[Core] Shared memory based object store for Multimodal data caching and IPC (#20452)

Signed-off-by: donglu <donglu@cohere.com>
This commit is contained in:
dongluw
2025-09-12 07:54:17 -07:00
committed by GitHub
parent 9f04d9d55f
commit a5b84f1cbf
17 changed files with 1487 additions and 27 deletions
+25
View File
@@ -0,0 +1,25 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from vllm.multimodal.cache import ShmObjectStoreReceiverCache
from vllm.v1.core.sched.output import SchedulerOutput
def get_and_update_mm_cache(
receiver_cache: ShmObjectStoreReceiverCache,
args: tuple[SchedulerOutput],
) -> None:
"""
For each MultiModalKwargsItem in SchedulerOutput, fetch from shared memory
cache as needed.
Args:
receiver_cache: The receiver cache to update.
args: According to the collective_rpc call of execute_model method in
executor, args is a tuple of only one SchedulerOutput element.
"""
scheduler_output = args[0]
for request_data in scheduler_output.scheduled_new_reqs:
for i in range(len(request_data.mm_kwargs)):
mm_input = request_data.mm_kwargs[i]
request_data.mm_kwargs[i] = \
receiver_cache.get_and_update_item(mm_input, None)