[New Model] TiRex Foundation Model

[daidahao]

This is a sub-issue of #2933.

Is your feature request related to a current problem? Please describe.
TiRex is a new and lightweight 35M time series foundation model based on xLSTM, a next-gen RNN architecture. Like previous foundation models supported in Darts, TiRex is a top scorer on public forecasting benchmarks such as GIFT-EVAL and fev. In addition, the official library tirex-ts has strong technical support and fast response time.

Supporting TiRex in Darts could bring meaningful performance and efficiency benefits to our users, while also boosting the model standing and usage. I have invited TiRex team for a Darts implementation (see NX-AI/tirex#29). They are interested and happy to contribute a GlobalForecastingModel-based implementation which wraps tirex-ts as the single source of truth.

I am now creating this issue to align on API design between TiRex team (led by @lukfischer @martinloretzzz) and Darts team (led by @dennisbader), and to host related discussions before TiRex submits an implementation PR.

Describe proposed solution
A clear and concise description of what the library should provide to solve missing functionality.

Describe potential alternatives
A clear and concise description of any alternative solutions or existing features that might solve it.

Additional context
Thank you to @lukfischer and TiRex team for the quick response and the generosity of offering code contributions.

[dennisbader]

Hi @daidahao, @lukfischer and thanks for opening this feature request and the interest. I've quickly scanned through TiRex and agree that it seems like a great model. Congrats for that :) 🚀

I do have some concerns however to add it to Darts:

• I'm a bit worried about the NXAI license and its commercial terms. We use the Apache-2.0 license which offers full / free commercial use. Not sure about the legalities there, and would like to avoid any potential risks.
• Regarding the implementation: Now that we've added the FoundationModel base model, I was rather thinking to implement new models built on top of it, rather than going through the GlobalForecastingModel. Through FoundationModel, we have the full Darts torch-capabilities (optimized backtesting, lightning support, ...).

I'm open for discussing this further though :)

[daidahao]

@dennisbader

Thank you for your comments. While we await TiRex team response, I would like to propose a solution to address your first concern on NXAI license:

To avoid users unknowingly accepting non-Apache-2.0 license, may I suggest a accept_license boolean parameter for TiRexModel, similar to sktime implementation here. By default, accept_license would be False, and users would need to explicitly set it to True to proceed--at which point the NXAI license could be displayed and TiRex enabled. What do you think?

As for any other requirements in the NXAI license, I’m not a legal expert, so I’d defer to the TiRex team on potential solutions that don’t conflict with Darts’ Apache-2.0 licensing.

[lukfischer]

Hi @dennisbader,

thank you for the thoughtful feedback and for taking the time to review TiRex. We appreciate it!

On the license concern:
We fully understand the sensitivity around license compatibility, especially given that Darts is distributed under Apache-2.0.

TiRex is released under the NXAI Community License, which is based on the Meta Llama 3 Community License. The intent is to make the model freely available for research, academic use, and small to mid-sized enterprises, while requiring a separate commercial discussion only for very large companies exceeding €100M annual revenue that incorporate NXAI material into a commercial product or service.

We recognize that this differs from a fully permissive Apache-style model license, and we want to ensure there is no ambiguity for Darts users.

The proposed accept_license flag from @daidahao is a reasonable and clean mechanism. We would be happy to implement and integrate something like that.
If you have specific preferences for how such acknowledgements should be handled within Darts, we are happy to align.

On the implementation via FoundationModel:
Regarding the architectural point: our initial intention was to use a thin GlobalForecastingModel wrapper around tirex-ts. The motivation was to keep TiRex as the single source of truth and avoid duplicating logic in Darts, thereby minimizing long-term maintenance overhead when new TiRex versions are released.

That said, your point about leveraging the FoundationModel abstraction is very valid. If the long-term direction in Darts is to consolidate foundation models under that base class, we are open to adapting accordingly.

Our main concern is to keep the wrapper minimal and avoid re-implementing TiRex functionality inside Darts. As long as the integration remains a thin adapter around tirex-ts, aligning with FoundationModel should not introduce additional maintenance burden on our side.

We are happy to iterate on both points before moving forward. Our goal is to integrate TiRex in a way that is legally clean and architecturally consistent with Darts.

Looking forward to your thoughts.

[dennisbader]

Hi all, and thanks for joining the discussion @lukfischer.

• Regarding the license concerns: The approach you suggested regarding making the user explicitely accept the license via an accept_license parameter sounds good to me.
• Regarding the model implementation: I'm okay with adding TiRex as optional dependency if we can base the wrapper implementation on the foundation model :) In short that means, the source implementation lies on TiRex side. Darts will take care of model instantiation, weights loading, torch related setup, data ingestion, ...
  For that we need access to the forward() method (or similar) that generates forecasts from the input batch

[lukfischer]

Sounds good! We already have a first implementation based on foundation model in our fork. Next is an internal PR review and then we would be ready to open a PR directly with you. Initially this will only cover forecasting. Later we can also add classification and regression based on TiRex if ok for you.
Thx for your support!

[dennisbader]

Thanks @lukfischer! Looking forward to the PR once your ready 🚀

Covering only the forecasting case for now sounds great! Regarding regression or classification: we don't yet handle these cases in Darts, and therefore shouldn't add it at the current time.

Time series classification / clustering is a big feature on our roadmap though. But it'll take a while until we get to that point (we need to properly design the solution).