@seanhogg/builderforce-memory-engine 2026.6.18

package/dist/model/mamba1_block.js

@@ -0,0 +1,333 @@
+/**
+ * mamba1_block.ts – Mamba-1 Mixer Block (S6 selective scan).
+ *
+ * Renamed from mamba_block.ts; MambaBlock is kept as a deprecated alias.
+ * Implements SequenceLayer so HybridMambaModel can iterate blocks generically.
+ */
+import { createComputePipeline, createBindGroup, createStorageBuffer, createEmptyStorageBuffer, createUniformBuffer, dispatchKernel, cdiv, } from '../utils/gpu_utils.js';
+import { SELECTIVE_SCAN_FORWARD_WGSL } from '../kernels/selective_scan.js';
+import { gaussianArray } from '../utils/rng.js';
+import { CONV1D_FORWARD_WGSL } from '../kernels/conv1d.js';
+import { LINEAR_FORWARD_WGSL } from '../kernels/linear_projection.js';
+import { ACTIVATIONS_WGSL } from '../kernels/activations.js';
+// ── Element-wise helper shaders (compiled once per pipeline) ─────────────────
+const MUL_SHADER = /* wgsl */ `
+@group(0) @binding(0) var<storage, read>       a : array<f32>;
+@group(0) @binding(1) var<storage, read>       b : array<f32>;
+@group(0) @binding(2) var<storage, read_write> c : array<f32>;
+@group(0) @binding(3) var<uniform>             n : u32;
+@compute @workgroup_size(256)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let i = gid.x;
+    if (i < n) { c[i] = a[i] * b[i]; }
+}
+`;
+const ADD_SHADER = /* wgsl */ `
+@group(0) @binding(0) var<storage, read>       a : array<f32>;
+@group(0) @binding(1) var<storage, read>       b : array<f32>;
+@group(0) @binding(2) var<storage, read_write> c : array<f32>;
+@group(0) @binding(3) var<uniform>             n : u32;
+@compute @workgroup_size(256)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let i = gid.x;
+    if (i < n) { c[i] = a[i] + b[i]; }
+}
+`;
+// ── Mamba1Block ───────────────────────────────────────────────────────────────
+export class Mamba1Block {
+    layerType = 'mamba1';
+    device;
+    config;
+    dInner;
+    dtRank;
+    wInProj;
+    bInProj;
+    wConv;
+    bConv;
+    wXProj;
+    bXProj;
+    wDtProj;
+    bDtProj;
+    A_log;
+    D_vec;
+    wOutProj;
+    bOutProj;
+    normWeight;
+    gpuWeights;
+    pipelines;
+    _wslaMode = false;
+    constructor(device, config) {
+        this.device = device;
+        this.config = {
+            dState: 16,
+            dConv: 4,
+            expand: 2,
+            biasConv: true,
+            dtRank: Math.ceil(config.dModel / 16),
+            ...config,
+        };
+        const { dModel, expand } = this.config;
+        this.dInner = expand * dModel;
+        this.dtRank = config.dtRank ?? Math.ceil(dModel / 16);
+        this.wInProj = new Float32Array(0);
+        this.bInProj = new Float32Array(0);
+        this.wConv = new Float32Array(0);
+        this.bConv = new Float32Array(0);
+        this.wXProj = new Float32Array(0);
+        this.bXProj = new Float32Array(0);
+        this.wDtProj = new Float32Array(0);
+        this.bDtProj = new Float32Array(0);
+        this.A_log = new Float32Array(0);
+        this.D_vec = new Float32Array(0);
+        this.wOutProj = new Float32Array(0);
+        this.bOutProj = new Float32Array(0);
+        this.normWeight = new Float32Array(0);
+        this.gpuWeights = {};
+        this.pipelines = {};
+        this._initWeights();
+        this._buildPipelines();
+    }
+    _initWeights() {
+        const { dModel, dState, dConv } = this.config;
+        const D = this.dInner;
+        const N = dState;
+        const K = dConv;
+        const R = this.dtRank;
+        const randn = (n, std = 0.02) => gaussianArray(n, std);
+        const zeros = (n) => new Float32Array(n);
+        const ones = (n) => new Float32Array(n).fill(1.0);
+        this.wInProj = randn(2 * D * dModel);
+        this.bInProj = zeros(2 * D);
+        this.wConv = randn(D * K, 0.01);
+        this.bConv = zeros(D);
+        this.wXProj = randn((R + 2 * N) * D, 0.01);
+        this.bXProj = zeros(R + 2 * N);
+        this.wDtProj = randn(D * R, 0.02);
+        this.bDtProj = zeros(D);
+        this.A_log = new Float32Array(D * N);
+        for (let d = 0; d < D; d++) {
+            for (let n = 0; n < N; n++) {
+                this.A_log[d * N + n] = Math.log(n + 1);
+            }
+        }
+        this.D_vec = ones(D);
+        this.wOutProj = randn(dModel * D, 0.02);
+        this.bOutProj = zeros(dModel);
+        this.normWeight = ones(dModel);
+        this._uploadWeightsToGPU();
+    }
+    _uploadWeightsToGPU() {
+        const d = this.device;
+        const mk = (arr) => createStorageBuffer(d, arr, true);
+        this.gpuWeights = {
+            wInProj: mk(this.wInProj),
+            bInProj: mk(this.bInProj),
+            wConv: mk(this.wConv),
+            bConv: mk(this.bConv),
+            wXProj: mk(this.wXProj),
+            bXProj: mk(this.bXProj),
+            wDtProj: mk(this.wDtProj),
+            bDtProj: mk(this.bDtProj),
+            A_log: mk(this.A_log),
+            D_vec: mk(this.D_vec),
+            wOutProj: mk(this.wOutProj),
+            bOutProj: mk(this.bOutProj),
+            normWeight: mk(this.normWeight),
+        };
+    }
+    _buildPipelines() {
+        const d = this.device;
+        this.pipelines = {
+            linear: createComputePipeline(d, LINEAR_FORWARD_WGSL, 'linear_forward'),
+            conv1d: createComputePipeline(d, CONV1D_FORWARD_WGSL, 'conv1d_forward'),
+            silu: createComputePipeline(d, ACTIVATIONS_WGSL, 'silu_forward'),
+            rmsnorm: createComputePipeline(d, ACTIVATIONS_WGSL, 'rmsnorm_forward'),
+            scan_fwd: createComputePipeline(d, SELECTIVE_SCAN_FORWARD_WGSL, 'forward_scan'),
+            scan_reduce: createComputePipeline(d, SELECTIVE_SCAN_FORWARD_WGSL, 'forward_reduce'),
+            elMul: createComputePipeline(d, MUL_SHADER, 'main'),
+            elAdd: createComputePipeline(d, ADD_SHADER, 'main'),
+        };
+    }
+    forward(xBuf, batch, seqLen) {
+        const d = this.device;
