[Bugfix] Respect VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY in prefetch offloader (#37699)

Signed-off-by: Yufeng He <40085740+he-yufeng@users.noreply.github.com>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Yufeng He
2026-04-14 20:43:29 -07:00
committed by GitHub
co-authored by Isotr0py
parent 3abf858443
commit bcc2306cef
4 changed files with 23 additions and 11 deletions
+3 -6
View File
@@ -10,9 +10,9 @@ from torch.func import functional_call
import vllm.envs as envs
from vllm.logger import init_logger
from vllm.model_executor.offloader.base import BaseOffloader
from vllm.model_executor.offloader.base import BaseOffloader, should_pin_memory
from vllm.utils.mem_utils import format_gib
from vllm.utils.platform_utils import is_pin_memory_available, is_uva_available
from vllm.utils.platform_utils import is_uva_available
from vllm.utils.torch_utils import get_accelerator_view_from_cpu_tensor
logger = init_logger(__name__)
@@ -43,10 +43,7 @@ class UVAOffloader(BaseOffloader):
self.cpu_offload_bytes = 0
self.cpu_offload_params = cpu_offload_params or set()
self.pin_memory = (
is_pin_memory_available()
and not envs.VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY
)
self.pin_memory = should_pin_memory()
self.uva_offloading = (
is_uva_available() and not envs.VLLM_WEIGHT_OFFLOADING_DISABLE_UVA
)