> Unlike GPT, which only uses encoder blocks, or BERT, which only uses decoder blocks

I think you meant the reverse, autoregressive model like GPT uses only the decoder while autoencoder like BERT only uses the encoder.