@seanhogg/builderforce-memory-engine 2026.7.12 → 2026.9.27

package/dist/training/mamba1_cpu.js

@@ -0,0 +1,316 @@
+/**
+ * mamba1_cpu.ts — exact CPU forward + backward for a Mamba-1 (S6) block.
+ *
+ * This is the REFERENCE implementation of the same maths the WGSL kernels run
+ * (`rmsnorm_forward`, `linear_forward`, `conv1d_forward`, `silu_forward`,
+ * `forward_scan` + `forward_reduce`). It exists so real per-parameter gradients
+ * can be computed — and, crucially, PROVEN correct by finite differences —
+ * without a WebGPU device. {@link MambaTrainer} consumes it as its gradient
+ * engine: the optimizer (AdamW + trust region + NaN guard) stays on the GPU,
+ * the gradients are exact.
+ *
+ * Every formula below is a line-for-line mirror of the corresponding kernel; if
+ * a kernel changes, this file changes with it and `tests/backprop.test.ts`
+ * catches the drift (the gradient check fails the moment forward and backward
+ * disagree).
+ *
+ * Layout conventions (identical to the GPU buffers):
+ *   x        (M, dModel)        M = batch * seqLen, row r = b * seqLen + t
+ *   wInProj  (2*dInner, dModel) row-major, Y = X @ W^T + b
+ *   wConv    (dInner, dConv)    depthwise, causal
+ *   wXProj   (dtRank + 2*dState, dInner)
+ *   wDtProj  (dInner, dtRank)
+ *   A_log    (dInner, dState)
+ *   wOutProj (dModel, dInner)
+ *   hCache   (B, L, dInner, dState)
+ */
+import { linearForward, linearBackward, rmsNormForward, rmsNormBackward, silu, siluGrad, softplus, sigmoid, RMSNORM_EPS, } from './cpu_ops.js';
+/** Parameter names of a Mamba-1 block, in `parameters()` order. */
+export const MAMBA1_PARAM_NAMES = [
+    'wInProj', 'bInProj', 'wConv', 'bConv', 'wXProj', 'bXProj',
+    'wDtProj', 'bDtProj', 'A_log', 'D_vec', 'wOutProj', 'bOutProj', 'normWeight',
+];
+/**
+ * Clamp bounds on `A_log` applied by `discretise_A`/`discretise_B` in
+ * `SELECTIVE_SCAN_FORWARD_WGSL`. Outside this range the kernel's output is
+ * constant in `A_log`, so the gradient is exactly zero — mirrored here so the
+ * analytic gradient matches the function the model actually computes.
+ */
+export const A_LOG_CLAMP_LO = -10;
+export const A_LOG_CLAMP_HI = 5;
+/** Allocate a zeroed gradient set matching `dims`. */
+export function zeroMamba1Grads(dims) {
+    const { dModel, dState: N, dConv: K, dInner: D, dtRank: R } = dims;
+    return {
+        wInProj: new Float32Array(2 * D * dModel),
+        bInProj: new Float32Array(2 * D),
+        wConv: new Float32Array(D * K),
+        bConv: new Float32Array(D),
+        wXProj: new Float32Array((R + 2 * N) * D),
+        bXProj: new Float32Array(R + 2 * N),
+        wDtProj: new Float32Array(D * R),
+        bDtProj: new Float32Array(D),
+        A_log: new Float32Array(D * N),
+        D_vec: new Float32Array(D),
+        wOutProj: new Float32Array(dModel * D),
+        bOutProj: new Float32Array(dModel),
+        normWeight: new Float32Array(dModel),
+    };
+}
+/**
+ * Exact CPU forward for one Mamba-1 block.
+ *
+ * Mirrors `Mamba1Block.forward` step for step:
+ *   RMSNorm → in_proj → split(x, z) → causal conv1d → SiLU → x_proj(Δ, B, C)
+ *   → dt_proj → selective scan → gate by SiLU(z) → out_proj → residual add.
+ */
+export function mamba1CpuForward(x, w, dims) {
+    const { dModel, dState: N, dConv: K, dInner: D, dtRank: R, batch: B, seqLen: L } = dims;
+    const M = B * L;
+    const P = R + 2 * N;
+    // 1. Pre-block RMSNorm.
+    const normOut = new Float32Array(M * dModel);
+    const normInv = new Float32Array(M);
+    rmsNormForward(x, w.normWeight, M, dModel, normOut, normInv);
+    // 2. Input projection → (x, z).
+    const inProj = new Float32Array(M * 2 * D);
+    linearForward(normOut, w.wInProj, w.bInProj, M, dModel, 2 * D, inProj);
+    // 3. COLUMN split: x = inProj[:, 0:D], z = inProj[:, D:2D].
+    const xConvIn = new Float32Array(M * D);
+    const z = new Float32Array(M * D);
+    for (let r = 0; r < M; r++) {
+        for (let d = 0; d < D; d++) {
+            xConvIn[r * D + d] = inProj[r * 2 * D + d];
+            z[r * D + d] = inProj[r * 2 * D + D + d];
+        }
+    }
+    // 4. Depthwise causal conv1d over the sequence axis.
+    const convOut = new Float32Array(M * D);
+    for (let b = 0; b < B; b++) {
+        for (let t = 0; t < L; t++) {
+            const base = (b * L + t) * D;
+            for (let d = 0; d < D; d++) {
+                let acc = w.bConv[d];
+                for (let k = 0; k < K; k++) {
+                    if (t >= k)
+                        acc += w.wConv[d * K + k] * xConvIn[(b * L + t - k) * D + d];
+                }
+                convOut[base + d] = acc;
+            }
+        }
+    }
+    // 5. SiLU.
+    const u = new Float32Array(M * D);
+    for (let i = 0; i < M * D; i++)
+        u[i] = silu(convOut[i]);
+    // 6. x_proj → Δ_raw, B, C (COLUMN slices of the (M, R+2N) projection).
+    const xProj = new Float32Array(M * P);
+    linearForward(u, w.wXProj, w.bXProj, M, D, P, xProj);
+    const dtRaw = new Float32Array(M * R);
+    const bRaw = new Float32Array(M * N);
+    const cRaw = new Float32Array(M * N);
+    for (let r = 0; r < M; r++) {
+        for (let i = 0; i < R; i++)
+            dtRaw[r * R + i] = xProj[r * P + i];
+        for (let n = 0; n < N; n++) {
+            bRaw[r * N + n] = xProj[r * P + R + n];
+            cRaw[r * N + n] = xProj[r * P + R + N + n];
+        }
+    }
+    // 7. dt_proj: expand Δ to the inner dimension.
+    const deltaFull = new Float32Array(M * D);
+    linearForward(dtRaw, w.wDtProj, w.bDtProj, M, R, D, deltaFull);
+    // 8. Selective scan (S6). ZOH discretisation, exactly as the kernel does it.
