FFPA: Extend FlashAttention-2 with Split-D, ~O(1) SRAM complexity for large headdim, 1.8x~3x↑🎉 vs SDPA.
INT8 Sparse Tensor Core GEMM for PyTorch — built for Windows
