What it is
DeepGEMM is a CUDA kernel library from DeepSeek that provides matrix multiplication kernels in FP8, FP4 and BF16, grouped GEMMs for mixture-of-experts models, fused MoE with overlapped communication, and attention scoring kernels for the DeepSeek V3.2 indexer. Kernels are JIT-compiled at runtime and called from Python via PyTorch on NVIDIA Hopper (SM90) and Blackwell (SM100) GPUs. Engineers building LLM training and inference stacks use it for fast low-precision math, and its small codebase also works as study material for GPU kernel optimization.