+    const dv = new Float32Array(M * D);
+    for (let i = 0; i < M * D; i++)
+        dv[i] = softplus(deltaFull[i]);
+    const aCont = new Float32Array(D * N);
+    for (let i = 0; i < D * N; i++) {
+        aCont[i] = -Math.exp(Math.min(A_LOG_CLAMP_HI, Math.max(A_LOG_CLAMP_LO, w.A_log[i])));
+    }
+    const aBar = new Float32Array(M * D * N);
+    const bBar = new Float32Array(M * D * N);
+    const h = new Float32Array(M * D * N);
+    const scanY = new Float32Array(M * D);
+    for (let b = 0; b < B; b++) {
+        for (let t = 0; t < L; t++) {
+            const r = b * L + t;
+            for (let d = 0; d < D; d++) {
+                const dvv = dv[r * D + d];
+                const uv = u[r * D + d];
+                let acc = w.D_vec[d] * uv;
+                for (let n = 0; n < N; n++) {
+                    const ac = aCont[d * N + n];
+                    const ab = Math.exp(dvv * ac);
+                    const bb = ((ab - 1) / ac) * bRaw[r * N + n];
+                    const idx = (r * D + d) * N + n;
+                    aBar[idx] = ab;
+                    bBar[idx] = bb;
+                    const hPrev = t === 0 ? 0 : h[((r - 1) * D + d) * N + n];
+                    const hv = ab * hPrev + bb * uv;
+                    h[idx] = hv;
+                    acc += cRaw[r * N + n] * hv;
+                }
+                scanY[r * D + d] = acc;
+            }
+        }
+    }
+    // 9. Gate: y ⊗ SiLU(z).
+    const siluZ = new Float32Array(M * D);
+    const gated = new Float32Array(M * D);
+    for (let i = 0; i < M * D; i++) {
+        siluZ[i] = silu(z[i]);
+        gated[i] = scanY[i] * siluZ[i];
+    }
+    // 10 + 11. Output projection, then the residual add.
+    const output = new Float32Array(M * dModel);
+    linearForward(gated, w.wOutProj, w.bOutProj, M, D, dModel, output);
+    for (let i = 0; i < M * dModel; i++)
+        output[i] = output[i] + x[i];
+    return {
+        output,
+        cache: { x, normInv, normOut, xConvIn, z, convOut, u, dtRaw, bRaw, cRaw, deltaFull, dv, aCont, aBar, bBar, h, scanY, siluZ, gated },
+    };
+}
+/**
+ * Exact CPU backward for one Mamba-1 block.
+ *
+ * @param dOut  dL/d(block output), (M, dModel)
+ * @param grads accumulator — parameter gradients are ADDED into it
+ * @returns dL/d(block input), (M, dModel)
+ */
+export function mamba1CpuBackward(dOut, w, cache, dims, grads) {
+    const { dModel, dState: N, dConv: K, dInner: D, dtRank: R, batch: B, seqLen: L } = dims;
+    const M = B * L;
+    const P = R + 2 * N;
+    // 11. Residual: the block input receives dOut directly, plus whatever flows
+    //     back through the mixer branch (accumulated at the very end).
+    const dx = new Float32Array(M * dModel);
+    dx.set(dOut);
+    // 10. out_proj.
+    const dGated = new Float32Array(M * D);
+    linearBackward(dOut, cache.gated, w.wOutProj, M, D, dModel, dGated, grads.wOutProj, grads.bOutProj);
+    // 9. Gate: gated = scanY * SiLU(z).
+    const dScanY = new Float32Array(M * D);
+    const dz = new Float32Array(M * D);
+    for (let i = 0; i < M * D; i++) {
+        dScanY[i] = dGated[i] * cache.siluZ[i];
+        dz[i] = dGated[i] * cache.scanY[i] * siluGrad(cache.z[i]);
+    }
+    // 8. Selective scan, walked backwards in time.
+    const du = new Float32Array(M * D);
+    const dBRaw = new Float32Array(M * N);
+    const dCRaw = new Float32Array(M * N);
+    const dDv = new Float32Array(M * D); // dL/d softplus(Δ)
+    const dACont = new Float32Array(D * N);
+    // dhNext[d*N+n] carries dL/dh_t propagated from step t+1 for the current batch row.
+    const dhNext = new Float32Array(D * N);
+    for (let b = 0; b < B; b++) {
+        dhNext.fill(0);
+        for (let t = L - 1; t >= 0; t--) {
+            const r = b * L + t;
+            for (let d = 0; d < D; d++) {
+                const dY = dScanY[r * D + d];
+                const uv = cache.u[r * D + d];
+                const dvv = cache.dv[r * D + d];
+                // Skip connection: y_t[d] += D_vec[d] * u_t[d].
+                grads.D_vec[d] = grads.D_vec[d] + dY * uv;
+                du[r * D + d] = du[r * D + d] + dY * w.D_vec[d];
+                let ddvAcc = 0;
+                for (let n = 0; n < N; n++) {
+                    const idx = (r * D + d) * N + n;
+                    const hv = cache.h[idx];
+                    const ab = cache.aBar[idx];
+                    const bb = cache.bBar[idx];
+                    const ac = cache.aCont[d * N + n];
+                    const cv = cache.cRaw[r * N + n];
+                    // y_t[d] = Σ_n C_t[n] * h_t[d,n]
+                    dCRaw[r * N + n] = dCRaw[r * N + n] + dY * hv;
+                    const dh = dY * cv + dhNext[d * N + n];
+                    // h_t = a_bar * h_{t-1} + b_bar * u_t
+                    const hPrev = t === 0 ? 0 : cache.h[((r - 1) * D + d) * N + n];
+                    const dABar = dh * hPrev;
+                    const dBBar = dh * uv;
+                    du[r * D + d] = du[r * D + d] + dh * bb;
+                    dhNext[d * N + n] = dh * ab; // flows to step t-1
+                    // a_bar = exp(dv * a_cont);  b_bar = (a_bar - 1) / a_cont * B_raw
+                    const bv = cache.bRaw[r * N + n];
+                    ddvAcc += dABar * ac * ab + dBBar * ab * bv;
+                    dACont[d * N + n] = dACont[d * N + n]
+                        + dABar * dvv * ab
+                        + dBBar * bv * ((dvv * ab * ac - (ab - 1)) / (ac * ac));
+                    dBRaw[r * N + n] = dBRaw[r * N + n] + dBBar * ((ab - 1) / ac);
+                }
+                dDv[r * D + d] = dDv[r * D + d] + ddvAcc;
+            }
+        }
+    }
+    // a_cont = -exp(clamp(A_log)) ⇒ d a_cont / d A_log = a_cont inside the clamp,
+    // and exactly 0 outside it (the kernel's output is constant there).
+    for (let i = 0; i < D * N; i++) {
+        const al = w.A_log[i];
+        if (al <= A_LOG_CLAMP_LO || al >= A_LOG_CLAMP_HI)
+            continue;
+        grads.A_log[i] = grads.A_log[i] + dACont[i] * cache.aCont[i];
+    }
+    // softplus: dv = softplus(deltaFull)
+    const dDeltaFull = new Float32Array(M * D);
+    for (let i = 0; i < M * D; i++)
+        dDeltaFull[i] = dDv[i] * sigmoid(cache.deltaFull[i]);
+    // 7. dt_proj.
+    const dDtRaw = new Float32Array(M * R);
+    linearBackward(dDeltaFull, cache.dtRaw, w.wDtProj, M, R, D, dDtRaw, grads.wDtProj, grads.bDtProj);
+    // 6. x_proj — re-assemble the (M, P) column slices, then one linear backward.
