| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Implementation of Key-Replicated Cache (see docs/configuration/optimization.md).""" |
| 4 | |
| 5 | from typing_extensions import override |
| 6 | |
| 7 | from vllm.config import ModelConfig |
| 8 | from vllm.utils.cache import CacheInfo |
| 9 | |
| 10 | from ..inputs import MultiModalKwargsItem |
| 11 | from .base import ( |
| 12 | BaseMultiModalProcessorCache, |
| 13 | BaseMultiModalReceiverCache, |
| 14 | MultiModalCache, |
| 15 | MultiModalCacheMissError, |
| 16 | MultiModalProcessorCacheInItem, |
| 17 | MultiModalProcessorCacheItemMetadata, |
| 18 | MultiModalProcessorCacheOutItem, |
| 19 | ) |
| 20 | |
| 21 | |
| 22 | class LruKeyReplicatedSenderCache(BaseMultiModalProcessorCache): |
| 23 | """The cache which is used on P0 when LRU caching is enabled. |
| 24 | |
| 25 | How to update each item: |
| 26 | |
| 27 | - If the item is already in the cache, clear the input to avoid |
| 28 | unnecessary IPC. |
| 29 | |
| 30 | - If the item is not in the cache, store the metadata of that item so |
| 31 | that the eviction policy remains the same as the cache on P1, |
| 32 | and return the input. |
| 33 | By only storing the metadata, we avoid keeping the data itself in |
| 34 | memory inside P0. |
| 35 | """ |
| 36 | |
| 37 | def __init__(self, model_config: ModelConfig) -> None: |
| 38 | super().__init__() |
| 39 | |
| 40 | mm_config = model_config.get_multimodal_config() |
| 41 | |
| 42 | self._cache = MultiModalCache.get_lru_cache( |
| 43 | mm_config.mm_processor_cache_gb, |
| 44 | MultiModalProcessorCacheItemMetadata, |
| 45 | ) |
| 46 | |
| 47 | @override |
| 48 | def is_cached_item(self, mm_hash: str) -> bool: |
| 49 | return mm_hash in self._cache |
| 50 | |
| 51 | @override |
| 52 | def get_and_update_item( |
| 53 | self, |
| 54 | mm_item: MultiModalProcessorCacheInItem, |
| 55 | mm_hash: str, |
| 56 | ) -> MultiModalProcessorCacheOutItem: |
| 57 | if (cached_item := self._cache.get(mm_hash)) is not None: |
| 58 | return None, cached_item.prompt_updates |
| 59 | |
| 60 | assert mm_item is not None, f"Expected a cached item for {mm_hash=}" |
| 61 | |
| 62 | self.cache_if_fits( |
| 63 | self._cache, mm_hash, MultiModalProcessorCacheItemMetadata(*mm_item) |
| 64 | ) |
| 65 | return mm_item |
| 66 | |
| 67 | @override |
| 68 | def touch_sender_cache_item(self, mm_hash: str) -> None: |
| 69 | self._cache.touch(mm_hash) |
| 70 | |
| 71 | @override |
| 72 | def clear_cache(self) -> None: |
| 73 | self._cache.clear() |
| 74 | |
| 75 | @override |
| 76 | def make_stats(self, *, delta: bool = False) -> CacheInfo: |
| 77 | return self._cache.stat(delta=delta) |
| 78 | |
| 79 | @override |
| 80 | def invalidate(self, mm_hash: str) -> None: |
| 81 | # Drop our stale shadow entry so the next request for this hash re-sends the |
| 82 | # data and repopulates P1 (see MultiModalCacheMissError). |
| 83 | self._cache.pop(mm_hash, None) |
| 84 | |
| 85 | |
| 86 | class LruKeyReplicatedReceiverCache(BaseMultiModalReceiverCache): |
| 87 | """The cache which is used on P1 when LRU caching is enabled. |
| 88 | |
| 89 | How to update each item: |
| 90 | |
| 91 | - If the caller sent tensor data, store it (replacing any cached item |
| 92 | under the same key) and return that data. P0 can miss after independent |
| 93 | LRU eviction and resend a different item for the same identity. |
| 94 | - If the caller sent no data and the item is cached, return the cached item. |
| 95 | - If the caller sent no data and the item is not cached, raise |
| 96 | `MultiModalCacheMissError`. |
| 97 | """ |
| 98 | |
| 99 | def __init__(self, model_config: ModelConfig) -> None: |
| 100 | super().__init__() |
| 101 | |
| 102 | mm_config = model_config.get_multimodal_config() |
| 103 | |
| 104 | self._cache = MultiModalCache.get_lru_cache( |
| 105 | mm_config.mm_processor_cache_gb, |
| 106 | MultiModalKwargsItem, |
| 107 | ) |
| 108 | |
| 109 | @override |
| 110 | def get_and_update_item( |
| 111 | self, |
| 112 | mm_item: MultiModalKwargsItem | None, |
| 113 | mm_hash: str, |
| 114 | ) -> MultiModalKwargsItem: |
| 115 | if mm_item is not None: |
| 116 | # P0 sent a payload. Never keep a stale cached tensor under this |
| 117 | # identity: a later P0 hit would pair new placeholders with the old |
| 118 | # item and crash the engine. Drop first so a too-large replacement |
| 119 | # cannot leave the previous entry in place. |
| 120 | self._cache.pop(mm_hash, None) |
| 121 | self.cache_if_fits(self._cache, mm_hash, mm_item) |
| 122 | return mm_item |
| 123 | |
| 124 | if (cached_item := self._cache.get(mm_hash)) is not None: |
| 125 | return cached_item |
| 126 | |
| 127 | # No data and not cached here: P0 sent data=None trusting its shadow, but |
| 128 | # the P0/P1 caches have drifted. Raise a retryable error (not assert) so the |
| 129 | # engine can have P0 drop the stale entry and the client resend the data. |
| 130 | raise MultiModalCacheMissError([mm_hash]) |
| 131 | |
| 132 | @override |
| 133 | def touch_receiver_cache_item( |
| 134 | self, |
| 135 | mm_hash: str, |
| 136 | mm_item: MultiModalKwargsItem | None = None, |
| 137 | ) -> None: |
| 138 | self._cache.touch(mm_hash) |
| 139 | |
| 140 | @override |
| 141 | def clear_cache(self) -> None: |
| 142 | self._cache.clear() |