreSpec that can be used by a worker to locate and load |
| 268 | the actual offloaded KV data. |
| 269 | |
| 270 | """ |
| 271 | pass |
| 272 | |
| 273 | def touch(self, keys: Collection[OffloadKey], req_context: ReqContext): |
| 274 | """Mark the given blocks as recently used. |
| 275 | This could in practice mean moving them to the end of an LRU list. |
| 276 | |
| 277 | Args: |
| 278 | keys: the keys identifying the blocks. |
| 279 | req_context: per-request context (e.g. kv_transfer_params). |
| 280 | |
| 281 | """ |
| 282 | return |
| 283 | |
| 284 | def complete_load(self, keys: Collection[OffloadKey], req_context: ReqContext): |
| 285 | """Marks previous blocks that were prepared to load as done loading. |
| 286 | |
| 287 | Args: |
| 288 | keys: the keys identifying the blocks. |
| 289 | req_context: per-request context (e.g. kv_transfer_params). |
| 290 | |
| 291 | """ |
| 292 | return |
| 293 | |
| 294 | @abstractmethod |
| 295 | def prepare_store( |
| 296 | self, |
| 297 | keys: Collection[OffloadKey], |
| 298 | req_context: ReqContext, |
| 299 | ) -> PrepareStoreOutput | None: |
| 300 | """Prepare the given blocks to be offloaded. |
| 301 | The given blocks will be protected from eviction until |
| 302 | complete_store is called. |
| 303 | |
| 304 | Args: |
| 305 | keys: the keys identifying the blocks. |
| 306 | req_context: per-request context (e.g. kv_transfer_params). |
| 307 | |
| 308 | Returns: |
| 309 | A PrepareStoreOutput indicating which blocks need storing, |
| 310 | where to store them (LoadStoreSpec), and list of blocks that |
| 311 | were evicted as a result. |
| 312 | None is returned if the blocks cannot be stored. |
| 313 | |
| 314 | """ |
| 315 | pass |
| 316 | |
| 317 | def complete_store( |
| 318 | self, |
| 319 | keys: Collection[OffloadKey], |
| 320 | req_context: ReqContext, |
| 321 | success: bool = True, |
| 322 | ): |
| 323 | """Marks blocks which were previously prepared to be stored, as stored. |
| 324 | Following this call, the blocks become loadable. |
| 325 | If success is False, blocks that were not marked as stored will be |
| 326 | removed. |
| 327 | |
| 328 | Args: |
| 329 | keys: the keys identifying the blocks. |
| 330 | req_context: per-request context (e.g. kv_transfer_params). |
| 331 | success: whether the blocks were stored successfully. |
| 332 | |
| 333 | """ |
| 334 | return |
| 335 | |
| 336 | @abstractmethod |
| 337 | def on_new_request(self, req_context: ReqContext) -> RequestOffloadingContext: |
| 338 | """Called when a new request is first seen by the scheduler. |
| 339 | |
| 340 | Returns a RequestOffloadingContext indicating how this request's |
| 341 | blocks should be offloaded. |
| 342 | |
| 343 | Args: |
| 344 | req_context: per-request context. |
| 345 | |
| 346 | """ |
| 347 | pass |
| 348 | |
| 349 | def on_request_finished(self, req_context: ReqContext) -> None: |
| 350 | """Called when a request has finished. |
| 351 | |
| 352 | By the time this is called, the scheduler will issue no more |
| 353 | submit-side calls for this request, such as prepare_store() and |
| 354 | prepare_load(). Completion callbacks for already-submitted transfers |
| 355 | (complete_store() and complete_load()) may still arrive afterward. |
| 356 | |
| 357 | This hook does NOT imply the data has been persisted. Asynchronous |
| 358 | transfers already submitted for this request may still be in flight. |
| 359 | Managers that cascade to lower tiers should delay those tiers' |
| 360 | on_request_finished() calls until no more lower-tier submit calls can |
| 361 | be issued for this request. |
| 362 | |
| 363 | Args: |
