thu-ml / SLA

SLA: Beyond Sparsity in Diffusion Transformers via Fine-Tunable Sparse–Linear Attention