how to do fair-generate on two GPU separately?

[tjhd1475]

❓ Questions and Help

What is your question?

how to do two fair-generate on two GPU separately?

Code

CUDA_VISIBLE_DEVICES=0 nohup fairseq-generate data-bin/wmt19_en_de_random_del \ --path checkpoints/fconv_wmt_en_de_2/checkpoint_best1.pt \ --num-workers 1 \ --scoring bert_score \ --source-lang en --target-lang de \ --results-path wmt19_en_de/data/trans_result/random_del/ \ --gen-subset train \ --beam 5 --remove-bpe &>/dev/null &

CUDA_VISIBLE_DEVICES=1 nohup fairseq-generate data-bin/wmt19_en_de_random_swap \ --path checkpoints/fconv_wmt_en_de_2/checkpoint_best1.pt \ --num-workers 1 \ --scoring bert_score \ --source-lang en --target-lang de \ --results-path wmt19_en_de/data/trans_result/random_swap/ \ --gen-subset train \ --beam 5 --remove-bpe &>/dev/null &

Tried

i tried to use two gpus with this script but it turned to use only one gpu .

What's your environment?

• fairseq Version (e.g., 1.0 or main):0.12.0
• PyTorch Version (e.g., 1.0):1.13.0
• OS (e.g., Linux):linux
• How you installed fairseq (pip, source): source
• Python version:3.9.7
• CUDA/cuDNN version:11.4
• GPU: A100*2

[gmryu]

What do you mean by "it only uses one gpu"?

They will not merge into one excution. But they should generate at the same time.

You may try to use his script in his question: https://github.com/facebookresearch/fairseq/issues/4478

[gwenzek]

the CUDA_VISIBLE_DEVICES=1 should be on the same line than the fairseq-generate otherwise it's not correctly passed to the subprocess.

[tjhd1475]

@gmryu i mean two seperate fairseq-generate run on one gpu rather than two, #4478 is useful ,thanks!

[tjhd1475]

@gwenzek it works! thanks a lot