text2token/models/t2umbart.py

from fairseq.models.bart.model import mbart_large_architecture
from fairseq.models import register_model_architecture


@register_model_architecture("bart", "t2umbart_large")
def t2umbart_large_architecture(args):
    args.no_scale_embedding = False

    args.encoder_learned_pos = False
    args.decoder_learned_pos = False

    args.encoder_normalize_before = True
    args.decoder_normalize_before = True

    args.share_decoder_input_output_embed = True
    args.share_all_embeddings = False

    args.layernorm_embedding = False
    args.no_token_positional_embeddings = False

    mbart_large_architecture(args)