cleanup: remove adapter commons (#25045)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
Simon Mo
2025-09-17 16:46:29 +00:00
committed by GitHub
co-authored by Jee Jee Li
parent 4b946d693e
commit 4aa8c7b047
11 changed files with 89 additions and 330 deletions
+27 -17
View File
@@ -6,11 +6,6 @@ from typing import Any, Literal, Optional, Union
import torch
from vllm.adapter_commons.utils import (add_adapter_worker,
apply_adapters_worker,
list_adapters_worker,
set_active_adapters_worker)
from vllm.adapter_commons.worker_manager import AbstractWorkerManager
from vllm.config.lora import LoRAConfig
from vllm.logger import init_logger
from vllm.lora.models import (LoRAModel, LoRAModelManager,
@@ -22,7 +17,7 @@ from vllm.lora.utils import get_adapter_absolute_path
logger = init_logger(__name__)
class WorkerLoRAManager(AbstractWorkerManager):
class WorkerLoRAManager:
"""WorkerLoRAManager that manages LoRA models on the worker side.
Every request, the requested LoRAs will be loaded (unless they are already
@@ -51,7 +46,7 @@ class WorkerLoRAManager(AbstractWorkerManager):
self.vocab_size = vocab_size
self.lora_config = lora_config
self.max_position_embeddings = max_position_embeddings
super().__init__(device)
self.device = device
# Lazily initialized by create_lora_manager.
self._adapter_manager: LoRAModelManager
@@ -164,19 +159,34 @@ class WorkerLoRAManager(AbstractWorkerManager):
def set_active_adapters(self, requests: set[Any],
mapping: Optional[Any]) -> None:
set_active_adapters_worker(requests, mapping, self._apply_adapters,
self._adapter_manager.set_adapter_mapping)
self._apply_adapters(requests)
if mapping is not None:
self._adapter_manager.set_adapter_mapping(mapping)
def _apply_adapters(self, adapter_requests: set[Any]) -> None:
apply_adapters_worker(adapter_requests, self.list_adapters,
self._adapter_manager.adapter_slots,
self.remove_adapter, self.add_adapter)
existing_adapters = self.list_adapters()
models_map = {
adapter_request.adapter_id: adapter_request
for adapter_request in adapter_requests if adapter_request
}
if len(models_map) > self._adapter_manager.adapter_slots:
raise RuntimeError(
f"Number of requested models ({len(models_map)}) is greater "
"than the number of GPU model slots "
f"({self._adapter_manager.adapter_slots}).")
requested_ids = set(models_map)
for adapter_id in existing_adapters - requested_ids:
self.remove_adapter(adapter_id)
for adapter_id in requested_ids - existing_adapters:
self.add_adapter(models_map[adapter_id])
def add_adapter(self, adapter_request: Any) -> bool:
return add_adapter_worker(adapter_request, self.list_adapters,
self._load_adapter,
self._adapter_manager.add_adapter,
self._adapter_manager.activate_adapter)
if adapter_request.adapter_id in self.list_adapters():
return False
loaded_adapter = self._load_adapter(adapter_request)
loaded = self._adapter_manager.add_adapter(loaded_adapter)
self._adapter_manager.activate_adapter(loaded_adapter.id)
return loaded
def remove_adapter(self, adapter_id: int) -> bool:
return self._adapter_manager.remove_adapter(adapter_id)
@@ -185,7 +195,7 @@ class WorkerLoRAManager(AbstractWorkerManager):
self._adapter_manager.remove_all_adapters()
def list_adapters(self) -> set[int]:
return list_adapters_worker(self._adapter_manager.list_adapters)
return set(self._adapter_manager.list_adapters())
class LRUCacheWorkerLoRAManager(WorkerLoRAManager):