+    const dXProj = new Float32Array(M * P);
+    for (let r = 0; r < M; r++) {
+        for (let i = 0; i < R; i++)
+            dXProj[r * P + i] = dDtRaw[r * R + i];
+        for (let n = 0; n < N; n++) {
+            dXProj[r * P + R + n] = dBRaw[r * N + n];
+            dXProj[r * P + R + N + n] = dCRaw[r * N + n];
+        }
+    }
+    const duFromXProj = new Float32Array(M * D);
+    linearBackward(dXProj, cache.u, w.wXProj, M, D, P, duFromXProj, grads.wXProj, grads.bXProj);
+    for (let i = 0; i < M * D; i++)
+        du[i] = du[i] + duFromXProj[i];
+    // 5. SiLU.
+    const dConvOut = new Float32Array(M * D);
+    for (let i = 0; i < M * D; i++)
+        dConvOut[i] = du[i] * siluGrad(cache.convOut[i]);
+    // 4. Depthwise causal conv1d.
+    const dXConvIn = new Float32Array(M * D);
+    for (let b = 0; b < B; b++) {
+        for (let t = 0; t < L; t++) {
+            const r = b * L + t;
+            for (let d = 0; d < D; d++) {
+                const g = dConvOut[r * D + d];
+                if (g === 0)
+                    continue;
+                grads.bConv[d] = grads.bConv[d] + g;
+                for (let k = 0; k < K; k++) {
+                    if (t < k)
+                        continue;
+                    const src = (b * L + t - k) * D + d;
+                    grads.wConv[d * K + k] = grads.wConv[d * K + k] + g * cache.xConvIn[src];
+                    dXConvIn[src] = dXConvIn[src] + g * w.wConv[d * K + k];
+                }
+            }
+        }
+    }
+    // 3. Un-split the (x, z) columns.
+    const dInProj = new Float32Array(M * 2 * D);
+    for (let r = 0; r < M; r++) {
+        for (let d = 0; d < D; d++) {
+            dInProj[r * 2 * D + d] = dXConvIn[r * D + d];
+            dInProj[r * 2 * D + D + d] = dz[r * D + d];
+        }
+    }
+    // 2. in_proj.
+    const dNormOut = new Float32Array(M * dModel);
+    linearBackward(dInProj, cache.normOut, w.wInProj, M, dModel, 2 * D, dNormOut, grads.wInProj, grads.bInProj);
+    // 1. RMSNorm — adds into the residual path already sitting in `dx`.
+    rmsNormBackward(dNormOut, cache.x, w.normWeight, cache.normInv, M, dModel, dx, grads.normWeight, RMSNORM_EPS);
+    return dx;
+}
+//# sourceMappingURL=mamba1_cpu.js.map

package/dist/training/mamba1_cpu.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"mamba1_cpu.js","sourceRoot":"","sources":["../../src/training/mamba1_cpu.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;GAyBG;AAEH,OAAO,EACH,aAAa,EACb,cAAc,EACd,cAAc,EACd,eAAe,EACf,IAAI,EACJ,QAAQ,EACR,QAAQ,EACR,OAAO,EACP,WAAW,GACd,MAAM,cAAc,CAAC;AAoBtB,mEAAmE;AACnE,MAAM,CAAC,MAAM,kBAAkB,GAAG;IAC9B,SAAS,EAAE,SAAS,EAAE,OAAO,EAAE,OAAO,EAAE,QAAQ,EAAE,QAAQ;IAC1D,SAAS,EAAE,SAAS,EAAE,OAAO,EAAE,OAAO,EAAE,UAAU,EAAE,UAAU,EAAE,YAAY;CACtB,CAAC;AA2C3D;;;;;GAKG;AACH,MAAM,CAAC,MAAM,cAAc,GAAG,CAAC,EAAE,CAAC;AAClC,MAAM,CAAC,MAAM,cAAc,GAAG,CAAC,CAAC;AAEhC,sDAAsD;AACtD,MAAM,UAAU,eAAe,CAAC,IAAmB;IAC/C,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,GAAG,IAAI,CAAC;IACnE,OAAO;QACH,OAAO,EAAK,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,GAAG,MAAM,CAAC;QAC5C,OAAO,EAAK,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC;QACnC,KAAK,EAAO,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC;QACnC,KAAK,EAAO,IAAI,YAAY,CAAC,CAAC,CAAC;QAC/B,MAAM,EAAM,IAAI,YAAY,CAAC,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC;QAC7C,MAAM,EAAM,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;QACvC,OAAO,EAAK,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC;QACnC,OAAO,EAAK,IAAI,YAAY,CAAC,CAAC,CAAC;QAC/B,KAAK,EAAO,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC;QACnC,KAAK,EAAO,IAAI,YAAY,CAAC,CAAC,CAAC;QAC/B,QAAQ,EAAI,IAAI,YAAY,CAAC,MAAM,GAAG,CAAC,CAAC;QACxC,QAAQ,EAAI,IAAI,YAAY,CAAC,MAAM,CAAC;QACpC,UAAU,EAAE,IAAI,YAAY,CAAC,MAAM,CAAC;KACvC,CAAC;AACN,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,gBAAgB,CAC5B,CAAe,EACf,CAAmB,EACnB,IAAmB;IAEnB,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,GAAG,IAAI,CAAC;IACxF,MAAM,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;IAChB,MAAM,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;IAEpB,wBAAwB;IACxB,MAAM,OAAO,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,MAAM,CAAC,CAAC;IAC7C,MAAM,OAAO,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;IACpC,cAAc,CAAC,CAAC,EAAE,CAAC,CAAC,UAAU,EAAE,CAAC,EAAE,MAAM,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC;IAE7D,gCAAgC;IAChC,MAAM,MAAM,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC;IAC3C,aAAa,CAAC,OAAO,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,GAAG,CAAC,EAAE,MAAM,CAAC,CAAC;IAEvE,4DAA4D;IAC5D,MAAM,OAAO,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACxC,MAAM,CAAC,GAAS,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACxC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QACzB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YACzB,OAAO,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,MAAM,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;YAC5C,CAAC,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAS,MAAM,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;QACpD,CAAC;IACL,CAAC;IAED,qDAAqD;IACrD,MAAM,OAAO,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACxC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QACzB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YACzB,MAAM,IAAI,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC;YAC7B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBACzB,IAAI,GAAG,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,CAAE,CAAC;gBACtB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;oBACzB,IAAI,CAAC,IAAI,CAAC;wBAAE,GAAG,IAAI,CAAC,CAAC,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,GAAG,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;gBAC/E,CAAC;gBACD,OAAO,CAAC,IAAI,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC;YAC5B,CAAC;QACL,CAAC;IACL,CAAC;IAED,WAAW;IACX,MAAM,CAAC,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IAClC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;QAAE,CAAC,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,OAAO,CAAC,CAAC,CAAE,CAAC,CAAC;IAEzD,uEAAuE;IACvE,MAAM,KAAK,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACtC,aAAa,CAAC,CAAC,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,KAAK,CAAC,CAAC;IACrD,MAAM,KAAK,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACtC,MAAM,IAAI,GAAI,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACtC,MAAM,IAAI,GAAI,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACtC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QACzB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;YAAE,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;QACjE,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YACzB,IAAI,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;YACxC,IAAI,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;QAChD,CAAC;IACL,CAAC;IAED,+CAA+C;IAC/C,MAAM,SAAS,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