+        const { dModel, dState, dConv } = this.config;
+        const D = this.dInner;
+        const N = dState;
+        const B = batch;
+        const L = seqLen;
+        const M = B * L;
+        const R = this.dtRank;
+        const cache = {};
+        // 1. Pre-block RMSNorm
+        const normOut = createEmptyStorageBuffer(d, M * dModel * 4, true);
+        const normInv = createEmptyStorageBuffer(d, M * 4, true);
+        cache.normInv = normInv;
+        cache.normIn = xBuf;
+        {
+            const params = new ArrayBuffer(16);
+            new Uint32Array(params, 0, 2).set([M, dModel]);
+            new Float32Array(params, 8, 1).set([1e-6]);
+            const pBuf = createUniformBuffer(d, params);
+            const bg = createBindGroup(d, this.pipelines['rmsnorm'], [pBuf, xBuf, this.gpuWeights['normWeight'], normOut, normInv]);
+            dispatchKernel(d, this.pipelines['rmsnorm'], bg, [cdiv(M, 64), 1, 1]);
+        }
+        // 2. Input projection → x and z
+        const inProjOut = createEmptyStorageBuffer(d, M * 2 * D * 4, true);
+        cache.normOut = normOut;
+        {
+            const params = new Uint32Array([M, dModel, 2 * D]).buffer;
+            const pBuf = createUniformBuffer(d, params);
+            const bg = createBindGroup(d, this.pipelines['linear'], [pBuf, normOut, this.gpuWeights['wInProj'], this.gpuWeights['bInProj'], inProjOut]);
+            dispatchKernel(d, this.pipelines['linear'], bg, [cdiv(M, 16), cdiv(2 * D, 16), 1]);
+        }
+        // 3. Split into x and z
+        const xConvIn = createEmptyStorageBuffer(d, M * D * 4, true);
+        const zBuf = createEmptyStorageBuffer(d, M * D * 4, true);
+        {
+            const enc = d.createCommandEncoder();
+            enc.copyBufferToBuffer(inProjOut, 0, xConvIn, 0, M * D * 4);
+            enc.copyBufferToBuffer(inProjOut, M * D * 4, zBuf, 0, M * D * 4);
+            d.queue.submit([enc.finish()]);
+        }
+        inProjOut.destroy();
+        cache.zBuf = zBuf;
+        cache.xConvIn = xConvIn;
+        // 4. Causal conv1d on x
+        const convOut = createEmptyStorageBuffer(d, M * D * 4, true);
+        cache.convOut = convOut;
+        {
+            const params = new Uint32Array([L, D, dConv, B]).buffer;
+            const pBuf = createUniformBuffer(d, params);
+            const bg = createBindGroup(d, this.pipelines['conv1d'], [pBuf, xConvIn, this.gpuWeights['wConv'], this.gpuWeights['bConv'], convOut]);
+            dispatchKernel(d, this.pipelines['conv1d'], bg, [cdiv(L, 16), cdiv(D, 16), B]);
+        }
+        // 5. SiLU activation
+        const siluOut = createEmptyStorageBuffer(d, M * D * 4, true);
+        cache.siluOut = siluOut;
+        {
+            const params = new Uint32Array([M * D]).buffer;
+            const pBuf = createUniformBuffer(d, params);
+            const bg = createBindGroup(d, this.pipelines['silu'], [pBuf, convOut, siluOut]);
+            dispatchKernel(d, this.pipelines['silu'], bg, [cdiv(M * D, 256), 1, 1]);
+        }
+        // 6. x_proj → Δ (dtRaw), B, C
+        const xProjOut = createEmptyStorageBuffer(d, M * (R + 2 * N) * 4, true);
+        {
+            const params = new Uint32Array([M, D, R + 2 * N]).buffer;
+            const pBuf = createUniformBuffer(d, params);
+            const bg = createBindGroup(d, this.pipelines['linear'], [pBuf, siluOut, this.gpuWeights['wXProj'], this.gpuWeights['bXProj'], xProjOut]);
+            dispatchKernel(d, this.pipelines['linear'], bg, [cdiv(M, 16), cdiv(R + 2 * N, 16), 1]);
+        }
+        const dtRaw = createEmptyStorageBuffer(d, M * R * 4, true);
+        const B_raw = createEmptyStorageBuffer(d, B * L * N * 4, true);
+        const C_raw = createEmptyStorageBuffer(d, B * L * N * 4, true);
+        {
+            const enc = d.createCommandEncoder();
+            enc.copyBufferToBuffer(xProjOut, 0, dtRaw, 0, M * R * 4);
+            enc.copyBufferToBuffer(xProjOut, M * R * 4, B_raw, 0, B * L * N * 4);
+            enc.copyBufferToBuffer(xProjOut, M * (R + N) * 4, C_raw, 0, B * L * N * 4);
+            d.queue.submit([enc.finish()]);
+        }
+        xProjOut.destroy();
+        cache.B_raw = B_raw;
+        cache.C_raw = C_raw;
+        // 7. dt_proj: expand Δ to full dim
+        const deltaFull = createEmptyStorageBuffer(d, M * D * 4, true);
+        cache.deltaFull = deltaFull;
+        {
+            const params = new Uint32Array([M, R, D]).buffer;
+            const pBuf = createUniformBuffer(d, params);
+            const bg = createBindGroup(d, this.pipelines['linear'], [pBuf, dtRaw, this.gpuWeights['wDtProj'], this.gpuWeights['bDtProj'], deltaFull]);
+            dispatchKernel(d, this.pipelines['linear'], bg, [cdiv(M, 16), cdiv(D, 16), 1]);
+        }
+        dtRaw.destroy();
+        // 8. Selective scan (S6)
+        const scanY = createEmptyStorageBuffer(d, B * L * D * 4, true);
+        const hCache = createEmptyStorageBuffer(d, 2 * B * L * D * N * 4, true);
+        cache.hCache = hCache;
+        {
+            const params = new Uint32Array([L, N, D, B]).buffer;
+            const pBuf = createUniformBuffer(d, params);
+            const bg1 = createBindGroup(d, this.pipelines['scan_fwd'], [pBuf, siluOut, deltaFull, this.gpuWeights['A_log'], B_raw, C_raw,
+                this.gpuWeights['D_vec'], scanY, hCache]);
+            dispatchKernel(d, this.pipelines['scan_fwd'], bg1, [cdiv(D, 8), cdiv(N, 8), B]);
+            const bg2 = createBindGroup(d, this.pipelines['scan_reduce'], [pBuf, siluOut, deltaFull, this.gpuWeights['A_log'], B_raw, C_raw,
+                this.gpuWeights['D_vec'], scanY, hCache]);
+            dispatchKernel(d, this.pipelines['scan_reduce'], bg2, [cdiv(L, 64), D, B]);
+        }
+        // 9. Gate: y ⊗ SiLU(z)
+        const siluZ = createEmptyStorageBuffer(d, M * D * 4, true);
+        const gatedOut = createEmptyStorageBuffer(d, M * D * 4, true);
+        {
