@seanhogg/builderforce-memory-engine 2026.7.11 → 2026.7.14

package/dist/optim/adamw.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"adamw.d.ts","sourceRoot":"","sources":["../../src/optim/adamw.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,MAAM,WAAW,UAAU;IACzB,IAAI,EAAE,YAAY,CAAC;CACpB;AAED,MAAM,WAAW,WAAW;IAC1B,UAAU,IAAI,UAAU,EAAE,CAAC;IAC3B,SAAS,IAAI,UAAU,EAAE,CAAC;CAC3B;AAED,MAAM,WAAW,YAAY;IAC3B,EAAE,CAAC,EAAE,MAAM,CAAC;IACZ,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,GAAG,CAAC,EAAE,MAAM,CAAC;IACb,WAAW,CAAC,EAAE,MAAM,CAAC;CACtB;AAED,qBAAa,KAAK;IAOd,OAAO,CAAC,QAAQ,CAAC,MAAM;IANzB,OAAO,CAAC,QAAQ,CAAC,CAAC,CAAsB;IACxC,OAAO,CAAC,QAAQ,CAAC,CAAC,CAAsB;IACxC,OAAO,CAAC,CAAC,CAAK;IACd,OAAO,CAAC,QAAQ,CAAC,GAAG,CAAyB;gBAG1B,MAAM,EAAE,WAAW,EACpC,OAAO,GAAE,YAAiB;IAe5B,mEAAmE;IACnE,IAAI,IAAI,IAAI;CAsBb"}
+{"version":3,"file":"adamw.d.ts","sourceRoot":"","sources":["../../src/optim/adamw.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,MAAM,WAAW,UAAU;IACzB,IAAI,EAAE,YAAY,CAAC;CACpB;AAED,MAAM,WAAW,WAAW;IAC1B,UAAU,IAAI,UAAU,EAAE,CAAC;IAC3B,SAAS,IAAI,UAAU,EAAE,CAAC;CAC3B;AAED;;;;;;;;;;;GAWG;AACH,MAAM,WAAW,SAAS;IACxB,kCAAkC;IAClC,KAAK,EAAE,MAAM,CAAC;IACd,8BAA8B;IAC9B,KAAK,EAAE,MAAM,CAAC;CACf;AAED,MAAM,WAAW,YAAY;IAC3B,EAAE,CAAC,EAAE,MAAM,CAAC;IACZ,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,GAAG,CAAC,EAAE,MAAM,CAAC;IACb,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,4EAA4E;IAC5E,KAAK,CAAC,EAAE,SAAS,CAAC;CACnB;AAED,qBAAa,KAAK;IASd,OAAO,CAAC,QAAQ,CAAC,MAAM;IAPzB,OAAO,CAAC,QAAQ,CAAC,CAAC,CAA+B;IACjD,OAAO,CAAC,QAAQ,CAAC,CAAC,CAA+B;IACjD,OAAO,CAAC,CAAC,CAAK;IACd,OAAO,CAAC,QAAQ,CAAC,GAAG,CAAwC;IAC5D,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAmB;gBAGtB,MAAM,EAAE,WAAW,EACpC,OAAO,GAAE,YAAiB;IA0B5B,kEAAkE;IAClE,OAAO,CAAC,KAAK;IAIb,wFAAwF;IACxF,IAAI,IAAI,IAAI;IAwBZ,qFAAqF;IACrF,UAAU,IAAI,MAAM;CAMrB"}

package/dist/optim/adamw.js

@@ -7,10 +7,12 @@
  */
 export class AdamW {
     target;
+    // Per-parameter moments, or null for tensors this shard does not own (no memory).
