bentoml serve GPU Memory usage

[bruce99kang]

While using bentoml serving an onnx model.
Using code below only, gpu mem usage was around 1500MB.

But, when I built this by

bentoml serve --api-workers 1

It seemed to be using all of my GPUs instead of using 1 GPU.
Even if I specified CUDA_VISIBLE_DEVICES=1, the GPU Mem usage took around 6000GB for a GPU.
Are there any files needs to be modified?