A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Plugin] Make plugin group clear (#26757)
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
wangxiyuan committed
Oct 14, 2025 at 15:49 UTC
577d498212022f95dc3a59746b1da1c6ed23eaba
3 files changed
+11
-4
vllm/platforms/__init__.py
+2
-2
@@ -6,7 +6,7 @@ from itertools import chain
6
from typing import TYPE_CHECKING
7
8
from vllm import envs
9
-from vllm.plugins import load_plugins_by_group
9
+from vllm.plugins import PLATFORM_PLUGINS_GROUP, load_plugins_by_group
10
from vllm.utils import resolve_obj_by_qualname, supports_xccl
11
12
from .interface import CpuArchEnum, Platform, PlatformEnum
@@ -188,7 +188,7 @@ builtin_platform_plugins = {
188
189
190
def resolve_current_platform_cls_qualname() -> str:
191
- platform_plugins = load_plugins_by_group("vllm.platform_plugins")
191
+ platform_plugins = load_plugins_by_group(PLATFORM_PLUGINS_GROUP)
192
193
activated_plugins = []
194
vllm/plugins/__init__.py
+7
@@ -9,7 +9,14 @@ import vllm.envs as envs
9
10
logger = logging.getLogger(__name__)
11
12
+# Default plugins group will be loaded in all processes(process0, engine core
13
+# process and worker processes)
14
DEFAULT_PLUGINS_GROUP = "vllm.general_plugins"
15
+# IO processor plugins group will be loaded in process0 only
16
+IO_PROCESSOR_PLUGINS_GROUP = "vllm.io_processor_plugins"
17
+# Platform plugins group will be loaded in all processes when
18
+# `vllm.platforms.current_platform` is called and the value not initialized,
19
+PLATFORM_PLUGINS_GROUP = "vllm.platform_plugins"
20
21
# make sure one process only loads plugins once
22
plugins_loaded = False
vllm/plugins/io_processors/__init__.py
+2
-2
@@ -4,7 +4,7 @@
4
import logging
5
6
from vllm.config import VllmConfig
7
-from vllm.plugins import load_plugins_by_group
7
+from vllm.plugins import IO_PROCESSOR_PLUGINS_GROUP, load_plugins_by_group
8
from vllm.plugins.io_processors.interface import IOProcessor
9
from vllm.utils import resolve_obj_by_qualname
10
@@ -37,7 +37,7 @@ def get_io_processor(
37
38
# Load all installed plugin in the group
39
multimodal_data_processor_plugins = load_plugins_by_group(
40
- "vllm.io_processor_plugins"
40
+ IO_PROCESSOR_PLUGINS_GROUP
41
)
42
43
loadable_plugins = {}