Duplicated process and memory leakage for evaluation process in all_gather

[SangbumChoi]

System Info

For all accelerate version

Information

• [ ] The official example scripts
• [X] My own modified scripts

Tasks

• [ ] One of the scripts in the examples/ folder of Accelerate or an officially supported no_trainer script in the examples folder of the transformers repo (such as run_no_trainer_glue.py)
• [X] My own task or dataset (give details below)

Reproduction

Related issue https://github.com/huggingface/transformers/issues/15466 https://github.com/huggingface/transformers/pull/28769/files

Expected behavior

https://github.com/huggingface/accelerate/blob/55136b8dc4a1f5bf8a33f38f25b279debdabcc00/src/accelerate/utils/operations.py#L353

All the accelerate gather function is stricted to all_gather. However, there are also the way of using gather in main process to calculate the evaluation process. If we use all_gather for the evaluation process and pass it to cpu it will cost n times (n is number of process). However we only require to gather the distributed variable to one place to calculate.

What do you think about this?

https://github.com/facebookresearch/detectron2/blob/ebe8b45437f86395352ab13402ba45b75b4d1ddb/detectron2/utils/comm.py#L188

[SangbumChoi]

https://github.com/huggingface/accelerate/issues/2898

[muellerzr]

@SangbumChoi definitely open to trying out something more efficient! Best case scenario we have a flag to use all_gather instead, and default to this new method as part of the func. Would you like to take a stab at a PR?

[github-actions[bot]]

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.