+            const nBuf = createUniformBuffer(d, new Uint32Array([M * D]).buffer);
+            const bgZ = createBindGroup(d, this.pipelines['silu'], [nBuf, zBuf, siluZ]);
+            dispatchKernel(d, this.pipelines['silu'], bgZ, [cdiv(M * D, 256), 1, 1]);
+            const nBuf2 = createUniformBuffer(d, new Uint32Array([M * D]).buffer);
+            const bgMul = createBindGroup(d, this.pipelines['elMul'], [scanY, siluZ, gatedOut, nBuf2]);
+            dispatchKernel(d, this.pipelines['elMul'], bgMul, [cdiv(M * D, 256), 1, 1]);
+        }
+        siluZ.destroy();
+        scanY.destroy();
+        // 10. Output projection
+        const outProjOut = createEmptyStorageBuffer(d, M * dModel * 4, true);
+        {
+            const params = new Uint32Array([M, D, dModel]).buffer;
+            const pBuf = createUniformBuffer(d, params);
+            const bg = createBindGroup(d, this.pipelines['linear'], [pBuf, gatedOut, this.gpuWeights['wOutProj'], this.gpuWeights['bOutProj'], outProjOut]);
+            dispatchKernel(d, this.pipelines['linear'], bg, [cdiv(M, 16), cdiv(dModel, 16), 1]);
+        }
+        gatedOut.destroy();
+        // 11. Residual add
+        const output = createEmptyStorageBuffer(d, M * dModel * 4, true);
+        {
+            const nBuf = createUniformBuffer(d, new Uint32Array([M * dModel]).buffer);
+            const bg = createBindGroup(d, this.pipelines['elAdd'], [outProjOut, xBuf, output, nBuf]);
+            dispatchKernel(d, this.pipelines['elAdd'], bg, [cdiv(M * dModel, 256), 1, 1]);
+        }
+        outProjOut.destroy();
+        return { output, cache };
+    }
+    parameters() {
+        const { dModel, dState, dConv } = this.config;
+        const D = this.dInner;
+        const N = dState;
+        const K = dConv;
+        const R = this.dtRank;
+        return [
+            { buf: this.gpuWeights['wInProj'], numel: 2 * D * dModel, name: 'wInProj' },
+            { buf: this.gpuWeights['bInProj'], numel: 2 * D, name: 'bInProj' },
+            { buf: this.gpuWeights['wConv'], numel: D * K, name: 'wConv' },
+            { buf: this.gpuWeights['bConv'], numel: D, name: 'bConv' },
+            { buf: this.gpuWeights['wXProj'], numel: (R + 2 * N) * D, name: 'wXProj' },
+            { buf: this.gpuWeights['bXProj'], numel: R + 2 * N, name: 'bXProj' },
+            { buf: this.gpuWeights['wDtProj'], numel: D * R, name: 'wDtProj' },
+            { buf: this.gpuWeights['bDtProj'], numel: D, name: 'bDtProj' },
+            { buf: this.gpuWeights['A_log'], numel: D * N, name: 'A_log' },
+            { buf: this.gpuWeights['D_vec'], numel: D, name: 'D_vec' },
+            { buf: this.gpuWeights['wOutProj'], numel: dModel * D, name: 'wOutProj' },
+            { buf: this.gpuWeights['bOutProj'], numel: dModel, name: 'bOutProj' },
+            { buf: this.gpuWeights['normWeight'], numel: dModel, name: 'normWeight' },
+        ];
+    }
+    getTrainableParams() {
+        if (this._wslaMode) {
+            return [
+                { buf: this.gpuWeights['wXProj'], numel: this.wXProj.length, name: 'wXProj' },
+                { buf: this.gpuWeights['bXProj'], numel: this.bXProj.length, name: 'bXProj' },
+            ];
+        }
+        return this.parameters();
+    }
+    setWSLAMode(enabled) {
+        this._wslaMode = enabled;
+    }
+    destroy() {
+        for (const buf of Object.values(this.gpuWeights)) {
+            buf.destroy();
+        }
+        this.gpuWeights = {};
+    }
+}
+// Deprecated alias — kept until mambacode.js 3.0.0
+export { Mamba1Block as MambaBlock };
+//# sourceMappingURL=mamba1_block.js.map

package/dist/model/mamba1_block.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"mamba1_block.js","sourceRoot":"","sources":["../../src/model/mamba1_block.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,EACH,qBAAqB,EACrB,eAAe,EACf,mBAAmB,EACnB,wBAAwB,EACxB,mBAAmB,EACnB,cAAc,EACd,IAAI,GACP,MAAM,uBAAuB,CAAC;AAE/B,OAAO,EAAE,2BAA2B,EAAE,MAAO,8BAA8B,CAAC;AAC5E,OAAO,EAAE,aAAa,EAAE,MAAM,iBAAiB,CAAC;AAChD,OAAO,EAAE,mBAAmB,EAAE,MAAe,sBAAsB,CAAC;AACpE,OAAO,EAAE,mBAAmB,EAAE,MAAe,iCAAiC,CAAC;AAC/E,OAAO,EAAE,gBAAgB,EAAE,MAAkB,2BAA2B,CAAC;AAmCzE,gFAAgF;AAEhF,MAAM,UAAU,GAAG,UAAU,CAAA;;;;;;;;;;CAU5B,CAAC;AAEF,MAAM,UAAU,GAAG,UAAU,CAAA;;;;;;;;;;CAU5B,CAAC;AAEF,iFAAiF;AAEjF,MAAM,OAAO,WAAW;IACX,SAAS,GAAG,QAAiB,CAAC;IAEvC,MAAM,CAAc;IACpB,MAAM,CAAgC;IACtC,MAAM,CAAW;IACjB,MAAM,CAAW;IAEjB,OAAO,CAAkB;IACzB,OAAO,CAAkB;IACzB,KAAK,CAAoB;IACzB,KAAK,CAAoB;IACzB,MAAM,CAAmB;IACzB,MAAM,CAAmB;IACzB,OAAO,CAAkB;IACzB,OAAO,CAAkB;IACzB,KAAK,CAAoB;IACzB,KAAK,CAAoB;IACzB,QAAQ,CAAiB;IACzB,QAAQ,CAAiB;IACzB,UAAU,CAAe;IAEzB,UAAU,CAA6B;IACvC,SAAS,CAAuC;IAExC,SAAS,GAAG,KAAK,CAAC;IAE1B,YAAY,MAAiB,EAAE,MAAyB;QACpD,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;QACrB,IAAI,CAAC,MAAM,GAAG;YACV,MAAM,EAAI,EAAE;YACZ,KAAK,EAAK,CAAC;YACX,MAAM,EAAI,CAAC;YACX,QAAQ,EAAE,IAAI;YACd,MAAM,EAAI,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,MAAM,GAAG,EAAE,CAAC;YACvC,GAAG,MAAM;SACmB,CAAC;QAEjC,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,GAAG,IAAI,CAAC,MAAM,CAAC;QACvC,IAAI,CAAC,MAAM,GAAG,MAAM,GAAG,MAAM,CAAC;QAC9B,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC,MAAM,IAAI,IAAI,CAAC,IAAI,CAAC,MAAM,GAAG,EAAE,CAAC,CAAC;QAEtD,IAAI,CAAC,OAAO,GAAM,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QACtC,IAAI,CAAC,OAAO,GAAM,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QACtC,IAAI,CAAC,KAAK,GAAQ,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QACtC,IAAI,CAAC,KAAK,GAAQ,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QACtC,IAAI,CAAC,MAAM,GAAO,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QACtC,IAAI,CAAC,MAAM,GAAO,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QACtC,IAAI,CAAC,OAAO,GAAM,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QACtC,IAAI,CAAC,OAAO,GAAM,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QACtC,IAAI,CAAC,KAAK,GAAQ,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QACtC,IAAI,CAAC,KAAK,GAAQ,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