Device-graph-based execution is the norm for modern serving stacks, and Modular’s inference stack (MAX) is no exception. Like every widely-used stack, MAX currently builds its device graph by capturing a runtime command stream and replaying it: the record & replay functionality offered by CUDA and HIP. This carries significant downsides, and it leaves out accelerators that expose device-graph or command-buffer construction APIs but no record/replay equivalent. This talk presents Modular’s compiler-first alternative, built on its MLIR & LLVM based graph compiler: rather than capturing a graph at runtime, the compiler emits direct calls to the target’s graph-construction APIs. We will cover how kernels declare their device-graph capabilities, the front-end IR and middle-end dialects used to construct device graphs, and the Mojo-level APIs for interacting with them.