/**
 * WebGPU Tensor Operations
 * GPU-accelerated tensor operations using compute shaders
 */
import { Tensor } from '../tensor.js';
export declare const relu: (x: Tensor) => Promise<Tensor>;
export declare const silu: (x: Tensor) => Promise<Tensor>;
export declare const gelu: (x: Tensor) => Promise<Tensor>;
export declare const neg: (x: Tensor) => Promise<Tensor>;
export declare const exp: (x: Tensor) => Promise<Tensor>;
export declare const sqrt: (x: Tensor) => Promise<Tensor>;
export declare const rsqrt: (x: Tensor) => Promise<Tensor>;
export declare const tanh: (x: Tensor) => Promise<Tensor>;
export declare const add: (a: Tensor, b: Tensor) => Promise<Tensor>;
export declare const sub: (a: Tensor, b: Tensor) => Promise<Tensor>;
export declare const mul: (a: Tensor, b: Tensor) => Promise<Tensor>;
export declare const div: (a: Tensor, b: Tensor) => Promise<Tensor>;
export declare const addScalar: (x: Tensor, s: number) => Promise<Tensor>;
export declare const mulScalar: (x: Tensor, s: number) => Promise<Tensor>;
/**
 * Broadcast add: input[seqLen, dim] + bias[dim] -> output[seqLen, dim]
 * Adds bias vector to each row of the input matrix
 */
export declare function broadcastAdd(input: Tensor, bias: Tensor): Promise<Tensor>;
/**
 * Matrix multiplication: C = A @ B
 * A: [M, K], B: [K, N] -> C: [M, N]
 */
export declare function matmul(a: Tensor, b: Tensor): Promise<Tensor>;
export declare const sum: (x: Tensor) => Promise<number>;
export declare const max: (x: Tensor) => Promise<number>;
/**
 * Compute mean of tensor
 */
export declare function mean(x: Tensor): Promise<number>;
/**
 * Softmax along the last dimension
 */
export declare function softmax(x: Tensor): Promise<Tensor>;
/**
 * Reset all cached pipelines (useful after shader updates)
 */
export declare function resetOpsPipelines(): void;
/**
 * Embedding lookup: maps token indices to embedding vectors
 * embeddings: [vocabSize, hiddenSize] or [hiddenSize, vocabSize] (GGUF format)
 * indices: [seqLen] (token IDs)
 * vocabSize: optional, helps detect transposed format
 * returns: [seqLen, hiddenSize]
 */
export declare function embeddingLookup(embeddings: Tensor, indices: number[], vocabSizeHint?: number): Promise<Tensor>;
/**
 * GPU-accelerated causal attention with GQA support
 *
 * @param Q - Query tensor [numQPos, numHeads * headDim]
 * @param K - Key tensor [numKeys, numKVHeads * headDim]
 * @param V - Value tensor [numKeys, numKVHeads * headDim]
 * @param numHeads - Number of query heads
 * @param numKVHeads - Number of KV heads (for GQA)
 * @param headDim - Dimension per head
 * @param startPos - Starting position for queries (for incremental inference)
 * @returns Output tensor [numQPos, numHeads * headDim]
 */
export declare function causalAttention(Q: Tensor, K: Tensor, V: Tensor, numHeads: number, numKVHeads: number, headDim: number, startPos?: number): Promise<Tensor>;
/**
 * Extract the last row of a 2D tensor on GPU
 * input: [rows, cols] -> output: [1, cols]
 * This avoids transferring the entire tensor to CPU just to get the last row
 */
export declare function sliceLastRow(input: Tensor): Promise<Tensor>;
/**
 * Transpose a 2D tensor
 * input: [rows, cols] -> output: [cols, rows]
 */
export declare function transpose(input: Tensor): Promise<Tensor>;
//# sourceMappingURL=index.d.ts.map