[MM] Move supports_multimodal_inputs and cache out of registry (#57913)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-09-21 07:43:07 -07:00
committed by GitHub
parent 7488b5780f
commit c8824a1e42
36 changed files with 784 additions and 707 deletions
+26 -23
View File
@@ -10,20 +10,23 @@ import torch
from vllm.config import ModelConfig, ParallelConfig, VllmConfig
from vllm.config.multimodal import MultiModalConfig
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.cache import (
BaseMultiModalProcessorCache,
BaseMultiModalReceiverCache,
MultiModalCache,
LruKeyReplicatedReceiverCache,
LruKeyReplicatedSenderCache,
MultiModalCacheMissError,
MultiModalProcessorOnlyCache,
ShmObjectStoreReceiverCache,
ShmObjectStoreSenderCache,
engine_receiver_cache_from_config,
processor_cache_from_config,
)
from vllm.multimodal.cache.base import (
MultiModalCache,
MultiModalProcessorCacheInItem,
MultiModalProcessorCacheItem,
MultiModalProcessorCacheItemMetadata,
MultiModalProcessorOnlyCache,
MultiModalProcessorSenderCache,
MultiModalReceiverCache,
ShmObjectStoreReceiverCache,
ShmObjectStoreSenderCache,
)
from vllm.multimodal.hasher import MultiModalHasher
from vllm.multimodal.inputs import (
@@ -137,10 +140,10 @@ def _compare_caches(
n_iter: int = 100,
seed: int = 0,
):
cache_0_p0 = MULTIMODAL_REGISTRY.processor_cache_from_config(config_0)
cache_0_p1 = MULTIMODAL_REGISTRY.engine_receiver_cache_from_config(config_0)
cache_1_p0 = MULTIMODAL_REGISTRY.processor_cache_from_config(config_1)
cache_1_p1 = MULTIMODAL_REGISTRY.engine_receiver_cache_from_config(config_1)
cache_0_p0 = processor_cache_from_config(config_0)
cache_0_p1 = engine_receiver_cache_from_config(config_0)
cache_1_p0 = processor_cache_from_config(config_1)
cache_1_p1 = engine_receiver_cache_from_config(config_1)
cache_size_gb = max(
config_0.model_config.multimodal_config.mm_processor_cache_gb,
@@ -275,8 +278,8 @@ def test_oversized_item_is_served_uncached():
assert p0_only.get_and_update_item((small, []), "small")[0] is small
assert p0_only.is_cached_item("small")
p0 = MultiModalProcessorSenderCache(model_config) # type: ignore[arg-type]
p1 = MultiModalReceiverCache(model_config) # type: ignore[arg-type]
p0 = LruKeyReplicatedSenderCache(model_config) # type: ignore[arg-type]
p1 = LruKeyReplicatedReceiverCache(model_config) # type: ignore[arg-type]
assert p0.get_and_update_item((item, []), "big")[0] is item
assert not p0.is_cached_item("big")
assert p1.get_and_update_item(item, "big") is item
@@ -296,8 +299,8 @@ def test_mm_cache_miss_raises_and_recovers():
model or network.
"""
model_config = _StubModelConfig(mm_processor_cache_gb=1)
p0 = MultiModalProcessorSenderCache(model_config) # type: ignore[arg-type]
p1 = MultiModalReceiverCache(model_config) # type: ignore[arg-type]
p0 = LruKeyReplicatedSenderCache(model_config) # type: ignore[arg-type]
p1 = LruKeyReplicatedReceiverCache(model_config) # type: ignore[arg-type]
mm_hash = "image_A"
item = MultiModalKwargsItem.dummy(nbytes=64)
@@ -338,7 +341,7 @@ def test_receiver_cache_replaces_stale_item_when_payload_resent():
and kills EngineCore. Prefer the fresh payload and keep it for later hits.
"""
model_config = _StubModelConfig(mm_processor_cache_gb=1)
p1 = MultiModalReceiverCache(model_config) # type: ignore[arg-type]
p1 = LruKeyReplicatedReceiverCache(model_config) # type: ignore[arg-type]
small = MultiModalKwargsItem.dummy(nbytes=64)
large = MultiModalKwargsItem.dummy(nbytes=256)
mm_hash = "shared-id"
@@ -351,7 +354,7 @@ def test_receiver_cache_replaces_stale_item_when_payload_resent():
def test_receiver_cache_features_keep_resent_payload():
"""EngineCore updates features through get_and_update_features."""
model_config = _StubModelConfig(mm_processor_cache_gb=1)
p1 = MultiModalReceiverCache(model_config) # type: ignore[arg-type]
p1 = LruKeyReplicatedReceiverCache(model_config) # type: ignore[arg-type]
small = MultiModalKwargsItem.dummy(nbytes=64)
large = MultiModalKwargsItem.dummy(nbytes=256)
mm_hash = "shared-id"
@@ -384,7 +387,7 @@ def test_mm_cache_miss_batches_all_drifted_hashes():
non-drifted items in the same request are still ingested.
"""
model_config = _StubModelConfig(mm_processor_cache_gb=1)
p1 = MultiModalReceiverCache(model_config) # type: ignore[arg-type]
p1 = LruKeyReplicatedReceiverCache(model_config) # type: ignore[arg-type]
item = MultiModalKwargsItem.dummy(nbytes=64)
def _feature(
@@ -582,8 +585,8 @@ def test_cache_eviction_lru_cache():
model="llava-hf/llava-onevision-qwen2-0.5b-ov-hf",
mm_processor_cache_gb=6 / GiB_bytes,
)
sender_cache = MultiModalProcessorSenderCache(model_config)
receiver_cache = MultiModalReceiverCache(model_config)
sender_cache = LruKeyReplicatedSenderCache(model_config)
receiver_cache = LruKeyReplicatedReceiverCache(model_config)
_run_test_cache_eviction_lru(sender_cache, receiver_cache, base_item_size=1)
@@ -775,7 +778,7 @@ def test_processor_cache_shared_across_loras():
model="llava-hf/llava-onevision-qwen2-0.5b-ov-hf",
mm_processor_cache_gb=1,
)
receiver_cache = MultiModalReceiverCache(model_config)
receiver_cache = LruKeyReplicatedReceiverCache(model_config)
base_mm_hash = "image_hash_abc123"
lora_a_identifier = f"12345:{base_mm_hash}"
@@ -827,8 +830,8 @@ async def test_release_kv_cache_resends_mm_payload(use_async, release_error):
)
model_config = ctx.model_config
sender = MultiModalProcessorSenderCache(model_config)
receiver = MultiModalReceiverCache(model_config)
sender = LruKeyReplicatedSenderCache(model_config)
receiver = LruKeyReplicatedReceiverCache(model_config)
item = _dummy_item({"pixel_values": 16})
mm_hash = "image_A"
payload, _ = sender.get_and_update_item((item, []), mm_hash)