Reactant.jl – Julia execution code on GPUs (github.com)

Reactant.jl is a new Julia project that traces Julia functions, lowers them to MLIR, applies optimizations (including EnzymeMLIR for automatic differentiation), and emits XLA-backed executables for CPU/GPU/TPU. It operates as a tracing system: you convert standard Julia arrays or nested data structures into ConcreteRArray buffers (Reactant.ConcreteRArray or via Reactant.to_rarray), compile with @compile, and run optimized kernels on accelerators by selecting a backend (e.g., Reactant.set_default_backend("gpu")). During compilation values inside ConcreteRArrays are represented as TracedRArrays, and compiled code is targeted through the MLIR→XLA stack (no CUDA.jl required to hit GPUs). This is significant because it makes a wide class of existing Julia numerical code available for aggressive ML-style optimizations and autodiff with minimal developer change. Important technical caveats: Reactant currently traces control flow and removes type instabilities, so compiled functions assume the same control-flow path, types and sizes as observed at trace time; updates to ordinary non-ConcreteRArray state (counters, conditional branches driven by non-RArray data) are not captured. The project is under active development and may change its API/semantics (possibly moving toward a partial source rewriter), so early adopters and package integrators should expect instability while exploring integration and capabilities.