Relax Version Bounds to enable `sktime` integration

[benHeid]

Hi thank you for sharing the weights and the model code.

We are currently trying to interface timesfm in sktime. Thereby, we are struggling with the version bounds, since this is leading to a lot of conflicts with other packages and makes it very difficult to install it. Thus, I wonder if it would be possible, to relax the version bounds so that it is easier to interface in sktime.

Issue in sktime: https://github.com/sktime/sktime/issues/6408

[rajatsen91]

Hi @benHeid, which of the version bounds are problematic for your integration ? Do you want us to have a version for python >= 3.11 ?

[fkiraly]

Just to add a comment on this (this is fkiraly, also from sktime), there is a more general statement:

• in fact, all of the pins are problematic for integration
• and not just with sktime, but with anything.

The short story is a bit of a paradox:

• pins are great for ensuring stability of a solution that no one depends on, e.g., deployed pipelines in prod with elevated stability requirements
• but (too many) pins make a package basically unuseable as a dependency, especially together with other packages

Why: imagine there are two packages that pin everything and that you want to use. If even a single of these pins differs in version, the requirements are no longer satisfiable, i.e., the pip install resolution will fail. As a corollary, updates of pins or versions will also lead to breakages for your users.

The tl;dr is: any pin in your requirements is a significant problem for your users. I would request to change all pins to version ranges, optimally only lower bounds. A more defensive alternative is loose upper bounds (e.g., at next minor or next major - but note that this will work only if you update these regularly!).

[fkiraly]

Btw, some examples of the consequence of pinning can already be seen here:

https://github.com/google-research/timesfm/issues/101 https://github.com/google-research/timesfm/issues/100

As a rule of thumb that I like to apply mentally, for each user that complains there are 100s to 1000s that do not...

[rajatsen91]

Does our pytorch release work for your use case ?

[benHeid]

@geetu040, you worked on the TimesFM. Can you answer the question if the new release work for our use-case?