Open MLIR Meeting 12/2/2021: end-to-end flow for ML Compiler

This Thursday (9am California Time, 16:00 UTC ), @stephenneuendorffer proposed that we have a discussion on the topic of end-to-end compilation flow for ML. At the moment there exists no end-to-end machine learning flow upstream, at least not in a packaged and ready-to-use way. Compared to other ML frameworks (Like XLA or TVM), MLIR has primarily focused on compiler building blocks. The lack of an in-tree flow is also a significant barrier to entry for compiler builders because we each have to figure out how all the pieces together to make a working system. Recently, several projects in-tree (TOSA) and out-of-tree (Torch-MLIR, IREE) have made significant progress and we seem to have most of the pieces to build a complete flow. We’ll discuss what might a community end-to-end ML flow would look like and strategize about how to build it.

As usual the information to join the meeting:
‪+1 218-301-8485‬ PIN: ‪255 745‬#

I’ll also update this thread with slides (if any) and recording after the meeting.


Nice discussion this morning, probably the first of many! Here are the slides and the recording.

Now the main thing is to not lose the momentum again on this track of work, and as such we concluded on the following action items:

  • We’ll gather the folks interested in driving this forward to form a working group.
  • We’ll meet weekly or bi-weekly, with an open-meeting.
  • The first goal for the workgroup will be to collaborate on a doc to describe the landscape around MLIR & ML Compilation and describe a few options moving forward.
  • We’ll also try to get consensus on a “problem statement” to address and on some possible milestone we should focus on as first steps.

I’ll follow-up with this in the coming days!