Comment by dist-epoch

Doesn't decode also need to stream in the whole of the model weights, thus very I/O heavy?