Reverse engineering NVIDIA SASS instruction dictionary, kernel audits and pattern recognition across GPU architectures.
FlashKDA: high-performance Kimi Delta Attention kernels
FlashKDA: memory-efficient KDA kernels for training and decode
DeepSelect: TopK kernels for DeepSeek Sparse Attention (DSA) and Samplers