NVIDIA/cutlass ? reverse-engineered prompt

Build me a CUDA C++ library for high performance linear algebra, centered on fast matrix multiplication and related kernels.

I want it to let me write and tune GEMM style operations for different data types, including mixed precision, and target modern NVIDIA GPUs efficiently. It should include reusable templates and abstractions for building kernels, plus a Python based DSL for writing kernels more easily without losing performance. Please include a few practical examples, clear build instructions, and tests so I can see it working end to end.

If you need to, look up the current CUDA and NVIDIA docs online while you implement it, especially for any low level GPU details or compiler setup. Make the project feel like something serious users could extend for custom kernels, not just a demo.