| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import itertools |
| 5 | from collections.abc import Sequence |
| 6 | from dataclasses import dataclass |
| 7 | from typing import Literal, overload |
| 8 | |
| 9 | from vllm.distributed.kv_events import BlockStored, KVCacheEvent |
| 10 | from vllm.logger import init_logger |
| 11 | from vllm.utils.math_utils import cdiv |
| 12 | from vllm.v1.core.kv_cache_coordinator import ( |
| 13 | HybridKVCacheCoordinator, |
| 14 | get_kv_cache_coordinator, |
| 15 | ) |
| 16 | from vllm.v1.core.kv_cache_metrics import KVCacheMetricsCollector |
| 17 | from vllm.v1.core.kv_cache_utils import KVCacheBlock, KVCacheBlockCopy |
| 18 | from vllm.v1.core.single_type_kv_cache_manager import MambaManager |
| 19 | from vllm.v1.kv_cache_interface import ( |
| 20 | AttentionSpec, |
| 21 | CrossAttentionSpec, |
| 22 | EncoderOnlyAttentionSpec, |
| 23 | KVCacheConfig, |
| 24 | MambaSpec, |
| 25 | get_kv_cache_spec_kind, |
| 26 | get_kv_cache_spec_sliding_window, |
| 27 | ) |
| 28 | from vllm.v1.metrics.stats import PrefixCacheStats |
| 29 | from vllm.v1.request import Request, RequestStatus |
| 30 | |
| 31 | logger = init_logger(__name__) |
| 32 | |
| 33 | |
| 34 | @dataclass |
| 35 | class KVCacheBlocks: |
| 36 | """The allocation result of KVCacheManager, work as the interface between |
| 37 | Scheduler and KVCacheManager, to hide KVCacheManager's internal data |
| 38 | structure from the Scheduler. |
| 39 | """ |
| 40 | |
| 41 | blocks: tuple[Sequence[KVCacheBlock], ...] |
| 42 | """ |
| 43 | `blocks[i][j]` refers to the i-th kv_cache_group |
| 44 | and the j-th block of tokens.We don't use block of |
| 45 | tokens as the outer dimension because it assumes all |
| 46 | kv_cache_groups have the same number of blocks, which is true for now but |
| 47 | will be broken if we want to give different block_size to different |
| 48 | kv_cache_groups in the future. |
| 49 | |
| 50 | Each single type KVCacheBlocks could be represented as: |
| 51 | - list[KVCacheBlock] for more than one KVCacheBlock |
| 52 | - an empty tuple for requests without KVCacheBlock |
| 53 | (a precomputed KVCacheBlocks is in KVCacheManager to avoid GC overhead) |
| 54 | """ |
| 55 | |
| 56 | def __add__(self, other: "KVCacheBlocks") -> "KVCacheBlocks": |
| 57 | """Adds two KVCacheBlocks instances.""" |
| 58 | return KVCacheBlocks( |
| 59 | tuple( |
| 60 | list(itertools.chain(blk1, blk2)) |
| 61 | for blk1, blk2 in zip(self.blocks, other.blocks) |
| 62 | ) |
| 63 | ) |
| 64 | |
| 65 | @overload |
| 66 | def get_block_ids( |
| 67 | self, |
| 68 | allow_none: Literal[False] = False, |
| 69 | *, |
| 70 | group_ids: Sequence[int] | None = None, |
| 71 | ) -> tuple[list[int], ...]: ... |
| 72 | |
| 73 | @overload |
| 74 | def get_block_ids( |
| 75 | self, |
| 76 | allow_none: Literal[True] = True, |
| 77 | *, |
| 78 | group_ids: Sequence[int] | None = None, |
| 79 | ) -> tuple[list[int], ...] | None: ... |
| 80 | |
| 81 | def get_block_ids( |
| 82 | self, |
| 83 | allow_none: bool = False, |
| 84 | *, |
| 85 | group_ids: Sequence[int] | None = None, |
| 86 | ) -> tuple[list[int], ...] | None: |
| 87 | """Converts the KVCacheBlocks instance to block_ids. |
| 88 | |
| 89 | Args: |
| 90 | allow_none: Return None when every selected group is empty. |
| 91 | group_ids: KV cache groups to include. Includes all groups by default. |
| 92 | |
| 93 | Returns: |
| 94 | tuple[list[int], ...]: A tuple of lists where: |
| 95 | - the outer tuple corresponds to KV cache groups |
| 96 | - each inner list contains the block_ids of the blocks in that |
| 97 | group |
| 98 | |
| 99 | """ |
| 100 | groups = ( |
| 101 | self.blocks |
| 102 | if group_ids is None |
| 103 | else tuple(self.blocks[group_id] for group_id in group_ids) |
| 104 | ) |
| 105 | if allow_none and all(len(group) == 0 for group in groups): |
| 106 | return None |
| 107 | return tuple([blk.block_id for blk in group] for group in groups) |
| 108 | |
| 109 | def get_unhashed_block_ids(self) -> list[int]: |
| 110 | """Get block_ids of unhashed blocks from KVCacheBlocks instance.""" |
| 111 | assert len(self.blocks) == 1, "Only one group is supported" |
| 112 | return [block.block_id for block in self.blocks[0] if block.block_hash is None] |
| 113 | |
| 114 | def get_unhashed_block_ids_all_groups(self) -> list[list[int]]: |
| 115 | """Get block_ids of unhashed blocks from KVCacheBlocks instance.""" |
| 116 | # Skip padding blocks. |
| 117 | return [ |
| 118 | [ |
| 119 | block.block_id |
| 120 | for block in group |
| 121 | if block.block_hash is None and not block.is_null |
| 122 | ] |
| 123 | for group in self.blocks |
| 124 | ] |
| 125 | |
| 126 | def new_empty(self) -> "KVCacheBlocks": |
| 127 | """Creates a new KVCacheBlocks instance with no blocks.""" |
| 128 | return KVCacheBlocks(tuple(() for _ in range(len(self.blocks)))) |
| 129 | |
| 130 | |
| 131 | class KVCacheManager: |
| 132 | def __init__( |
| 133 | self, |
| 134 | kv_cache_config: KVCacheConfig, |
| 135 | max_model_len: int, |
| 136 | scheduler_block_size: int, |
| 137 | hash_block_size: int, |
| 138 | max_in_flight_tokens: int | None = None, |
| 139 | enable_caching: bool = True, |
| 140 | use_eagle: bool = False, |
| 141 | num_prefill_lookahead: int = 0, |
| 142 | log_stats: bool = False, |
| 143 | enable_kv_cache_events: bool = False, |
