| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import heapq |
| 4 | from collections.abc import Iterable, Sequence |
| 5 | |
| 6 | from typing_extensions import override |
| 7 | |
| 8 | from vllm.v1.kv_offload.base import OffloadKey, ReqContext |
| 9 | from vllm.v1.kv_offload.cpu.policies.base import ( |
| 10 | CachePolicy, |
| 11 | ChunkStatus, |
| 12 | order_request_keys, |
| 13 | ) |
| 14 | |
| 15 | |
| 16 | class LRUCachePolicy(CachePolicy): |
| 17 | """LRU caching policy with logical recency independent of transfer pinning. |
| 18 | |
| 19 | Evictable chunks live in a lazy-invalidating min-heap. A chunk's recency |
| 20 | can therefore be updated while it is pinned; when it later becomes |
| 21 | evictable, it enters the heap with the order assigned by the request |
| 22 | rather than its transfer-completion time. |
| 23 | |
| 24 | A use is indicated by, |
| 25 | - First time the key is added (store). |
| 26 | - A request-scoped access. |
| 27 | """ |
| 28 | |
| 29 | def __init__(self, cache_capacity: int): |
| 30 | super().__init__(cache_capacity) |
| 31 | self.chunks: dict[OffloadKey, ChunkStatus] = {} |
| 32 | self._ranks: dict[OffloadKey, int] = {} |
| 33 | self._evictable: set[OffloadKey] = set() |
| 34 | self._heap: list[tuple[int, OffloadKey]] = [] |
| 35 | self._next_rank = 0 |
| 36 | |
| 37 | def _assign_new_rank(self, key: OffloadKey) -> bool: |
| 38 | if key not in self.chunks: |
| 39 | return False |
| 40 | self._next_rank += 1 |
| 41 | self._ranks[key] = self._next_rank |
| 42 | return key in self._evictable |
| 43 | |
| 44 | def _push_evictable(self, key: OffloadKey) -> None: |
| 45 | heapq.heappush( |
| 46 | self._heap, |
| 47 | (self._ranks[key], key), |
| 48 | ) |
| 49 | |
| 50 | def _is_current(self, entry: tuple[int, OffloadKey]) -> bool: |
| 51 | rank, key = entry |
| 52 | return key in self._evictable and self._ranks.get(key) == rank |
| 53 | |
| 54 | def _maybe_compact_heap(self) -> None: |
| 55 | if len(self._heap) < max(64, 2 * len(self._evictable)): |
| 56 | return |
| 57 | self._heap = [(self._ranks[key], key) for key in self._evictable] |
| 58 | heapq.heapify(self._heap) |
| 59 | |
| 60 | def _update_recency(self, keys: Iterable[OffloadKey]) -> None: |
| 61 | updated_evictable = [key for key in keys if self._assign_new_rank(key)] |
| 62 | # Rebuilding is linear and substantially cheaper than k heap pushes |
| 63 | # for a long prefix. Small updates retain the incremental path. |
| 64 | if len(updated_evictable) >= max(64, len(self._evictable) // 4): |
| 65 | self._heap = [(self._ranks[key], key) for key in self._evictable] |
| 66 | heapq.heapify(self._heap) |
| 67 | else: |
| 68 | for key in updated_evictable: |
| 69 | self._push_evictable(key) |
| 70 | self._maybe_compact_heap() |
| 71 | |
| 72 | @override |
| 73 | def get(self, key: OffloadKey) -> ChunkStatus | None: |
| 74 | return self.chunks.get(key) |
| 75 | |
| 76 | @override |
| 77 | def insert(self, key: OffloadKey, chunk: ChunkStatus) -> None: |
| 78 | self.chunks[key] = chunk |
| 79 | self._next_rank += 1 |
| 80 | self._ranks[key] = self._next_rank |
| 81 | if chunk.ref_cnt == 0: |
| 82 | self._evictable.add(key) |
| 83 | self._push_evictable(key) |
