[misc][cuda] use nvml to avoid accidentally cuda initialization (#6007)

tests/quantization/utils.py

@@ -1,6 +1,7 @@
 import torch
 
 from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
+from vllm.utils import get_device_capability_stateless
 
 
 def is_quant_method_supported(quant_method: str) -> bool:
@@ -8,7 +9,7 @@ def is_quant_method_supported(quant_method: str) -> bool:
     if not torch.cuda.is_available():
         return False
 
-    capability = torch.cuda.get_device_capability()
+    capability = get_device_capability_stateless()
     capability = capability[0] * 10 + capability[1]
     return (capability >=
             QUANTIZATION_METHODS[quant_method].get_min_capability())