Examples

Each subdirectory of examples/ is a self-contained training recipe. Recipes are nvsubquadratic.lazy_config.LazyConfig trees describing the network, datamodule, Lightning wrapper, and trainer; running them is python -m experiments.run --config <path>.

The active experimental roadmap (priorities, owners, status) lives at examples/overview_tracker.md.

Classification

MNIST / SMNIST

examples/mnist_classification/ covers MNIST with both attention and Hyena baselines, plus a small CCNN backbone. examples/smnist_classification/ covers sequential MNIST (1D input).

ImageNet

examples/imagenet_classification/ ships seven CCNN configs (Hyena / Hyena-circular / attention, with and without augmentation, plus tiny variants for laptop sanity checks). Representative entry points: ccnn_7_512_hyena.py, ccnn_7_512_attention.py.

TinyImageNet: ViT-5

examples/vit5_imagenet/ is the ViT-5 baseline suite (v1–v5) with its own TRACKER.md.

UCF101

examples/ucf101_classification/ covers video classification with both sequence- and clip-mode datamodules.

Diffusion

MNIST

examples/mnist_diffusion/ is a small DDPM/JiT diffusion sanity-check.

ImageNet

examples/imagenet_diffusion/ is the full ImageNet diffusion setup. See its README for the JiT vs Hyena-vs-attention comparison.

Spatial recall

examples/spatial_recall_1d/, spatial_recall_2d/, spatial_recall_3d/, and the newer spatial_recall_v2/ are synthetic recall benchmarks measuring how well each mixer (Hyena, attention, Mamba, CKConv) routes information across long-range spatial/sequence positions. See spatial_recall_v2/TRACKER.md for the v2 task suite.

Benchmarks

examples/vit_b_benchmark/ holds the throughput-comparison configs used to produce the numbers in Benchmarks.

Scientific

The Well

examples/well/ covers The Well PDE benchmark; see its README for sub-datasets and baselines.