File size: 537 Bytes
5fa1a76
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
export TASK_NAME=mrpc
python examples/pytorch/text-classification/run_glue.py \
  --model_name_or_path google-bert/bert-base-cased \
  --task_name $TASK_NAME \
- --use_mps_device \
  --do_train \
  --do_eval \
  --max_seq_length 128 \
  --per_device_train_batch_size 32 \
  --learning_rate 2e-5 \
  --num_train_epochs 3 \
  --output_dir /tmp/$TASK_NAME/ \
  --overwrite_output_dir

Backends for distributed setups like gloo and nccl are not supported by the mps device which means you can only train on a single GPU with the MPS backend.