You might accumulate gradients when running your Huggingface model, which typically leads to out-of-memory errors after some iterations. If you use it for inference only, do

with torch.no_grad():
    # The code where you apply the model