Inappropriate reduce operation of "num_input_tokens_seen" is prone to get training stuck.

[YouliangHUANG]

System Info

Trivial

Who can help?

@pacman100

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

See src/transformers/trainer.py line 1870
self.state.num_input_tokens_seen += self.accelerator.gather(inputs[main_input_name]).numel()

The length of "inputs[main_input_name]" is not guaranteed to be the same when using ddp, which may make the training process hang. Besides, in a distributed setup, it costs a lot to gather the WHOLE input tensors on different workers. It is better to call .numel() first and then .gather().

Ref: Stuck when using model.generate() and acclerator.gather() in the distributed setting

Expected behavior

Fix:
input_device = inputs[main_input_name].device
self.state.num_input_tokens_seen += torch.sum(self.accelerator.gather(torch.tensor(inputs[main_input_name].numel(), device=input_device, dtype=torch.int64))).item()

[thincal]

  File "/opt/conda/lib/python3.10/site-packages/transformers/trainer.py", line 1851, in _inner_training_loop
    self.state.num_input_tokens_seen += torch.sum(self.accelerator.gather(
  File "/opt/conda/lib/python3.10/site-packages/accelerate/accelerator.py", line 2159, in gather
    return gather(tensor)
  File "/opt/conda/lib/python3.10/site-packages/accelerate/utils/operations.py", line 344, in wrapper
    return function(*args, **kwargs)
  File "/opt/conda/lib/python3.10/site-packages/accelerate/utils/operations.py", line 405, in gather
    return _gpu_gather(tensor)
  File "/opt/conda/lib/python3.10/site-packages/accelerate/utils/operations.py", line 324, in _gpu_gather
    return recursively_apply(_gpu_gather_one, tensor, error_on_other_type=True)
  File "/opt/conda/lib/python3.10/site-packages/accelerate/utils/operations.py", line 129, in recursively_apply
    return func(data, *args, **kwargs)
  File "/opt/conda/lib/python3.10/site-packages/accelerate/utils/operations.py", line 321, in _gpu_gather_one
    torch.distributed.all_gather(output_tensors, tensor)
  File "/opt/conda/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 47, in wrapper
    return func(*args, **kwargs)
  File "/opt/conda/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 2806, in all_gather
    work = default_pg.allgather([tensor_list], [tensor])
RuntimeError: No backend type associated with device type cpu

Patched transformer with above hotfix, it seems the error still happened. Could you help have a look ? thanks.

[YouliangHUANG]

  File "/opt/conda/lib/python3.10/site-packages/transformers/trainer.py", line 1851, in _inner_training_loop
    self.state.num_input_tokens_seen += torch.sum(self.accelerator.gather(
  File "/opt/conda/lib/python3.10/site-packages/accelerate/accelerator.py", line 2159, in gather
    return gather(tensor)
  File "/opt/conda/lib/python3.10/site-packages/accelerate/utils/operations.py", line 344, in wrapper
    return function(*args, **kwargs)
  File "/opt/conda/lib/python3.10/site-packages/accelerate/utils/operations.py", line 405, in gather
    return _gpu_gather(tensor)
  File "/opt/conda/lib/python3.10/site-packages/accelerate/utils/operations.py", line 324, in _gpu_gather
    return recursively_apply(_gpu_gather_one, tensor, error_on_other_type=True)
  File "/opt/conda/lib/python3.10/site-packages/accelerate/utils/operations.py", line 129, in recursively_apply
    return func(data, *args, **kwargs)
  File "/opt/conda/lib/python3.10/site-packages/accelerate/utils/operations.py", line 321, in _gpu_gather_one
    torch.distributed.all_gather(output_tensors, tensor)
  File "/opt/conda/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 47, in wrapper
    return func(*args, **kwargs)
  File "/opt/conda/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 2806, in all_gather
    work = default_pg.allgather([tensor_list], [tensor])
RuntimeError: No backend type associated with device type cpu

Patched transformer with above hotfix, it seems the error still happened. Could you help have a look ? thanks.

