FlagGems is a backend-neutral kernel library for LLMs implemented in the Triton language, developed as part of the FlagOS open-source ecosystem. It provides optimized Triton kernels for common LLM operations like attention, softmax, layer normalization, and gemm, targeting efficient execution across different AI accelerators. The library aims to unlock hardware performance and reduce the cost of running AI workloads by providing drop-in replacements for standard PyTorch operations.