torch.fx.passes.regional_inductor.regional_inductor

torch.fx.passes.regional_inductor.regional_inductor(gm, *example_args)

Scoops out inductor marked regions and compiles them with inductor.

Inductor options should be provided via the annotation API:

with fx_traceback.annotate(
    {
        "compile_with_inductor": {
            "inductor_configs": {
                "max_autotune": True,
                "triton.cudagraphs": False,
            }
        }
    }
):
    ...

Warning

This API is experimental and is NOT backward-compatible.

Return type:

GraphModule