QACtC,IAAI,CAAC,QAAQ,GAAK,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QACtC,IAAI,CAAC,QAAQ,GAAK,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QACtC,IAAI,CAAC,UAAU,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QACtC,IAAI,CAAC,UAAU,GAAG,EAAE,CAAC;QACrB,IAAI,CAAC,SAAS,GAAI,EAAE,CAAC;QAErB,IAAI,CAAC,YAAY,EAAE,CAAC;QACpB,IAAI,CAAC,eAAe,EAAE,CAAC;IAC3B,CAAC;IAEO,YAAY;QAChB,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,GAAG,IAAI,CAAC,MAAM,CAAC;QAC9C,MAAM,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC;QACtB,MAAM,CAAC,GAAG,MAAM,CAAC;QACjB,MAAM,CAAC,GAAG,KAAK,CAAC;QAChB,MAAM,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC;QAEtB,MAAM,KAAK,GAAG,CAAC,CAAS,EAAE,GAAG,GAAG,IAAI,EAAgB,EAAE,CAAC,aAAa,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC;QAE7E,MAAM,KAAK,GAAG,CAAC,CAAS,EAAgB,EAAE,CAAC,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QAC/D,MAAM,IAAI,GAAI,CAAC,CAAS,EAAgB,EAAE,CAAC,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;QAEzE,IAAI,CAAC,OAAO,GAAI,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,MAAM,CAAC,CAAC;QACtC,IAAI,CAAC,OAAO,GAAI,KAAK,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;QAC7B,IAAI,CAAC,KAAK,GAAM,KAAK,CAAC,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QACnC,IAAI,CAAC,KAAK,GAAM,KAAK,CAAC,CAAC,CAAC,CAAC;QACzB,IAAI,CAAC,MAAM,GAAK,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QAC7C,IAAI,CAAC,MAAM,GAAK,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC;QACjC,IAAI,CAAC,OAAO,GAAI,KAAK,CAAC,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QACnC,IAAI,CAAC,OAAO,GAAI,KAAK,CAAC,CAAC,CAAC,CAAC;QAEzB,IAAI,CAAC,KAAK,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;QACrC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YACzB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBACzB,IAAI,CAAC,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;YAC5C,CAAC;QACL,CAAC;QAED,IAAI,CAAC,KAAK,GAAO,IAAI,CAAC,CAAC,CAAC,CAAC;QACzB,IAAI,CAAC,QAAQ,GAAI,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QACzC,IAAI,CAAC,QAAQ,GAAI,KAAK,CAAC,MAAM,CAAC,CAAC;QAC/B,IAAI,CAAC,UAAU,GAAG,IAAI,CAAC,MAAM,CAAC,CAAC;QAE/B,IAAI,CAAC,mBAAmB,EAAE,CAAC;IAC/B,CAAC;IAEO,mBAAmB;QACvB,MAAM,CAAC,GAAI,IAAI,CAAC,MAAM,CAAC;QACvB,MAAM,EAAE,GAAG,CAAC,GAAiB,EAAa,EAAE,CAAC,mBAAmB,CAAC,CAAC,EAAE,GAAG,EAAE,IAAI,CAAC,CAAC;QAE/E,IAAI,CAAC,UAAU,GAAG;YACd,OAAO,EAAK,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC;YAC5B,OAAO,EAAK,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC;YAC5B,KAAK,EAAO,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC;YAC1B,KAAK,EAAO,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC;YAC1B,MAAM,EAAM,EAAE,CAAC,IAAI,CAAC,MAAM,CAAC;YAC3B,MAAM,EAAM,EAAE,CAAC,IAAI,CAAC,MAAM,CAAC;YAC3B,OAAO,EAAK,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC;YAC5B,OAAO,EAAK,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC;YAC5B,KAAK,EAAO,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC;YAC1B,KAAK,EAAO,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC;YAC1B,QAAQ,EAAI,EAAE,CAAC,IAAI,CAAC,QAAQ,CAAC;YAC7B,QAAQ,EAAI,EAAE,CAAC,IAAI,CAAC,QAAQ,CAAC;YAC7B,UAAU,EAAE,EAAE,CAAC,IAAI,CAAC,UAAU,CAAC;SAClC,CAAC;IACN,CAAC;IAEO,eAAe;QACnB,MAAM,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC;QACtB,IAAI,CAAC,SAAS,GAAG;YACb,MAAM,EAAQ,qBAAqB,CAAC,CAAC,EAAE,mBAAmB,EAAW,gBAAgB,CAAC;YACtF,MAAM,EAAQ,qBAAqB,CAAC,CAAC,EAAE,mBAAmB,EAAW,gBAAgB,CAAC;YACtF,IAAI,EAAU,qBAAqB,CAAC,CAAC,EAAE,gBAAgB,EAAc,cAAc,CAAC;YACpF,OAAO,EAAO,qBAAqB,CAAC,CAAC,EAAE,gBAAgB,EAAc,iBAAiB,CAAC;YACvF,QAAQ,EAAM,qBAAqB,CAAC,CAAC,EAAE,2BAA2B,EAAG,cAAc,CAAC;YACpF,WAAW,EAAG,qBAAqB,CAAC,CAAC,EAAE,2BAA2B,EAAG,gBAAgB,CAAC;YACtF,KAAK,EAAS,qBAAqB,CAAC,CAAC,EAAE,UAAU,EAAE,MAAM,CAAC;YAC1D,KAAK,EAAS,qBAAqB,CAAC,CAAC,EAAE,UAAU,EAAE,MAAM,CAAC;SAC7D,CAAC;IACN,CAAC;IAED,OAAO,CAAC,IAAe,EAAE,KAAa,EAAE,MAAc;QAClD,MAAM,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC;QACtB,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,GAAG,IAAI,CAAC,MAAM,CAAC;QAC9C,MAAM,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC;QACtB,MAAM,CAAC,GAAG,MAAM,CAAC;QACjB,MAAM,CAAC,GAAG,KAAK,CAAC;QAChB,MAAM,CAAC,GAAG,MAAM,CAAC;QACjB,MAAM,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;QAChB,MAAM,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC;QAEtB,MAAM,KAAK,GAAG,EAAgB,CAAC;QAE/B,uBAAuB;QACvB,MAAM,OAAO,GAAG,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,MAAM,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QAClE,MAAM,OAAO,GAAG,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QACzD,KAAK,CAAC,OAAO,GAAG,OAAO,CAAC;QACxB,KAAK,CAAC,MAAM,GAAI,IAAI,CAAC;QACrB,CAAC;YACG,MAAM,MAAM,GAAG,IAAI,WAAW,CAAC,EAAE,CAAC,CAAC;YACnC,IAAI,WAAW,CAAC,MAAM,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC,CAAC;YAC/C,IAAI,YAAY,CAAC,MAAM,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC;YAC3C,MAAM,IAAI,GAAG,mBAAmB,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC;YAC5C,MAAM,EAAE,GAAG,eAAe,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,SAAS,CAAE,EACpD,CAAC,IAAI,EAAE,IAAI,EAAE,IAAI,CAAC,UAAU,CAAC,YAAY,CAAE,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC,CAAC;YACpE,cAAc,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,SAAS,CAAE,EAAE,EAAE,EAAE,CAAC,IAAI,CAAC,CAAC,EAAE,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;QAC3E,CAAC;QAED,gCAAgC;QAChC,MAAM,SAAS,GAAG,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QACnE,KAAK,CAAC,OAAO,GAAG,OAAO,CAAC;QACxB,CAAC;YACG,MAAM,MAAM,GAAG,IAAI,WAAW,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC;YAC1D,MAAM,IAAI,GAAK,mBAAmB,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC;YAC9C,MAAM,EAAE,GAAG,eAAe,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAE,EACnD,CAAC,IAAI,EAAE,OAAO,EAAE,IAAI,CAAC,UAAU,CAAC,SAAS,CAAE,EAAE,IAAI,CAAC,UAAU,CAAC,SAAS,CAAE,EAAE,SAAS,CAAC,CAAC,CAAC;YAC1F,cAAc,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAE,EAAE,EAAE,EAAE,CAAC,IAAI,CAAC,CAAC,EAAE,EAAE,CAAC,EAAE,IAAI,CAAC,CAAC,GAAG,CAAC,EAAE,