Monroe
Copyright (c) 2026 Blazej Banaszewski

This product includes software adapted from the third-party projects listed
below. Both are used under the MIT License, and their original copyright
notices are reproduced here as that license requires.

--------------------------------------------------------------------------------
LibMTL
https://github.com/median-research-group/LibMTL
MIT License — Copyright (c) 2021 median-research-group

The entire multi-task weighting module (monroe/mtl/) is derived from LibMTL.
This includes the AbsWeighting base interface (monroe/mtl/abstract_weighting.py)
and every weighting strategy: EW, UW, RLW, DWA, and STCH.

LibMTL cites the original paper and reference implementation for each individual
strategy in that strategy's class docstring; those citations are retained
verbatim in the corresponding files here.

--------------------------------------------------------------------------------
GRIT — Graph Inductive Biases in Transformers without Message Passing
https://github.com/LiamMa/GRIT
MIT License — Copyright (c) 2023 Liheng Ma

The encoder (monroe/model/grit.py) is a derivative work of the official GRIT
implementation, including its attention computation
(MultiHeadAttentionLayerGritSparse) and transformer layer (GritTransformerLayer).
Monroe extends it with molecule-specific components — RBF edge-distance
expansion, learned fills for missing node features, stereo edge features, and
multi-head attention pooling — but the core architecture is GRIT's.

GRIT is introduced in "Graph Inductive Biases in Transformers without Message
Passing" (ICML 2023), Ma et al.
