Browse Teach Curate

Structured State-Space Duality Between Linear Recurrence and Masked Attention

an SSM with a scalar-times-identity transition is exactly masked attention under a 1-semiseparable mask, so the same operator can train as a parallel matmul and decode as a constant-memory loop — this dissolves the either-or the curation opened with

jeremy Video

Mamba State Space Models in PyTorch Explaining Mamba 2 Duality Training and Inference Efficiency

State space duality is a mathematical principle establishing that specific state space models with scalar transition matrices are rigorously equivalent to masked attention mechanisms. This theorem al…