| 364 | req_context: per-request context. |
| 365 | |
| 366 | """ |
| 367 | return |
| 368 | |
| 369 | def take_events(self) -> Iterable[OffloadingEvent]: |
| 370 | """Take the offloading events from the manager. |
| 371 | |
| 372 | A tier manager emits only events for storage state it owns. A |
| 373 | composing manager may aggregate child event streams, but should not |
| 374 | synthesize events on behalf of a child tier. |
| 375 | |
| 376 | Yields: |
| 377 | New OffloadingEvents collected since the last call. |
| 378 | |
| 379 | """ |
| 380 | return () |
| 381 | |
| 382 | def on_schedule_end(self, context: ScheduleEndContext) -> None: |
| 383 | """Called once at the end of each scheduler step. |
| 384 | |
| 385 | Managers may override this to flush deferred work accumulated |
| 386 | during the step (e.g., batched promotions). |
| 387 | """ |
| 388 | return |
| 389 | |
| 390 | def has_pending_work(self) -> bool: |
| 391 | """Whether this manager needs the engine to keep stepping. |
| 392 | |
| 393 | While True, on_schedule_end() and get_finished_jobs() continue |
| 394 | to be called even when no requests are scheduled. |
| 395 | """ |
| 396 | return False |
| 397 | |
| 398 | def reset_cache(self) -> None: |
| 399 | """Evict all tracked blocks and reset internal state.""" |
| 400 | return |
| 401 | |
| 402 | def get_stats(self) -> "OffloadingConnectorStats | None": |
| 403 | """Return collected metrics since last call, or None if disabled.""" |
| 404 | return None |
| 405 | |
| 406 | def shutdown(self) -> None: |
| 407 | """Shutdown the manager and release any resources.""" |
| 408 | return |
| 409 | |
| 410 | |
| 411 | class BlockIDsLoadStoreSpec(LoadStoreSpec, ABC): |
| 412 | """Spec for loading/storing KV blocks from given block numbers. |
| 413 | |
| 414 | Subclass semantics differ: GPULoadStoreSpec.block_ids are GPU block |
| 415 | indices; CPULoadStoreSpec.block_ids are CPU cache chunk indices. |
| 416 | """ |
| 417 | |
| 418 | def __init__(self, block_ids: list[int]): |
| 419 | self.block_ids = np.array(block_ids, dtype=np.int32) |
| 420 | |
| 421 | def __repr__(self) -> str: |
| 422 | return repr(self.block_ids) |
| 423 | |
| 424 | |
| 425 | class GPULoadStoreSpec(BlockIDsLoadStoreSpec): |
| 426 | """Spec for loading/storing a KV block to GPU memory. |
| 427 | |
| 428 | If there are multiple KV groups, the blocks are expected to be |
| 429 | ordered by the group index. |
| 430 | In that case, group_sizes[i] determines the number of blocks |
| 431 | per the i-th KV group, and thus sum(group_sizes) == len(block_ids). |
| 432 | group_sizes=None indicates a single KV group. |
| 433 | |
| 434 | If block_indices is given, each group (determined by group_sizes) of block IDs |
| 435 | will correspond to logically contiguous blocks, e.g. blocks 5-10 of a some request. |
| 436 | block_indices[i] will represent the block index of the first block in group #i. |
| 437 | Thus, len(block_indices) == len(group_sizes) = number of KV cache groups. |
| 438 | This information is required in order to support off/loading from offloaded blocks |
| 439 | which are larger than GPU blocks. |
| 440 | In such cases, the first GPU block per each group may be unaligned to the offloaded |
| 441 | block size, and so knowing block_indices[i] allows the worker to correctly |
| 442 | skip part of the first matching offloaded block. |
| 443 | """ |
| 444 | |
| 445 | def __init__( |
| 446 | self, |
| 447 | block_ids: list[int], |
| 448 | group_sizes: Sequence[int], |
| 449 | block_indices: Sequence[int], |
| 450 | ): |
| 451 | super().__init__(block_ids) |