     m = [];
     v = [];
     t = 0;
     opt;
+    shard;
     constructor(target, options = {}) {
         this.target = target;
         this.opt = {
@@ -20,12 +22,28 @@ export class AdamW {
             eps: options.eps ?? 1e-8,
             weightDecay: options.weightDecay ?? 0,
         };
-        for (const p of target.parameters()) {
-            this.m.push(new Float32Array(p.data.length));
-            this.v.push(new Float32Array(p.data.length));
+        const shard = options.shard ?? null;
+        if (shard && (shard.count <= 0 || shard.index < 0 || shard.index >= shard.count)) {
+            throw new Error(`AdamW: invalid shard ${shard.index}/${shard.count}`);
         }
+        this.shard = shard;
+        const params = target.parameters();
+        for (let p = 0; p < params.length; p++) {
+            if (this._owns(p)) {
+                this.m.push(new Float32Array(params[p].data.length));
+                this.v.push(new Float32Array(params[p].data.length));
+            }
+            else {
+                this.m.push(null);
+                this.v.push(null);
+            }
+        }
+    }
+    /** Whether this shard owns (and updates) parameter tensor `p`. */
+    _owns(p) {
+        return this.shard === null || p % this.shard.count === this.shard.index;
     }
-    /** One optimiser step from the currently-accumulated gradients. */
+    /** One optimiser step from the currently-accumulated gradients (owned tensors only). */
     step() {
         this.t++;
         const { lr, beta1, beta2, eps, weightDecay } = this.opt;
@@ -34,6 +52,8 @@ export class AdamW {
         const bc1 = 1 - Math.pow(beta1, this.t);
         const bc2 = 1 - Math.pow(beta2, this.t);
         for (let p = 0; p < params.length; p++) {
+            if (!this._owns(p))
+                continue;
             const w = params[p].data;
             const g = grads[p].data;
             const m = this.m[p];
@@ -48,5 +68,16 @@ export class AdamW {
             }
         }
     }
+    /** Bytes of optimizer state this shard holds (m + v) — 1/count of the full state. */
+    stateBytes() {
+        let n = 0;
+        for (const arr of this.m)
+            if (arr)
+                n += arr.byteLength;
+        for (const arr of this.v)
+            if (arr)
+                n += arr.byteLength;
+        return n;
+    }
 }
 //# sourceMappingURL=adamw.js.map

package/dist/optim/adamw.js.map

@@ -1 +1 @@
-{"version":3,"file":"adamw.js","sourceRoot":"","sources":["../../src/optim/adamw.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAmBH,MAAM,OAAO,KAAK;IAOG;IANF,CAAC,GAAmB,EAAE,CAAC;IACvB,CAAC,GAAmB,EAAE,CAAC;IAChC,CAAC,GAAG,CAAC,CAAC;IACG,GAAG,CAAyB;IAE7C,YACmB,MAAmB,EACpC,UAAwB,EAAE;QADT,WAAM,GAAN,MAAM,CAAa;QAGpC,IAAI,CAAC,GAAG,GAAG;YACT,EAAE,EAAE,OAAO,CAAC,EAAE,IAAI,IAAI;YACtB,KAAK,EAAE,OAAO,CAAC,KAAK,IAAI,GAAG;YAC3B,KAAK,EAAE,OAAO,CAAC,KAAK,IAAI,KAAK;YAC7B,GAAG,EAAE,OAAO,CAAC,GAAG,IAAI,IAAI;YACxB,WAAW,EAAE,OAAO,CAAC,WAAW,IAAI,CAAC;SACtC,CAAC;QACF,KAAK,MAAM,CAAC,IAAI,MAAM,CAAC,UAAU,EAAE,EAAE,CAAC;YACpC,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,YAAY,CAAC,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC;YAC7C,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,YAAY,CAAC,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC;QAC/C,CAAC;IACH,CAAC;IAED,mEAAmE;IACnE,IAAI;QACF,IAAI,CAAC,CAAC,EAAE,CAAC;QACT,MAAM,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,WAAW,EAAE,GAAG,IAAI,CAAC,GAAG,CAAC;QACxD,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,CAAC,UAAU,EAAE,CAAC;QACxC,MAAM,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC,SAAS,EAAE,CAAC;QACtC,MAAM