[Question] 2026 Roadmap and Development Scope for TorchTitan

[BIGBALLON]

I would appreciate any insights you could share about the 2026 roadmap and development scope for TorchTitan. As a user and contributor, understanding the project’s future direction and planned features would be very helpful.

Additionally, is there a plan to train a fully functional and usable MoE (Mixture of Experts) model?

[tianyu-l]

PyTorch will publish the overall 2026 roadmap soon, which will have a section for torchtitan. Please stay tuned.

Meanwhile, happy to hear what the community think we should work on, if you could share!

Additionally, is there a plan to train a fully functional and usable MoE (Mixture of Experts) model?

This sounds interesting proposal -- may I ask what's the motivation for this request?

[BIGBALLON]

Hi， @tianyu-l

This sounds interesting proposal -- may I ask what's the motivation for this request?

Assuming the goal is to train a complete, usable model, this would bring several clear benefits:

Demonstrate real production readiness
It would show that TorchTitan goes beyond experimental setups and is capable of training end-to-end, production-grade models.

Drive wider adoption
Being able to train a fully usable model would make TorchTitan more attractive to teams working on large-scale training, helping it reach a broader user base.

Strengthen the community
As more users adopt TorchTitan for real training workloads, they’re more likely to contribute improvements, optimizations, and best practices back to the project.

Overall, this would significantly improve TorchTitan’s value as a practical, production-oriented training framework.

[tianyu-l]

cc @aditvenk @dcci

[githubsgi]

@tianyu-l, has the 2026H1 roadmap for TorchTItan finalized yet ?

[aditvenk]

@githubsgi Meta PyTorch H1 2026 roadmaps are available at https://dev-discuss.pytorch.org/t/meta-pytorch-team-2026-h1-roadmaps/3300
The distributed one is linked there and lists TorchTitan goals.

[githubsgi]

@aditvenk and @tianyu-l , thanks. Wondering if PR's/RFC's have already started landing for "compiler-first approaches" , "automatic parallelization", "DeepEP-style for communication deduplication DeepSeek-V3", "Shared model definition between Titan and vLLM", etc ? If those do not exist as of now, is there any place where more details can be found ?