Rajib Lohia developed and integrated CDNA3 MFMA support for the flash attention MMA kernel in both the llama.cpp and ggml repositories, targeting MI300X (gfx942) GPUs. Using CUDA and C++, Rajib implemented FP16 MFMA intrinsic paths and optimized dispatch logic to handle various head sizes, replacing macros with constexpr warp sizing for improved code maintainability. The work addressed Q loading and stride handling for non-power-of-2 heads, resulting in throughput gains of 7% to 39% on large input batches. All 2480 flash attention tests passed, demonstrating robust performance optimization and correctness for large-context model inference workloads.