Experimental APIs and Backends

FlashInfer distinguishes stable functionality from experimental functionality intended for fast-moving work — client-GPU (e.g. SM12x) kernels, new operations from the latest models, or highly specialized kernels for specific problem sizes:

  • an experimental API is a public interface that may change or disappear;

  • an experimental backend is an implementation that is not yet ready for stable support.

Experimental APIs are marked with a warning banner in their documentation. A stable API exposes an experimental backend by name; it routes to one automatically only when the user opts in (below).

Opting in to experimental features

Using experimental functionality is always an explicit opt-in. Two forms need no environment variable:

  • calling an API marked experimental (@flashinfer_experimental_api);

  • passing an experimental backend by name to a stable API, e.g. backend="sm12x_cute".

Both emit an ExperimentalWarning once (per API, or per API/backend pair).

Automatic selection is gated. With backend="auto" — including the dispatch heuristics and autotuning behind it — a stable API considers only stable backends unless you set:

export FLASHINFER_ALLOW_EXPERIMENTAL_AUTO_BACKENDS=1

With the variable set, experimental backends join the automatic candidate list. Without it, a call whose only viable backends are experimental fails with an error that names the variable and suggests an explicit backend=.

User contract

Experimental APIs and backends provide no compatibility or long-term support guarantees and may change or be removed without deprecation. They are intended primarily for main-branch users, community containers, and local framework integrations, and should not be enabled by default in supported framework releases.

Each experimental feature’s documentation identifies whether the API, the backend, or both are experimental; the supported use cases and limitations; and whether the feature participates in automatic routing once enabled.

Experimental features are JIT-only: they are not included in the flashinfer-jit-cache / flashinfer-cubin pre-built packages.

For contributors

The full policy — placement rules, admission criteria, ownership and lifecycle requirements, and the graduation checklist — lives in flashinfer/experimental/README.md in the repository, with a summary in CONTRIBUTING.md.