[misc][cuda] use nvml to avoid accidentally cuda initialization (#6007)
This commit is contained in:
@@ -11,6 +11,7 @@ from vllm.model_executor.layers.linear import (LinearBase, LinearMethodBase,
|
||||
set_weight_attrs)
|
||||
from vllm.model_executor.layers.quantization.base_config import (
|
||||
QuantizationConfig)
|
||||
from vllm.utils import get_device_capability_stateless
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
@@ -165,7 +166,7 @@ class GPTQMarlinConfig(QuantizationConfig):
|
||||
return False
|
||||
|
||||
# If the capability of the device is too low, cannot convert.
|
||||
major, minor = torch.cuda.get_device_capability()
|
||||
major, minor = get_device_capability_stateless()
|
||||
device_capability = major * 10 + minor
|
||||
if device_capability < cls.get_min_capability():
|
||||
return False
|
||||
|
||||
Reference in New Issue
Block a user