| 144 | dcp_world_size: int = 1, |
| 145 | pcp_world_size: int = 1, |
| 146 | metrics_collector: KVCacheMetricsCollector | None = None, |
| 147 | watermark: float = 0.0, |
| 148 | enable_mamba_shared_prefix_checkpoint: bool = False, |
| 149 | ) -> None: |
| 150 | self.max_model_len = max_model_len |
| 151 | # When unset, fall back to `max_model_len` so the recycling-aware cap |
| 152 | # collapses to the prior (uncapped) admission behavior. The scheduler |
| 153 | # always supplies the real value at runtime. |
| 154 | if max_in_flight_tokens is None: |
| 155 | max_in_flight_tokens = max_model_len |
| 156 | |
| 157 | self.enable_caching = enable_caching |
| 158 | self.enable_kv_cache_events = enable_kv_cache_events |
| 159 | self.use_eagle = use_eagle |
| 160 | self.log_stats = log_stats |
| 161 | self.metrics_collector = metrics_collector |
| 162 | # FIXME: make prefix cache stats conditional on log_stats. We still need |
| 163 | # this comment because when the log stats is enabled there are still |
| 164 | # potential configs we could expose in the future. |
| 165 | self.prefix_cache_stats = PrefixCacheStats() if log_stats else None |
| 166 | |
| 167 | self.coordinator = get_kv_cache_coordinator( |
| 168 | kv_cache_config=kv_cache_config, |
| 169 | max_model_len=self.max_model_len, |
| 170 | max_in_flight_tokens=max_in_flight_tokens, |
| 171 | use_eagle=self.use_eagle, |
| 172 | enable_caching=self.enable_caching, |
| 173 | enable_kv_cache_events=enable_kv_cache_events, |
| 174 | dcp_world_size=dcp_world_size, |
| 175 | pcp_world_size=pcp_world_size, |
| 176 | scheduler_block_size=scheduler_block_size, |
| 177 | hash_block_size=hash_block_size, |
| 178 | metrics_collector=self.metrics_collector, |
| 179 | num_prefill_lookahead=num_prefill_lookahead, |
| 180 | ) |
| 181 | # One predicate, read by both sides of the feature, so the scheduler |
| 182 | # cannot end a chunk at a junction the manager would refuse -- a refused |
| 183 | # junction costs a forward pass and displaces the block-boundary stop. |
| 184 | # Multi-module MTP is excluded because ``cache_blocks`` then hands the |
| 185 | # manager ``num_computed - num_reprefillable`` rather than the chunk end. |
| 186 | self.mamba_shared_prefix_checkpoint = ( |
| 187 | enable_mamba_shared_prefix_checkpoint |
| 188 | and bool(self.coordinator.eagle_group_ids) |
| 189 | and self.coordinator.enable_partial_hash_hits |
| 190 | and self.coordinator.num_reprefillable_tokens == 0 |
| 191 | ) |
| 192 | if self.mamba_shared_prefix_checkpoint: |
| 193 | for manager in self.coordinator.single_type_managers: |
| 194 | if isinstance(manager, MambaManager): |
| 195 | manager.shared_prefix_checkpoint = True |
| 196 | self.num_kv_cache_groups = len(kv_cache_config.kv_cache_groups) |
| 197 | self.block_pool = self.coordinator.block_pool |
| 198 | self.retained_hit_group_ids = tuple( |
| 199 | manager.kv_cache_group_id |
| 200 | for manager in self.coordinator.single_type_managers |
| 201 | if manager.retains_longer_hit |
| 202 | ) |
| 203 | self.kv_cache_config = kv_cache_config |
| 204 | |
| 205 | # Watermark: minimum number of KV cache blocks to keep free when |
| 206 | # admitting waiting/preempted requests, to avoid frequent preemptions. |
| 207 | assert watermark >= 0.0, "watermark must be non-negative" |
| 208 | self.watermark_blocks = int(watermark * kv_cache_config.num_blocks) |
| 209 | self.kv_cache_event_metadata = tuple( |
| 210 | ( |
| 211 | get_kv_cache_spec_kind(group.kv_cache_spec).value, |
| 212 | get_kv_cache_spec_sliding_window(group.kv_cache_spec), |
| 213 | ) |
| 214 | for group in kv_cache_config.kv_cache_groups |
| 215 | ) |
| 216 | |
| 217 | # Pre-constructed KVCacheBlocks with no blocks, callers should use this |
| 218 | # via create_kv_cache_blocks instead of creating new ones to avoid GC |
| 219 | # overhead. |
| 220 | # |
| 221 | # We use nested tuples to ensure the empty KVCacheBlocks is immutable. |
| 222 | self.empty_kv_cache_blocks = KVCacheBlocks( |
| 223 | tuple(() for _ in range(self.num_kv_cache_groups)) |
| 224 | ) |
| 225 | |
| 226 | @property |
| 227 | def usage(self) -> float: |
| 228 | """Get the KV cache usage. |
| 229 | |
| 230 | Returns: |
| 231 | The KV cache usage (between 0.0 and 1.0). |
| 232 | |
| 233 | """ |
| 234 | return self.block_pool.get_usage() |
| 235 | |
| 236 | def make_prefix_cache_stats(self) -> PrefixCacheStats | None: |
| 237 | """Get (and reset) the prefix cache stats. |
| 238 | |
| 239 | Returns: |
| 240 | The current prefix caching stats, or None if logging is disabled. |
| 241 | |
| 242 | """ |
| 243 | if not self.log_stats: |
| 244 | return None |
| 245 | stats = self.prefix_cache_stats |
| 246 | self.prefix_cache_stats = PrefixCacheStats() |
| 247 | return stats |
| 248 | |
| 249 | def prefix_cache_lookup_enabled(self, request: Request) -> bool: |
| 250 | """Whether a local prefix cache lookup may be run for this request.""" |
| 251 | return self.enable_caching and not request.skip_reading_prefix_cache |
| 252 | |
| 253 | def record_prefix_cache_stats(self, request: Request, num_hits: int) -> None: |
| 254 | # Don't count a request that skipped the cache lookup. |
| 255 | if not self.log_stats or not self.prefix_cache_lookup_enabled(request): |
| 256 | return |
| 257 | assert self.prefix_cache_stats is not None |
| 258 | self.prefix_cache_stats.record( |
