FlashMLA

FlashMLA is DeepSeek's open-source library of optimized attention kernels for accelerating DeepSeek model inference/training workloads, including dense and sparse MLA/MHA kernels for prefill and decoding on NVIDIA GPU architectures.