IAC1C,aAAa,CAAC,KAAK,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,SAAS,CAAC,CAAC;IAE/D,6EAA6E;IAC7E,MAAM,EAAE,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACnC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;QAAE,EAAE,CAAC,CAAC,CAAC,GAAG,QAAQ,CAAC,SAAS,CAAC,CAAC,CAAE,CAAC,CAAC;IAEhE,MAAM,KAAK,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACtC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC7B,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,cAAc,EAAE,IAAI,CAAC,GAAG,CAAC,cAAc,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,CAAE,CAAC,CAAC,CAAC,CAAC;IAC1F,CAAC;IAED,MAAM,IAAI,GAAI,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC;IAC1C,MAAM,IAAI,GAAI,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC;IAC1C,MAAM,CAAC,GAAO,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC;IAC1C,MAAM,KAAK,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IAEtC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QACzB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YACzB,MAAM,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;YACpB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBACzB,MAAM,GAAG,GAAG,EAAE,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;gBAC3B,MAAM,EAAE,GAAI,CAAC,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;gBAC1B,IAAI,GAAG,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,CAAE,GAAG,EAAE,CAAC;gBAC3B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;oBACzB,MAAM,EAAE,GAAG,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;oBAC7B,MAAM,EAAE,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,GAAG,EAAE,CAAC,CAAC;oBAC9B,MAAM,EAAE,GAAG,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,GAAG,EAAE,CAAC,GAAG,IAAI,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;oBAC9C,MAAM,GAAG,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;oBAChC,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,CAAC;oBACf,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,CAAC;oBACf,MAAM,KAAK,GAAG,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;oBAC1D,MAAM,EAAE,GAAG,EAAE,GAAG,KAAK,GAAG,EAAE,GAAG,EAAE,CAAC;oBAChC,CAAC,CAAC,GAAG,CAAC,GAAG,EAAE,CAAC;oBACZ,GAAG,IAAI,IAAI,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,GAAG,EAAE,CAAC;gBACjC,CAAC;gBACD,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC;YAC3B,CAAC;QACL,CAAC;IACL,CAAC;IAED,wBAAwB;IACxB,MAAM,KAAK,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACtC,MAAM,KAAK,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACtC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC7B,KAAK,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC,CAAC;QACvB,KAAK,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,CAAE,GAAG,KAAK,CAAC,CAAC,CAAE,CAAC;IACrC,CAAC;IAED,qDAAqD;IACrD,MAAM,MAAM,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,MAAM,CAAC,CAAC;IAC5C,aAAa,CAAC,KAAK,EAAE,CAAC,CAAC,QAAQ,EAAE,CAAC,CAAC,QAAQ,EAAE,CAAC,EAAE,CAAC,EAAE,MAAM,EAAE,MAAM,CAAC,CAAC;IACnE,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,MAAM,EAAE,CAAC,EAAE;QAAE,MAAM,CAAC,CAAC,CAAC,GAAG,MAAM,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC;IAEpE,OAAO;QACH,MAAM;QACN,KAAK,EAAE,EAAE,CAAC,EAAE,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,CAAC,EAAE,OAAO,EAAE,CAAC,EAAE,KAAK,EAAE,IAAI,EAAE,IAAI,EAAE,SAAS,EAAE,EAAE,EAAE,KAAK,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE;KACtI,CAAC;AACN,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,iBAAiB,CAC7B,IAAkB,EAClB,CAAmB,EACnB,KAAqB,EACrB,IAAmB,EACnB,KAAqB;IAErB,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,GAAG,IAAI,CAAC;IACxF,MAAM,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;IAChB,MAAM,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;IAEpB,4EAA4E;IAC5E,mEAAmE;IACnE,MAAM,EAAE,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,MAAM,CAAC,CAAC;IACxC,EAAE,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;IAEb,gBAAgB;IAChB,MAAM,MAAM,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACvC,cAAc,CAAC,IAAI,EAAE,KAAK,CAAC,KAAK,EAAE,CAAC,CAAC,QAAQ,EAAE,CAAC,EAAE,CAAC,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,CAAC,QAAQ,EAAE,KAAK,CAAC,QAAQ,CAAC,CAAC;IAEpG,oCAAoC;IACpC,MAAM,MAAM,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACvC,MAAM,EAAE,GAAO,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACvC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC7B,MAAM,CAAC,CAAC,CAAC,GAAG,MAAM,CAAC,CAAC,CAAE,GAAG,KAAK,CAAC,KAAK,CAAC,CAAC,CAAE,CAAC;QACzC,EAAE,CAAC,CAAC,CAAC,GAAO,MAAM,CAAC,CAAC,CAAE,GAAG,KAAK,CAAC,KAAK,CAAC,CAAC,CAAE,GAAG,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC,CAAC;IACrE,CAAC;IAED,+CAA+C;IAC/C,MAAM,EAAE,GAAM,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACtC,MAAM,KAAK,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACtC,MAAM,KAAK,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACtC,MAAM,GAAG,GAAK,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAM,mBAAmB;IAC/D,MAAM,MAAM,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IAEvC,oFAAoF;IACpF,MAAM,MAAM,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACvC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QACzB,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QACf,KAAK,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC9B,MAAM,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;YACpB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBACzB,MAAM,EAAE,GAAI,MAAM,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;gBAC/B,MAAM,EAAE,GAAI,KAAK,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;gBAChC,MAAM,GAAG,GAAG,KAAK,CAAC,EAAE,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;gBACjC,gDAAgD;gBAChD,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC,KAAK,CAAC,CAAC,CAAE,GAAG,EAAE,GAAG,EAAE,CAAC;gBAC3C,EAAE,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,GAAG,EAAE,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,CAAE,CAAC;gBAElD,IAAI,MAAM,GAAG,CAAC,CAAC;gBACf,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;oBACzB,MAAM,GAAG,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;oBAChC,MAAM,EAAE,GAAI,KAAK,CAAC,CAAC,CAAC,GAAG,CAAE,CAAC;oBAC1B,MAAM,EAAE,GAAI,KAAK,CAAC,IAAI,CAAC,GAAG,CAAE,CAAC;oBAC7B,MAAM,EAAE,GAAI,KAAK,CAAC,IAAI,CAAC,GAAG,CAAE,CAAC;oBAC7B,MAAM,EAAE,GAAI,KAAK,CAAC,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;oBACpC,MAAM,EAAE,GAAI,KAAK,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;oBAEnC,iCAAiC;oBACjC,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,GAAG,EAAE,GAAG,EAAE,CAAC;oBAC/C,MAAM,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,MAAM,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;oBAExC,sCAAsC;oBACtC,MAAM,KAAK,GAAG,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;oBAChE,MAAM,KAAK,GAAG,EAAE,GAAG,KAAK,CAAC;oBACzB,MAAM,KAAK,GAAG,EAAE,GAAG,EAAE,CAAC;oBACtB,EAAE,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,GAAG,EAAE,GAAG,EAAE,CAAC;oBACzC,MAAM,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,GAAG,EAAE,CAAC,CAAG,oBAAoB;oBAEnD,kEAAkE;oBAClE,MAAM,EAAE,GAAG,KAAK,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;oBAClC,MAAM,IAAI,KAAK,GAAG,EAAE,GAAG,EAAE,GAAG,KAAK,GAAG,EAAE,GAAG,EAAE,CAAC;oBAC5C,MAAM,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,MAAM,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE;0BAChC,KAAK,GAAG,GAAG,GAAG,EAAE;0BAChB,KAAK,GAAG,EAAE,GAAG,CAAC,CAAC,GAAG,GAAG,EAAE,GAAG,EAAE,GAAG,CAAC,EAAE,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,EAAE,GAAG,EAAE,CAAC,CAAC,CAAC;oBAC5D,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,GAAG,KAAK,GAAG,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,GAAG,EAAE,CAAC,CAAC;gBACnE,CAAC;gBACD,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,GAAG,MAAM,CAAC;YAC9C,CAAC;QACL,CAAC;IACL,CAAC;IAED,8EAA8E;IAC9E,oEAAoE;IACpE,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC7B,MAAM,EAAE,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,CAAE,CAAC;QACvB,IAAI,EAAE,IAAI,cAAc,IAAI,EAAE,IAAI,cAAc;YAAE,SAAS;QAC3D,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC,KAAK,CAAC,CAAC,CAAE,GAAG,MAAM,CAAC,CAAC,CAAE,GAAG,KAAK,CAAC,KAAK,CAAC,CAAC,CAAE,CAAC;IACpE,CAAC;IAED,qCAAqC;IACrC,MAAM,UAAU,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IAC3C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;QAAE,UAAU,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,CAAE,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,CAAE,CAAC,CAAC;IAEvF,cAAc;IACd,MAAM,MAAM,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACvC,cAAc,CAAC,UAAU,EAAE,KAAK,CAAC,KAAK,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,MAAM,EAAE,KAAK,CAAC,OAAO,EAAE,KAAK,CAAC,OAAO,CAAC,CAAC;IAElG,8EAA8E;IAC9E,MAAM,MAAM,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACvC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QACzB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;YAAE,MAAM,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,MAAM,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;QACnE,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YACzB,MAAM,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAO,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;YAC9C,MAAM,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;QAClD,CAAC;IACL,CAAC;IACD,MAAM,WAAW,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IAC5C,cAAc,CAAC,MAAM,EAAE,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,WAAW,EAAE,KAAK,CAAC,MAAM,EAAE,KAAK,CAAC,MAAM,CAAC,CAAC;IAC5F,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;QAAE,EAAE,CAAC,CAAC,CAAC,GAAG,EAAE,CAAC,CAAC,CAAE,GAAG,WAAW,CAAC,CAAC,CAAE,CAAC;IAEjE,WAAW;IACX,MAAM,QAAQ,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACzC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;QAAE,QAAQ,CAAC,CAAC,CAAC,GAAG,EAAE,CAAC,CAAC,CAAE,GAAG,QAAQ,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAE,CAAC,CAAC;IAEnF,8BAA8B;IAC9B,MAAM,QAAQ,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACzC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QACzB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YACzB,MAAM,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;YACpB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBACzB,MAAM,CAAC,GAAG,QAAQ,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;gBAC/B,IAAI,CAAC,KAAK,CAAC;oBAAE,SAAS;gBACtB,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC,KAAK,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC;gBACrC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;oBACzB,IAAI,CAAC,GAAG,CAAC;wBAAE,SAAS;oBACpB,MAAM,GAAG,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;oBACpC,KAAK,CAAC,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,GAAG,CAAC,GAAG,KAAK,CAAC,OAAO,CAAC,GAAG,CAAE,CAAC;oBAC3E,QAAQ,CAAC,GAAG,CAAC,GAAG,QAAQ,CAAC,GAAG,CAAE,GAAG,CAAC,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;gBAC7D,CAAC;YACL,CAAC;QACL,CAAC;IACL,CAAC;IAED,kCAAkC;IAClC,MAAM,OAAO,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC;IAC5C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QACzB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YACzB,OAAO,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAO,QAAQ,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;YAClD,OAAO,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAE,CAAC;QAChD,CAAC;IACL,CAAC;IAED,cAAc;IACd,MAAM,QAAQ,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,MAAM,CAAC,CAAC;IAC9C,cAAc,CAAC,OAAO,EAAE,KAAK,CAAC,OAAO,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,GAAG,CAAC,EAAE,QAAQ,EAAE,KAAK,CAAC,OAAO,EAAE,KAAK,CAAC,OAAO,CAAC,CAAC;IAE5G,oEAAoE;IACpE,eAAe,CAAC,QAAQ,EAAE,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,UAAU,EAAE,KAAK,CAAC,OAAO,EAAE,CAAC,EAAE,MAAM,EAAE,EAAE,EAAE,KAAK,CAAC,UAAU,EAAE,WAAW,CAAC,CAAC;IAE9G,OAAO,EAAE,CAAC;AACd,CAAC"}

package/dist/training/mixed_precision.d.ts

@@ -0,0 +1,67 @@
+/**
+ * mixed_precision.ts — mixed-precision training discipline (AMP) for the CPU
+ * reference, faithful to the cookbook's recipe:
+ *
+ *   • fp16 compute, fp32 MASTER weights. The model's Float32Array params ARE the
+ *     master copy; the forward runs on an fp16-rounded view, so tiny weight
+ *     updates still accumulate in fp32 instead of vanishing.
+ *   • Dynamic LOSS SCALING. fp16 gradients underflow to zero below ~6e-5; scaling
+ *     the loss by a large factor S lifts them into fp16 range, then we unscale by
+ *     S before the optimiser step. On overflow (inf/nan) we skip the step and back
+ *     the scale off; after a clean streak we grow it. This is the mechanism that
+ *     makes fp16 training numerically stable.
+ *
+ * The payoff is memory/throughput: activations and gradients at half precision.
+ * On CPU we can't reclaim the bytes, so this module models the NUMERICS exactly
+ * (fp16 rounding + loss scaling) — the property the WGSL path will inherit — and
+ * its tests prove a small gradient survives WITH scaling and underflows WITHOUT.
+ */
+/** Round a value to what fp16 storage would preserve (half-precision simulation). */
+export declare function roundFp16(x: number): number;
+/** Round every element of a copy to fp16 precision (leaves the input untouched). */
+export declare function fp16View(a: Float32Array): Float32Array;
+export interface LossScalerOptions {
+    /** Initial scale. Default 2^14 = 16384. */
+    initScale?: number;
+    /** Multiply the scale by this after `growthInterval` clean steps. Default 2. */
+    growthFactor?: number;
+    /** Multiply the scale by this on overflow. Default 0.5. */
+    backoffFactor?: number;
+    /** Clean steps between growth attempts. Default 200. */
+    growthInterval?: number;
+    /** Floor for the scale so it can't collapse to zero. Default 1. */
+    minScale?: number;
+}
+/**
+ * Dynamic loss scaler — the standard AMP controller. Scale the loss/grads up
+ * before backward, {@link check} them for overflow, {@link unscale} the good ones,
+ * and {@link update} the scale based on whether this step overflowed.