| 259 | num_tokens=request.num_tokens, |
| 260 | num_hits=num_hits, |
| 261 | preempted=request.num_preemptions > 0, |
| 262 | ) |
| 263 | |
| 264 | def get_computed_blocks(self, request: Request) -> tuple[KVCacheBlocks, int, int]: |
| 265 | """Get the computed (cached) blocks for the request. |
| 266 | Note that the computed blocks must be full. |
| 267 | |
| 268 | Args: |
| 269 | request: The request to get the computed blocks. |
| 270 | |
| 271 | Returns: |
| 272 | A tuple containing: |
| 273 | - A list of blocks that are computed for the request. |
| 274 | - The number of computed tokens. |
| 275 | - ``shared_prefix_boundary``: the block-aligned token position of |
| 276 | a shared prefix that a sparse-retention group (Mamba / sliding |
| 277 | window) has not cached yet (Marconi-style APC), or 0 if none. |
| 278 | Pinned so sparse prefix-cache retention does not drop |
| 279 | the junction and defeat cross-request reuse. |
| 280 | |
| 281 | """ |
| 282 | # We skip finding the prefix cache hit when prefix caching is |
| 283 | # disabled or the request is marked as skipping kv cache read |
| 284 | # (which happens when the request requires prompt logprobs |
| 285 | # or calls a pooling model with all pooling). |
| 286 | if not self.prefix_cache_lookup_enabled(request): |
| 287 | return self.empty_kv_cache_blocks, 0, 0 |
| 288 | |
| 289 | # NOTE: When all tokens hit the cache, we must recompute the last token |
| 290 | # to obtain logits. Thus, set max_cache_hit_length to prompt_length - 1. |
| 291 | # This can trigger recomputation of an entire block, rather than just |
| 292 | # the single last token, because allocate_slots() requires |
| 293 | # num_computed_tokens to be block-size aligned. Removing this limitation |
| 294 | # could slightly improve performance in the future. |
| 295 | max_cache_hit_length = request.num_tokens - 1 |
| 296 | computed_blocks, num_new_computed_tokens, num_uncached = ( |
| 297 | self.coordinator.find_longest_cache_hit( |
| 298 | request.block_hashes, max_cache_hit_length |
| 299 | ) |
| 300 | ) |
| 301 | |
| 302 | # When kv_cache_report_mode is "full", emit BlockStored events |
| 303 | # for the reused prefix cache blocks so that external consumers |
| 304 | # (e.g. gateway) can learn about them. |
| 305 | if ( |
| 306 | num_new_computed_tokens > 0 |
| 307 | and self.enable_kv_cache_events |
| 308 | and getattr(request, "kv_cache_report_mode", "incremental") == "full" |
| 309 | ): |
| 310 | self.coordinator.emit_cached_block_events(request, computed_blocks) |
| 311 | |
| 312 | # The junction to pin is where the lagging sparse-retention group stops |
| 313 | # (``num_new_computed_tokens``) plus the uncached shared prefix -- i.e. |
| 314 | # the longest single-group hit. Sub-block gaps are left to the mask, |
| 315 | # which floors to the alignment boundary (a no-op there). |
| 316 | shared_prefix_boundary = ( |
| 317 | num_new_computed_tokens + num_uncached if num_uncached else 0 |
| 318 | ) |
| 319 | |
| 320 | blocks = self.create_kv_cache_blocks(computed_blocks) |
| 321 | return blocks, num_new_computed_tokens, shared_prefix_boundary |
| 322 | |
| 323 | def get_computed_blocks_for_connector( |
| 324 | self, request: Request |
| 325 | ) -> tuple[KVCacheBlocks, int, int, bool]: |
| 326 | """Local prefix-cache lookup for a request scheduled with a KV connector. |
| 327 | |
| 328 | Hybrid (Mamba + full-attention) models can have per-group prefix hits |
| 329 | diverge under block pressure: the full-attention tail may be evicted |
| 330 | while a deeper Mamba state survives, or vice versa. Report the |
| 331 | full-attention hit as the local prefix - the connector transfers the |
| 332 | remaining suffix and the Mamba state is transferred unconditionally by |
| 333 | nixl's ``_apply_prefix_caching`` - and flag when that hit ran deeper |
| 334 | than a lagging group. Such a hit only has a valid Mamba state at its |
| 335 | boundary if the connector supplies it, so the caller must fall back to |
| 336 | ``get_computed_blocks`` to reconcile when no external tokens are found. |
| 337 | |
| 338 | Non-hybrid models and already-convergent hits use ``get_computed_blocks``. |
| 339 | |
| 340 | Returns: |
| 341 | The ``get_computed_blocks`` triple (blocks, number of local computed |
| 342 | tokens, shared-prefix boundary) plus ``hit_diverged``. |
| 343 | |
| 344 | """ |
| 345 | coordinator = self.coordinator |
| 346 | if not ( |
| 347 | self.kv_cache_config.has_mamba_layers |
| 348 | and isinstance(coordinator, HybridKVCacheCoordinator) |
| 349 | and coordinator.full_attention_group_id is not None |
| 350 | ): |
| 351 | return *self.get_computed_blocks(request), False |
| 352 | |
| 353 | if not self.prefix_cache_lookup_enabled(request): |
| 354 | return self.empty_kv_cache_blocks, 0, 0, False |
| 355 | |
| 356 | fa_group_id = coordinator.full_attention_group_id |
| 357 | computed, per_group_hits = coordinator.find_longest_cache_hit_per_group( |
| 358 | request.block_hashes, request.num_tokens - 1 |
| 359 | ) |
| 360 | if any(hit > per_group_hits[fa_group_id] for hit in per_group_hits): |
| 361 | # A lagging group hit deeper than full attention means its |
| 362 | # full-attention blocks were evicted; use the reconciled boundary |
| 363 | # that every group agrees on. |
| 364 | return *self.get_computed_blocks(request), False |
| 365 | |