| 452 | assert sum(group_sizes) == len(block_ids) |
| 453 | assert len(block_indices) == len(group_sizes) |
| 454 | self.group_sizes: Sequence[int] = group_sizes |
| 455 | self.block_indices: Sequence[int] = block_indices |
| 456 | |
| 457 | |
| 458 | @dataclass |
| 459 | class CanonicalKVCacheTensor: |
| 460 | """A canonicalized KV cache tensor whose first dimension is num_blocks. |
| 461 | |
| 462 | With standardized layouts (RFC #42082) num_blocks is always the leading |
| 463 | logical dimension. |
| 464 | """ |
| 465 | |
| 466 | # The KV cache tensor with shape (num_blocks, ...) |
| 467 | tensor: torch.Tensor |
| 468 | # The (possibly padded) page size per block in bytes |
| 469 | page_size_bytes: int |
| 470 | |
| 471 | |
| 472 | @dataclass(frozen=True) |
| 473 | class CopyRun: |
| 474 | """A strided byte correspondence between this worker's physical page and |
| 475 | a canonical page: for i in range(num_fragments), fragment i spans |
| 476 | [local_offset + i * local_stride, +fragment_size) in the worker's page and |
| 477 | [canonical_offset + i * canonical_stride, +fragment_size) canonically.""" |
| 478 | |
| 479 | local_offset: int |
| 480 | canonical_offset: int |
| 481 | fragment_size: int |
| 482 | num_fragments: int |
| 483 | local_stride: int |
| 484 | canonical_stride: int |
| 485 | |
| 486 | |
| 487 | @dataclass(frozen=True) |
| 488 | class CanonicalPageMapping: |
| 489 | """How this worker's page maps into a canonical (parallelism-free) page. |
| 490 | In-process only, never serialized. Runs cover the full local page in both |
| 491 | directions; ranks holding identical bytes take turns writing them. |
| 492 | """ |
| 493 | |
| 494 | # Size of the canonical page in bytes |
| 495 | canonical_page_size_bytes: int |
| 496 | # Size of this worker's (un-padded) page in bytes |
| 497 | local_page_size_bytes: int |
| 498 | # Byte correspondences between this worker's page and a canonical page |
| 499 | runs: tuple[CopyRun, ...] |
| 500 | # Number of ranks holding these exact bytes |
| 501 | num_writers: int |
| 502 | # This worker's index among those ranks |
| 503 | writer_index: int |
| 504 | # Canonical bytes identical under any parallel config with this block span |
| 505 | parallelism_agnostic: bool |
| 506 | |
| 507 | def is_writer(self, block_id: int) -> bool: |
| 508 | """Whether this worker stores the canonical page of the given block. |
| 509 | Rotating by block spreads writes across ranks holding identical bytes. |
| 510 | """ |
| 511 | return block_id % self.num_writers == self.writer_index |
| 512 | |
| 513 | |
| 514 | @dataclass |
| 515 | class CanonicalKVCacheRef: |
| 516 | """Per-layer (or group of layers) reference to a specific (by index) |
| 517 | CanonicalKVCacheTensor and records the un-padded page size used by that layer. |
| 518 | """ |
| 519 | |
| 520 | # Index into the list of CanonicalKVCacheTensor objects |
| 521 | tensor_idx: int |
| 522 | # The un-padded page size per block in bytes |
| 523 | page_size_bytes: int |
| 524 | # How this worker's page maps into a canonical page; None = uncertified |
| 525 | mapping: CanonicalPageMapping | None = None |
| 526 | |
| 527 | |
| 528 | @dataclass |
| 529 | class CanonicalKVCaches: |
| 530 | """Canonicalized block-level representation of the KV caches. |
| 531 | |
| 532 | Composed of: |
| 533 | - Unique list of KV cache data tensors, |
| 534 | each with shape (num_blocks, page_size_in_bytes) and int8 dtype. |
| 535 | - Per-group data references of the tensors. |
| 536 | i.e. how each KV cache group maps to the tensors. |
| 537 | """ |
| 538 | |
| 539 | # Ordered list of unique block tensors, each with shape |
