# Multi-GPU collaborators wanted

The manuscript deliberately stops at a small CPU-friendly feasibility check. This repository turns the next scientific question into a reproducible scaling benchmark.

We are looking for contributors with **2, 4, 8, 16+ GPUs and multi-node systems**, especially H100/H200, B100/B200, GB200/GB300, MI300X-class, or comparable accelerators. Useful contributions include:

- strong- and weak-scaling runs at 128/256/512 resolution;
- larger low-rank transition ranks (32/64/128+);
- long-horizon rollout extensions and stability measurements;
- Moving MNIST, DMControl pixels, Habitat, Atari-like, or physical-video adapters;
- matched-compute Dreamer/PlaNet, Transformer, Koopman, and neural-operator baselines;
- profiling of throughput, peak memory, communication overhead, and operator kernels.

## Minimal contribution

1. Fork/clone the repository and install `requirements.txt`.
2. Run `./scripts/smoke_test.sh`.
3. Run, for example, `GPUS=8 RES=256 BATCH=16 ./scripts/run_scaling_sweep.sh`.
4. Add the generated JSON files and `results/leaderboard.csv` under a hardware-specific folder.
5. Open a PR with GPU model/count, CUDA/ROCm, PyTorch version, node topology, command, and any code changes.

Please do not report only the best run. Repeated seeds and failed/OOM configurations are scientifically useful.