+ */
+export declare class DynamicLossScaler {
+    private _scale;
+    private readonly growthFactor;
+    private readonly backoffFactor;
+    private readonly growthInterval;
+    private readonly minScale;
+    private cleanStreak;
+    private _overflows;
+    private _skipped;
+    constructor(opts?: LossScalerOptions);
+    get scale(): number;
+    get overflows(): number;
+    get skipped(): number;
+    /** True if any gradient is inf/nan (an fp16 overflow) — such a step must be skipped. */
+    check(grads: {
+        data: Float32Array;
+    }[]): boolean;
+    /** Divide gradients back down by the current scale, in place. */
+    unscale(grads: {
+        data: Float32Array;
+    }[]): void;
+    /**
+     * Advance the scale controller. Pass whether this step overflowed. Returns
+     * `true` if the optimiser step should PROCEED, `false` if it must be skipped.
+     */
+    update(overflow: boolean): boolean;
+}
+//# sourceMappingURL=mixed_precision.d.ts.map

package/dist/training/mixed_precision.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"mixed_precision.d.ts","sourceRoot":"","sources":["../../src/training/mixed_precision.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAIH,qFAAqF;AACrF,wBAAgB,SAAS,CAAC,CAAC,EAAE,MAAM,GAAG,MAAM,CAE3C;AAED,oFAAoF;AACpF,wBAAgB,QAAQ,CAAC,CAAC,EAAE,YAAY,GAAG,YAAY,CAItD;AAED,MAAM,WAAW,iBAAiB;IAChC,2CAA2C;IAC3C,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,gFAAgF;IAChF,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,2DAA2D;IAC3D,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,wDAAwD;IACxD,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,mEAAmE;IACnE,QAAQ,CAAC,EAAE,MAAM,CAAC;CACnB;AAED;;;;GAIG;AACH,qBAAa,iBAAiB;IAC5B,OAAO,CAAC,MAAM,CAAS;IACvB,OAAO,CAAC,QAAQ,CAAC,YAAY,CAAS;IACtC,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAS;IACvC,OAAO,CAAC,QAAQ,CAAC,cAAc,CAAS;IACxC,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAS;IAClC,OAAO,CAAC,WAAW,CAAK;IACxB,OAAO,CAAC,UAAU,CAAK;IACvB,OAAO,CAAC,QAAQ,CAAK;gBAET,IAAI,GAAE,iBAAsB;IAQxC,IAAI,KAAK,IAAI,MAAM,CAElB;IACD,IAAI,SAAS,IAAI,MAAM,CAEtB;IACD,IAAI,OAAO,IAAI,MAAM,CAEpB;IAED,wFAAwF;IACxF,KAAK,CAAC,KAAK,EAAE;QAAE,IAAI,EAAE,YAAY,CAAA;KAAE,EAAE,GAAG,OAAO;IAW/C,iEAAiE;IACjE,OAAO,CAAC,KAAK,EAAE;QAAE,IAAI,EAAE,YAAY,CAAA;KAAE,EAAE,GAAG,IAAI;IAQ9C;;;OAGG;IACH,MAAM,CAAC,QAAQ,EAAE,OAAO,GAAG,OAAO;CAcnC"}

package/dist/training/mixed_precision.js

@@ -0,0 +1,101 @@
+/**
+ * mixed_precision.ts — mixed-precision training discipline (AMP) for the CPU
+ * reference, faithful to the cookbook's recipe:
+ *
+ *   • fp16 compute, fp32 MASTER weights. The model's Float32Array params ARE the
+ *     master copy; the forward runs on an fp16-rounded view, so tiny weight
+ *     updates still accumulate in fp32 instead of vanishing.
+ *   • Dynamic LOSS SCALING. fp16 gradients underflow to zero below ~6e-5; scaling
+ *     the loss by a large factor S lifts them into fp16 range, then we unscale by
+ *     S before the optimiser step. On overflow (inf/nan) we skip the step and back
+ *     the scale off; after a clean streak we grow it. This is the mechanism that
+ *     makes fp16 training numerically stable.
+ *
+ * The payoff is memory/throughput: activations and gradients at half precision.
+ * On CPU we can't reclaim the bytes, so this module models the NUMERICS exactly
+ * (fp16 rounding + loss scaling) — the property the WGSL path will inherit — and
+ * its tests prove a small gradient survives WITH scaling and underflows WITHOUT.
+ */
+import { floatToFp16, fp16ToFloat } from "../utils/quantization.js";
+/** Round a value to what fp16 storage would preserve (half-precision simulation). */
+export function roundFp16(x) {
+    return fp16ToFloat(floatToFp16(x));
+}
+/** Round every element of a copy to fp16 precision (leaves the input untouched). */
+export function fp16View(a) {
+    const out = new Float32Array(a.length);
+    for (let i = 0; i < a.length; i++)
+        out[i] = roundFp16(a[i]);
+    return out;
+}
+/**
+ * Dynamic loss scaler — the standard AMP controller. Scale the loss/grads up
+ * before backward, {@link check} them for overflow, {@link unscale} the good ones,
+ * and {@link update} the scale based on whether this step overflowed.
+ */
+export class DynamicLossScaler {
+    _scale;
+    growthFactor;
+    backoffFactor;
+    growthInterval;
+    minScale;
+    cleanStreak = 0;
+    _overflows = 0;
+    _skipped = 0;
+    constructor(opts = {}) {
+        this._scale = opts.initScale ?? 16384;
+        this.growthFactor = opts.growthFactor ?? 2;
+        this.backoffFactor = opts.backoffFactor ?? 0.5;
+        this.growthInterval = opts.growthInterval ?? 200;
+        this.minScale = opts.minScale ?? 1;
+    }
+    get scale() {
+        return this._scale;
+    }
+    get overflows() {
+        return this._overflows;
+    }
+    get skipped() {
+        return this._skipped;
+    }
+    /** True if any gradient is inf/nan (an fp16 overflow) — such a step must be skipped. */
+    check(grads) {
+        for (const g of grads) {
+            const a = g.data;
+            for (let i = 0; i < a.length; i++) {
+                const x = a[i];
+                if (!Number.isFinite(x))
+                    return true;
+            }
+        }
+        return false;
+    }
+    /** Divide gradients back down by the current scale, in place. */
+    unscale(grads) {
+        const inv = 1 / this._scale;
+        for (const g of grads) {
+            const a = g.data;
+            for (let i = 0; i < a.length; i++)
+                a[i] = a[i] * inv;
+        }
+    }
+    /**
+     * Advance the scale controller. Pass whether this step overflowed. Returns
+     * `true` if the optimiser step should PROCEED, `false` if it must be skipped.
