6 projects
hiera-optim
Drop-in throughput and memory optimisations for FAIR Hiera. 0.2: graph-safe MAE for torch.compile(reduce-overhead) — 2.37x on Hiera-Base GH200 over eager.
siglip-kernel
Fused Triton kernels for memory-efficient SigLIP training
neuroencoder
EEG model embeddings - distilled EPI-250k
flash-eeg
GPU-accelerated electrophysiology (EEG, iEEG, LFP) transforms for large batch jobs
output-shape
A very lightweight and minimalistic output shape examiner of layers and models.
greenscreen
A package for checking the energy efficiency of large models while training
