/**
 * Single prefill attention WGSL kernel
 *
 * Flash Attention for prefill phase (multiple query tokens)
 * Uses online softmax and supports causal masking
 */
export interface PrefillAttentionConfig {
    num_qo_heads: number;
    num_kv_heads: number;
    head_dim: number;
    qo_len: number;
    kv_len: number;
    sm_scale: number;
    causal: boolean;
    dtype: 'f32' | 'f16';
}
export declare function generatePrefillAttentionShader(config: PrefillAttentionConfig): string;