+     */
+    update(overflow) {
+        if (overflow) {
+            this._overflows++;
+            this._skipped++;
+            this._scale = Math.max(this.minScale, this._scale * this.backoffFactor);
+            this.cleanStreak = 0;
+            return false;
+        }
+        if (++this.cleanStreak >= this.growthInterval) {
+            this._scale *= this.growthFactor;
+            this.cleanStreak = 0;
+        }
+        return true;
+    }
+}
+//# sourceMappingURL=mixed_precision.js.map

package/dist/training/mixed_precision.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"mixed_precision.js","sourceRoot":"","sources":["../../src/training/mixed_precision.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAEH,OAAO,EAAE,WAAW,EAAE,WAAW,EAAE,MAAM,0BAA0B,CAAC;AAEpE,qFAAqF;AACrF,MAAM,UAAU,SAAS,CAAC,CAAS;IACjC,OAAO,WAAW,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC;AACrC,CAAC;AAED,oFAAoF;AACpF,MAAM,UAAU,QAAQ,CAAC,CAAe;IACtC,MAAM,GAAG,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC;IACvC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE;QAAE,GAAG,CAAC,CAAC,CAAC,GAAG,SAAS,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC,CAAC;IAC7D,OAAO,GAAG,CAAC;AACb,CAAC;AAeD;;;;GAIG;AACH,MAAM,OAAO,iBAAiB;IACpB,MAAM,CAAS;IACN,YAAY,CAAS;IACrB,aAAa,CAAS;IACtB,cAAc,CAAS;IACvB,QAAQ,CAAS;IAC1B,WAAW,GAAG,CAAC,CAAC;IAChB,UAAU,GAAG,CAAC,CAAC;IACf,QAAQ,GAAG,CAAC,CAAC;IAErB,YAAY,OAA0B,EAAE;QACtC,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC,SAAS,IAAI,KAAK,CAAC;QACtC,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC,YAAY,IAAI,CAAC,CAAC;QAC3C,IAAI,CAAC,aAAa,GAAG,IAAI,CAAC,aAAa,IAAI,GAAG,CAAC;QAC/C,IAAI,CAAC,cAAc,GAAG,IAAI,CAAC,cAAc,IAAI,GAAG,CAAC;QACjD,IAAI,CAAC,QAAQ,GAAG,IAAI,CAAC,QAAQ,IAAI,CAAC,CAAC;IACrC,CAAC;IAED,IAAI,KAAK;QACP,OAAO,IAAI,CAAC,MAAM,CAAC;IACrB,CAAC;IACD,IAAI,SAAS;QACX,OAAO,IAAI,CAAC,UAAU,CAAC;IACzB,CAAC;IACD,IAAI,OAAO;QACT,OAAO,IAAI,CAAC,QAAQ,CAAC;IACvB,CAAC;IAED,wFAAwF;IACxF,KAAK,CAAC,KAA+B;QACnC,KAAK,MAAM,CAAC,IAAI,KAAK,EAAE,CAAC;YACtB,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC;YACjB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;gBAClC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC;gBAChB,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC;oBAAE,OAAO,IAAI,CAAC;YACvC,CAAC;QACH,CAAC;QACD,OAAO,KAAK,CAAC;IACf,CAAC;IAED,iEAAiE;IACjE,OAAO,CAAC,KAA+B;QACrC,MAAM,GAAG,GAAG,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC;QAC5B,KAAK,MAAM,CAAC,IAAI,KAAK,EAAE,CAAC;YACtB,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC;YACjB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE;gBAAE,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAE,GAAG,GAAG,CAAC;QACxD,CAAC;IACH,CAAC;IAED;;;OAGG;IACH,MAAM,CAAC,QAAiB;QACtB,IAAI,QAAQ,EAAE,CAAC;YACb,IAAI,CAAC,UAAU,EAAE,CAAC;YAClB,IAAI,CAAC,QAAQ,EAAE,CAAC;YAChB,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,QAAQ,EAAE,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC,aAAa,CAAC,CAAC;YACxE,IAAI,CAAC,WAAW,GAAG,CAAC,CAAC;YACrB,OAAO,KAAK,CAAC;QACf,CAAC;QACD,IAAI,EAAE,IAAI,CAAC,WAAW,IAAI,IAAI,CAAC,cAAc,EAAE,CAAC;YAC9C,IAAI,CAAC,MAAM,IAAI,IAAI,CAAC,YAAY,CAAC;YACjC,IAAI,CAAC,WAAW,GAAG,CAAC,CAAC;QACvB,CAAC;QACD,OAAO,IAAI,CAAC;IACd,CAAC;CACF"}

package/dist/training/model_cpu.d.ts

@@ -0,0 +1,100 @@
+/**
+ * model_cpu.ts — the exact CPU forward + backward for a whole
+ * {@link HybridMambaModel}: token embedding → layer stack → final RMSNorm →
+ * tied LM head → cross-entropy.
+ *
+ * This is the gradient engine {@link MambaTrainer} uses. Before it existed the
+ * trainer handed the SAME `dLogits` buffer to AdamW for every parameter, so no
+ * parameter ever received its own gradient — `learn()` was safe but inert. Now
+ * the backward pass is real, and `tests/backprop.test.ts` proves it against
+ * finite differences without needing a GPU.
+ *
+ * Parameter names match {@link HybridMambaModel.parameters} exactly
+ * (`embedding`, `layer{i}.{name}`, `final_norm`), which is how the trainer
+ * pairs each parameter with its OWN gradient.
+ */
+import { type Mamba1CpuWeights, type Mamba1CpuGrads } from './mamba1_cpu.js';
+/** Layer types the CPU gradient engine can differentiate today. */
+export declare const CPU_GRADIENT_LAYER_TYPES: readonly ["mamba1"];
+export type CpuGradientLayerType = (typeof CPU_GRADIENT_LAYER_TYPES)[number];
+export interface CpuModelDims {
+    vocabSize: number;
+    dModel: number;
+    dState: number;
+    dConv: number;
+    dInner: number;
+    dtRank: number;
+    numLayers: number;
+}
+/** Every tensor the CPU model reads, mirrored out of the GPU buffers. */
+export interface CpuModelWeights {
+    embedding: Float32Array;
+    finalNorm: Float32Array;
+    lmHeadBias: Float32Array;
+    layers: Mamba1CpuWeights[];
+}
+export interface CpuModelGrads {
+    embedding: Float32Array;
+    finalNorm: Float32Array;
+    lmHeadBias: Float32Array;
+    layers: Mamba1CpuGrads[];
+}
+/** Knobs on one backward pass. */
+export interface CpuBackwardOptions {
+    /**
+     * Recompute each layer's activations during the backward pass instead of
+     * retaining every layer's cache from the forward. Trades one extra forward per
+     * layer for O(1) instead of O(layers) retained activations; the gradients are
+     * numerically identical (`tests/peft_trainer.test.ts` asserts bit equality).
+     */
+    activationCheckpointing?: boolean;
+}
+/** Loss + the gradient of every parameter, keyed by `parameters()` name. */
+export interface CpuBackwardResult {
+    loss: number;
+    grads: Map<string, Float32Array>;
+}
+export declare function zeroCpuModelGrads(dims: CpuModelDims): CpuModelGrads;
+/** Flatten a {@link CpuModelGrads} into the `parameters()` naming the optimizer uses. */
+export declare function toNamedGrads(g: CpuModelGrads): Map<string, Float32Array>;
+/** The model's tensors keyed the same way — the inverse of {@link cpuWeightsFromNamed}. */
+export declare function namedWeights(w: CpuModelWeights): Map<string, Float32Array>;
+/**
+ * Assemble a {@link CpuModelWeights} from tensors keyed by `parameters()` name —
+ * how {@link MambaTrainer} mirrors the GPU buffers into the gradient engine.