| 84 | |
| 85 | @override |
| 86 | def remove(self, key: OffloadKey) -> None: |
| 87 | del self.chunks[key] |
| 88 | self._ranks.pop(key, None) |
| 89 | self._evictable.discard(key) |
| 90 | |
| 91 | @override |
| 92 | def touch(self, keys: Iterable[OffloadKey], req_context: ReqContext) -> None: |
| 93 | self._update_recency(reversed(list(keys))) |
| 94 | |
| 95 | @override |
| 96 | def on_request_finished( |
| 97 | self, |
| 98 | key_groups: Sequence[Sequence[OffloadKey]], |
| 99 | insertion_only_keys: set[OffloadKey], |
| 100 | reused_keys: set[OffloadKey], |
| 101 | req_context: ReqContext, |
| 102 | ) -> None: |
| 103 | del insertion_only_keys, reused_keys |
| 104 | self._update_recency(reversed(order_request_keys(key_groups, req_context))) |
| 105 | |
| 106 | @override |
| 107 | def clear(self) -> None: |
| 108 | self.chunks.clear() |
| 109 | self._ranks.clear() |
| 110 | self._evictable.clear() |
| 111 | self._heap.clear() |
| 112 | self._next_rank = 0 |
| 113 | |
| 114 | @override |
| 115 | def evict( |
| 116 | self, n: int, protected: set[OffloadKey] |
| 117 | ) -> list[tuple[OffloadKey, ChunkStatus]] | None: |
| 118 | if n == 0: |
| 119 | return [] |
| 120 | |
| 121 | selected: list[tuple[tuple[int, OffloadKey], ChunkStatus]] = [] |
| 122 | selected_keys: set[OffloadKey] = set() |
| 123 | deferred: list[tuple[int, OffloadKey]] = [] |
| 124 | while self._heap and len(selected) < n: |
| 125 | entry = heapq.heappop(self._heap) |
| 126 | if not self._is_current(entry): |
| 127 | continue |
| 128 | key = entry[1] |
| 129 | # Re-pinning without a recency change can leave an equivalent |
| 130 | # lazy entry behind. Select each cache key at most once. |
| 131 | if key in selected_keys: |
| 132 | continue |
| 133 | if key in protected: |
| 134 | deferred.append(entry) |
| 135 | continue |
| 136 | chunk = self.chunks[key] |
| 137 | assert chunk.ref_cnt == 0 |
| 138 | selected.append((entry, chunk)) |
| 139 | selected_keys.add(key) |
| 140 | |
| 141 | if len(selected) < n: |
| 142 | for entry, _ in selected: |
| 143 | heapq.heappush(self._heap, entry) |
| 144 | for entry in deferred: |
| 145 | heapq.heappush(self._heap, entry) |
| 146 | return None |
| 147 | |
| 148 | for entry in deferred: |
| 149 | heapq.heappush(self._heap, entry) |
| 150 | |
| 151 | candidates: list[tuple[OffloadKey, ChunkStatus]] = [] |
| 152 | for entry, chunk in selected: |
| 153 | key = entry[1] |
| 154 | self._evictable.remove(key) |
| 155 | del self.chunks[key] |
| 156 | del self._ranks[key] |
| 157 | candidates.append((key, chunk)) |
| 158 | self._maybe_compact_heap() |
| 159 | return candidates |
| 160 | |
| 161 | @override |
| 162 | def mark_evictable(self, key: OffloadKey) -> None: |
| 163 | # chunks can become evictable when, |
| 164 | # store completes - i.e. ref_cnt -1 -> 0 # not in evictable list |
| 165 | # all loads complete - i.e ref_cnt 1 -> 0 # not in evictable list |
| 166 | self._evictable.add(key) |
| 167 | self._push_evictable(key) |
| 168 | self._maybe_compact_heap() |
| 169 | |
| 170 | @override |
| 171 | def mark_non_evictable(self, key: OffloadKey) -> None: |
| 172 | # key must have been evictable before it was pinned. |
| 173 | self._evictable.remove(key) |