[examples] run_glue_deebert.py distrbuted fails

[stas00]

I'm working on making the tests work under multiple gpus and run into and this one that proved to be stubborn, for some reason it doesn't work under any DP scheme. I don't know anything about this script, To reproduce:

Note - you need at least 2 gpus:

Actually it fails with 1 gpu too (just change to --nproc_per_node=1)

python -m torch.distributed.launch --nproc_per_node=2 examples/research_projects/deebert/run_glue_deebert.py --model_type roberta --model_name_or_path roberta-base --task_name MRPC --do_train --do_eval --do_lower_case --data_dir ./tests/fixtures/tests_samples/MRPC/ --max_seq_length 128 --per_gpu_eval_batch_size=1 --per_gpu_train_batch_size=8 --learning_rate 2e-4 --num_train_epochs 3 --overwrite_output_dir --seed 42 --output_dir ./examples/deebert/saved_models/roberta-base/MRPC/two_stage --plot_data_dir ./examples/deebert/results/ --save_steps 0 --overwrite_cache --eval_after_first_stage

W reducer.cpp:1084] Warning: find_unused_parameters=True was specified in DDP constructor, but did not find any unused parameters in the forward pass. This flag results in an extra traversal of the autograd graph every iteration,  which can adversely affect performance. If your model indeed never has any unused parameters in the forward pass, consider turning this flag off. Note that this warning may be a false positive if your model has flow control causing later iterations to have unused parameters. (function operator())
Traceback (most recent call last):
  File "examples/research_projects/deebert/run_glue_deebert.py", line 730, in <module>
    main()
  File "examples/research_projects/deebert/run_glue_deebert.py", line 645, in main
    global_step, tr_loss = train(args, train_dataset, model, tokenizer)
  File "examples/research_projects/deebert/run_glue_deebert.py", line 176, in train
    outputs = model(**inputs)
  File "/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/torch/nn/modules/module.py", line 872, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/torch/nn/parallel/distributed.py", line 705, in forward
    if self.reducer._rebuild_buckets():
RuntimeError: Expected to have finished reduction in the prior iteration before starting a new one. This error indicates that your module has parameters that were not used in producing loss. Since `find_unused_parameters=True` is enabled, this likely  means that not all `forward` outputs participate in computing loss. You can fix this by making sure all `forward` function outputs participate in calculating loss. 
If you already have done the above, then the distributed data parallel module wasn't able to locate the output tensors in the return value of your module's `forward` function. Please include the loss function and the structure of the return value of `forward` of your module when reporting this issue (e.g. list, dict, iterable).
[W reducer.cpp:1084] Warning: find_unused_parameters=True was specified in DDP constructor, but did not find any unused parameters in the forward pass. This flag results in an extra traversal of the autograd graph every iteration,  which can adversely affect performance. If your model indeed never has any unused parameters in the forward pass, consider turning this flag off. Note that this warning may be a false positive if your model has flow control causing later iterations to have unused parameters. (function operator())
Iteration: 100%1/1 [00:00<00:00,  1.83it/s]
Iteration:   0%|                                                                                                                           | 0/1 [00:00<?, ?it/s]
Epoch:  33                                                                      | 1/3 [00:00<00:01,  1.82it/s]
Traceback (most recent call last):
  File "examples/research_projects/deebert/run_glue_deebert.py", line 730, in <module>
    main()
  File "examples/research_projects/deebert/run_glue_deebert.py", line 645, in main
    global_step, tr_loss = train(args, train_dataset, model, tokenizer)
  File "examples/research_projects/deebert/run_glue_deebert.py", line 176, in train
    outputs = model(**inputs)
  File "/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/torch/nn/modules/module.py", line 872, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/torch/nn/parallel/distributed.py", line 705, in forward
    if self.reducer._rebuild_buckets():
RuntimeError: Expected to have finished reduction in the prior iteration before starting a new one. This error indicates that your module has parameters that were not used in producing loss. Since `find_unused_parameters=True` is enabled, this likely  means that not all `forward` outputs participate in computing loss. You can fix this by making sure all `forward` function outputs participate in calculating loss.
If you already have done the above, then the distributed data parallel module wasn't able to locate the output tensors in the return value of your module's `forward` function. Please include the loss function and the structure of the return value of `forward` of your module when reporting this issue (e.g. list, dict, iterable).
Killing subprocess 2242528
Killing subprocess 2242529
Traceback (most recent call last):
  File "/home/stas/anaconda3/envs/main-38/lib/python3.8/runpy.py", line 194, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/home/stas/anaconda3/envs/main-38/lib/python3.8/runpy.py", line 87, in _run_code
    exec(code, run_globals)
  File "/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/torch/distributed/launch.py", line 340, in <module>
    main()
  File "/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/torch/distributed/launch.py", line 326, in main
    sigkill_handler(signal.SIGTERM, None)  # not coming back
  File "/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/torch/distributed/launch.py", line 301, in sigkill_handler
    raise subprocess.CalledProcessError(returncode=last_return_code, cmd=cmd)
subprocess.CalledProcessError: Command '['/home/stas/anaconda3/envs/main-38/bin/python', '-u', 'examples/research_projects/deebert/run_glue_deebert.py', '--local_rank=1', '--model_type', 'roberta', '--model_name_or_path', 'roberta-base', '--task_name', 'MRPC', '--do_train', '--do_eval', '--do_lower_case', '--data_dir', './tests/fixtures/tests_samples/MRPC/', '--max_seq_length', '128', '--per_gpu_eval_batch_size=1', '--per_gpu_train_batch_size=8', '--learning_rate', '2e-4', '--num_train_epochs', '3', '--overwrite_output_dir', '--seed', '42', '--output_dir', './examples/deebert/saved_models/roberta-base/MRPC/two_stage', '--plot_data_dir', './examples/deebert/results/', '--save_steps', '0', '--overwrite_cache', '--eval_after_first_stage']' returned non-zero exit status 1.

@LysandreJik

[LysandreJik]

Pinging @JetRunner

[JetRunner]

@stas00 Well it is just not designed for DP or DDP. DeeBERT is for accelerating inference with bs=1 (especially on CPU). I don't believe it should support DP.

[JetRunner]

But yes theoretically it can support multi-GPU training but I'm not sure if it's necessary?

[stas00]

That's good enough for me, I will leave it at 0 or 1-gpu - no problem - thank you for elaborating about the needs of this example, @JetRunner!