[Metrics] Log multi-modal cache stats and fix reset (#26285)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-10-10 01:45:55 -07:00
committed by GitHub
parent 6f0f570c43
commit ad430a67ca
25 changed files with 586 additions and 235 deletions
+2 -10
View File
@@ -12,11 +12,8 @@ import torch.distributed as dist
import vllm.envs as envs
from vllm.executor.executor_base import ExecutorBase
from vllm.logger import init_logger
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.cache import worker_receiver_cache_from_config
from vllm.utils import get_distributed_init_method, get_ip, get_open_port, run_method
from vllm.v1.engine import ReconfigureDistributedRequest, ReconfigureRankType
from vllm.v1.executor.utils import get_and_update_mm_cache
from vllm.v1.outputs import AsyncModelRunnerOutput
from vllm.v1.worker.worker_base import WorkerWrapperBase
@@ -30,16 +27,13 @@ class UniProcExecutor(ExecutorBase):
"""Initialize the worker and load the model."""
self.driver_worker = WorkerWrapperBase(vllm_config=self.vllm_config, rpc_rank=0)
distributed_init_method, rank, local_rank = self._distributed_args()
is_driver_worker = True
kwargs = dict(
vllm_config=self.vllm_config,
local_rank=local_rank,
rank=rank,
distributed_init_method=distributed_init_method,
is_driver_worker=is_driver_worker,
)
self.mm_receiver_cache = worker_receiver_cache_from_config(
self.vllm_config, MULTIMODAL_REGISTRY, Lock()
is_driver_worker=True,
shared_worker_lock=Lock(),
)
self.async_output_thread: Optional[ThreadPoolExecutor] = None
@@ -74,8 +68,6 @@ class UniProcExecutor(ExecutorBase):
) -> list[Any]:
if kwargs is None:
kwargs = {}
if self.mm_receiver_cache is not None and method == "execute_model":
get_and_update_mm_cache(self.mm_receiver_cache, args)
if not non_block:
return [run_method(self.driver_worker, method, args, kwargs)]