| 366 | num_local = per_group_hits[fa_group_id] |
| 367 | blocks = self.create_kv_cache_blocks(computed) |
| 368 | # Per-group lookups do not detect an uncached shared prefix (boundary 0). |
| 369 | return blocks, num_local, 0, min(per_group_hits) < num_local |
| 370 | |
| 371 | def allocate_slots( |
| 372 | self, |
| 373 | request: Request, |
| 374 | num_new_tokens: int, |
| 375 | num_new_computed_tokens: int = 0, |
| 376 | new_computed_blocks: KVCacheBlocks | None = None, |
| 377 | num_lookahead_tokens: int = 0, |
| 378 | num_external_computed_tokens: int = 0, |
| 379 | delay_cache_blocks: bool = False, |
| 380 | num_encoder_tokens: int = 0, |
| 381 | full_sequence_must_fit: bool = False, |
| 382 | reserved_blocks: int = 0, |
| 383 | has_scheduled_reqs: bool = True, |
| 384 | ) -> KVCacheBlocks | None: |
| 385 | """Add slots for a request with new tokens to append. |
| 386 | |
| 387 | Args: |
| 388 | request: The request to allocate slots. |
| 389 | num_new_tokens: The number of new tokens to be allocated and computed. |
| 390 | num_new_computed_tokens: The number of new computed tokens just |
| 391 | hitting the prefix caching, excluding external tokens. |
| 392 | new_computed_blocks: The cached blocks for the above new computed |
| 393 | tokens, grouped as a tuple by kv cache groups. |
| 394 | num_lookahead_tokens: The number of speculative tokens to allocate. |
| 395 | This is used by spec decode proposers with kv-cache such |
| 396 | as eagle. |
| 397 | num_external_computed_tokens: The number of tokens that their |
| 398 | KV caches are not cached by vLLM but cached by the connector. |
| 399 | delay_cache_blocks: Whether to skip caching the blocks. This is |
| 400 | used by P/D when allocating blocks used in a KV transfer |
| 401 | which will complete in a future step. |
| 402 | num_encoder_tokens: The number of encoder tokens to allocate for |
| 403 | cross-attention in encoder-decoder models(e.g., Whisper). |
| 404 | For decoder-only models, this should be 0. |
| 405 | full_sequence_must_fit: Only allocate blocks if the KV cache has enough |
| 406 | free blocks to hold the full sequence, accounting for prefix cache hits |
| 407 | and sliding window. Used as an admission gate to prevent over-admitting |
| 408 | requests when chunked prefill would otherwise only check the first chunk |
| 409 | reserved_blocks: Number of free blocks that must be left available for |
| 410 | other in-flight sequences to complete. The actual allocation is only |
| 411 | made if it fits within (free blocks - reserved_blocks). Used to gate |
| 412 | async KV-connector loads so their initial allocation cannot consume |
| 413 | blocks an already in-flight (prefilling) sequence is relying on. |
| 414 | has_scheduled_reqs: Whether any requests are already scheduled to run |
| 415 | this step, controls whether watermark is applied. |
| 416 | |
| 417 | Blocks layout: |
| 418 | ``` |
| 419 | ---------------------------------------------------------------------- |
| 420 | | < comp > | < new_comp > | < ext_comp > | < new > | < lookahead > | |
| 421 | ---------------------------------------------------------------------- |
| 422 | | < to be computed > | |
| 423 | ---------------------------------------------------------------------- |
| 424 | | < to be allocated > | |
| 425 | ---------------------------------------------------------------------- |
| 426 | | < to be cached (roughly, | |
| 427 | | details below)> | |
| 428 | ---------------------------------------------------------------------- |
| 429 | | Prefix-cached tokens from either vLLM | |
| 430 | | or connector. Can be safely removed if | |
| 431 | | they are outside sliding window. | |
| 432 | ---------------------------------------------------------------------- |
| 433 | | < cached by vLLM > | not cached by | |
| 434 | | vLLM, but | |
| 435 | | ref_cnt | ref_cnt not | cached by | |
| 436 | | increased| increased yet| connector | |
| 437 | ---------------------------------------------------------------------- |
| 438 | ``` |
| 439 | |
| 440 | Abbreviations: |
| 441 | |
| 442 | ``` |
| 443 | comp = request.num_computed_tokens |
| 444 | new_comp = num_new_computed_tokens |
| 445 | = len(new_computed_blocks) * block_size |
| 446 | ext_comp = num_external_computed_tokens, cached by the connector |
| 447 | new = num_new_tokens, including unverified draft tokens |
| 448 | lookahead = num_lookahead_tokens |
| 449 | ``` |
| 450 | |
| 451 | NOTE: for new tokens which include both verified and unverified draft |
| 452 | tokens, we only cache the verified tokens (by capping the number at |
| 453 | `request.num_tokens`). |
| 454 | |
| 455 | The allocation has three stages: |
| 456 | - Free unnecessary blocks in `comp` and check |
| 457 | if we have sufficient free blocks (return None if not). |
| 458 | - Handle prefix tokens (`comp + new_comp + ext_comp`): |
| 459 | - Free unnecessary blocks (e.g. outside sliding window) |
| 460 | - Allocate new blocks for `ext_comp` tokens inside |
| 461 | sliding window |
| 462 | - Allocate new blocks for tokens to be computed (`new + lookahead`) |
| 463 | |
| 464 | Returns: |
| 465 | A list of new allocated blocks. |
| 466 | |
| 467 | """ |
| 468 | # A step may need no slots of its own while still adopting computed |
| 469 | # tokens: an async KV load, or a chunk that ends inside the replayed |