,GAAG,GAAG,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,IAAI,CAAC,CAAC,CAAC,CAAC;QACxC,MAAM,GAAG,GAAG,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,IAAI,CAAC,CAAC,CAAC,CAAC;QACxC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACvC,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC,CAAE,CAAC,IAAI,CAAC;YAC1B,MAAM,CAAC,GAAG,KAAK,CAAC,CAAC,CAAE,CAAC,IAAI,CAAC;YACzB,MAAM,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC;YACrB,MAAM,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC;YACrB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;gBAClC,MAAM,EAAE,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC;gBACjB,CAAC,CAAC,CAAC,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,GAAG,EAAE,CAAC;gBACxC,CAAC,CAAC,CAAC,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,GAAG,EAAE,GAAG,EAAE,CAAC;gBAC7C,MAAM,EAAE,GAAG,CAAC,CAAC,CAAC,CAAE,GAAG,GAAG,CAAC;gBACvB,MAAM,EAAE,GAAG,CAAC,CAAC,CAAC,CAAE,GAAG,GAAG,CAAC;gBACvB,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAE,GAAG,EAAE,GAAG,CAAC,EAAE,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,GAAG,GAAG,CAAC,GAAG,WAAW,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC,CAAC;YACzE,CAAC;QACH,CAAC;IACH,CAAC;CACF"}
+{"version":3,"file":"adamw.js","sourceRoot":"","sources":["../../src/optim/adamw.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAwCH,MAAM,OAAO,KAAK;IASG;IARnB,kFAAkF;IACjE,CAAC,GAA4B,EAAE,CAAC;IAChC,CAAC,GAA4B,EAAE,CAAC;IACzC,CAAC,GAAG,CAAC,CAAC;IACG,GAAG,CAAwC;IAC3C,KAAK,CAAmB;IAEzC,YACmB,MAAmB,EACpC,UAAwB,EAAE;QADT,WAAM,GAAN,MAAM,CAAa;QAGpC,IAAI,CAAC,GAAG,GAAG;YACT,EAAE,EAAE,OAAO,CAAC,EAAE,IAAI,IAAI;YACtB,KAAK,EAAE,OAAO,CAAC,KAAK,IAAI,GAAG;YAC3B,KAAK,EAAE,OAAO,CAAC,KAAK,IAAI,KAAK;YAC7B,GAAG,EAAE,OAAO,CAAC,GAAG,IAAI,IAAI;YACxB,WAAW,EAAE,OAAO,CAAC,WAAW,IAAI,CAAC;SACtC,CAAC;QACF,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,IAAI,IAAI,CAAC;QACpC,IAAI,KAAK,IAAI,CAAC,KAAK,CAAC,KAAK,IAAI,CAAC,IAAI,KAAK,CAAC,KAAK,GAAG,CAAC,IAAI,KAAK,CAAC,KAAK,IAAI,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC;YACjF,MAAM,IAAI,KAAK,CAAC,wBAAwB,KAAK,CAAC,KAAK,IAAI,KAAK,CAAC,KAAK,EAAE,CAAC,CAAC;QACxE,CAAC;QACD,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,MAAM,MAAM,GAAG,MAAM,CAAC,UAAU,EAAE,CAAC;QACnC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACvC,IAAI,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC;gBAClB,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,YAAY,CAAC,MAAM,CAAC,CAAC,CAAE,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC;gBACtD,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,YAAY,CAAC,MAAM,CAAC,CAAC,CAAE,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC;YACxD,CAAC;iBAAM,CAAC;gBACN,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;gBAClB,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;YACpB,CAAC;QACH,CAAC;IACH,CAAC;IAED,kEAAkE;IAC1D,KAAK,CAAC,CAAS;QACrB,OAAO,IAAI,CAAC,KAAK,KAAK,IAAI,IAAI,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,KAAK,KAAK,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC;IAC1E,CAAC;IAED,wFAAwF;IACxF,IAAI;QACF,IAAI,CAAC,CAAC,EAAE,CAAC;QACT,MAAM,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,WAAW,EAAE,GAAG,IAAI,CAAC,GAAG,CAAC;QACxD,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,CAAC,UAAU,EAAE,CAAC;QACxC,MAAM,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC,SAAS,EAAE,CAAC;QACtC,MAAM,GAAG,GAAG,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,IAAI,CAAC,CAAC,CAAC,CAAC;QACxC,MAAM,GAAG,GAAG,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,IAAI,CAAC,CAAC,CAAC,CAAC;QACxC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACvC,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC;gBAAE,SAAS;YAC7B,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC,CAAE,CAAC,IAAI,CAAC;YAC1B,MAAM,CAAC,GAAG,KAAK,CAAC,CAAC,CAAE,CAAC,IAAI,CAAC;YACzB,MAAM,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC;YACrB,MAAM,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC;YACrB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;gBAClC,MAAM,EAAE,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC;gBACjB,CAAC,CAAC,CAAC,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,GAAG,EAAE,CAAC;gBACxC,CAAC,CAAC,CAAC,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,GAAG,EAAE,GAAG,EAAE,CAAC;gBAC7C,MAAM,EAAE,GAAG,CAAC,CAAC,CAAC,CAAE,GAAG,GAAG,CAAC;gBACvB,MAAM,EAAE,GAAG,CAAC,CAAC,CAAC,CAAE,GAAG,GAAG,CAAC;gBACvB,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAE,GAAG,EAAE,GAAG,CAAC,EAAE,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,GAAG,GAAG,CAAC,GAAG,WAAW,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC,CAAC;YACzE,CAAC;QACH,CAAC;IACH,CAAC;IAED,qFAAqF;IACrF,UAAU;QACR,IAAI,CAAC,GAAG,CAAC,CAAC;QACV,KAAK,MAAM,GAAG,IAAI,IAAI,CAAC,CAAC;YAAE,IAAI,GAAG;gBAAE,CAAC,IAAI,GAAG,CAAC,UAAU,CAAC;QACvD,KAAK,MAAM,GAAG,IAAI,IAAI,CAAC,CAAC;YAAE,IAAI,GAAG;gBAAE,CAAC,IAAI,GAAG,CAAC,UAAU,CAAC;QACvD,OAAO,CAAC,CAAC;IACX,CAAC;CACF"}

package/dist/training/lora.d.ts

@@ -0,0 +1,153 @@
+/**
+ * lora.ts — Low-Rank Adaptation (LoRA) + QLoRA for the EvermindLM CPU reference.
+ *
+ * The Meta llama-cookbook's central PEFT idea: freeze the base weights and train
+ * a tiny low-rank delta ΔW = (α/r)·B·A on top. Three payoffs, all realised here:
+ *
+ *   • Cheap — you train r·(rows+cols) params instead of rows·cols. For the tied
+ *     embedding (vocab×dModel) at rank 8 that is orders of magnitude smaller.
+ *   • Composable — the adapter serialises to a few KB (see {@link LoRAAdapter.serialize}),
+ *     so a persona / tenant / project is an MB-scale artifact you swap at load
+ *     time, not a full checkpoint.
+ *   • Forgetting-safe — the base never moves, so an adapter cannot catastrophically
+ *     overwrite the pretrained model (the exact property WSLA approximates with a
+ *     trust region; LoRA gets it structurally).
+ *
+ * QLoRA adds one more: the frozen base is held QUANTIZED (fp16 or int8) and
+ * dequantized on the fly for the merged forward, while the small adapter trains
+ * in fp32. On a single constrained device (our WebGPU target) the frozen base is
+ * where most bytes live, so quantizing it is the biggest memory unlock.
+ *
+ * Pure CPU, exact gradients (finite-difference checked in tests). The WGSL path
+ * is a future acceleration with the same shapes — same contract as EvermindLM.
+ */
+import { type AdamWOptions } from "../optim/adamw.js";
+/** How the frozen base matrix is stored. `none` = fp32 (plain LoRA); the others = QLoRA. */
+export type BaseQuant = "none" | "fp16" | "int8";
+export interface LoRAConfig {
+    /** Low-rank bottleneck. Higher = more capacity, larger adapter. Default 8. */
+    rank?: number;
+    /** LoRA scaling; the delta is scaled by alpha/rank. Default = rank (unit scale). */
+    alpha?: number;
+    /** Deterministic init seed for the A matrix. */
+    seed?: number;
+}
+/**
+ * A low-rank adapter over one base matrix of shape [rows × cols] (row-major).
+ *
+ *   ΔW[i,j] = (α/r) · Σ_k B[i,k]·A[k,j]          B: [rows×r]  A: [r×cols]
+ *
+ * Standard LoRA init: A ~ N(0, σ²), B = 0, so ΔW starts at exactly zero and the
+ * adapted model equals the base until training moves it. Only A and B are
+ * trainable ({@link parameters}/{@link gradients} expose them to {@link AdamW});
+ * the base is supplied by the caller and never mutated here.
