fla-dispatch-backends

Create a BaseBackend subclass under the operation's backends package, set backend_type, package_name, env_var, default_enable, and priority, implement the verifier and backend function, then register it in the operation's backends __init__.py and add dispatch tests.

Dispatch tries registered backends sorted by priority where lower means higher priority. A backend is used only when is_available() and is_enabled() are both true and its verifier returns (True, None); otherwise the next backend or the original implementation runs.

Yes, set FLA_DISABLE_BACKEND_DISPATCH=1 to bypass the dispatch decorator entirely. The decorated function remains the semantic fallback, so behavior stays identical to the default implementation.

Verifiers must be cheap, deterministic, and side-effect free, returning (True, None) or (False, reason). They must not mutate registries, environment variables, tensors, RNG state, or caches, and should check dtype, shape, mode, and package requirements.

The dispatch wrapper is marked with torch.compiler.disable, so backend selection must happen outside compiled graphs. Compiled work belongs inside the selected backend implementation to keep dispatch torch.compile friendly.