Stage 1 can only be run on one gpu card 0

[Beilong-Tang]

When i run stage 1

bash encoding_decoding.sh --stage 1 --batch_size 16 --num_workers 4 --gpu_devices "0" \
  --model_dir exp/${model_name} --bit_width 16000 \
  --wav_scp input_wav.scp  --out_dir outputs/codecs/

It seems that the gpu_devices can only be 0. Other number or array than 0 will give a cuda ordinal error.

[ZhengRachel]

It seems that by removing the os.environ["CUDA_VISIBLE_DEVICES"] = gpuid line in the scripts will solve the problem. There is a conflict while specifying cuda when using os.environ["CUDA_VISIBLE_DEVICES"] = gpuid and torch.cuda.set_device(int(gpuid)) at the same time if cuda is not cuda:0.