RT-Lynx: Putting the GEMM Sparsity In a Right Way for Diffusion Models

Hugging Face Daily Papers Papers
diffusion-models sparsity activation-sparsity gemm acceleration transformers icml-2026

Summary

RT-Lynx proposes using activation sparsity instead of weight sparsity to accelerate diffusion models, achieving up to 1.55× linear-layer speedup while maintaining generation quality, and is accepted at ICML 2026.

Diffusion Transformers (DiT) achieve strong performance in image generation but incur substantial inference costs. While prior work has reduced this cost via quantization and distillation, semi-structured sparsity, which can nearly halve FLOPs, remains underexplored. A key reason is that most existing approaches focus on weight sparsification, and pruning 50% of the weights can remove critical model capacity and degrade generation quality. Our study, however, shows that DiT activations are intrinsically sparse and significantly more robust to N:M semi-structured sparsification than weights. Motivated by this observation, we advocate a paradigm shift from weight sparsification to activation sparsification. We propose RT-Lynx, which applies N:M sparsification to activations and incorporates error-compensation techniques to mitigate accuracy loss. We further implement highly optimized CUDA kernels tailored to this setting, achieving up to a 1.55x speedup on average in linear layers. Extensive experiments across multiple diffusion models demonstrate that our method preserves the generation quality of the original models while substantially accelerating inference.
Original Article
View Cached Full Text

Cached at: 05/27/26, 02:47 AM

Paper page - RT-Lynx: Putting the GEMM Sparsity In a Right Way for Diffusion Models

Source: https://huggingface.co/papers/2605.26632 👋 Hi everyone! We’re excited to share our ICML 2026 workRT-Lynx: Putting GEMM Sparsity in the Right Place for Diffusion Models.

Semi-structured sparsity has the potential to nearly halve GEMM FLOPs, but applying it to diffusion models remains challenging: conventional weight sparsification often removes critical generative capacity and causes visible quality degradation.

We revisit this problem and find that, unlike weights, DiT activations are intrinsically sparse and significantly more robust to 2:4 semi-structured sparsity. This suggests that activation sparsity is a better target than weight sparsity for accelerating Diffusion Transformers.Based on this observation, we propose RT-Lynx, which shifts the sparsification target from weights to activations. It combines online activation sparsification with norm-based compensation and a lightweight LoRA branch to recover fine-grained visual details.To make this practically efficient, we further design optimized CUDA kernels that fuse sparsification, compression, and sparse Tensor Core computation into a unified inference pipeline.

Across Qwen-Image, FLUX.1-dev, and Z-Image, RT-Lynx preserves generation quality while achieving around 1.2× end-to-end speedup and up to 1.55× average linear-layer acceleration.

We hope this work highlights activation sparsity as a more suitable and hardware-friendly direction for accelerating modern Diffusion Transformers. Feedback is very welcome!

Similar Articles

Supportive Token Revealing for Fast Diffusion Language Model Decoding

arXiv cs.CL

This paper proposes AXON, a training-free module that improves the quality-latency trade-off of discrete diffusion language model decoding by intelligently selecting 'anchor' tokens to reveal first, using attention, uncertainty, and confidence signals to support subsequent denoising steps. Experiments on reasoning and code-generation benchmarks show AXON reduces function evaluations while maintaining or improving accuracy.

Bug or Feature^2: Weight Drift, Activation Sparsity, and Spikes

Hugging Face Daily Papers

This paper formally proves that training neural networks with asymmetric activation functions like ReLU, GELU, or SiLU causes weights to drift negative, leading to up to 90% activation sparsity. It also shows that squared activations like ReLU² improve performance but cause activation spikes, which can be fixed by clipping, with GELU² achieving the best validation loss.