[V0 deprecation] Remove VLLM_USE_V1 usage in most modules (#27955)
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
@@ -12,7 +12,6 @@ from torch.func import functional_call
|
||||
from transformers import PretrainedConfig
|
||||
from typing_extensions import deprecated
|
||||
|
||||
import vllm.envs as envs
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.distributed import (
|
||||
get_tensor_model_parallel_rank,
|
||||
@@ -576,11 +575,8 @@ def maybe_offload_to_cpu(module: torch.nn.Module) -> torch.nn.Module:
|
||||
pin_memory = is_pin_memory_available()
|
||||
uva_available = is_uva_available()
|
||||
|
||||
if envs.VLLM_USE_V1:
|
||||
assert uva_available, "V1 CPU offloading requires uva (pin memory) support"
|
||||
uva_offloading = True
|
||||
else:
|
||||
uva_offloading = False
|
||||
assert uva_available, "V1 CPU offloading requires uva (pin memory) support"
|
||||
uva_offloading = True
|
||||
|
||||
# offload parameters to CPU
|
||||
# use pin_memory if possible, which helps cudagraph capture speed
|
||||
|
||||
Reference in New Issue
Block a user