tile-ai
Plan for TileLang integration with TVM-FFI
Hi TileLang team, I’m a developer from the @PaddlePaddle and I’d like to ask about TileLang’s plans for integrating TVM-FFI.
I noticed the mention in #960 regarding future TVM-FFI integration. Since TVM-FFI (now separated from TVM) focuses on DLPack-based tensor exchange and provides a clear FFI interface for DSLs and external kernel implementations, it can greatly simplify interoperability between frameworks. I have a couple of questions:
- Do you have a more specific timeline or roadmap for adding TVM-FFI support to TileLang? I couldn’t find a related item in the roadmap issue (#79).
- After TVM-FFI is supported, are there plans to make TileLang framework-agnostic so it can be used out-of-the-box with more frameworks (for example, PaddlePaddle), and make PyTorch an optional dependency? I’ve noticed many current PyTorch-based usages in the codebase could be replaced by TVM-FFI.
Thanks for your work on TileLang — we’d be glad to help with testing or building integration prototypes if that’s useful.
@SigureMo Thanks for your advice. We do have a plan, but for now, it would be better for us not to rely directly on TVM FFI, since TVM now depends on tvm-ffi. We plan to integrate tvm-ffi in our next update, once TVM has a new release.
But due to certain reasons, we can move this matter forward, and perhaps complete it before Oct. 21.
@SigureMo I think this pr can be closed, we have now fully integrated with tvm ffi and the performance looks amazing.
@LeiWang1999 Thank you so much for your work! I have already validated TileLang’s basic execution flow under PaddlePaddle. To make it easier for our internal R&D teams, I’m temporarily maintaining a fork at https://github.com/PFCCLab/tilelang-paddle and publishing separate packages from there. If there’s an opportunity, we’d love to help make PaddlePaddle one of the officially supported TileLang frameworks. Looking forward to further collaboration with you!