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;QACxF,CAAC;QAED,wBAAwB;QACxB,MAAM,OAAO,GAAG,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QAC7D,MAAM,IAAI,GAAM,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QAC7D,CAAC;YACG,MAAM,GAAG,GAAG,CAAC,CAAC,oBAAoB,EAAE,CAAC;YACrC,GAAG,CAAC,kBAAkB,CAAC,SAAS,EAAE,CAAC,EAAU,OAAO,EAAE,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC;YACpE,GAAG,CAAC,kBAAkB,CAAC,SAAS,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,IAAI,EAAK,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC;YACpE,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,GAAG,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC;QACnC,CAAC;QACD,SAAS,CAAC,OAAO,EAAE,CAAC;QACpB,KAAK,CAAC,IAAI,GAAM,IAAI,CAAC;QACrB,KAAK,CAAC,OAAO,GAAG,OAAO,CAAC;QAExB,wBAAwB;QACxB,MAAM,OAAO,GAAG,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QAC7D,KAAK,CAAC,OAAO,GAAG,OAAO,CAAC;QACxB,CAAC;YACG,MAAM,MAAM,GAAG,IAAI,WAAW,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC;YACxD,MAAM,IAAI,GAAK,mBAAmB,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC;YAC9C,MAAM,EAAE,GAAG,eAAe,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAE,EACnD,CAAC,IAAI,EAAE,OAAO,EAAE,IAAI,CAAC,UAAU,CAAC,OAAO,CAAE,EAAE,IAAI,CAAC,UAAU,CAAC,OAAO,CAAE,EAAE,OAAO,CAAC,CAAC,CAAC;YACpF,cAAc,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAE,EAAE,EAAE,EAAE,CAAC,IAAI,CAAC,CAAC,EAAE,EAAE,CAAC,EAAE,IAAI,CAAC,CAAC,EAAE,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;QACpF,CAAC;QAED,qBAAqB;QACrB,MAAM,OAAO,GAAG,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QAC7D,KAAK,CAAC,OAAO,GAAG,OAAO,CAAC;QACxB,CAAC;YACG,MAAM,MAAM,GAAG,IAAI,WAAW,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC;YAC/C,MAAM,IAAI,GAAK,mBAAmB,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC;YAC9C,MAAM,EAAE,GAAG,eAAe,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,MAAM,CAAE,EACjD,CAAC,IAAI,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC,CAAC;YAC9B,cAAc,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,MAAM,CAAE,EAAE,EAAE,EAAE,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,EAAE,GAAG,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;QAC7E,CAAC;QAED,8BAA8B;QAC9B,MAAM,QAAQ,GAAG,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QACxE,CAAC;YACG,MAAM,MAAM,GAAG,IAAI,WAAW,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC;YACzD,MAAM,IAAI,GAAK,mBAAmB,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC;YAC9C,MAAM,EAAE,GAAG,eAAe,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAE,EACnD,CAAC,IAAI,EAAE,OAAO,EAAE,IAAI,CAAC,UAAU,CAAC,QAAQ,CAAE,EAAE,IAAI,CAAC,UAAU,CAAC,QAAQ,CAAE,EAAE,QAAQ,CAAC,CAAC,CAAC;YACvF,cAAc,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAE,EAAE,EAAE,EAAE,CAAC,IAAI,CAAC,CAAC,EAAE,EAAE,CAAC,EAAE,IAAI,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;QAC5F,CAAC;QAED,MAAM,KAAK,GAAG,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QAC3D,MAAM,KAAK,GAAG,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QAC/D,MAAM,KAAK,GAAG,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QAC/D,CAAC;YACG,MAAM,GAAG,GAAG,CAAC,CAAC,oBAAoB,EAAE,CAAC;YACrC,GAAG,CAAC,kBAAkB,CAAC,QAAQ,EAAE,CAAC,EAAgB,KAAK,EAAE,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC;YACvE,GAAG,CAAC,kBAAkB,CAAC,QAAQ,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAQ,KAAK,EAAE,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC;YAC3E,GAAG,CAAC,kBAAkB,CAAC,QAAQ,EAAE,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC;YAC3E,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,GAAG,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC;QACnC,CAAC;QACD,QAAQ,CAAC,OAAO,EAAE,CAAC;QACnB,KAAK,CAAC,KAAK,GAAG,KAAK,CAAC;QACpB,KAAK,CAAC,KAAK,GAAG,KAAK,CAAC;QAEpB,mCAAmC;QACnC,MAAM,SAAS,GAAG,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QAC/D,KAAK,CAAC,SAAS,GAAG,SAAS,CAAC;QAC5B,CAAC;YACG,MAAM,MAAM,GAAG,IAAI,WAAW,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC;YACjD,MAAM,IAAI,GAAK,mBAAmB,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC;YAC9C,MAAM,EAAE,GAAG,eAAe,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAE,EACnD,CAAC,IAAI,EAAE,KAAK,EAAE,IAAI,CAAC,UAAU,CAAC,SAAS,CAAE,EAAE,IAAI,CAAC,UAAU,CAAC,SAAS,CAAE,EAAE,SAAS,CAAC,CAAC,CAAC;YACxF,cAAc,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAE,EAAE,EAAE,EAAE,CAAC,IAAI,CAAC,CAAC,EAAE,EAAE,CAAC,EAAE,IAAI,CAAC,CAAC,EAAE,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;QACpF,CAAC;QACD,KAAK,CAAC,OAAO,EAAE,CAAC;QAEhB,yBAAyB;QACzB,MAAM,KAAK,GAAI,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QAChE,MAAM,MAAM,GAAG,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QACxE,KAAK,CAAC,MAAM,GAAG,MAAM,CAAC;QACtB,CAAC;YACG,MAAM,MAAM,GAAG,IAAI,WAAW,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC;YACpD,MAAM,IAAI,GAAK,mBAAmB,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC;YAE9C,MAAM,GAAG,GAAG,eAAe,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,UAAU,CAAE,EACtD,CAAC,IAAI,EAAE,OAAO,EAAE,SAAS,EAAE,IAAI,CAAC,UAAU,CAAC,OAAO,CAAE,EAAE,KAAK,EAAE,KAAK;gBACjE,IAAI,CAAC,UAAU,CAAC,OAAO,CAAE,EAAE,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC;YAChD,cAAc,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,UAAU,CAAE,EAAE,GAAG,EAAE,CAAC,IAAI,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;YAEjF,MAAM,GAAG,GAAG,eAAe,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,aAAa,CAAE,EACzD,CAAC,IAAI,EAAE,OAAO,EAAE,SAAS,EAAE,IAAI,CAAC,UAAU,CAAC,OAAO,CAAE,EAAE,KAAK,EAAE,KAAK;gBACjE,IAAI,CAAC,UAAU,CAAC,OAAO,CAAE,EAAE,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC;YAChD,cAAc,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,aAAa,CAAE,EAAE,GAAG,EAAE,CAAC,IAAI,CAAC,CAAC,EAAE,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;QAChF,CAAC;QAED,uBAAuB;QACvB,MAAM,KAAK,GAAM,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QAC9D,MAAM,QAAQ,GAAG,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QAC9D,CAAC;YACG,MAAM,IAAI,GAAG,mBAAmB,CAAC,CAAC,EAAE,IAAI,WAAW,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC;YACrE,MAAM,GAAG,GAAI,eAAe,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,MAAM,CAAE,EACnD,CAAC,IAAI,EAAE,IAAI,EAAE,KAAK,CAAC,CAAC,CAAC;YACzB,cAAc,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,MAAM,CAAE,EAAE,GAAG,EAAE,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,EAAE,GAAG,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;YAE1E,MAAM,KAAK,GAAG,mBAAmB,CAAC,CAAC,EAAE,IAAI,WAAW,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC;YACtE,MAAM,KAAK,GAAG,eAAe,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,OAAO,CAAE,EACrD,CAAC,KAAK,EAAE,KAAK,EAAE,QAAQ,EAAE,KAAK,CAAC,CAAC,CAAC;YACrC,cAAc,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,OAAO,CAAE,EAAE,KAAK,EAAE,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,EAAE,GAAG,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;QACjF,CAAC;QACD,KAAK,CAAC,OAAO,EAAE,CAAC;QAChB,KAAK,CAAC,OAAO,EAAE,CAAC;QAEhB,wBAAwB;QACxB,MAAM,UAAU,GAAG,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,MAAM,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QACrE,CAAC;YACG,MAAM,MAAM,GAAG,IAAI,WAAW,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC;YACtD,MAAM,IAAI,GAAK,mBAAmB,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC;YAC9C,MAAM,EAAE,GAAG,eAAe,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAE,EACnD,CAAC,IAAI,EAAE,QAAQ,EAAE,IAAI,CAAC,UAAU,CAAC,UAAU,CAAE,EAAE,IAAI,CAAC,UAAU,CAAC,UAAU,CAAE,EAAE,UAAU,CAAC,CAAC,CAAC;YAC9F,cAAc,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAE,EAAE,EAAE,EAAE,CAAC,IAAI,CAAC,CAAC,EAAE,EAAE,CAAC,EAAE,IAAI,CAAC,MAAM,EAAE,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;QACzF,CAAC;QACD,QAAQ,CAAC,OAAO,EAAE,CAAC;QAEnB,mBAAmB;QACnB,MAAM,MAAM,GAAG,wBAAwB,CAAC,CAAC,EAAE,CAAC,GAAG,MAAM,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;QACjE,CAAC;YACG,MAAM,IAAI,GAAG,mBAAmB,CAAC,CAAC,EAAE,IAAI,WAAW,CAAC,CAAC,CAAC,GAAG,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC;YAC1E,MAAM,EAAE,GAAK,eAAe,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,OAAO,CAAE,EACpD,CAAC,UAAU,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,CAAC,CAAC,CAAC;YACtC,cAAc,CAAC,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,OAAO,CAAE,EAAE,EAAE,EAAE,CAAC,IAAI,CAAC,CAAC,GAAG,MAAM,EAAE,GAAG,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;QACnF,CAAC;QACD,UAAU,CAAC,OAAO,EAAE,CAAC;QAErB,OAAO,EAAE,MAAM,EAAE,KAAK,EAAE,CAAC;IAC7B,CAAC;IAED,UAAU;QACN,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,GAAG,IAAI,CAAC,MAAM,CAAC;QAC9C,MAAM,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC;QACtB,MAAM,CAAC,GAAG,MAAM,CAAC;QACjB,MAAM,CAAC,GAAG,KAAK,CAAC;QAChB,MAAM,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC;QAEtB,OAAO;YACH,EAAE,GAAG,EAAE,IAAI,CAAC,UAAU,CAAC,SAAS,CAAE,EAAK,KAAK,EAAE,CAAC,GAAG,CAAC,GAAG,MAAM,EAAI,IAAI,EAAE,SAAS,EAAK;YACpF,EAAE,GAAG,EAAE,IAAI,CAAC,UAAU,CAAC,SAAS,CAAE,EAAK,KAAK,EAAE,CAAC,GAAG,CAAC,EAAa,IAAI,EAAE,SAAS,EAAK;YACpF,EAAE,GAAG,EAAE,IAAI,CAAC,UAAU,CAAC,OAAO,CAAE,EAAO,KAAK,EAAE,CAAC,GAAG,CAAC,EAAa,IAAI,EAAE,OAAO,EAAO;YACpF,EAAE,GAAG,EAAE,IAAI,CAAC,UAAU,CAAC,OAAO,CAAE,EAAO,KAAK,EAAE,CAAC,EAAiB,IAAI,EAAE,OAAO,EAAO;YACpF,EAAE,GAAG,EAAE,IAAI,CAAC,UAAU,CAAC,QAAQ,CAAE,EAAM,KAAK,EAAE,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,EAAE,IAAI,EAAE,QAAQ,EAAK;YAClF,EAAE,GAAG,EAAE,IAAI,CAAC,UAAU,CAAC,QAAQ,CAAE,EAAM,KAAK,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAQ,IAAI,EAAE,QAAQ,EAAK;YAClF,EAAE,GAAG,EAAE,IAAI,CAAC,UAAU,CAAC,SAAS,CAAE,EAAK,KAAK,EAAE,CAAC,GAAG,CAAC,EAAa,IAAI,EAAE,SAAS,EAAI;YACnF,EAAE,GAAG,EAAE,IAAI,CAAC,UAAU,CAAC,SAAS,CAAE,EAAK,KAAK,EAAE,CAAC,EAAiB,IAAI,EAAE,SAAS,EAAI;YACnF,EAAE,GAAG,EAAE,IAAI,CAAC,UAAU,CAAC,OAAO,CAAE,EAAO,KAAK,EAAE,CAAC,GAAG,CAAC,EAAa,IAAI,EAAE,OAAO,EAAM;YACnF,EAAE,GAAG,EAAE,IAAI,CAAC,UAAU,CAAC,OAAO,CAAE,EAAO,KAAK,EAAE,CAAC,EAAiB,IAAI,EAAE,OAAO,EAAM;YACnF,EAAE,GAAG,EAAE,IAAI,CAAC,UAAU,CAAC,UAAU,CAAE,EAAI,KAAK,EAAE,MAAM,GAAG,CAAC,EAAQ,IAAI,EAAE,UAAU,EAAG;YACnF,EAAE,GAAG,EAAE,IAAI,CAAC,UAAU,CAAC,UAAU,CAAE,EAAI,KAAK,EAAE,MAAM,EAAY,IAAI,EAAE,UAAU,EAAG;YACnF,EAAE,GAAG,EAAE,IAAI,CAAC,UAAU,CAAC,YAAY,CAAE,EAAE,KAAK,EAAE,MAAM,EAAY,IAAI,EAAE,YAAY,EAAC;SACtF,CAAC;IACN,CAAC;IAED,kBAAkB;QACd,IAAI,IAAI,CAAC,SAAS,EAAE,CAAC;YACjB,OAAO;gBACH,EAAE,GAAG,EAAE,IAAI,CAAC,UAAU,CAAC,QAAQ,CAAE,EAAE,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,MAAM,EAAE,IAAI,EAAE,QAAQ,EAAE;gBAC9E,EAAE,GAAG,EAAE,IAAI,CAAC,UAAU,CAAC,QAAQ,CAAE,EAAE,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,MAAM,EAAE,IAAI,EAAE,QAAQ,EAAE;aACjF,CAAC;QACN,CAAC;QACD,OAAO,IAAI,CAAC,UAAU,EAAE,CAAC;IAC7B,CAAC;IAED,WAAW,CAAC,OAAgB;QACxB,IAAI,CAAC,SAAS,GAAG,OAAO,CAAC;IAC7B,CAAC;IAED,OAAO;QACH,KAAK,MAAM,GAAG,IAAI,MAAM,CAAC,MAAM,CAAC,IAAI,CAAC,UAAU,CAAC,EAAE,CAAC;YAC/C,GAAG,CAAC,OAAO,EAAE,CAAC;QAClB,CAAC;QACD,IAAI,CAAC,UAAU,GAAG,EAAE,CAAC;IACzB,CAAC;CACJ;AAED,mDAAmD;AACnD,OAAO,EAAE,WAAW,IAAI,UAAU,EAAE,CAAC"}

package/dist/model/mamba2_block.d.ts

@@ -0,0 +1,44 @@
+/**
+ * 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

package/dist/model/mamba2_block.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"mamba2_block.d.ts","sourceRoot":"","sources":["../../src/model/mamba2_block.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAkBH,OAAO,KAAK,EAAE,aAAa,EAAE,kBAAkB,EAAE,UAAU,EAAE,MAAM,qBAAqB,CAAC;AAEzF,MAAM,WAAW,iBAAiB;IAC9B,MAAM,EAAK,MAAM,CAAC;IAClB,MAAM,EAAK,MAAM,CAAC;IAClB,KAAK,EAAM,MAAM,CAAC;IAClB,MAAM,EAAK,MAAM,CAAC;IAClB,MAAM,EAAK,MAAM,CAAC;IAClB,OAAO,EAAI,MAAM,CAAC;IAClB,QAAQ,EAAG,MAAM,CAAC;CACrB;AAED,MAAM,WAAW,WAAW;IACxB,UAAU,EAAG,SAAS,CAAC;CAC1B;AAcD,qBAAa,WAAY,YAAW,aAAa;IAC7C,QAAQ,CAAC,SAAS,EAAG,QAAQ,CAAU;IAEvC,MAAM,EAAG,SAAS,CAAC;IACnB,MAAM,EAAG,QAAQ,CAAC,iBAAiB,CAAC,CAAC;IACrC,MAAM,EAAG,MAAM,CAAC;IAChB,KAAK,EAAI,MAAM,CAAC;IAEhB,UAAU,EAAG,MAAM,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;IACvC,SAAS,EAAI,MAAM,CAAC,MAAM,EAAE,kBAAkB,CAAC,CAAC;IAEhD,OAAO,CAAC,SAAS,CAAS;gBAEd,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,iBAAiB;IAwBxD,OAAO,CAAC,YAAY;IA8BpB,OAAO,CAAC,eAAe;IAWvB,OAAO,CAAC,IAAI,EAAE,SAAS,EAAE,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,GAAG,kBAAkB;IA6I3E,UAAU,IAAI,UAAU,EAAE;IAsB1B,kBAAkB,IAAI,UAAU,EAAE;IAYlC,WAAW,CAAC,OAAO,EAAE,OAAO,GAAG,IAAI;IAInC,OAAO,IAAI,IAAI;CAIlB"}

