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

[v1] Move block_hashes from KVCacheManager to Request.block_hashes (#19728)

Browse Source 
Signed-off-by: Or Ozeri <oro@il.ibm.com>
This commit is contained in:
Or Ozeri
2025-08-16 02:52:52 +03:00
committed by GitHub
parent b9dc9d2607
commit c280066f9d
19 changed files with 381 additions and 335 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
vllm/v1/engine/core.py
View File

@@ -25,9 +25,11 @@ from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.tasks import POOLING_TASKS, SupportedTask
from vllm.transformers_utils.config import (
maybe_register_config_serialize_by_value)
from vllm.utils import (decorate_logs, make_zmq_socket,
from vllm.utils import (decorate_logs, get_hash_fn_by_name, make_zmq_socket,
resolve_obj_by_qualname, set_process_title)
from vllm.v1.core.kv_cache_utils import (get_kv_cache_config,
from vllm.v1.core.kv_cache_utils import (BlockHash, get_kv_cache_config,
get_request_block_hasher,
init_none_hash,
unify_kv_cache_configs)
from vllm.v1.core.sched.interface import SchedulerInterface
from vllm.v1.core.sched.output import SchedulerOutput
@@ -140,6 +142,19 @@ class EngineCore:
self.batch_queue_size)
self.batch_queue = queue.Queue(self.batch_queue_size)
self.request_block_hasher: Optional[Callable[[Request],
list[BlockHash]]] = None
if (self.vllm_config.cache_config.enable_prefix_caching
or self.scheduler.get_kv_connector() is not None):
block_size = vllm_config.cache_config.block_size
caching_hash_fn = get_hash_fn_by_name(
vllm_config.cache_config.prefix_caching_hash_algo)
init_none_hash(caching_hash_fn)
self.request_block_hasher = get_request_block_hasher(
block_size, caching_hash_fn)
def _initialize_kv_caches(
self, vllm_config: VllmConfig) -> tuple[int, int, KVCacheConfig]:
start = time.time()
@@ -417,7 +432,8 @@ class EngineCore:
request.mm_kwargs = self.mm_input_cache_server.get_and_update(
request.mm_kwargs, request.mm_hashes)
req = Request.from_engine_core_request(request)
req = Request.from_engine_core_request(request,
self.request_block_hasher)
if req.use_structured_output:
# Note on thread safety: no race condition.
# `grammar_init` is only invoked in input processing thread. For