Plug-and-Play Spiking Operators: Breaking the Nonlinearity Bottleneck in Spiking Transformers

TL;DR

This paper introduces a plug-and-play framework for implementing nonlinear operators in spiking neural networks, enabling compatibility with large language models without fine-tuning.

Contribution

It proposes a modular, spike-friendly approximation method for nonlinearities in Transformers, supporting common functions without additional training.

Findings

Less than 1% accuracy drop on LLM tasks when replacing nonlinear operators

Supports nonlinearities like Softmax, SiLU, and normalization in SNNs

Decomposes nonlinear computations into primitives using LIF neurons and bit-shift scaling

Abstract

ANN-to-SNN conversion offers a practical, training-free route to spiking large language models. However, current pipelines primarily focus on spike-driven realizations for Transformer linear-algebra operations, while providing limited support for key nonlinear operators. This gap limits compatibility with neuromorphic-style execution constraints, where such nonlinearities typically require division, exponentiation, or norm computations that are not naturally supported by standard leaky integrate-and-fire dynamics. To solve this problem, we propose a plug-and-play framework that implements spike-friendly approximations for Transformer nonlinearities and integrates into existing ANN-to-SNN pipelines. Our method decomposes these nonlinear computations into three recurring primitives -- division, exponentiation, and $\ell_2$ norms -- and realizes them via population computation using LIF neuron groups, combined with lightweight bit-shift scaling to avoid floating-point arithmetic. By composing these primitives as modular operator blocks, our framework supports common Transformer nonlinearities (e.g., Softmax, SiLU, and normalization) without any fine-tuning. Experiments on a range of LLMs Transformers show that selectively replacing the targeted nonlinear operators incurs less than a $1\%$ accuracy drop across all evaluated tasks.