/**
 * mamba2_block.ts – Mamba-2 Mixer Block (Structured State Space Duality).
 *
 * Key differences from Mamba-1:
 *   - Multi-head SSM with scalar A per head
 *   - Single fused in_proj (no separate dt_proj expansion)
 *   - SSD (chunked) scan replaces S6 selective scan
 *   - Inner RMSNorm on scan output instead of SiLU gate
 *   - No separate z gate
 *
 * Implements SequenceLayer.
 */
import type { SequenceLayer, LayerForwardResult, LayerParam } from './sequence_layer.js';
export interface Mamba2BlockConfig {
    dModel: number;
    dState: number;
    dConv: number;
    expand: number;
    nHeads: number;
    nGroups: number;
    chunkLen: number;
}
export interface Mamba2Cache {
    stateCarry: GPUBuffer;
}
export declare class Mamba2Block implements SequenceLayer {
    readonly layerType: "mamba2";
    device: GPUDevice;
    config: Required<Mamba2BlockConfig>;
    dInner: number;
    dHead: number;
    gpuWeights: Record<string, GPUBuffer>;
    pipelines: Record<string, GPUComputePipeline>;
    private _wslaMode;
    constructor(device: GPUDevice, config: Mamba2BlockConfig);
    private _initWeights;
    private _buildPipelines;
    forward(xBuf: GPUBuffer, batch: number, seqLen: number): LayerForwardResult;
    parameters(): LayerParam[];
    getTrainableParams(): LayerParam[];
    setWSLAMode(enabled: boolean): void;
    destroy(): void;
}
//# sourceMappingURL=mamba2_block.d.ts.map