【[22星]AttentionEngine:一个统一的框架,用于定制注意力机制,包括Transformer注意力和线性注意力。亮点:1. 只需80行代码即可定义softmax注意力;2. 自动生成高性能的设备代码;3. 支持NVIDIA H100等多款设备】
'AttentionEngine is a unified framework to customize attention, including transformer attention and linear attention. AttentionEngine provides users with pythonic interface to define customized attention flexibly and automatically generate device code with high performance. For example, user can define softmax attention with only 80 lines of code and get the optimized fused kernel automatically.'
GitHub: github.com/microsoft/AttentionEngine
#注意力机制# #高性能计算# #AI框架# #AI创造营#
'AttentionEngine is a unified framework to customize attention, including transformer attention and linear attention. AttentionEngine provides users with pythonic interface to define customized attention flexibly and automatically generate device code with high performance. For example, user can define softmax attention with only 80 lines of code and get the optimized fused kernel automatically.'
GitHub: github.com/microsoft/AttentionEngine
#注意力机制# #高性能计算# #AI框架# #AI创造营#