Mingfei Ma developed high-performance backend optimizations for deep learning inference across the Mintplex-Labs/whisper.cpp and bytedance-iaas/sglang repositories. He introduced an Intel AMX backend to ggml in whisper.cpp, enabling efficient matrix multiplication for quantized data types on compatible hardware through low-level C++ and template metaprogramming. In sglang, Mingfei optimized native CPU kernels for activation functions, batch matrix multiplication, and attention mechanisms, leveraging AVX and parallel computing to improve throughput for CPU-bound workloads. He further enhanced GEMM kernels and implemented BRGEMM support for int8 and fp8, refactoring thread management to maximize CPU utilization and inference performance in production environments.