+ */
+export declare class LoRAAdapter {
+    readonly rows: number;
+    readonly cols: number;
+    readonly rank: number;
+    readonly alpha: number;
+    readonly scale: number;
+    /** [rows × rank], zero-initialised. */
+    readonly B: Float32Array;
+    /** [rank × cols], gaussian-initialised. */
+    readonly A: Float32Array;
+    private readonly gB;
+    private readonly gA;
+    constructor(rows: number, cols: number, config?: LoRAConfig);
+    /** Trainable count — the LoRA saving vs a full rows·cols matrix. */
+    numParams(): number;
+    /** Materialise ΔW = scale·B·A as a flat [rows·cols] row-major array. */
+    delta(): Float32Array;
+    /** Base + ΔW (new array; `base` is not mutated). */
+    applyTo(base: Float32Array): Float32Array;
+    /**
+     * Project dL/dW (flat [rows·cols], the gradient the base matrix WOULD receive)
+     * onto the adapter, accumulating dL/dA and dL/dB. Exact chain rule for
+     * ΔW = scale·B·A:  gA = scale·Bᵀ·G,  gB = scale·G·Aᵀ.
+     */
+    accumulateGradient(gW: Float32Array): void;
+    parameters(): {
+        data: Float32Array;
+    }[];
+    gradients(): {
+        data: Float32Array;
+    }[];
+    zeroGrad(): void;
+    /** Compact self-describing adapter blob (magic, rows, cols, rank, alpha, B, A). */
+    serialize(): ArrayBuffer;
+    static deserialize(buffer: ArrayBuffer): LoRAAdapter;
+}
+/** Quantize a base matrix for QLoRA storage; returns a dequantized view + byte cost. */
+export declare function quantizeBase(base: Float32Array, mode: BaseQuant): {
+    view: Float32Array;
+    bytes: number;
+};
+import { EvermindLM } from "../lm/evermind_lm.js";
+export interface LoRAFitOptions extends AdamWOptions {
+    epochs?: number;
+    /**
+     * Gradient accumulation: average the adapter gradient over this many sequences
+     * (micro-batches) before each optimiser step. Lets a memory-constrained device
+     * train at a larger *effective* batch. Default 1 (step per sequence).
+     */
+    accumSteps?: number;
+}
+/**
+ * LoRA / QLoRA fine-tuning of an {@link EvermindLM} through its tied token
+ * embedding — the dominant parameter (vocab×dModel) and the natural adapter
+ * target (input lookup and output head share it). The base model is frozen; only
+ * the {@link LoRAAdapter} trains.
+ *
+ * QLoRA: pass `baseQuant: "fp16" | "int8"` and the frozen base is stored
+ * quantized and dequantized on the fly for the merged forward, so the resident
+ * base costs half (fp16) or a quarter (int8) of the bytes while the adapter
+ * trains full-precision.
+ *
+ * The adapter is the shippable artifact: {@link serializeAdapter} emits a few KB
+ * you swap per persona/tenant/project, versus rewriting the whole checkpoint.
+ */
+export declare class EvermindLMLoRA {
+    private readonly model;
+    readonly adapter: LoRAAdapter;
+    readonly baseQuant: BaseQuant;
+    /** Frozen base embedding as used in the merged forward (dequantized under QLoRA). */
+    private readonly frozenBase;
+    private readonly baseBytesStored;
+    private readonly rows;
+    private readonly cols;
+    constructor(model: EvermindLM, config?: LoRAConfig & {
+        baseQuant?: BaseQuant;
+    });
+    /** The frozen base model this adapter rides on. */
+    get baseModel(): EvermindLM;
+    /** Effective embedding used for training/generation: frozenBase + adapter delta. */
+    mergedEmb(): Float32Array;
+    /**
+     * One forward+backward on the merged weights, projecting the base-embedding
+     * gradient onto the adapter. The underlying model's own weights are left
+     * exactly as they were (base frozen); adapter gradients ACCUMULATE (caller
+     * zeroes between optimiser windows).
