Enzyme

Why we included this project

Enzyme produces gradients for code you did not write, working at the compiler level instead of asking you to port kernels into a framework. It plugs into LLVM and MLIR and differentiates optimized native code directly, which is how it reaches performance on par with hand-tuned AD tools. That matters for teams doing differentiable programming in C, C++, Rust, or Julia who need derivatives of legacy solvers, physics engines, or custom numerical kernels that PyTorch and TensorFlow never see. It is not an end-user ML framework: expect a build-from-source setup against your LLVM toolchain, with Homebrew, Spack, and Nix packages available, rather than a pip install and a Python API. If your work sits in HPC or performance-sensitive scientific computing, this is one of the few practical routes to efficient gradients without abandoning native code.