📄 中文摘要
FlexAttention 现已在 Hopper 和 Blackwell GPU 上集成了 FlashAttention-4 后端。PyTorch 添加了对 CuTeDSL 分数/掩码修改函数的自动生成支持，并实现了 FlashAttention-4 的 JIT 实例化，以便于用户自定义。这一进展显著提升了模型的计算效率和灵活性，使得在处理大规模数据时能够更快速地进行训练和推理，满足了日益增长的计算需求。
FlexAttention + FlashAttention-4: 快速与灵活
出处: FlexAttention + FlashAttention-4: Fast and Flexible
发布: 2026年3月5日
📄 English Summary
FlexAttention + FlashAttention-4: Fast and Flexible
FlexAttention has now integrated a FlashAttention-4 backend on Hopper and Blackwell GPUs. PyTorch has added support for the automatic generation of CuTeDSL score/mask modification functions and JIT-instantiated FlashAttention-4 for user customization. This advancement significantly enhances the computational efficiency and flexibility of models, enabling faster training and inference when handling large-scale data, thus meeting the increasing computational demands.