| 470 | # range of a hit (SWA bounded replay). |
| 471 | if ( |
| 472 | num_new_tokens == 0 |
| 473 | and num_external_computed_tokens == 0 |
| 474 | and num_new_computed_tokens == 0 |
| 475 | ): |
| 476 | raise ValueError( |
| 477 | "num_new_tokens must be greater than 0 when there are no " |
| 478 | "computed tokens to adopt" |
| 479 | ) |
| 480 | |
| 481 | if new_computed_blocks is not None: |
| 482 | new_computed_block_list = new_computed_blocks.blocks |
| 483 | if self.retained_hit_group_ids: |
| 484 | # Only retain the prefix covered by the local and external hits. |
| 485 | reused = num_new_computed_tokens + num_external_computed_tokens |
| 486 | groups = list(new_computed_block_list) |
| 487 | for group_id in self.retained_hit_group_ids: |
| 488 | manager = self.coordinator.single_type_managers[group_id] |
| 489 | groups[group_id] = groups[group_id][ |
| 490 | : cdiv(reused, manager.block_size) |
| 491 | ] |
| 492 | new_computed_block_list = tuple(groups) |
| 493 | else: |
| 494 | new_computed_block_list = self.empty_kv_cache_blocks.blocks |
| 495 | |
| 496 | # The number of computed tokens is the number of computed tokens plus |
| 497 | # the new prefix caching hits |
| 498 | num_local_computed_tokens = ( |
| 499 | request.num_computed_tokens + num_new_computed_tokens |
| 500 | ) |
| 501 | total_computed_tokens = min( |
| 502 | num_local_computed_tokens + num_external_computed_tokens, |
| 503 | self.max_model_len, |
| 504 | ) |
| 505 | |
| 506 | watermark_blocks = 0 |
| 507 | # The watermark is applied to waiting/preempted requests only, and only |
| 508 | # when there's at least one request already scheduled. |
| 509 | if has_scheduled_reqs and request.status in ( |
| 510 | RequestStatus.WAITING, |
| 511 | RequestStatus.PREEMPTED, |
| 512 | ): |
| 513 | watermark_blocks = self.watermark_blocks |
| 514 | |
| 515 | # Matches the scheduler's own prefill boundary: `num_tokens - 1` |
| 516 | # extends it to resumed requests replaying their output tokens. |
| 517 | prefill_end = max(request.num_prompt_tokens, request.num_tokens - 1) |
| 518 | |
| 519 | if full_sequence_must_fit: |
| 520 | # First check and fail if the full request sequence won't fit. |
| 521 | full_num_tokens = min(request.num_tokens, self.max_model_len) |
| 522 | |
| 523 | num_blocks_to_allocate = self.coordinator.get_num_blocks_to_allocate( |
| 524 | request_id=request.request_id, |
| 525 | num_tokens=full_num_tokens, |
| 526 | new_computed_blocks=new_computed_block_list, |
| 527 | num_encoder_tokens=num_encoder_tokens, |
| 528 | total_computed_tokens=total_computed_tokens, |
| 529 | num_local_computed_tokens=num_local_computed_tokens, |
| 530 | num_tokens_main_model=full_num_tokens, |
| 531 | apply_admission_cap=True, |
| 532 | prefill_end=prefill_end, |
| 533 | ) |
| 534 | required_blocks = num_blocks_to_allocate + watermark_blocks |
| 535 | if required_blocks > self.block_pool.get_num_free_blocks(): |
| 536 | return None |
| 537 | |
| 538 | num_tokens_main_model = total_computed_tokens + num_new_tokens |
| 539 | num_tokens_need_slot = min( |
| 540 | num_tokens_main_model + num_lookahead_tokens, self.max_model_len |
| 541 | ) |
| 542 | |
| 543 | # Free the blocks that are skipped during the attention computation |
| 544 | # (e.g., tokens outside the sliding window). |
| 545 | # We can do this even if we cannot schedule this request due to |
| 546 | # insufficient free blocks. |
| 547 | # Should call this function before allocating new blocks to reduce |
| 548 | # the number of evicted blocks. |
| 549 | # Free on the processed-token basis: in-flight steps' attention windows |
| 550 | # still read blocks below the optimistic boundary, and rejected spec |
| 551 | # tokens can roll it back. |
| 552 | self.coordinator.remove_skipped_blocks( |
| 553 | request.request_id, |
| 554 | max(0, total_computed_tokens - request.num_in_flight_tokens), |
| 555 | num_prompt_tokens=request.num_prompt_tokens, |
| 556 | ) |
| 557 | |
| 558 | num_blocks_to_allocate = self.coordinator.get_num_blocks_to_allocate( |
| 559 | request_id=request.request_id, |
| 560 | num_tokens=num_tokens_need_slot, |
| 561 | new_computed_blocks=new_computed_block_list, |
| 562 | num_encoder_tokens=num_encoder_tokens, |
| 563 | total_computed_tokens=num_local_computed_tokens |
| 564 | + num_external_computed_tokens, |
| 565 | num_local_computed_tokens=num_local_computed_tokens, |
| 566 | num_tokens_main_model=num_tokens_main_model, |
| 567 | prefill_end=prefill_end, |
| 568 | ) |
| 569 | |
| 570 | # Keep `reserved_blocks` free for other in-flight sequences, and an |
| 571 | # additional watermark of headroom for waiting/preempted admissions. |
| 572 | available_blocks = self.block_pool.get_num_free_blocks() - reserved_blocks |
| 573 | required_blocks = num_blocks_to_allocate + watermark_blocks |
| 574 | if required_blocks > available_blocks: |
| 575 | # Cannot allocate new blocks |
| 576 | return None |
| 577 | |
| 578 | if ( |
| 579 | new_computed_block_list is not self.empty_kv_cache_blocks.blocks |
| 580 | or num_external_computed_tokens > 0 |
| 581 | ): |
| 582 | # Append the new computed blocks to the request blocks until now to |
| 583 | # avoid the case where the new blocks cannot be allocated. |
| 584 | self.coordinator.allocate_new_computed_blocks( |
| 585 | request_id=request.request_id, |
| 586 | new_computed_blocks=new_computed_block_list, |