package/dist/model/mamba2_block.js

@@ -0,0 +1,252 @@
+/**
+ * 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 { createComputePipeline, createBindGroup, createStorageBuffer, createEmptyStorageBuffer, createUniformBuffer, dispatchKernel, cdiv, } from '../utils/gpu_utils.js';
+import { SSD_FORWARD_WGSL } from '../kernels/ssd.js';
+import { gaussianArray } from '../utils/rng.js';
+import { CONV1D_FORWARD_WGSL } from '../kernels/conv1d.js';
+import { LINEAR_FORWARD_WGSL } from '../kernels/linear_projection.js';
+import { ACTIVATIONS_WGSL } from '../kernels/activations.js';
+const ADD_SHADER = /* wgsl */ `
+@group(0) @binding(0) var<storage, read>       a : array<f32>;
+@group(0) @binding(1) var<storage, read>       b : array<f32>;
+@group(0) @binding(2) var<storage, read_write> c : array<f32>;
+@group(0) @binding(3) var<uniform>             n : u32;
+@compute @workgroup_size(256)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let i = gid.x;
+    if (i < n) { c[i] = a[i] + b[i]; }
+}
+`;
+export class Mamba2Block {
+    layerType = 'mamba2';
+    device;
+    config;
+    dInner;
+    dHead;
+    gpuWeights;
+    pipelines;
+    _wslaMode = false;
+    constructor(device, config) {
+        this.device = device;
+        this.config = {
+            ...{ dState: 16, dConv: 4, expand: 2, nGroups: 1, chunkLen: 256 },
+            ...config,
+        };
+        const { dModel, expand, nHeads } = this.config;
+        this.dInner = expand * dModel;
+        this.dHead = this.dInner / nHeads;
+        if (this.dInner % nHeads !== 0) {
+            throw new Error(`Mamba2Block: dInner (${this.dInner}) must be divisible by nHeads (${nHeads}).`);
+        }
+        this.gpuWeights = {};
+        this.pipelines = {};
+        this._initWeights();
+        this._buildPipelines();
+    }
+    _initWeights() {
+        const { dModel, dState, dConv, nHeads, nGroups } = this.config;
+        const D = this.dInner;
+        const N = dState;
+        const K = dConv;
+        const H = nHeads;
+        const G = nGroups;
+        const randn = (n, std = 0.02) => gaussianArray(n, std);
+        const zeros = (n) => new Float32Array(n);
+        const ones = (n) => new Float32Array(n).fill(1.0);
+        // wInProj: (D_inner + 2*n_groups*N + H, D_model) — no bias per Mamba-2 spec
+        const inProjRows = D + 2 * G * N + H;
+        const mk = (arr) => createStorageBuffer(this.device, arr, true);
+        this.gpuWeights = {
+            wInProj: mk(randn(inProjRows * dModel)),
+            wConv: mk(randn((D + 2 * G * N) * K, 0.01)),
+            bConv: mk(zeros(D + 2 * G * N)),
+            A_log: mk(new Float32Array(H).fill(Math.log(1.0))),
+            dt_bias: mk(zeros(H)),
+            D_vec: mk(ones(H)),
+            wOutProj: mk(randn(dModel * D, 0.02)),
+            normWeight: mk(ones(D)), // inner RMSNorm
+            preNormWeight: mk(ones(dModel)), // pre-block RMSNorm
+        };
+    }
+    _buildPipelines() {
+        const d = this.device;
+        this.pipelines = {
+            linear: createComputePipeline(d, LINEAR_FORWARD_WGSL, 'linear_forward'),
+            conv1d: createComputePipeline(d, CONV1D_FORWARD_WGSL, 'conv1d_forward'),
+            rmsnorm: createComputePipeline(d, ACTIVATIONS_WGSL, 'rmsnorm_forward'),
+            ssd_fwd: createComputePipeline(d, SSD_FORWARD_WGSL, 'ssd_chunk_forward'),
+            elAdd: createComputePipeline(d, ADD_SHADER, 'main'),
+        };
+    }
+    forward(xBuf, batch, seqLen) {
+        const d = this.device;
+        const { dModel, dState, dConv, nHeads, nGroups, chunkLen } = this.config;
+        const D = this.dInner;
+        const N = dState;
+        const K = dConv;
+        const H = nHeads;
+        const G = nGroups;
+        const dh = this.dHead;
+        const B = batch;
+        const L = seqLen;
+        const M = B * L;
+        const convD = D + 2 * G * N; // channels for conv (x, B_proj, C_proj)
+        const numChunks = Math.ceil(L / chunkLen);
+        // 1. Pre-block RMSNorm
+        const normOut = createEmptyStorageBuffer(d, M * dModel * 4, true);
+        const normInv = createEmptyStorageBuffer(d, M * 4, true);
+        {
+            const params = new ArrayBuffer(16);
+            new Uint32Array(params, 0, 2).set([M, dModel]);
+            new Float32Array(params, 8, 1).set([1e-6]);
+            const pBuf = createUniformBuffer(d, params);
+            const bg = createBindGroup(d, this.pipelines['rmsnorm'], [pBuf, xBuf, this.gpuWeights['preNormWeight'], normOut, normInv]);
+            dispatchKernel(d, this.pipelines['rmsnorm'], bg, [cdiv(M, 64), 1, 1]);
+        }
+        normInv.destroy();
+        // 2. Fused in_proj → [x (D), B_proj (G*N), C_proj (G*N), dt (H)]
+        const inProjRows = D + 2 * G * N + H;
+        const inProjOut = createEmptyStorageBuffer(d, M * inProjRows * 4, true);
+        {
+            const params = new Uint32Array([M, dModel, inProjRows]).buffer;
+            const pBuf = createUniformBuffer(d, params);
+            // wInProj has no bias — pass a zero-filled buffer
+            const zeroBias = createStorageBuffer(d, new Float32Array(inProjRows), true);
+            const bg = createBindGroup(d, this.pipelines['linear'], [pBuf, normOut, this.gpuWeights['wInProj'], zeroBias, inProjOut]);
+            dispatchKernel(d, this.pipelines['linear'], bg, [cdiv(M, 16), cdiv(inProjRows, 16), 1]);
+            zeroBias.destroy();
+        }
+        normOut.destroy();
+        // Split: xConv [D+2GN], dt [H]
+        const xConvBuf = createEmptyStorageBuffer(d, M * convD * 4, true);
+        const dtBuf = createEmptyStorageBuffer(d, M * H * 4, true);
+        {
+            const enc = d.createCommandEncoder();
+            enc.copyBufferToBuffer(inProjOut, 0, xConvBuf, 0, M * convD * 4);
