VRVijay Ram Enagantiinvijay-ram.hashnode.dev·Aug 31 · 9 min readReproducing Sequential AttentionTL;DR Implemented Google's Sequential Attention feature-selection algorithm (ICLR 2023) from scratch in PyTorch (seq-attention/), faithful to Algorithm 1: a softmax mask over per-feature attention lo10
VRVijay Ram Enagantiinvijay-ram.hashnode.dev·Apr 2 · 16 min readTurboQuant by GoogleImplementing an ICLR 2026 paper on KV cache compression, discovering the gap between theory and practice, and building something that actually works. Idea The idea to try and build a justified clone 10