| 587 | num_local_computed_tokens=num_local_computed_tokens, |
| 588 | num_external_computed_tokens=num_external_computed_tokens, |
| 589 | ) |
| 590 | |
| 591 | new_blocks = self.coordinator.allocate_new_blocks( |
| 592 | request.request_id, |
| 593 | num_tokens_need_slot, |
| 594 | num_tokens_main_model, |
| 595 | num_encoder_tokens, |
| 596 | ) |
| 597 | |
| 598 | # P/D: delay caching blocks if we have to recv from |
| 599 | # remote. Update state for locally cached blocks. |
| 600 | if delay_cache_blocks: |
| 601 | return self.create_kv_cache_blocks(new_blocks) |
| 602 | |
| 603 | # NOTE(woosuk): We want to commit (cache) up to num_local_computed_tokens |
| 604 | # + num_external_computed_tokens + num_new_tokens, but must exclude |
| 605 | # "non-committable" tokens (e.g., draft tokens that could be rejected). |
| 606 | # Therefore, we cap the number at `request.num_tokens`, ensuring only |
| 607 | # "finalized" tokens are cached. |
| 608 | num_tokens_to_cache = min( |
| 609 | total_computed_tokens + num_new_tokens, |
| 610 | request.num_tokens, |
| 611 | ) |
| 612 | self.coordinator.cache_blocks(request, num_tokens_to_cache) |
| 613 | |
| 614 | return self.create_kv_cache_blocks(new_blocks) |
| 615 | |
| 616 | def free(self, request: Request) -> None: |
| 617 | """Free the blocks allocated for the request. |
| 618 | We free the blocks in reverse order so that the tail blocks are evicted |
| 619 | first when caching is enabled. |
| 620 | |
| 621 | Args: |
| 622 | request: The request to free the blocks. |
| 623 | |
| 624 | """ |
| 625 | self.coordinator.free(request.request_id) |
| 626 | |
| 627 | def remove_skipped_blocks( |
| 628 | self, |
| 629 | request_id: str, |
| 630 | processed_computed_tokens: int, |
| 631 | num_prompt_tokens: int | None = None, |
| 632 | ) -> None: |
| 633 | """Remove the blocks that are no longer needed from `blocks` and replace |
| 634 | the removed blocks with null_block. |
| 635 | |
| 636 | Args: |
| 637 | request_id: The request ID. |
| 638 | processed_computed_tokens: Computed-token prefix length covering |
| 639 | fully processed and committed tokens only (safe to free). |
| 640 | num_prompt_tokens: Optional prompt length for R-SWA gap eviction. |
| 641 | |
| 642 | """ |
| 643 | self.coordinator.remove_skipped_blocks( |
| 644 | request_id, processed_computed_tokens, num_prompt_tokens |
| 645 | ) |
| 646 | |
| 647 | def pop_blocks_for_free(self, request: Request) -> list[KVCacheBlock]: |
| 648 | """Pop the request's bookkeeping and return its blocks without |
| 649 | returning them to the block pool. The caller must eventually free |
| 650 | them in reverse order (so that tail blocks are evicted first). |
| 651 | |
| 652 | Args: |
| 653 | request: The request to pop the blocks for. |
| 654 | |
| 655 | Returns: |
| 656 | The request's blocks in allocation order. |
| 657 | |
| 658 | """ |
| 659 | return self.coordinator.pop_blocks_for_free(request.request_id) |
| 660 | |
| 661 | def evict_blocks(self, block_ids: set[int]) -> None: |
| 662 | """Evict blocks from the prefix cache by their block IDs. |
| 663 | |
| 664 | Args: |
| 665 | block_ids: Set of block IDs to evict from cache. |
| 666 | |
| 667 | """ |
| 668 | self.block_pool.evict_blocks(block_ids) |
| 669 | |
| 670 | def reset_prefix_cache(self) -> bool: |
| 671 | """Reset prefix cache. This function may be used in RLHF |
| 672 | flows to invalidate prefix caching after the weights are updated, |
| 673 | or used for resetting prefix caching status for benchmarking. |
| 674 | |
| 675 | Returns: |
| 676 | bool: True if the prefix cache is successfully reset, |
| 677 | False otherwise. |
| 678 | |
| 679 | """ |
| 680 | if not self.coordinator.reset_prefix_cache(): |
| 681 | return False |
| 682 | if self.log_stats: |
| 683 | assert self.prefix_cache_stats is not None |
| 684 | self.prefix_cache_stats.reset = True |
| 685 | return True |
| 686 | |
| 687 | def get_num_common_prefix_blocks(self, running_request_id: str) -> list[int]: |
| 688 | """Calculate the number of common prefix blocks for each kv cache group. |
| 689 | |
| 690 | The function selects a running request and iterates through its blocks. |
| 691 | A block is considered a common prefix block if ALL requests with |
| 692 | allocated KV cache share it (i.e., ref_cnt equals the number of entries |
| 693 | in req_to_blocks). |
| 694 | |
| 695 | NOTE(woosuk): The number of requests with allocated KV cache is **greater |
| 696 | than or equal to** the number of requests scheduled in the current step. |
| 697 | This is because having allocated KV cache only indicates that: |
| 698 | 1. The request has not yet finished, and |
| 699 | 2. The request holds its blocks unfreed. |
| 700 | |
| 701 | While all scheduled requests must have allocated KV cache, the inverse |
| 702 | is not necessarily true. There may be requests with allocated KV cache |
| 703 | that are not scheduled in the current step. |
| 704 | |
| 705 | This can result in an edge case where the number of common prefix blocks |
| 706 | is 0, even though all scheduled requests share a common prefix. This |
| 707 | occurs because there may be unscheduled requests that do not share the |
| 708 | common prefix. Currently, this case cannot be easily detected, so the |
| 709 | function returns 0 in such cases. |
| 710 | |
| 711 | Args: |