+            enc.copyBufferToBuffer(inProjOut, M * convD * 4, dtBuf, 0, M * H * 4);
+            d.queue.submit([enc.finish()]);
+        }
+        inProjOut.destroy();
+        // 3. Causal conv1d over x + B_proj + C_proj (fused, convD channels)
+        const convOut = createEmptyStorageBuffer(d, M * convD * 4, true);
+        {
+            const params = new Uint32Array([L, convD, K, B, 1]).buffer;
+            const pBuf = createUniformBuffer(d, params);
+            const bg = createBindGroup(d, this.pipelines['conv1d'], [pBuf, xConvBuf, this.gpuWeights['wConv'], this.gpuWeights['bConv'], convOut]);
+            dispatchKernel(d, this.pipelines['conv1d'], bg, [cdiv(L, 16), cdiv(convD, 16), B]);
+        }
+        xConvBuf.destroy();
+        // Split conv output: x [D], B_proj [G*N], C_proj [G*N]
+        const xSsdBuf = createEmptyStorageBuffer(d, M * D * 4, true);
+        const bProjBuf = createEmptyStorageBuffer(d, M * G * N * 4, true);
+        const cProjBuf = createEmptyStorageBuffer(d, M * G * N * 4, true);
+        {
+            const enc = d.createCommandEncoder();
+            enc.copyBufferToBuffer(convOut, 0, xSsdBuf, 0, M * D * 4);
+            enc.copyBufferToBuffer(convOut, M * D * 4, bProjBuf, 0, M * G * N * 4);
+            enc.copyBufferToBuffer(convOut, M * (D + G * N) * 4, cProjBuf, 0, M * G * N * 4);
+            d.queue.submit([enc.finish()]);
+        }
+        convOut.destroy();
+        // 4. SSD scan
+        // state_carry: [numChunks+1, B, H, N, dHead]
+        const stateCarry = createEmptyStorageBuffer(d, (numChunks + 1) * B * H * N * dh * 4, true);
+        const ssdOut = createEmptyStorageBuffer(d, M * D * 4, true);
+        {
+            const ssdParams = new Uint32Array([L, D, H, dh, G, N, chunkLen, numChunks, B]).buffer;
+            const pBuf = createUniformBuffer(d, ssdParams);
+            const bg = createBindGroup(d, this.pipelines['ssd_fwd'], [pBuf, xSsdBuf, bProjBuf, cProjBuf, dtBuf,
+                this.gpuWeights['A_log'], this.gpuWeights['dt_bias'],
+                this.gpuWeights['D_vec'], ssdOut, stateCarry]);
+            dispatchKernel(d, this.pipelines['ssd_fwd'], bg, [numChunks, H, B]);
+        }
+        xSsdBuf.destroy();
+        bProjBuf.destroy();
+        cProjBuf.destroy();
+        dtBuf.destroy();
+        // 5. Inner RMSNorm on scan output
+        const innerNormOut = createEmptyStorageBuffer(d, M * D * 4, true);
+        const innerNormInv = createEmptyStorageBuffer(d, M * 4, true);
+        {
+            const params = new ArrayBuffer(16);
+            new Uint32Array(params, 0, 2).set([M, D]);
+            new Float32Array(params, 8, 1).set([1e-6]);
+            const pBuf = createUniformBuffer(d, params);
+            const bg = createBindGroup(d, this.pipelines['rmsnorm'], [pBuf, ssdOut, this.gpuWeights['normWeight'], innerNormOut, innerNormInv]);
+            dispatchKernel(d, this.pipelines['rmsnorm'], bg, [cdiv(M, 64), 1, 1]);
+        }
+        ssdOut.destroy();
+        innerNormInv.destroy();
+        // 6. Output projection
+        const outProjOut = createEmptyStorageBuffer(d, M * dModel * 4, true);
+        {
+            const params = new Uint32Array([M, D, dModel]).buffer;
+            const pBuf = createUniformBuffer(d, params);
+            const zeroBias = createStorageBuffer(d, new Float32Array(dModel), true);
+            const bg = createBindGroup(d, this.pipelines['linear'], [pBuf, innerNormOut, this.gpuWeights['wOutProj'], zeroBias, outProjOut]);
+            dispatchKernel(d, this.pipelines['linear'], bg, [cdiv(M, 16), cdiv(dModel, 16), 1]);
+            zeroBias.destroy();
+        }
+        innerNormOut.destroy();
+        // 7. Residual add
+        const output = createEmptyStorageBuffer(d, M * dModel * 4, true);
+        {
+            const nBuf = createUniformBuffer(d, new Uint32Array([M * dModel]).buffer);
+            const bg = createBindGroup(d, this.pipelines['elAdd'], [outProjOut, xBuf, output, nBuf]);
+            dispatchKernel(d, this.pipelines['elAdd'], bg, [cdiv(M * dModel, 256), 1, 1]);
+        }
+        outProjOut.destroy();
+        const cache = { stateCarry };
+        return { output, cache };
+    }
+    parameters() {
+        const { dModel, dState, dConv, nHeads, nGroups } = this.config;
+        const D = this.dInner;
+        const N = dState;
+        const K = dConv;
+        const H = nHeads;
+        const G = nGroups;
+        const convD = D + 2 * G * N;
+        return [
+            { buf: this.gpuWeights['wInProj'], numel: (D + 2 * G * N + H) * dModel, name: 'wInProj' },
+            { buf: this.gpuWeights['wConv'], numel: convD * K, name: 'wConv' },
+            { buf: this.gpuWeights['bConv'], numel: convD, name: 'bConv' },
+            { buf: this.gpuWeights['A_log'], numel: H, name: 'A_log' },
+            { buf: this.gpuWeights['dt_bias'], numel: H, name: 'dt_bias' },
+            { buf: this.gpuWeights['D_vec'], numel: H, name: 'D_vec' },
+            { buf: this.gpuWeights['wOutProj'], numel: dModel * D, name: 'wOutProj' },
+            { buf: this.gpuWeights['normWeight'], numel: D, name: 'normWeight' },
+            { buf: this.gpuWeights['preNormWeight'], numel: dModel, name: 'preNormWeight' },
+        ];
+    }
+    getTrainableParams() {
+        if (this._wslaMode) {
+            // WSLA: train only B/C rows of wInProj (the selective projection part)
+            return [
+                { buf: this.gpuWeights['wInProj'],
+                    numel: (this.config.nGroups * this.config.dState * 2) * this.config.dModel,
+                    name: 'wInProj_BC' },
+            ];
+        }
+        return this.parameters();
+    }
+    setWSLAMode(enabled) {
+        this._wslaMode = enabled;
+    }
+    destroy() {
+        for (const buf of Object.values(this.gpuWeights))
+            buf.destroy();
+        this.gpuWeights = {};
+    }
+}
+//# sourceMappingURL=mamba2_block.js.map