Training model with GPU pipeline - RuntimeError Expected all tensors to be on the same device

[babangain]

I am trying to train a model with pipeline

echo `date`
exp_dir=$1                             
model_arch=${2:-"transformer_18_18"}    
pretrained_ckpt=$3               


fairseq-train $exp_dir/final_bin \
--max-source-positions=256 \
--max-target-positions=256 \
--source-lang=SRC \
--target-lang=TGT \
--max-update=1000000 \
--save-interval-updates=1000 \
--arch=$model_arch \
--activation-fn gelu --fp16 \
--criterion=label_smoothed_cross_entropy \
--label-smoothing=0.1 \
--optimizer adam \
--adam-betas "(0.9, 0.98)" \
--lr-scheduler=inverse_sqrt \
--clip-norm 1.0 \
--warmup-init-lr 1e-07 \
--lr 3e-5 \
--warmup-updates 2000 \
--dropout 0.2 \
--save-dir $exp_dir/model \
--keep-last-epochs 5 \
--keep-interval-updates 3 \
--patience 10 \
--skip-invalid-size-inputs-valid-test \
--user-dir model_configs \
--update-freq=1 \
--distributed-world-size 1 \
--num-workers 1 \
--max-tokens 256 \
--eval-bleu \
--eval-bleu-args "{\"beam\": 1, \"lenpen\": 1.0, \"max_len_a\": 1.2, \"max_len_b\": 10}" \
--eval-bleu-detok moses \
--eval-bleu-remove-bpe sentencepiece \
--eval-bleu-print-samples \
--best-checkpoint-metric bleu \
--maximize-best-checkpoint-metric \
--restore-file $pretrained_ckpt \
--reset-lr-scheduler \
--reset-meters \
--reset-dataloader \
--reset-optimizer \
--task translation \
--distributed-no-spawn \
--pipeline-model-parallel  --ddp-backend "no_c10d" --pipeline-balance '[18,18]' --pipeline-devices '[0,1]'  \
--pipeline-encoder-balance '[18]' \
--pipeline-encoder-devices '[0]' \
--pipeline-decoder-balance '[18]' \
--pipeline-decoder-devices '[1]'

This is inspired from https://github.com/facebookresearch/fairseq/tree/main/examples/m2m_100#generation-for-the-12b-model However, I am receiving

Traceback (most recent call last):                                                                                                                                    
  File "/Data/baban/anaconda3/envs/it2/bin/fairseq-train", line 8, in <module>
    sys.exit(cli_main())
  File "/Data/baban/scripts/it2/fairseq/fairseq_cli/train.py", line 574, in cli_main
    distributed_utils.call_main(cfg, main)
  File "/Data/baban/scripts/it2/fairseq/fairseq/distributed/utils.py", line 404, in call_main
    main(cfg, **kwargs)
  File "/Data/baban/scripts/it2/fairseq/fairseq_cli/train.py", line 205, in main
    valid_losses, should_stop = train(cfg, trainer, task, epoch_itr)
  File "/Data/baban/anaconda3/envs/it2/lib/python3.8/contextlib.py", line 75, in inner
    return func(*args, **kwds)
  File "/Data/baban/scripts/it2/fairseq/fairseq_cli/train.py", line 331, in train
    log_output = trainer.train_step(samples)
  File "/Data/baban/anaconda3/envs/it2/lib/python3.8/contextlib.py", line 75, in inner
    return func(*args, **kwds)
  File "/Data/baban/scripts/it2/fairseq/fairseq/trainer.py", line 868, in train_step
    raise e
  File "/Data/baban/scripts/it2/fairseq/fairseq/trainer.py", line 843, in train_step
    loss, sample_size_i, logging_output = self.task.train_step(
  File "/Data/baban/scripts/it2/fairseq/fairseq/tasks/fairseq_task.py", line 532, in train_step
    loss, sample_size, logging_output = criterion(model, sample)
  File "/Data/baban/anaconda3/envs/it2/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
  File "/Data/baban/scripts/it2/fairseq/fairseq/criterions/label_smoothed_cross_entropy.py", line 80, in forward
    net_output = model(**sample["net_input"])
  File "/Data/baban/anaconda3/envs/it2/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
  File "/Data/baban/scripts/it2/fairseq/fairseq/models/transformer/transformer_base.py", line 164, in forward
    encoder_out = self.encoder(
  File "/Data/baban/anaconda3/envs/it2/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
  File "/Data/baban/scripts/it2/fairseq/fairseq/models/transformer/transformer_encoder.py", line 166, in forward
    return self.forward_scriptable(
  File "/Data/baban/scripts/it2/fairseq/fairseq/models/transformer/transformer_encoder.py", line 216, in forward_scriptable
    x, encoder_embedding = self.forward_embedding(src_tokens, token_embeddings)
  File "/Data/baban/scripts/it2/fairseq/fairseq/models/transformer/transformer_encoder.py", line 125, in forward_embedding
    token_embedding = self.embed_tokens(src_tokens)
  File "/Data/baban/anaconda3/envs/it2/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
  File "/Data/baban/anaconda3/envs/it2/lib/python3.8/site-packages/torch/nn/modules/sparse.py", line 162, in forward
    return F.embedding(
  File "/Data/baban/anaconda3/envs/it2/lib/python3.8/site-packages/torch/nn/functional.py", line 2210, in embedding
    return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper_CUDA__index_select)

I am using two GPUs for training. Any help regarding this will be really appreciated.

[wangpichao]

Have you figured out how to solve this problem? I have the same issue. Thank you. @babangain

[yugaljain1999]

@wangpichao @babangain Were you guys able to resolve this? I am also facing this issue. Thanks