from transformers import NougatProcessor
from optimum.onnxruntime import ORTModelForVision2Seq

model_name = 'pszemraj/nougat-small-onnx'
processor = NougatProcessor.from_pretrained(model_name)
model = ORTModelForVision2Seq.from_pretrained(
    model_name,
    provider="CPUExecutionProvider", # 'CUDAExecutionProvider' for gpu 
    use_merged=False,
    use_io_binding=True
)

on colab CPU-only (at time of writing) you may get CuPy errors, to solve this uninstall it:

pip uninstall cupy-cuda11x -y

how do da inference?

See here or this basic notebook I uploaded. It seems ONNX brings CPU inference times to 'feasible' - it took ~15 mins for Attention is All You Meme on Colab free CPU runtime.