[ENH] Add LightTS model implementation (v2)

[Sylver-Icy]

Reference Issues/PRs

#2126

What does this implement?

This PR adds an implementation of the LightTS model for the PyTorch Forecasting v2 model framework.
Implemented the LightTS architecture based on the original paper.
The implementation follows the same design and integration patterns used by existing models such as DLinear.

Did you add any tests for the change?

Yes.

Added tests/test_models/test_lightts_v2.py which includes tests for:

• Model initialization with different hyperparameters
• Forward pass prediction shape
• Quantile loss output shape
• Univariate forecasting behavior

Any other comments?

The implementation aims to stay consistent with existing v2 model implementations (e.g., DLinear) in structure and integration patterns.

Feedback on architectural alignment with the original LightTS design is welcome.

[codecov]

Codecov Report

❌ Patch coverage is 89.54248% with 16 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (main@1952984). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
pytorch_forecasting/models/lightts/_lightts_v2.py	88.23%	12 Missing ⚠️
pytorch_forecasting/layers/_blocks/_ie_block.py	84.00%	4 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2135   +/-   ##
=======================================
  Coverage        ?   86.63%           
=======================================
  Files           ?      169           
  Lines           ?     9884           
  Branches        ?        0           
=======================================
  Hits            ?     8563           
  Misses          ?     1321           
  Partials        ?        0           

Flag	Coverage Δ
cpu	86.63% <89.54%> (?)
pytest	86.63% <89.54%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[phoeenniixx]

Thanks!
Please add docstrings to each class and atleast to all the public methods.

[Sylver-Icy]

Thanks for the feedback! I’ll add docstrings
For the _get_test_datamodule_from part I’ll check the implementation in timexer and adjust/remove it if needed.

And good point about _IEBlock moving it to the layers module makes sense. I’ll update the structure accordingly.

[PranavBhatP]

LGTM!

[Sylver-Icy]

@phoeenniixx I kept _IEBlock inside the lightts module for now since it’s currently a private helper used only by lightts. In this repo it looks like layers is mainly used for reusable/shared components across multiple models

since no other model is using it yet, I thought keeping it local would keep things a bit simpler for now xd
If you’d still prefer it moved to layers (or another shared location) just lmk I’m happy to update the structure

[PranavBhatP]

since no other model is using it yet, I thought keeping it local would keep things a bit simpler for now xd
If you’d still prefer it moved to layers (or another shared location) just lmk I’m happy to update the structure

please move it to the layers folders, all layer modules are meant to be a part of it regardless of whether it is shared or not.

[phoeenniixx]

Added some doubts and suggestions

[phoeenniixx]

I have a doubt, if n_quantiles is 0 then what will be the shape here? I think for point prediction losses the shape should be (batch,time,1) - as much as I remember. @PranavBhatP please correct me if I am wrong here

[Sylver-Icy]

n_quantiles cannot actually be 0 in this implementation. It is either
None or
len(self.loss.quantiles) when using QuantileLoss

[PranavBhatP]

I think for point prediction losses the shape should be (batch,time,1) - as much as I remember.

No, point prediction losses simply expect (batch_size, timesteps) - a 2d tensor.

[phoeenniixx]

i meant the output format of the forward it was 3D no? When using a PP loss