Practical overview
When it fits
Use RelGT when developing or studying predictive models for multi-table relational data, or when preparing to reproduce its RelBench experiments.
Who it is for
- Researchers studying relational deep learning and graph transformers
- Machine-learning practitioners building predictive models from connected tables
- Users reproducing RelBench experiments
Documented uses
- Build predictive models on multi-table relational databases.
- Represent relational entities as heterogeneous temporal graphs for learning.
- Explore attention over sampled neighborhoods alongside global learnable centroids.
- Reproduce RelBench experiments using the provided scripts.
Requirements & limits
The repository is public and lists the MIT license.
- The implementation is Python-based and documents a Linux environment using Python 3.12, PyTorch with CUDA and PyTorch Geometric.
- The experiment scripts target an eight-GPU NVIDIA A100 setup but can be adjusted; no specific alternative configuration is validated here.
- The architecture combines sampled-subgraph local attention with global attention to learnable centroids.
- The repository notes that its PyTorch Distributed Data Parallel support may not be fully optimized.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up RelGT’s documented experiment environment
Create the Python and graph-learning environment described by the repository as preparation for RelBench experiments.
Before you start
- Linux environment
- Micromamba and uv installed as described in the README
- Python 3.12
- PyTorch with CUDA 12.1 and the corresponding PyTorch Geometric dependencies
Steps
Create and activate a Micromamba environment named gt with Python 3.12.
Install PyTorch, torchvision, torchaudio and pytorch-cuda 12.1 using the PyTorch and NVIDIA channels.
Install uv, the listed PyTorch Geometric dependencies and pyg.
Install the additional packages listed in the README, including RelBench and pytorch_frame[full].
Expected outcome
A software environment configured according to the repository’s setup instructions, ready for the documented experiment stage.
Guide limitations
- The README’s experiment scripts are configured for an eight-GPU NVIDIA A100 server and may need adjustment for another environment.
- The setup instructions alone do not run experiments or establish compatibility with a different GPU configuration.
Select a RelBench experiment script
Choose a repository script for a specific RelBench experiment after setting up the environment.
Before you start
- RelGT repository
- The environment prerequisites listed in the README
Steps
Review the available scripts under expts/ to identify the specific experiment.
Use the corresponding expts/run-.sh script as the README’s documented experiment entry point.
Expected outcome
A selected script for the intended RelBench experiment.
Guide limitations
- The README gives the script pattern but does not document a specific experiment command or result in the supplied material.
- Scripts may need adjustment for user environments; the listed configuration targets eight NVIDIA A100 GPUs.
- Distributed Data Parallel is supported in the main file but may not be fully optimized.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 15 October 2026
- Original website title
- snap-stanford/relgt · Official homepage title · retained in its source language