apache/tvm ? reverse-engineered prompt

Build me an open source machine learning compiler framework that helps people take a model and turn it into something that runs efficiently on different hardware.

I want it to be Python first so it’s easy to customize, and I want the system to support a clean workflow for importing models, optimizing them, and producing small deployable packages for things like GPUs and other device targets. Please include a simple way to experiment with transformations in Python, and make sure the project has good docs, examples, and getting started instructions so someone new can try it without a lot of setup pain.

If there are current best practices or docs that would help, look them up online as you build. Focus on making the compiler framework feel practical for real world machine learning work, not just a research demo.