add bnb support for Ascend NPU

[statelesshz]

What does this PR do?

This PR can make bnb nf4 QLoRA out of the box on Ascend NPUs. 🤗

Related PR: bitsandbytes-foundation/bitsandbytes#1422

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[amyeroberts]

cc @younesbelkada @SunMarc

[younesbelkada]

Thanks @statelesshz !
Let us know also with @Titus-von-Koeller once you have a WIP branch on bitsandbytes as well to add that in the multi-backend ongoing effort

[Titus-von-Koeller]

Hey all!

I'm only aware of a backend stub that was added on the multi-backend-refactor branch by @statelesshz, no implementation logic. There wasn't much context given on the PR at the time, but since the implementation was yet to come, for me that was ok as is; as an incremental step.

@statelesshz Could you please elaborate a bit more what you mean by the following statement? Where does your implementation live right now?

internally we first added NPU support based on bitsandbytes-0.42.0

Generally, there've been quite a lot of changes in 0.43.0 and even 0.43.1 (the setup logic), so it would be really advisable to already take those changes into account.

Regarding the timeline: we'll go into public alpha/beta testing of the Intel/AMD backends, finalize the refactoring (integrating with the Torch dispatcher through the torch.library API) in the next weeks and do a full release soon after.

[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.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[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.

[statelesshz]

@younesbelkada @Titus-von-Koeller Good day! I have submitted a PR to add nf4 quant/dequant support to Ascend NPU. The corresponding end-to-end test results are as follows:
bitsandbytes-foundation/bitsandbytes#1422 (comment)
PTAL 🤗

[SunMarc]

+1 on @matthewdouglas comments ! LGTM otherwise

[statelesshz]

@SunMarc I have deleted the changes of "Bnb4BitHfQuantizer". But CI has an error when executing check_repository_consistency. The specific error information is as follows:

Traceback (most recent call last):
  File "/root/transformers/utils/check_copies.py", line 1110, in <module>
    check_copies(args.fix_and_overwrite, args.file)
  File "/root/transformers/utils/check_copies.py", line 860, in check_copies
    raise Exception(
Exception: Found the following copy inconsistencies:
- src/transformers/quantizers/quantizer_bnb_4bit.py: copy does not match quantizers.quantizer_bnb_8bit.Bnb8BitHfQuantizer.update_device_map at line 271
Run `make fix-copies` or `python utils/check_copies.py --fix_and_overwrite` to fix them.

I tried to fix this problem locally based on the error message. After executing python utils/check_copies.py --fix_and_overwrite, it seemed that all the changes in this PR were rolled back. The specific result is shown in the figure below.

Any guidance would be appreciated.

[SunMarc]

Hey @statelesshz , this happens because this method have the following comment above:
# Copied from transformers.quantizers.quantizer_bnb_8bit.Bnb8BitHfQuantizer.update_device_map. So you just need to delete that line !

[SunMarc]

LGTM !

[statelesshz]

Hey @statelesshz , this happens because this method have the following comment above: # Copied from transformers.quantizers.quantizer_bnb_8bit.Bnb8BitHfQuantizer.update_device_map. So you just need to delete that line !

Hey @SunMarc, thanks for your suggestion, it worked. CI reported that there was one test case failing, I think it's not caused by this PR.

[statelesshz]

The failing test case can pass in my local environment:

cc @SunMarc @younesbelkada

[SunMarc]

The failing test case can pass in my local environment:

Indeed this is not related. Can you try to rebase the PR ? I've also pinged Arthur for a final check before merging

[statelesshz]

@SunMarc done! 🤗

[statelesshz]

Hey @ArthurZucker, could you help take a look at this PR, thanks :-)

[SlightwindSec]

Hey @ArthurZucker , this PR addresses a key issue I'm also facing, and I'm hoping it can be merged soon. Thanks!