srimanachanta/ fold-attention
Attention kernels for NVIDIA Hopper that fix the softmax reference before scanning the KV cache. Gives faster long-context decode and a bit-identical deterministic backward pass.
Attention kernels for NVIDIA Hopper that fix the softmax reference before scanning the KV cache. Gives faster long-context decode and a bit-identical deterministic backward pass.
GPU finite element solver for transcranial magnetic stimulation (TMS) electric fields in tetrahedral head models. Supports single-placement simulation, placement sweeps, conductivity uncertainty analysis, and coil-placement optimization on NVIDIA GPUs.
Low-cost 3D-printed prosthetic hand with 15 degrees of freedom, controlled by a transformer model that classifies surface EMG signals. Runs on an Orange Pi with custom I2C servo drivers. Presented at ISEF 2024.
Brain tumor segmentation on BraTS 2019, 2020, and 2021. Combines region-of-interest cropping, an attention-gated convolutional autoencoder, and energy-based voxel uncertainty estimates.