Removed in Mamba-3: exponential-trapezoidal discretisation induces an implicit data-dependent convolution of size two on the state input, so the explicit short causal convolution is no longer needed.

Learn more →

Why study this historical topic?

Removed in Mamba-3: exponential-trapezoidal discretisation induces an implicit data-dependent convolution of size two on the state input, so the explicit short causal convolution is no longer needed.

The Short Causal Convolution in Early Mamba Blocks

a fixture of the Mamba-1 and Mamba-2 blocks that students still copy from tutorials, and a clean example of an architectural patch being absorbed into the mathematics

This Concept is waiting for its first lesson!

a fixture of the Mamba-1 and Mamba-2 blocks that students still copy from tutorials, and a clean example of an architectural patch being absorbed into the mathematics

Are you a teacher? Sign in to start contributing.