MLIR Area Team Meeting Minutes - 2025-03-13

This was our first meeting after forming the area team, still figuring out what the community expects from us. Thank you everyone for attending and asking questions. We’re off to a good start!

The notes document is public, updated every meeting.

Here are the notes from this meeting:

  • Gathering opinions of the community about what we should be doing and how.
  • Clarity on how the community is changing: will there be any change or still “business as usual”?
    • Mostly reactive, waiting for people to ask for help, connect people to experts
    • Like “code owner”, but instead of PRs, RFCs: make sure discussions reach conclusions
  • How do we solve real conflicts (no consensus reached after long time, big disagreements)
    • Instead of making decisions, reach a compromise design, and propose that as a solution
    • Thinking longer term, propose multi-step process, work, re-evaluate, decide next step
    • Create parallel prototypes, see what breaks, then take decisions
  • Can we help newcomers start working with MLIR?
    • Tutorials? Better documentation?
    • Linalg group is looking at rationale, docs, information
    • What about core? Core is more stable, but still needs some care
    • Clearer core rationale would help other areas to be more well defined
    • Crowd-sorcing would help (docs, videos, tutorials)
  • Crazy idea: use some LLM bot to help newcomers
    • Parsing docs, etc.
    • How do we handle hallucination (this is highly technical)
    • Collection of good prompts to guide models to give reasonable answers
    • Free vs. paid models vary hugely in quality
  • Can we have steps to reproduce output on docs?
    • Some automation to generate the docs with actual tools
    • Similar to “Jupiter Notebooks” but simpler
    • Table gen with the back-ends, showing the actual output
    • MLIR with transforms, showing the transformed IR
    • Test cases that ignore text but look at code and CHECK output
  • How do we know which pass would have an impact on a particular op/file
    • A (Python) tool that looks at all alternatives (PR)
  • Operation interfaces documentation are missing
    • Table-gen does generate docs, but they’re not linked to their actual usage
    • Generated docs may not be included in other more global docs
    • Need to check if they’re included in the right place
  • Proposals for the future meetings
    • People facing problems could come and present their problems
    • This can serve as a starting point for decision making, consensus forming

@ftynse @jpienaar @matthias-springer

6 Likes