LowRank-SSM: Hardware-Software Co-Design for Rank-Reduced Mamba Acceleration on FPGA
LowRank-SSM is a hardware-software co-design framework that decomposes the input and output projection weights via post-training truncated SVD and introduces a greedy bandwise rank-allocation algorithm that searches for the per-band rank vector that minimizes weight storage while respecting a user-specified accuracy co...