Backend Extension API

Backend loading

tvm.backend discovers and loads target-owned Python semantics. Depending on the target, loading a backend registers its TVMScript namespaces, tile-dispatch implementations, target tags, compilation-pipeline entry points, and code-generation support.

Backend-owned Python modules and load hooks.

tvm.backend.is_loaded(name: str) → bool

Return whether a backend has been loaded.

tvm.backend.load(name: str) → None

Load a backend’s Python registration hooks.

Loading is idempotent. A backend package must live at tvm.backend.<name> and expose register_backend().

CUDA registration

tvm.backend.cuda.register_backend()

Register CUDA-owned Python semantics.

tvm.backend.cuda.script_namespace(**kwargs)

Return the CUDA TVMScript namespace object.

tvm.backend.cuda.script_namespaces(**_)

Return CUDA-owned TVMScript namespaces.

Trainium registration

tvm.backend.trn.register_backend()

Register Trainium-owned Python semantics.

tvm.backend.trn.script_namespace(op_wrapper=None)

Return the Trainium TVMScript namespace object.

tvm.backend.trn.script_namespaces(op_wrapper=None, **_)

Return Trainium-owned TVMScript namespaces.

These functions are backend integration points. Kernel-facing CUDA APIs are documented in CUDA Authoring and Support APIs and Direct PTX Instructions; Trainium APIs are documented in Trainium Authoring and Support APIs. See Backend Internals for module ownership and registration side effects.