/**
 * @octomil/browser — Inference engine
 *
 * Wraps ONNX Runtime Web to create an inference session from a model
 * buffer, auto-detect the best execution provider (WebGPU > WASM),
 * and run forward passes.
 */
import type { Backend, NamedTensors, PredictOutput } from "../../../types.js";
import type { ModelRuntime } from "../../core/model-runtime.js";
export type { ModelRuntime } from "../../core/model-runtime.js";
export declare class InferenceEngine implements ModelRuntime {
    private session;
    private ortModule;
    private resolvedBackend;
    /**
     * Create an ONNX Runtime session from the given model bytes.
     *
     * @param modelData  Raw ONNX model ArrayBuffer.
     * @param backend    Requested backend (`"webgpu"`, `"wasm"`, or `undefined` for auto).
     */
    createSession(modelData: ArrayBuffer, backend?: Backend): Promise<void>;
    /**
     * Run inference and return the output tensors plus timing info.
     */
    run(inputs: NamedTensors): Promise<PredictOutput>;
    /** Names of the model's input tensors. */
    get inputNames(): readonly string[];
    /** Names of the model's output tensors. */
    get outputNames(): readonly string[];
    /** The backend that was actually used after negotiation. */
    get activeBackend(): Backend | null;
    /** Release WASM / WebGPU resources. */
    dispose(): void;
    /** Check if ONNX Runtime Web can be loaded in this environment. */
    isAvailable(): Promise<boolean>;
    private loadOrt;
    private resolveProvider;
    private detectWebGPU;
    private ensureSession;
    private convertOutputs;
    /**
     * Best-effort extraction of `label` / `score` / `scores` from the
     * first output tensor — only if it looks like a classification head.
     */
    private extractConvenience;
}
//# sourceMappingURL=engine.d.ts.map