Loading docs/getting_started.rst +2 −0 Original line number Diff line number Diff line Loading @@ -193,10 +193,12 @@ Alternatively you can manually start one process per GPU: > DATA=... # path to the preprocessed dataset, must be visible from all nodes > HOST_PORT=master.example.com:9218 # one of the hosts used by the job > RANK=... # the rank of this process, from 0 to 127 in case of 128 GPUs > LOCAL_RANK=... # the local rank of this process, from 0 to 7 in case of 8 GPUs per machine > python train.py $DATA \ --distributed-world-size 128 \ --distributed-init-method 'tcp://$HOST_PORT' \ --distributed-rank $RANK \ --device-id $LOCAL_RANK \ --force-anneal 50 --lr-scheduler fixed --max-epoch 55 \ --arch fconv_wmt_en_fr --optimizer nag --lr 0.1,4 --max-tokens 3000 \ --clip-norm 0.1 --dropout 0.1 --criterion label_smoothed_cross_entropy \ Loading Loading
docs/getting_started.rst +2 −0 Original line number Diff line number Diff line Loading @@ -193,10 +193,12 @@ Alternatively you can manually start one process per GPU: > DATA=... # path to the preprocessed dataset, must be visible from all nodes > HOST_PORT=master.example.com:9218 # one of the hosts used by the job > RANK=... # the rank of this process, from 0 to 127 in case of 128 GPUs > LOCAL_RANK=... # the local rank of this process, from 0 to 7 in case of 8 GPUs per machine > python train.py $DATA \ --distributed-world-size 128 \ --distributed-init-method 'tcp://$HOST_PORT' \ --distributed-rank $RANK \ --device-id $LOCAL_RANK \ --force-anneal 50 --lr-scheduler fixed --max-epoch 55 \ --arch fconv_wmt_en_fr --optimizer nag --lr 0.1,4 --max-tokens 3000 \ --clip-norm 0.1 --dropout 0.1 --criterion label_smoothed_cross_entropy \ Loading
