Function compile_graph 

pub fn compile_graph<'a, L: Lowering<'a>>(
    graph: &Graph,
    lowering: &L,
    mode: LoweringMode,
    options: &Options,
    cache_dir: &str,
    force: bool,
) -> Result<CudaModel<'a>>

Ahead-of-time compile an already-traced graph to PTX files on disk, driven by an already-parsed Options (see Options::parse).

This is the same LlvmCompiler + CudaGraphCompiler sequence used for JIT compilation at runtime (e.g. in models/teeny-vision/examples/mnist.rs), just parameterized by CLI-driven options/cache_dir instead of hardcoded values. No live CUDA device/context is required — only .load() on the returned CudaModel needs one.

cache_dir is where compiled PTX is written/read. Passing the same directory a later run resolves via TEENYC_CACHE_DIR pre-warms that runtime JIT cache. The teenyc binary is resolved via teeny_compiler::compiler::find_teenyc, matching the existing JIT compile path.