| 712 | running_request_id: The request ID of any running request, used to |
| 713 | identify the common prefix blocks. |
| 714 | |
| 715 | Returns: |
| 716 | list[int]: The number of common prefix blocks for each kv cache |
| 717 | group. |
| 718 | |
| 719 | """ |
| 720 | return self.coordinator.get_num_common_prefix_blocks(running_request_id) |
| 721 | |
| 722 | def take_events(self) -> list[KVCacheEvent]: |
| 723 | """Take the KV cache events from the block pool. |
| 724 | |
| 725 | Returns: |
| 726 | A list of KV cache events. |
| 727 | |
| 728 | """ |
| 729 | events = self.block_pool.take_events() |
| 730 | for event in events: |
| 731 | if not isinstance(event, BlockStored): |
| 732 | continue |
| 733 | if event.group_idx is None: |
| 734 | continue |
| 735 | if event.group_idx < 0 or event.group_idx >= len( |
| 736 | self.kv_cache_event_metadata |
| 737 | ): |
| 738 | logger.warning( |
| 739 | "Group index `%s` not in KV cache metadata", event.group_idx |
| 740 | ) |
| 741 | continue |
| 742 | # Annotate here so BlockPool can keep emitting structural cache |
| 743 | # events without owning semantic KV cache spec metadata. |
| 744 | kind, sliding_window = self.kv_cache_event_metadata[event.group_idx] |
| 745 | event.kv_cache_spec_kind = kind |
| 746 | event.kv_cache_spec_sliding_window = sliding_window |
| 747 | return events |
| 748 | |
| 749 | def get_blocks(self, request_id: str) -> KVCacheBlocks: |
| 750 | """Get the blocks of a request.""" |
| 751 | return self.create_kv_cache_blocks(self.coordinator.get_blocks(request_id)) |
| 752 | |
| 753 | def get_block_ids(self, request_id: str) -> tuple[list[int], ...]: |
| 754 | """Get the block ids of a request.""" |
| 755 | return self.get_blocks(request_id).get_block_ids() |
| 756 | |
| 757 | def get_block_ids_for_computed_tokens( |
| 758 | self, |
| 759 | request_id: str, |
| 760 | num_computed_tokens: int, |
| 761 | ) -> tuple[list[int], ...]: |
| 762 | """Get block ids covering the request's computed tokens.""" |
| 763 | block_ids = self.get_block_ids(request_id) |
| 764 | clipped_block_ids: list[list[int]] = [] |
| 765 | for group, ids in zip(self.kv_cache_config.kv_cache_groups, block_ids): |
| 766 | spec = group.kv_cache_spec |
| 767 | if not isinstance(spec, AttentionSpec) or isinstance( |
| 768 | spec, (CrossAttentionSpec, EncoderOnlyAttentionSpec) |
| 769 | ): |
| 770 | clipped_block_ids.append(ids) |
| 771 | continue |
| 772 | |
| 773 | num_valid_blocks = cdiv(num_computed_tokens, spec.block_size) |
| 774 | clipped_block_ids.append(ids[:num_valid_blocks]) |
| 775 | return tuple(clipped_block_ids) |
| 776 | |
| 777 | def estimate_cached_tokens(self, request: Request) -> int: |
| 778 | """Estimate the number of tokens cached by the request.""" |
| 779 | cached_tokens: int | None = None |
| 780 | for group, blocks in zip( |
| 781 | self.kv_cache_config.kv_cache_groups, |
| 782 | self.get_blocks(request.request_id).blocks, |
| 783 | ): |
| 784 | if not group.kv_cache_spec.prefix_cacheable: |
| 785 | continue |
| 786 | if isinstance( |
| 787 | group.kv_cache_spec, |
| 788 | (CrossAttentionSpec, EncoderOnlyAttentionSpec), |
| 789 | ): |
| 790 | # Cross-attention and encoder-only groups are not prefix cached. |
| 791 | continue |
| 792 | |
| 793 | group_cached_tokens = 0 |
| 794 | for block in blocks: |
| 795 | group_cached_tokens = max( |
| 796 | group_cached_tokens, |
| 797 | block.block_hash_num_tokens or 0, |
| 798 | ) |
| 799 | |
| 800 | cached_tokens = ( |
| 801 | group_cached_tokens |
| 802 | if cached_tokens is None |
| 803 | else min(cached_tokens, group_cached_tokens) |
| 804 | ) |
| 805 | |
| 806 | return cached_tokens or 0 |
| 807 | |
| 808 | def cache_blocks(self, request: Request, num_computed_tokens: int) -> None: |
| 809 | """Cache the blocks for the request, if enabled. |
| 810 | |
| 811 | Args: |
| 812 | request: The request to cache the blocks. |
| 813 | num_computed_tokens: The number of computed tokens, including tokens |
| 814 | that are already cached and tokens to be cached. |
| 815 | |
| 816 | """ |
| 817 | if self.enable_caching: |
| 818 | self.coordinator.cache_blocks(request, num_computed_tokens) |
| 819 | |
| 820 | def create_kv_cache_blocks( |
| 821 | self, blocks: tuple[list[KVCacheBlock], ...] |
| 822 | ) -> KVCacheBlocks: |
| 823 | # Only create new KVCacheBlocks for non-empty blocks |
| 824 | return KVCacheBlocks(blocks) if any(blocks) else self.empty_kv_cache_blocks |
| 825 | |
| 826 | def truncate_computed_blocks( |
| 827 | self, blocks: KVCacheBlocks, num_computed_tokens: int |
| 828 | ) -> KVCacheBlocks: |
| 829 | """Return a lookup-result view truncated at an aligned token endpoint. |
| 830 | |
| 831 | An external hit can supply the final Mamba state even when the local |
| 832 | Mamba group ends before this endpoint. Other groups must cover it. |
| 833 | Pure slicing: refcounts are untouched and ``blocks`` is not mutated. |
| 834 | """ |
| 835 | truncated: list[list[KVCacheBlock]] = [] |
| 836 | for group_blocks, manager, group in zip( |
| 837 | blocks.blocks, |
| 838 | self.coordinator.single_type_managers, |
| 839 | self.kv_cache_config.kv_cache_groups, |
| 840 | strict=True, |
| 841 | ): |
| 842 | if not group.kv_cache_spec.prefix_cacheable: |
| 843 | # Scratch groups (e.g. the CSA compressor ring) hold a fixed |
| 844 | # block covering no token range, so a lookup result never |
