业务需要，需要TensorRT 模型转换的进程常驻，使用中发现每次build后会占据一个显存缓存，del也无法清除，后来找到了：
python - How to run TensorRT in multiple threads? - Stack Overflow
import pycuda.driver as cuda
from pycuda.tools import make_default_context
print("Create TensorRT Plan.")
cuda.init() # Initialize CUDA
ctx = make_default_context()
TRT_LOGGER = trt.Logger()
build_result = self.build_engine(TRT_LOGGER)
ctx.pop()
del ctx
