deepseek-ai/DeepGEMM ? reverse-engineered prompt

Build me a Python package for fast NVIDIA GPU matrix math, with a simple install and import experience.

I want it to let me run very fast GEMM style operations for FP8, FP4, and BF16, plus the extra fused pieces for MoE and the MQA scoring kernels the README talks about. It should work as a clean CUDA codebase with runtime JIT compilation, so I can install it without doing a full CUDA build first, and then import it from Python with useful helper functions and tests.

Please make the main focus performance and correctness on SM90 and SM100 GPUs, and include the build and install scripts so the project is easy to set up from a fresh clone. Add clear examples and tests that show the basic matrix multiply path and the grouped MoE cases working end to end. If you need anything current from NVIDIA or PyTorch docs, look it up online.