SageAttention gets autotuned block sizes

View on Reddit

Developer woct0rdho releases an autotuned block size feature for SageAttention, automatically optimizing performance for varying input sizes. The tool aims to improve attention efficiency, particularly for StableDiffusion models.