Nextlat: Train latent-prediction transformer research models

Understand the project

Practical overview

Use Nextlat when you want to investigate or reproduce latent-prediction transformer research using its code, configurations and benchmark workflows.

Researchers studying self-supervised latent prediction and compact world models.
Practitioners comparing transformer training objectives using the repository’s included baseline algorithms.
Readers seeking a code implementation and configured research workflow rather than a hosted model service.

Inspect the NextLat model implementation and compare its loss computation with included baseline algorithms.
Configure experiments using the repository’s YAML benchmark and hyperparameter files.
Train experiments through the single-GPU or documented Lightning Fabric/DDP workflow.
Track metrics through CSV files or Weights & Biases and follow documented post-training evaluation guidance.

Before you choose

The public repository is open source and lists the MIT license.

The repository requires PyTorch 2.6 or later; PyTorch 2.9 or later is needed for the Muon optimizer.
Training supports documented single-GPU and multi-GPU DDP routes; reproducing paper results used NVIDIA RTX A5000, H100 NVL and B200 GPUs.
Some datasets are downloaded from Hugging Face; setting HF_TOKEN is documented as a way to obtain higher rate limits and faster downloads.
Post-hoc Manhattan Taxi Rides and TinyStories evaluations require checkpoints saved during training.
The README cautions that torch.compile can yield inconsistent results on Path-Star and A5, and recommends compile: false for those benchmarks when reproducing results.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Prepare the Python environment with the repository’s declared requirements.

Install the listed dependencies with the command below.
Documented command
```
pip install -r requirements.txt
```
Use a PyTorch version meeting the repository’s stated minimum; select PyTorch 2.9 or later if using Muon.

The repository’s declared Python dependencies are installed, with the documented PyTorch version requirement observed.

This covers dependency installation only; it does not establish that training or evaluation will run successfully on a particular machine.
GPU, dataset and other runtime requirements are not fully specified in this setup guidance.

Evidence and transparency

Source record details

Source availability: Active
Sources checked: 14 October 2026
Original website title: JaydenTeoh/NextLat · Official homepage title · retained in its source language