Release Planning - Reviewing our Tracking Issues

Hello!

Following from the community meeting on the 18th of November, I’ve gone over the tracking issues we have in flight to try to see what we could get into the next release of TVM, which I’m optimistically hoping to take a branch for in January.

I’ve marked the owners on each tracking issue below, and would like to get a response as to whether we should plan to include these features either fully or partially - I really appreciate your support in figuring this out together. Please reply in this thread if we should aim for the release to include your listed work.

The hope is that we continue to generate releases regularly, so if your issue doesn’t make this release expect another one in just a few months, with the aim to continuously improve the process :smile_cat:

Candidates

These are tracking issues with a majority of changes available in pull requests or completed.

@ashutosh-arm

@elenkalda-arm

@Leo-arm

@Lunderberg

@thierry

@jiangjiajun

@areusch

@junrushao1994

@manupa-arm

@andrewzhaoluo

Me

1 Like

Remaining

These are tracking issues either lacking detail or without a majority of the work available in pull requests.

@matt-arm

@mbs-octoml

@leandron

@hjiang

@jroesch

@hhhfccz on GitHub, unsure our Discuss

CC @denise @areusch @junrushao1994

For my tracking issue https://github.com/apache/tvm/issues/8751

Currently, PaddlePaddle frontend has supported 115 operators, and these would covering the most common deep learning models, we can see the frontend is available in TVM release 0.8.

And for the remaining task in the tracking issue

  • control flow support
  • Quantized model supporting

I’m still working on these, but there are some risks to finish in January, I’ll update the progress here

For planned updates for the Arm Ethos-N, this is small and should complete easily; most is already done. separate to this, I also aim to have an update of the driver stack in place.

Regarding to https://github.com/apache/tvm/issues/9410, I will expand it with more operators we want to support for NPU and we definitely won’t get all of these done by January :sweat_smile: So I think the content of this tracking issue will be included partially.

Thanks Chris. #9327 is just the required(?) companion for https://github.com/apache/tvm-rfcs/pull/38 which is moribund at the moment pending, I think, more detail from me and probably some good discussion between you and I. I’ll update the issue to that effect.

I might complete writing a tutorial for https://github.com/apache/tvm/issues/8296 but otherwise no plans to tackle the major bullet points on the issue.

That being said things work pretty well for a majority of models right now.

@Mousius , thanks for the follow up, about issue 8596, I just updated this issue to mark some PR as related not belong to issue solution scope to avoid confusion, the current progress is 2 of 5 planed PR are done, one is in process of review, the whole solution may can not be fully done by Jan, this feature should be include partially.

@Mousius Thank you for checking in on these. For my two tracking issues (8903 and 8901), I won’t have the bandwidth to push them through by January, so these aren’t expected for the next release.

thanks @mousius, let’s include the Project API in the next release. The C Device API can be considered landed when we can do heterogeneous execution from AOT–I’m not sure yet if it will make an end-of-Jan release, but I suspect it would make an end-of-March release.