examples/huggingface failed

[yaxan]

When trying to run the examples I seem to always run into this error:

File "/home/ubuntu/.local/lib/python3.10/site-packages/transformers/utils/import_utils.py", line 1512, in _get_module raise RuntimeError( RuntimeError: Failed to import transformers.generation.utils because of the following error (look up to see its traceback): No module named 'pippy.IR'

[yaxan]

From what I can tell huggingface's accelerate has the following check:

if is_pippy_available(): from pippy.IR import Pipe, PipeSplitWrapper, annotate_split_points from pippy.PipelineStage import PipelineStage

which doesn't reflect the PRs that changed IR and PipelineStage to private

[kwen2501]

Hi thanks for the report.

Sorry that PiPPy is currently under refactorization and clean-up, as we are planning to migrate it into pytorch as torch.distributed.pipelining.
See here: https://github.com/pytorch/pytorch/tree/main/torch/distributed/pipelining
During the refactorization, unfortunately some APIs are privatized to following the more rigorous rule of pytorch.

If you are trying PiPPy through HuggingFace Accelerate, the current "stable" version is torchpippy's 0.2.0 binary release, which can be downloaded through pypi: https://pypi.org/project/torchpippy/

pip install torchpippy

[yaxan]

Hi thanks for the report.

Sorry that PiPPy is currently under refactorization and clean-up, as we are planning to migrate it into pytorch as torch.distributed.pipelining. See here: https://github.com/pytorch/pytorch/tree/main/torch/distributed/pipelining During the refactorization, unfortunately some APIs are privatized to following the more rigorous rule of pytorch.

If you are trying PiPPy through HuggingFace Accelerate, the current "stable" version is torchpippy's 0.2.0 binary release, which can be downloaded through pypi: https://pypi.org/project/torchpippy/

pip install torchpippy

Do you see foresee any issues in me opening a PR to update accelerate's inference.py file to

if is_pippy_available(): from pippy import Pipe, annotate_split_points, SplitPoint, PipelineStage
?

[kwen2501]

If the proposed line works for both the stable version and nightly version (seems so), then it should be fine.

We will also contact them to migrate to torch.distributed.pipelining too.

[yaxan]

If the proposed line works for both the stable version and nightly version (seems so), then it should be fine.

We will also contact them to migrate to torch.distributed.pipelining too.

I'd imagine that's probably the better long term solution than what I was going to do.
Bit of a question for my own understanding - is adding PipelineStage to torch.distributed.pipelining still in the works?

[kwen2501]

The migration is mostly done.
Our new documentation is here:
https://pytorch.org/docs/main/distributed.pipelining.html

[yaxan]

The migration is mostly done. Our new documentation is here: https://pytorch.org/docs/main/distributed.pipelining.html

Sounds great. I'll open a PR for Accelerate following the release of 2.4

[kwen2501]

Wow, looking forward to that. Thanks a lot!
Cc @muellerzr @SunMarc