What's next?

[enjustli]

Since Microsoft no longer maintains this project, what should we do next?

[enjustli]

@nhat-nguyen

[enjustli]

I’ve noticed that development on https://github.com/triton-lang/triton-cpu has stalled. If we could merge triton-shared into triton-cpu, it would be beneficial for both projects.
of course, this is just a potential suggestion.

[enjustli]

I still believe that triton-shared has significant value in industrial settings. As the ecosystem around vLLM/sglang + PyTorch + Triton becomes more mature, more ASIC vendors will need to integrate—or are already integrating—with Triton. The work done by triton-shared provides strong reference value for them. I myself am an NPU chip developer, and I very much agree with the approach taken by triton-shared. Converting Triton tensor<ptr> into the form ptr + offset : {int / tensor<int>} offers a lot of practical convenience.

[enjustli]

Or, can we try submitting this code repository to https://github.com/triton-lang, just like triton-cpu? @manbearian @nhat-nguyen @parsifal-47

[enjustli]

due to the Triton community is evolving so quickly, the matter of Triton-shared has become very urgent. Sorry to bother all of you. 🥹🥹🥹

[parsifal-47]

@enjustli, I can't speak for everybody, but to the best of my knowledge all of us are now are far from decision making organizations for this project. If you have enough people to support it at triton-cpu and licenses are compatible, why not.

[enjustli]

Frankly speaking, I do not have enough people 😂😂😂
Based on the current progress of the triton-cpu project, immediately starting the integration of triton-shared would be a huge challenge. Issues include keeping up with triton-cpu’s code updates, communicating with the triton-cpu developers, and whether they would even agree to the merge.

However, from a long-term development perspective, merging the two is a good option, since triton-shared can also run directly on CPUs.

We can first find a temporary solution that allows triton-shared to continue moving forward while we communicate with the triton-cpu developers to see if they are open to such an integration.

[enjustli]

I think the top priority is to keep triton-shared moving forward. Only then can we consider any future development.

[nhat-nguyen]

@enjustli We have a meeting with Microsoft this Thursday to figure out the project direction and will post any updates here

[enjustli]

@nhat-nguyen is there any update? 🍺

[nhat-nguyen]

We had a discussion between Meta and Qualcomm. Meta is going to continue development and maintenance of the project going forward. It might take a bit of time as we figure out the logistics required for creating a new fork.

[enjustli]

Is there any update? @nhat-nguyen

[nhat-nguyen]

Is there any update? @nhat-nguyen

We're still going through internal review, stay tuned!

[nhat-nguyen]

The new Meta-supported repository is now live at https://github.com/facebookincubator/triton-shared

[enjustli]

The new Meta-supported repository is now live at https://github.com/facebookincubator/triton-shared

I have seen new repo, could we preserve the Git commit history? this would be very helpful for ongoing and incremental development? @nhat-nguyen