import { B as BaseInference } from '../../factory-CTQWNTae.js';
export { I as InferenceOptions, T as Tokenizer, W as WASMInference, a as WebGPUInference, c as createInferenceEngine, d as detectBestBackend, g as getInferenceCapabilities, l as loadTokenizer } from '../../factory-CTQWNTae.js';
import { f as InferenceBackend, B as BaseStorage } from '../../base-storage-CfkDwX0r.js';

/**
 * SIMD-Optimized Inference Engine
 *
 * Uses WebAssembly SIMD for accelerated transformer inference.
 * 2.5x faster than pure JS implementation.
 */

declare class SIMDInference extends BaseInference {
    readonly backend: InferenceBackend;
    private simd;
    private heapF32;
    private heapOffset;
    private embeddings;
    private lmHead;
    private outputNorm;
    private currentLayer;
    private ropeCache;
    private kvCache;
    constructor(storage: BaseStorage, modelId: string);
    /**
     * Initialize SIMD backend
     */
    protected initializeBackend(): Promise<void>;
    /**
     * Load embedding and output projection weights
     */
    private loadEmbeddings;
    /**
     * Combine tensor chunks
     */
    private combineChunks;
    /**
     * Load a layer's weights
     */
    private loadLayer;
    private ensureMemory;
    private alloc;
    private resetHeap;
    private simdMatmul;
    private simdRmsNorm;
    private simdSilu;
    /**
     * Run forward pass with SIMD acceleration
     */
    protected forward(inputIds: number[], _position: number): Promise<Float32Array>;
    /**
     * Self-attention with RoPE and KV caching
     */
    private selfAttention;
    /**
     * Apply RoPE
     */
    private applyRoPE;
    private getRoPETrigonometry;
    /**
     * FFN with SwiGLU and SIMD
     */
    private ffn;
    /**
     * Reset KV cache
     */
    resetCache(): void;
}

/**
 * WebNN Inference Engine
 *
 * Uses WebNN capability detection and delegates tensor math to the SIMD
 * implementation until dedicated WebNN kernels are enabled.
 */

declare class WebNNInference extends SIMDInference {
    readonly backend: InferenceBackend;
    constructor(storage: BaseStorage, modelId: string);
    protected initializeBackend(): Promise<void>;
}

export { SIMDInference, WebNNInference };