FlashQLA

Alibaba Qwen introduced FlashQLA, a TileLang-based linear-attention kernel stack that reports 2–3x faster forward passes and 2x faster backward passes. The release gives edge and long-context deployments a new optimization lever below the model layer itself.