https://mer.run/posts/i-vit-integer-only-quantization-for-efficient-vision-transformer-inference/