export CUDA_VISIBLE_DEVICES=0,1,2,3

./conformer_ctc3/train.py \
  --world-size 4 \
  --num-epochs 25 \
  --start-epoch 1 \
  --exp-dir conformer_ctc3/exp \
  --full-libri 1 \
  --dynamic-chunk-training 0 \
  --causal-convolution 0 \
  --short-chunk-size 25 \
  --num-left-chunks 4 \
  --max-duration 300 \
  --delay-penalty 0 \
  --nnet-delay-penalty 0 \
  --master-port 12345