/**
 * Single decode attention WGSL kernel
 *
 * Flash Attention for decode phase (single token query)
 * Uses online softmax to avoid materializing full attention matrix
 */
export interface DecodeAttentionConfig {
    num_qo_heads: number;
    num_kv_heads: number;
    head_dim: number;
    seq_len: number;
    sm_scale: number;
    dtype: 'f32' | 'f16';
    useSubgroups: boolean;
}
export declare function generateDecodeAttentionShader(config: DecodeAttentionConfig): string;