add pytorch-xpu Dockerfile

[yao-matrix]

No description provided.

[ydshieh]

I am open to this. But do you think it's better to have this within transformers repository and build the images in our GitHub Actions?

I think it also works very well if this is in one of Intel XPU repository and build the image there ..?

[yao-matrix]

I think finally we'll integrate XPU into transformers in-tree CI pipeline, the prior work has enabled XPU w/ a good pass rate, integrating XPU into in-tree CI will enable us spot and fix issues in day-0.

[ydshieh]

Do you mean the CI for XPU will be triggered from huggingface/transformres GitHub Actions .. ?
(Do we, huggingface, have the access to XPU runners, via Github Actions)?

OK, ping me once this PR is ready.

[yao-matrix]

Do you mean the CI for XPU will be triggered from huggingface/transformres GitHub Actions .. ? (Do we, huggingface, have the access to XPU runners, via Github Actions)?

OK, ping me once this PR is ready.

We've shared some XPU devices(1 or 2) w/ HF through Intel Cloud, I suppose they are mainly on TGI and BNB now. And I suppose there are some work planned to make XPU available in HF datacenter. I suppose we'll finally let CI for XPU triggered from transformers github actions. That's why we are preparing PRs to make sure we prepare well from engineering side, :).

[ydshieh]

OK, thank you, I will leave you to maintain this :-) 🙏

I assume you tested this docker file could be built. Do we want to add the build to

.github/workflows/build-docker-images.yml

?

[yao-matrix]

OK, thank you, I will leave you to maintain this :-) 🙏

I assume you tested this docker file could be built. Do we want to add the build to

.github/workflows/build-docker-images.yml

?

Yes, this is the Dockerfile we are using daily and originally kept in our internal repo, it works. And i will maintain it and upgrade it to make it up-to-date.

for adding it to ".github/workflows/build-docker-images.yml", sure, let me do it.