/**
 * attention_block.ts – Causal Multi-Head Self-Attention Block.
 *
 * Intentionally simple for WebGPU — naive O(L²) tiled attention,
 * no Flash-Attention dependency. Suitable for hybrid (Jamba/Zamba) schedules
 * where a few attention layers interleave with SSM layers.
 *
 * Data flow:
 *   Input (B, L, D_model)
 *     └─ RMSNorm
 *     └─ wQKV → Q (B,L,H,dh), K (B,L,H,dh), V (B,L,H,dh)
 *     └─ causal attention scores = Q·Kᵀ / √dh  (masked)
 *     └─ softmax
 *     └─ weighted V sum
 *     └─ concat heads → wO → D_model
 *     └─ + residual
 *   [optional FFN sublayer]
 *
 * Implements SequenceLayer.
 */
import type { SequenceLayer, LayerForwardResult, LayerParam } from './sequence_layer.js';
export interface AttentionBlockConfig {
    dModel: number;
    nHeads: number;
    dHead?: number;
    hasFfn?: boolean;
    ffnMult?: number;
}
export interface AttentionCache {
    scores: GPUBuffer;
}
export declare class AttentionBlock implements SequenceLayer {
    readonly layerType: "attention";
    device: GPUDevice;
    config: Required<AttentionBlockConfig>;
    dHead: number;
    gpuWeights: Record<string, GPUBuffer>;
    pipelines: Record<string, GPUComputePipeline>;
    constructor(device: GPUDevice, config: AttentionBlockConfig);
    private _initWeights;
    private _buildPipelines;
    forward(xBuf: GPUBuffer, batch: number, seqLen: number): LayerForwardResult;
    parameters(): LayerParam[];
    getTrainableParams(): LayerParam[];
    setWSLAMode(_enabled: boolean): void;
    destroy(): void;
}
//# sourceMappingURL=attention_block.d.ts.map