I also encountered the same problem in the first place, and that's why I added a statement to assign the device using input_device = inputs[main_input_name].device.
As the original code works, assigning the new tensor to the same device should also work as it was. Can you double-check the device assigned to the tensor?

[pacman100]

Thank you @YouliangHUANG for the issue as well as the suggestion to fix it. It makes sense, it would be great if you want to open a PR with the suggested fix.

[thincal]

I also encountered the same problem in the first place, and that's why I added a statement to assign the device using input_device = inputs[main_input_name].device. As the original code works, assigning the new tensor to the same device should also work as it was. Can you double-check the device assigned to the tensor?

@YouliangHUANG

inputs[main_input_name].numel(): 1336
inputs[main_input_name].device: cpu

After applied the fix, same error happened (transformers==4.37.2):

  File "/opt/conda/lib/python3.10/site-packages/transformers/trainer.py", line 1855, in _inner_training_loop
    self.accelerator.gather(
  File "/opt/conda/lib/python3.10/site-packages/accelerate/accelerator.py", line 2161, in gather
    return gather(tensor)
  File "/opt/conda/lib/python3.10/site-packages/accelerate/utils/operations.py", line 376, in wrapper
    return function(*args, **kwargs)
  File "/opt/conda/lib/python3.10/site-packages/accelerate/utils/operations.py", line 437, in gather
    return _gpu_gather(tensor)
  File "/opt/conda/lib/python3.10/site-packages/accelerate/utils/operations.py", line 356, in _gpu_gather
    return recursively_apply(_gpu_gather_one, tensor, error_on_other_type=True)
  File "/opt/conda/lib/python3.10/site-packages/accelerate/utils/operations.py", line 135, in recursively_apply
    return func(data, *args, **kwargs)
  File "/opt/conda/lib/python3.10/site-packages/accelerate/utils/operations.py", line 353, in _gpu_gather_one
    torch.distributed.all_gather(output_tensors, tensor)
  File "/opt/conda/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 47, in wrapper
    return func(*args, **kwargs)
  File "/opt/conda/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 2806, in all_gather
    work = default_pg.allgather([tensor_list], [tensor])
RuntimeError: No backend type associated with device type cpu

[thincal]

it works now after force set the device as 'cuda', so it seems that original error is caused by the allgather op not supported in cpu device ?

[YouliangHUANG]

it works now after force set the device as 'cuda', so it seems that original error is caused by the allgather op not supported in cpu device ?

@thincal Please check your backend type, and refer to https://pytorch.org/docs/stable/distributed.html for more details.

[thincal]

it works now after force set the device as 'cuda', so it seems that original error is caused by the allgather op not supported in cpu device ?

@thincal Please check your backend type, and refer to https://pytorch.org/docs/stable/distributed.html for more details.

Yes, it's the nccl backend used, which doesn't support cpu device.

[thincal]

The length of "inputs[main_input_name]" is not guaranteed to be the same when using ddp, which may make the training process hang.

so what change is solving this problem ?

[YouliangHUANG]

The length of "inputs[main_input_name]" is not guaranteed to be the same when using ddp, which may make the training process hang.

so what change is solving this problem ?

torch.tensor(inputs[main_input_name].numel(), device=input_device, dtype=torch.int64)
@thincal This code will create a tensor with the size of 1, which records how many input tokens there are in the local worker. Therefore the tensor length is aligned and can be gathered through self.accelerator.gather and then sum into the total number.

[thincal]

The length of "inputs[main_input_name]" is not guaranteed to be the same when using ddp, which may make the training process hang.

so what change is solving this problem ?

torch.tensor(inputs[main_input_name].numel(), device=input_device, dtype=torch.int64) @thincal This code will create a tensor with the size of 1, which records how many input tokens there are in the local worker. Therefore the tensor length is aligned and can be gathered through self.accelerator.gather and then sum into the total number.

OK, that's great. But it seems that the device should be decided according to the ddp backend ?

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.