## Fine-tuning
```bash
PRETRAINED_MODEL=/path/to/checkpoint/model.pt
fairseq-train /path/to/binarized/data \
        --restore-file $PRETRAINED_MODEL  --reset-lr-scheduler --reset-optimizer --reset-dataloader \
        --task translation \
        --criterion label_smoothed_cross_entropy \
        --arch transformer_edge --encoder-layers 12 --decoder-ffn-embed-dim 128 --lora-r 32 --lora-r-shape 0 \
        --share-all-embeddings \
        --required-batch-size-multiple 8 \
        --optimizer adam \
        --adam-betas '(0.9,0.98)' \
        --adam-eps 1e-6 \
        --clip-norm 1.0 \
        --lr-scheduler polynomial_decay \
        --lr 0.00015 \
        --warmup-updates 8000 \
        --total-num-update 100000 \
        --max-update 100000 --max-epoch 1000 \
        --max-tokens 20000 \
        --update-freq 1 \
        --log-format simple \
        --log-interval 1000 \
        --save-interval-updates 5000 \
        --fp16 \
        --fp16-init-scale 4 \
        --fp16-scale-window 256 \
        --min-loss-scale 0.0001 \
        --seed 1
        --save-dir /path/to/save/checkpoints
        --ddp-backend legacy_ddp

fairseq-generate $data_bin \
    --path $save_dir/checkpoint_best.pt \
    --batch-size 64 --beam 5 --remove-bpe=sentencepiece

@article{ge2022edgeformer,
  title={EdgeFormer: A Parameter-Efficient Transformer for On-Device Seq2seq Generation},
  author={Ge, Tao and Wei, Furu},
  journal={arXiv preprint arXiv:2202.07959},
  year={2022}
}