/**
 * WebGPU Transformer Layers
 * GPU-accelerated implementations of transformer building blocks
 */
import { Tensor } from '../tensor.js';
/**
 * Layer Normalization
 */
export declare function layerNorm(input: Tensor, weight: Tensor, bias: Tensor, eps?: number): Promise<Tensor>;
/**
 * RMS Normalization (used in Llama)
 */
export declare function rmsNorm(input: Tensor, weight: Tensor, eps?: number): Promise<Tensor>;
/**
 * Apply Rotary Position Embeddings
 * @param input - Input tensor [seqLen, numHeads * headDim]
 * @param seqLen - Number of tokens being processed
 * @param numHeads - Number of attention heads
 * @param headDim - Dimension per head
 * @param base - RoPE base frequency (default 10000)
 * @param startPos - Starting position for incremental inference (default 0)
 */
export declare function applyRope(input: Tensor, seqLen: number, numHeads: number, headDim: number, base?: number, startPos?: number): Promise<Tensor>;
export interface AttentionConfig {
    numHeads: number;
    headDim: number;
    scale?: number;
    causal?: boolean;
}
/**
 * Multi-Head Attention
 *
 * Q, K, V: [batchSize, seqLen, numHeads * headDim]
 * Output: [batchSize, seqLen, numHeads * headDim]
 */
export declare function attention(q: Tensor, k: Tensor, v: Tensor, config: AttentionConfig): Promise<Tensor>;
/**
 * Feed-Forward Network with SiLU activation (Llama-style)
 *
 * gate = x @ W_gate
 * up = x @ W_up
 * output = (silu(gate) * up) @ W_down
 */
export declare function feedForward(x: Tensor, wGate: Tensor, wUp: Tensor, wDown: Tensor): Promise<Tensor>;
/**
 * Simple MLP with GELU activation
 */
export declare function mlp(x: Tensor, w1: Tensor, w2: Tensor): Promise<Tensor>;
/**
 * Reset all cached pipelines (useful after shader updates)
 */
export declare function resetLayersPipelines(): void;
//# sourceMappingURL=index.d.ts.map