SpikingJelly: An Open-Source Spiking Neural Network Deep Learning Framework Based on PyTorch

SpikingJelly is an open-source Spiking Neural Network (SNN) deep learning framework based on PyTorch, providing complete functionalities such as model construction, ANN-SNN conversion, CUDA/Triton acceleration, and neuromorphic dataset support. Recently, its research work on memory-efficient training was accepted by ICLR 2026, offering important support for the development of the SNN field.

As the third generation of neural networks, Spiking Neural Networks (SNNs) have advantages of high energy efficiency and strong biological interpretability, closely resembling the operating mechanism of biological nervous systems. Their energy consumption is significantly reduced on dedicated hardware. However, SNN development has long faced challenges like incomplete toolchains and complex learning algorithms—SpikingJelly was created to address these issues.

SpikingJelly uses syntax consistent with PyTorch and supports classic neuron models such as IF, LIF, and ParametricLIF. Its modular design allows users to quickly build SNNs. It also provides ANN-SNN conversion tools; through strategies like weight scaling and threshold adjustment, users can gain SNN energy efficiency advantages while maintaining performance.

The framework introduces the Triton backend for IFNode, LIFNode, etc., generating optimized GPU kernels to improve computational efficiency. The FlexSN tool can dynamically convert spiking neurons into Triton kernels. Additionally, it has built-in neuromorphic dataset support and a unified loading interface, making SNN experiments convenient for users.

The SpikingJelly team's paper 'Towards Lossless Memory-efficient Training of Spiking Neural Networks via Gradient Checkpointing and Spike Compression' was accepted by ICLR 2026. This method combines gradient checkpointing and spike compression techniques, integrated into the spikingjelly.activation_based.memopt module, which significantly reduces memory usage without performance loss.

The framework supports spiking self-attention mechanisms (SpikingSelfAttention, QKAttention) for complex sequence tasks; the nir_exchange module enables interoperability with NIR format; the op_counter tool helps analyze model computational complexity, enriching the SNN development ecosystem.

SpikingJelly uses a version strategy where odd-numbered versions are development versions (synchronized with repository updates) and even-numbered versions are stable versions (released on PyPI). The community is active, offering bilingual Chinese-English documentation, clear contribution guidelines, and code standards, accepting contributions from global developers.

SpikingJelly has great potential in edge computing, low-power AI, and neuromorphic chip development, facilitating research validation and engineering deployment. Future plans include optimizing Huawei NPU support and improving the Triton backend to become a bridge between SNN research and industrial applications.