biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Core] Use key-only cache for BaseMultiModalProcessor (#23018)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-08-27 14:19:13 +08:00
committed by GitHub
parent 8dbf6ed7be
commit 69244e67e6
29 changed files with 954 additions and 394 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
vllm/v1/engine/core.py
View File

@@ -22,6 +22,7 @@ from vllm.logger import init_logger
from vllm.logging_utils.dump_input import dump_engine_exception
from vllm.lora.request import LoRARequest
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.cache import receiver_cache_from_config
from vllm.tasks import POOLING_TASKS, SupportedTask
from vllm.transformers_utils.config import (
maybe_register_config_serialize_by_value)
@@ -38,7 +39,6 @@ from vllm.v1.engine import (EngineCoreOutputs, EngineCoreRequest,
EngineCoreRequestType,
ReconfigureDistributedRequest, ReconfigureRankType,
UtilityOutput, UtilityResult)
from vllm.v1.engine.mm_input_cache import MultiModalInputCacheServer
from vllm.v1.engine.utils import EngineHandshakeMetadata, EngineZmqAddresses
from vllm.v1.executor.abstract import Executor
from vllm.v1.kv_cache_interface import KVCacheConfig
@@ -128,8 +128,9 @@ class EngineCore:
)
self.use_spec_decode = vllm_config.speculative_config is not None
self.mm_input_cache_server = MultiModalInputCacheServer(
vllm_config.model_config, MULTIMODAL_REGISTRY)
self.mm_registry = mm_registry = MULTIMODAL_REGISTRY
self.mm_receiver_cache = receiver_cache_from_config(
vllm_config, mm_registry)
# Setup batch queue for pipeline parallelism.
# Batch queue for scheduled batches. This enables us to asynchronously
@@ -370,7 +371,8 @@ class EngineCore:
logger.warning("Resetting the multi-modal cache when requests are "
"in progress may lead to desynced internal caches.")
self.mm_input_cache_server.reset()
if self.mm_receiver_cache is not None:
self.mm_receiver_cache.clear_cache()
def reset_prefix_cache(self):
self.scheduler.reset_prefix_cache()
@@ -435,10 +437,11 @@ class EngineCore:
assert request.mm_kwargs is not None
# Note on thread safety: no race condition.
# `mm_input_cache_server` is reset at the end of LLMEngine init,
# `mm_receiver_cache` is reset at the end of LLMEngine init,
# and will only accessed in the input processing thread afterwards.
request.mm_kwargs = self.mm_input_cache_server.get_and_update(
request.mm_kwargs, request.mm_hashes)
if self.mm_receiver_cache is not None:
request.mm_kwargs = self.mm_receiver_cache.get_and_update(
request.mm_kwargs, request.mm_hashes)
req = Request.from_engine_core_request(request,
self.request_block_hasher)