Intel SYCL standard runtime support with oneAPI

[abhilash1910]

Firstly thanks @karpathy for creating the repository.
This is a discussion thread regarding a support /collaboration with standard SYCL runtime from Intel .
SYCL being a hardware agnostic language, can be useful for porting and running different kernels across different runtimes . Since there was a mention by @karpathy that this repository should be only Cuda specific, there is a proposal to have an official SYCL implementation from Intel (oneapi) .
Currently I will be heading the development of the SYCL /oneapi compiler backend from Intel for this project and wanted to know :

• Is it possible to have a SYCL branch from this repository which would be contributed and maintained by myself and Intel colleagues?
• If not, can the new repository for sycl oneapi support be linked in the readme of this repository ?
• How to best collaborate regarding unified SYCL runtime across different vendors for the kernels ? (Through devices from Intel etc)
• Best way to benchmark accelerator interfaces such as Nvidia cuda/cudnn vs sycl/onednn and blas for performance ?

I am open to suggestions and greatly appreciate the effort on this repository.

[abhilash1910]

Request : #579

[karpathy]

Hi @abhilash1910 . Great, excited to hear! I won't accept PRs into this repo but I'm very happy to link to yours in the "notable forks" section. Or answer any questions you might have or etc of course. Cheers, -A

[abhilash1910]

Thanks @karpathy , yes it would be best if you could link the repository : https://github.com/abhilash1910/llm.sycl/
Should I create a PR to add the repo link ?
We plan to have the same codebase inside intel/ as well and would ping/PR when that is done from our side.
Thanks again for the quick response, appreciated 👍🏻

[abhilash1910]

PR for readme update: #581

[abhilash1910]

Hi @karpathy , I had a doubt, will this repo only support gpt2 , or is there a plan to add on llama and other models as well ?
Also could you please help with #581 . Many thanks.
PS: Also an inprogress development work is ongoing for habana accelerators which I will link and let you know.

[abhilash1910]

Hi @karpathy could you help review #581 and #668 as both accelerator codes are in ready state for single card runs . More details will be followed on the separate forks. Includes Intel SYCL and Habana Gaudi accelerators.