XAMBA makes Mamba and Mamba-2 state-space models run up to 4.8x faster on an Intel AI PC NPU by replacing CumSum and ReduceSum with matrix operations and activations with piecewise linear approximations.
Efficiently Modeling Long Sequences with Structured State Spaces
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
XAMBA: Enabling Efficient State Space Models on Resource-Constrained Neural Processing Units
XAMBA makes Mamba and Mamba-2 state-space models run up to 4.8x faster on an Intel AI PC NPU by replacing CumSum and ReduceSum with matrix operations and activations with piecewise linear approximations.