+ * Throws on a missing or wrongly-sized tensor rather than training on garbage.
+ */
+export declare function cpuWeightsFromNamed(named: Map<string, Float32Array>, dims: CpuModelDims): CpuModelWeights;
+/** Logits only — the CPU mirror of `HybridMambaModel.forward`. */
+export declare function cpuModelForward(tokenIds: ArrayLike<number>, w: CpuModelWeights, dims: CpuModelDims, batch: number, seqLen: number): Float32Array;
+/** Mean next-token cross-entropy over the sequence — the trainer's loss. */
+export declare function cpuModelLoss(tokenIds: ArrayLike<number>, targets: ArrayLike<number>, w: CpuModelWeights, dims: CpuModelDims, batch: number, seqLen: number): number;
+/**
+ * Forward + backward. Returns the mean cross-entropy loss and the gradient of
+ * EVERY parameter, keyed by its `parameters()` name.
+ *
+ * The loss is averaged over the `batch * seqLen` predicted positions, so the
+ * gradients are the gradients of that mean — the quantity the trainer reports.
+ */
+export declare function cpuModelBackward(tokenIds: ArrayLike<number>, targets: ArrayLike<number>, w: CpuModelWeights, dims: CpuModelDims, batch: number, seqLen: number, opts?: CpuBackwardOptions): CpuBackwardResult;
+/**
+ * The 2-D shape of a parameter, or `null` when it is a vector (bias / norm gain
+ * / `D_vec`).
+ *
+ * Low-rank adaptation only means anything over a matrix, so this is the single
+ * place that says which parameters an adapter can cover — and it derives the
+ * shapes from `dims` rather than restating the block's layout, so it cannot
+ * drift from {@link cpuWeightsFromNamed}.
+ */
+export declare function paramMatrixShape(name: string, dims: CpuModelDims): [rows: number, cols: number] | null;
+/**
+ * Parameters a low-rank adapter covers by default: the four projections inside
+ * each block.
+ *
+ * Deliberately NOT the token embedding (by far the largest tensor, and tied to
+ * the output head — adapting it low-rank changes the head too) and NOT `A_log`
+ * (the state-transition decay, which WSLA freezes precisely because drifting it
+ * destabilises the SSM).
+ */
+export declare const DEFAULT_LORA_TARGETS: readonly ["wInProj", "wXProj", "wDtProj", "wOutProj"];
+//# sourceMappingURL=model_cpu.d.ts.map

package/dist/training/model_cpu.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"model_cpu.d.ts","sourceRoot":"","sources":["../../src/training/model_cpu.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAQH,OAAO,EAKH,KAAK,gBAAgB,EAGrB,KAAK,cAAc,EACtB,MAAM,iBAAiB,CAAC;AAGzB,mEAAmE;AACnE,eAAO,MAAM,wBAAwB,qBAAsB,CAAC;AAC5D,MAAM,MAAM,oBAAoB,GAAG,CAAC,OAAO,wBAAwB,CAAC,CAAC,MAAM,CAAC,CAAC;AAE7E,MAAM,WAAW,YAAY;IACzB,SAAS,EAAE,MAAM,CAAC;IAClB,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,SAAS,EAAE,MAAM,CAAC;CACrB;AAED,yEAAyE;AACzE,MAAM,WAAW,eAAe;IAC5B,SAAS,EAAI,YAAY,CAAC;IAC1B,SAAS,EAAI,YAAY,CAAC;IAC1B,UAAU,EAAG,YAAY,CAAC;IAC1B,MAAM,EAAO,gBAAgB,EAAE,CAAC;CACnC;AAED,MAAM,WAAW,aAAa;IAC1B,SAAS,EAAI,YAAY,CAAC;IAC1B,SAAS,EAAI,YAAY,CAAC;IAC1B,UAAU,EAAG,YAAY,CAAC;IAC1B,MAAM,EAAO,cAAc,EAAE,CAAC;CACjC;AAED,kCAAkC;AAClC,MAAM,WAAW,kBAAkB;IAC/B;;;;;OAKG;IACH,uBAAuB,CAAC,EAAE,OAAO,CAAC;CACrC;AAED,4EAA4E;AAC5E,MAAM,WAAW,iBAAiB;IAC9B,IAAI,EAAG,MAAM,CAAC;IACd,KAAK,EAAE,GAAG,CAAC,MAAM,EAAE,YAAY,CAAC,CAAC;CACpC;AAED,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,YAAY,GAAG,aAAa,CAWnE;AAED,yFAAyF;AACzF,wBAAgB,YAAY,CAAC,CAAC,EAAE,aAAa,GAAG,GAAG,CAAC,MAAM,EAAE,YAAY,CAAC,CAUxE;AAED,2FAA2F;AAC3F,wBAAgB,YAAY,CAAC,CAAC,EAAE,eAAe,GAAG,GAAG,CAAC,MAAM,EAAE,YAAY,CAAC,CAU1E;AAED;;;;GAIG;AACH,wBAAgB,mBAAmB,CAAC,KAAK,EAAE,GAAG,CAAC,MAAM,EAAE,YAAY,CAAC,EAAE,IAAI,EAAE,YAAY,GAAG,eAAe,CA2BzG;AA6DD,kEAAkE;AAClE,wBAAgB,eAAe,CAC3B,QAAQ,EAAE,SAAS,CAAC,MAAM,CAAC,EAC3B,CAAC,EAAE,eAAe,EAClB,IAAI,EAAE,YAAY,EAClB,KAAK,EAAE,MAAM,EACb,MAAM,EAAE,MAAM,GACf,YAAY,CAEd;AAED,4EAA4E;AAC5E,wBAAgB,YAAY,CACxB,QAAQ,EAAE,SAAS,CAAC,MAAM,CAAC,EAC3B,OAAO,EAAE,SAAS,CAAC,MAAM,CAAC,EAC1B,CAAC,EAAE,eAAe,EAClB,IAAI,EAAE,YAAY,EAClB,KAAK,EAAE,MAAM,EACb,MAAM,EAAE,MAAM,GACf,MAAM,CASR;AAED;;;;;;GAMG;AACH,wBAAgB,gBAAgB,CAC5B,QAAQ,EAAE,SAAS,CAAC,MAAM,CAAC,EAC3B,OAAO,EAAE,SAAS,CAAC,MAAM,CAAC,EAC1B,CAAC,EAAE,eAAe,EAClB,IAAI,EAAE,YAAY,EAClB,KAAK,EAAE,MAAM,EACb,MAAM,EAAE,MAAM,EACd,IAAI,GAAE,kBAAuB,GAC9B,iBAAiB,CAiEnB;AAID;;;;;;;;GAQG;AACH,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,YAAY,GAAG,CAAC,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,CAAC,GAAG,IAAI,CAatG;AAED;;;;;;;;GAQG;AACH,eAAO,MAAM,oBAAoB,uDAAwD,CAAC"}