video1.0<iframe src="https://www.loom.com/embed/0d491fee18cb459da598e6a99cf37185" frameborder="0" width="2320" height="1740" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>17402320Loomhttps://www.loom.com17402320https://cdn.loom.com/sessions/thumbnails/0d491fee18cb459da598e6a99cf37185-800a90483b565cb5.gif149.8239In this video, I showcase a compiler I've developed that can efficiently fuse Triton kernels, like FlashAttention, from Torch code. I demonstrate how the compiler can handle various operations, such as adding logic for soft cupping, inlining functions, and reasoning about shapes and sparsity. Additionally, I highlight the performance benchmarks, showing that our implementation maintains high efficiency compared to existing solutions like Torch compile and FlexAttention. I encourage viewers to check the readme for more details and benchmarks. Your feedback on this tool would be greatly appreciated as we continue to refine its capabilities.