| 845 | # carries computed blocks for them and their block size does |
| 846 | # not divide the endpoint. |
| 847 | assert not group_blocks |
| 848 | truncated.append([]) |
| 849 | continue |
| 850 | assert num_computed_tokens % manager.block_size == 0 |
| 851 | num_blocks = num_computed_tokens // manager.block_size |
| 852 | if isinstance(group.kv_cache_spec, MambaSpec): |
| 853 | num_blocks = min(num_blocks, len(group_blocks)) |
| 854 | else: |
| 855 | assert num_blocks <= len(group_blocks) |
| 856 | truncated.append(list(group_blocks[:num_blocks])) |
| 857 | return self.create_kv_cache_blocks(tuple(truncated)) |
| 858 | |
| 859 | def take_new_block_ids(self) -> list[int]: |
| 860 | """Drain and return new attention block IDs for zeroing.""" |
| 861 | ids: list[int] = [] |
| 862 | for mgr in self.coordinator.single_type_managers: |
| 863 | ids.extend(mgr.take_new_block_ids()) |
| 864 | return ids |
| 865 | |
| 866 | def get_zeroing_block_ids_in_range( |
| 867 | self, request_id: str, start_token: int, end_token: int |
| 868 | ) -> list[int]: |
| 869 | """The request's block ids covering [start_token, end_token), from |
| 870 | the groups whose new blocks are zeroed by the worker.""" |
| 871 | ids: list[int] = [] |
| 872 | for mgr in self.coordinator.single_type_managers: |
| 873 | if mgr.records_new_block_ids: |
| 874 | start_idx = start_token // mgr.block_size |
| 875 | end_idx = cdiv(end_token, mgr.block_size) |
| 876 | blocks = mgr.req_to_blocks[request_id] |
| 877 | ids.extend(blk.block_id for blk in blocks[start_idx:end_idx]) |
| 878 | return ids |
| 879 | |
| 880 | def record_blocks_for_zeroing(self, request_id: str, start_token: int) -> None: |
| 881 | """Re-record the request's blocks from start_token onwards for |
| 882 | zeroing, e.g. blocks a failed async KV load left unwritten. |
| 883 | |
| 884 | start_token must be block-aligned: zeroing a partially-valid block |
| 885 | would wipe its valid prefix. |
| 886 | """ |
| 887 | for mgr in self.coordinator.single_type_managers: |
| 888 | if mgr.records_new_block_ids: |
| 889 | assert start_token % mgr.block_size == 0 |
| 890 | start_idx = start_token // mgr.block_size |
| 891 | blocks = mgr.req_to_blocks[request_id] |
| 892 | mgr.new_block_ids.extend(blk.block_id for blk in blocks[start_idx:]) |
| 893 | |
| 894 | def take_kv_cache_block_copies( |
| 895 | self, |
| 896 | ) -> tuple[list[KVCacheBlockCopy], list[KVCacheBlock]]: |
| 897 | """Drain pending copies and return their retained endpoints.""" |
| 898 | pending_copies: list[tuple[KVCacheBlock, KVCacheBlock]] = [] |
| 899 | for mgr in self.coordinator.single_type_managers: |
| 900 | pending_copies.extend(mgr.take_pending_cow_copies()) |
| 901 | copies = [ |
| 902 | KVCacheBlockCopy( |
| 903 | src_block_id=source_block.block_id, |
| 904 | dst_block_id=cow_block.block_id, |
| 905 | ) |
| 906 | for source_block, cow_block in pending_copies |
| 907 | ] |
| 908 | retained_blocks = [block for pair in pending_copies for block in pair] |
| 909 | return copies, retained_blocks |
| 910 | |
| 911 | def take_boundary_state_offloads( |
| 912 | self, |
| 913 | ) -> dict[str, list[tuple[int, int, int]]]: |
| 914 | """Drain this step's boundary-state hand-offs for a KV connector. |
| 915 | |
| 916 | Returns ``{request_id: [(group_id, block_id, boundary_tokens), ...]}`` |
| 917 | for mamba "align" boundary states: the |
| 918 | request's committed boundary-state snapshots and, on a sub-block partial |
| 919 | hit, the CoW copy of its last-prompt-boundary state. Only mamba "align" |
| 920 | groups contribute; empty otherwise. A connector reads the referenced |
| 921 | blocks — never resolving them positionally — and offloads them so a |
| 922 | later request can hit that prefix. |
| 923 | """ |
| 924 | offloads: dict[str, list[tuple[int, int, int]]] = {} |
| 925 | for mgr in self.coordinator.single_type_managers: |
| 926 | for ( |
| 927 | req_id, |
| 928 | group_id, |
| 929 | block, |
| 930 | boundary_tokens, |
| 931 | ) in mgr.take_pending_boundary_state_offloads(): |
| 932 | offloads.setdefault(req_id, []).append( |
| 933 | (group_id, block.block_id, boundary_tokens) |
| 934 | ) |
| 935 | return offloads |
| 936 | |
| 937 | def finalize_partial_tail_offloads( |
| 938 | self, request: Request |
| 939 | ) -> list[tuple[int, int, int]]: |
| 940 | """Consume safe producer partial tails when a request finishes. |
| 941 | |
| 942 | A mamba align table block is a valid boundary source only if no later |
| 943 | token was forwarded. The connector pins and queues the exact table |
| 944 | block before request cleanup, then releases the pin when every worker |
| 945 | reports the store job complete. |
| 946 | """ |
| 947 | offloads: list[tuple[int, int, int]] = [] |
| 948 | for mgr in self.coordinator.single_type_managers: |
| 949 | finalized = mgr.finalize_partial_tail_offload( |
| 950 | request.request_id, |
| 951 | request.num_computed_tokens, |
| 952 | request.num_in_flight_tokens, |
| 953 | ) |
| 954 | if finalized is None: |
| 955 | continue |
| 956 | group_id, block, boundary_tokens = finalized |
| 957 | offloads.append((group_id, block.block_id, boundary_tokens)) |
| 958 | return offloads |
| 959 | |
| 960 | def new_step_starts(self) -> None: |
| 961 | """Notify the coordinator that a new step is starting.""" |
| 962 | self.coordinator.new_step_starts() |