Triton kernels for large models, including sparse attention. - View it on GitHub