All issues

Selective scan is recomputed during inference ? consider caching hidden states for autoregressive generation

#8

· eeshsaxena opened

on Jul 30, 2026

MambaConfig has no dropout field — training on small datasets will overfit

#7

· eeshsaxena opened

on Jul 16, 2026

Backward pass never updates grad_C — C matrix receives zero gradient during training

#6

· eeshsaxena opened

on Jul 16, 2026

parallel_scan down-sweep is incorrect — produces wrong outputs for power-of-2 lengths

#5

· eeshsaxena opened

on Jul 16, 2026

generate() re-processes full sequence every step — O(L^2) instead of O(L)

#4

· eeshsaxena opened

on Jul 16, 2026

C matrix in selective_scan is shared across d_model channels - incorrect per the paper

#3

· eeshsaxena opened

on Jul 16, 2026

A matrix uses S4 HiPPO initialization instead of Mamba initialization from the paper

#2

· eeshsaxena opened

on Jul 16, 2026

selective_scan runs a Python for-loop instead of parallel associative scan - defeats the purpose of Mamba

#1

· eeshsaxena opened

on Jul 16, 2026