Tecorigin sdaa accelerator

[siqi654321]

Description
This PR includes Tecorigin SDAA accelerator support.
With this PR, DeepSpeed supports SDAA as backend for training tasks.

[tjruwase]

@siqi654321, is this ready for review?

[siqi654321]

@tjruwase Yes, it's ready for review. And Tecorigin SDAA is a AI processor that support AI frameworks like PyTorch, etc. It‘s possible to run Transformers/Accelerate/DeepSpeed on SDAA to train foundation model. Website: http://www.tecorigin.com/

[tjruwase]

@siqi654321, please address the DCO requirement

[siqi654321]

@tjruwase The DCO problem has been solved.

[tjruwase]

@siqi654321, please consider doing the following as appropriate:

1. Reviewing https://www.deepspeed.ai/tutorials/accelerator-abstraction-interface/
2. Updating https://www.deepspeed.ai/tutorials/accelerator-setup-guide/
3. Updating https://github.com/deepspeedai/DeepSpeed?tab=readme-ov-file#contributed-hw-support
4. Updating https://github.com/deepspeedai/DeepSpeed?tab=readme-ov-file#build-pipeline-status

@loadams, did I miss anything?

[tjruwase]

@siqi654321, the following can help fix the Formatting issue
https://github.com/deepspeedai/DeepSpeed/blob/master/CONTRIBUTING.md#prerequisites

[siqi654321]

@tjruwase Thx for help. It seems that the formatting error has been fixed. I would also like to ask whether the additional documentation you mentioned is necessary. I see that some accelerators, such as mlu, do not provide these documents.

[tjruwase]

@tjruwase Thx for help. It seems that the formatting error has been fixed. I would also like to ask whether the additional documentation you mentioned is necessary.

@siqi654321, the additional documentation is completely optional. This PR will merge once CI completes.

[siqi654321]

@tjruwase I found that the xpu-max1100 test has failed, but it doesn't seem to be related to my changes. Could you help take a look at this issue?

[tjruwase]

@delock, this CI failure occurs sporadically but seems unrelated to the PRs. Can you help investigate? Thanks!

https://github.com/deepspeedai/DeepSpeed/actions/runs/13260306651/job/37017257486?pr=6903

[loadams]

@siqi654321, please consider doing the following as appropriate:

1. Reviewing https://www.deepspeed.ai/tutorials/accelerator-abstraction-interface/
2. Updating https://www.deepspeed.ai/tutorials/accelerator-setup-guide/
3. Updating https://github.com/deepspeedai/DeepSpeed?tab=readme-ov-file#contributed-hw-support
4. Updating https://github.com/deepspeedai/DeepSpeed?tab=readme-ov-file#build-pipeline-status

@loadams, did I miss anything?

I think this covers everything, the README and the accelerator setup guide being the most important.

[delock]

@delock, this CI failure occurs sporadically but seems unrelated to the PRs. Can you help investigate? Thanks!

https://github.com/deepspeedai/DeepSpeed/actions/runs/13260306651/job/37017257486?pr=6903

@tjruwase we are updating firmware of this server to see whether the random failure goes away, at this time please ignore this error. Thanks!

[siqi654321]

@tjruwase The tests "xpu-max1100" and "nv-torch-latest-v100" appear to be executed on the "xpu" and "cuda" accelerators respectively. However, I have confirmed that my modifications did not impact these two accelerators, which is quite puzzling. Is it possible to run the tests on a different CI machine? Or are there any other methods I can assist with to help troubleshoot the issue?

[loadams]

@tjruwase The tests "xpu-max1100" and "nv-torch-latest-v100" appear to be executed on the "xpu" and "cuda" accelerators respectively. However, I have confirmed that my modifications did not impact these two accelerators, which is quite puzzling. Is it possible to run the tests on a different CI machine? Or are there any other methods I can assist with to help troubleshoot the issue?

@siqi654321 - nothing you need to do to troubleshoot the issue, we will check on the tests and ensure they pass or will mark them as not required if needed. I think you could just make the changes to the README as discussed above.

[loadams]

@siqi654321 - did you want to update the README to indicate the SDAA is a supported accelerator?

[siqi654321]

@loadams Is it possible to merge the code first in order to quickly support our ecosystem users? We will consider adding documentation related to this accelerator in subsequent pull requests.