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

[Hybrid Allocator] Support Pipeline Parallel (#23974)

Browse Source 
Signed-off-by: Chen Zhang <zhangch99@outlook.com>
This commit is contained in:
Chen Zhang
2025-09-14 15:55:17 -07:00
committed by GitHub
parent 90f3f7d73e
commit 8e5cdcda4e
7 changed files with 472 additions and 235 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
vllm/v1/engine/core.py
View File

@@ -29,10 +29,9 @@ from vllm.transformers_utils.config import (
maybe_register_config_serialize_by_value)
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 (BlockHash, get_kv_cache_config,
from vllm.v1.core.kv_cache_utils import (BlockHash, get_kv_cache_configs,
get_request_block_hasher,
init_none_hash,
unify_kv_cache_configs)
init_none_hash)
from vllm.v1.core.sched.interface import SchedulerInterface
from vllm.v1.core.sched.output import SchedulerOutput
from vllm.v1.core.sched.scheduler import Scheduler as V1Scheduler
@@ -191,18 +190,9 @@ class EngineCore:
available_gpu_memory = [0] * len(kv_cache_specs)
assert len(kv_cache_specs) == len(available_gpu_memory)
# Get the kv cache tensor size
kv_cache_configs = [
get_kv_cache_config(vllm_config, kv_cache_spec_one_worker,
available_gpu_memory_one_worker)
for kv_cache_spec_one_worker, available_gpu_memory_one_worker in
zip(kv_cache_specs, available_gpu_memory)
]
# Since we use a shared centralized controller, we need the
# `kv_cache_config` to be consistent across all workers to make sure
# all the memory operators can be applied to all workers.
unify_kv_cache_configs(kv_cache_configs)
kv_cache_configs = get_kv_cache_configs(vllm_config, kv_cache_specs,
available_gpu_memory)
# All workers have the same kv_cache_config except layer names, so use
# an arbitrary one to initialize the scheduler.