| 540 | # (num_blocks, ...). |
| 541 | tensors: list[CanonicalKVCacheTensor] |
| 542 | # Per-KV-cache-group list of data references that map each layer |
| 543 | # in the group to the appropriate entry in the tensors list. |
| 544 | group_data_refs: list[list[CanonicalKVCacheRef]] |
| 545 | |
| 546 | |
| 547 | @dataclass |
| 548 | class TransferResult: |
| 549 | job_id: int |
| 550 | success: bool |
| 551 | transfer_size: int | None = None |
| 552 | transfer_time: float | None = None |
| 553 | |
| 554 | |
| 555 | class OffloadingWorker(ABC): |
| 556 | """Runs in the worker process. Performs async KV transfers for ONE |
| 557 | offloaded medium (e.g. CPU). Direction is explicit via submit_store / |
| 558 | submit_load, so there is no (src_medium, dst_medium) routing.""" |
| 559 | |
| 560 | @abstractmethod |
| 561 | def submit_store( |
| 562 | self, job_id: int, src_spec: GPULoadStoreSpec, dst_spec: LoadStoreSpec |
| 563 | ) -> bool: |
| 564 | """Async GPU -> offloaded medium.""" |
| 565 | |
| 566 | @abstractmethod |
| 567 | def submit_load( |
| 568 | self, job_id: int, src_spec: LoadStoreSpec, dst_spec: GPULoadStoreSpec |
| 569 | ) -> bool: |
| 570 | """Async offloaded medium -> GPU.""" |
| 571 | |
| 572 | @abstractmethod |
| 573 | def get_finished(self) -> list[TransferResult]: ... |
| 574 | |
| 575 | @abstractmethod |
| 576 | def wait(self, job_ids: set[int]) -> None: ... |
| 577 | |
| 578 | def shutdown(self) -> None: |
| 579 | return |
| 580 | |
| 581 | |
| 582 | class OffloadingSpec(ABC): |
| 583 | """Spec for an offloading connector""" |
| 584 | |
| 585 | @classmethod |
| 586 | def build_metric_definitions( |
| 587 | cls, extra_config: dict[str, Any] |
| 588 | ) -> dict[str, "OffloadingMetricMetadata"]: |
| 589 | """Return Prometheus metric definitions emitted by this spec.""" |
| 590 | return {} |
| 591 | |
| 592 | def __init__(self, config: OffloadingConfig): |
| 593 | self.config = config |
| 594 | self.extra_config = config.extra_config |
| 595 | self.replicated_layout: bool = False |
| 596 | self.kv_events_config = OffloadingKVEventsConfig( |
| 597 | enable_kv_cache_events=config.enable_kv_cache_events, |
| 598 | self_describing_kv_events=bool( |
| 599 | self.extra_config.get("self_describing_kv_events", False) |
| 600 | ), |
| 601 | ) |
| 602 | |
| 603 | # When True, only prompt (prefill) blocks are offloaded; decode-phase |
| 604 | # blocks (KV generated after the prompt) are skipped. Useful when prior |
| 605 | # turns' generated tokens are dropped before the next turn (e.g. |
| 606 | # reasoning models that strip thinking). |
| 607 | self.offload_prompt_only: bool = bool( |
| 608 | self.extra_config.get("offload_prompt_only", True) |
| 609 | ) |
| 610 | |
| 611 | self.tokens_per_block = tuple(group.tokens_per_block for group in config.groups) |
| 612 | self.tokens_per_hash = config.cache.tokens_per_hash |
| 613 | self.blocks_per_chunk = config.cache.blocks_per_chunk |
| 614 | |
| 615 | @abstractmethod |
| 616 | def get_manager(self) -> OffloadingManager: |
| 617 | """Get an OffloadingManager that will be used |
| 618 | by the scheduler-side offloading connector to track |
| 619 | offloaded blocks and manage evictions. |
| 620 | """ |
| 621 | pass |
| 622 | |
| 623 | @abstractmethod |
| 624 | def get_worker(self, kv_caches: CanonicalKVCaches) -> OffloadingWorker: |
| 625 | """Get an OffloadingWorker that handles async KV transfers for this spec. |
| 626 | |
| 627 | Args: |
| 628 | kv_caches: Canonicalized KV caches. |
| 629 | |
| 630 | Returns: |
| 631 | An OffloadingWorker instance for this medium. |
| 632 | |
| 633 | """ |
| 634 | pass |