Developed a fast exponential function (fexp) for AVX2 and AVX512 within the pytorch/pytorch repository to accelerate mixed-precision flash attention workloads. This work focused on optimizing exponential calculations by leveraging SIMD programming techniques and high-performance computing principles in C++. The implementation introduced SIMD-optimized methods for vectorized types, enabling efficient utilization of AVX2 and AVX512 instruction sets. As a result, the new approach delivered up to 20% performance gains in targeted operations. The contribution demonstrates depth in numerical methods and low-level C++ development, addressing a key computational bottleneck and enhancing the performance of core PyTorch functionality for advanced workloads.