+     */
+    private _accumulate;
+    /** Train the adapter (only) with AdamW + optional gradient accumulation. */
+    fit(sequences: number[][], opts?: LoRAFitOptions): number[];
+    generate(prompt: number[], opts: import("../lm/evermind_lm.js").LMGenerateOptions): number[];
+    generateText(prompt: string, codec: import("../lm/evermind_lm.js").TextCodec, opts: import("../lm/evermind_lm.js").LMGenerateOptions): string;
+    /** The shippable adapter artifact (a few KB). */
+    serializeAdapter(): ArrayBuffer;
+    /** Reconstruct a fine-tuned model from a base model + a serialized adapter. */
+    static loadAdapter(model: EvermindLM, adapterBuffer: ArrayBuffer, baseQuant?: BaseQuant): EvermindLMLoRA;
+    /** Bake the adapter into the base and return a standalone EVL0 checkpoint. */
+    merge(opts?: {
+        fp16?: boolean;
+    }): ArrayBuffer;
+    /** Byte cost of the trainable adapter vs the frozen base — the LoRA/QLoRA saving. */
+    footprint(): {
+        adapterBytes: number;
+        baseBytes: number;
+        trainableParams: number;
+        baseParams: number;
+    };
+}
+//# sourceMappingURL=lora.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"lora.d.ts","sourceRoot":"","sources":["../../src/training/lora.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAEH,OAAO,EAAS,KAAK,YAAY,EAAE,MAAM,mBAAmB,CAAC;AAS7D,4FAA4F;AAC5F,MAAM,MAAM,SAAS,GAAG,MAAM,GAAG,MAAM,GAAG,MAAM,CAAC;AAEjD,MAAM,WAAW,UAAU;IACzB,8EAA8E;IAC9E,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,oFAAoF;IACpF,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,gDAAgD;IAChD,IAAI,CAAC,EAAE,MAAM,CAAC;CACf;AAID;;;;;;;;;GASG;AACH,qBAAa,WAAW;IACtB,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,uCAAuC;IACvC,QAAQ,CAAC,CAAC,EAAE,YAAY,CAAC;IACzB,2CAA2C;IAC3C,QAAQ,CAAC,CAAC,EAAE,YAAY,CAAC;IACzB,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAe;IAClC,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAe;gBAEtB,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,EAAE,MAAM,GAAE,UAAe;IA0B/D,oEAAoE;IACpE,SAAS,IAAI,MAAM;IAInB,wEAAwE;IACxE,KAAK,IAAI,YAAY;IAgBrB,oDAAoD;IACpD,OAAO,CAAC,IAAI,EAAE,YAAY,GAAG,YAAY;IAUzC;;;;OAIG;IACH,kBAAkB,CAAC,EAAE,EAAE,YAAY,GAAG,IAAI;IAqB1C,UAAU,IAAI;QAAE,IAAI,EAAE,YAAY,CAAA;KAAE,EAAE;IAGtC,SAAS,IAAI;QAAE,IAAI,EAAE,YAAY,CAAA;KAAE,EAAE;IAGrC,QAAQ,IAAI,IAAI;IAKhB,mFAAmF;IACnF,SAAS,IAAI,WAAW;IAgBxB,MAAM,CAAC,WAAW,CAAC,MAAM,EAAE,WAAW,GAAG,WAAW;CAarD;AAED,wFAAwF;AACxF,wBAAgB,YAAY,CAAC,IAAI,EAAE,YAAY,EAAE,IAAI,EAAE,SAAS,GAAG;IAAE,IAAI,EAAE,YAAY,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,CAavG;AAED,OAAO,EAAE,UAAU,EAAE,MAAM,sBAAsB,CAAC;AAElD,MAAM,WAAW,cAAe,SAAQ,YAAY;IAClD,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB;;;;OAIG;IACH,UAAU,CAAC,EAAE,MAAM,CAAC;CACrB;AAED;;;;;;;;;;;;;GAaG;AACH,qBAAa,cAAc;IAUvB,OAAO,CAAC,QAAQ,CAAC,KAAK;IATxB,QAAQ,CAAC,OAAO,EAAE,WAAW,CAAC;IAC9B,QAAQ,CAAC,SAAS,EAAE,SAAS,CAAC;IAC9B,qFAAqF;IACrF,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAe;IAC1C,OAAO,CAAC,QAAQ,CAAC,eAAe,CAAS;IACzC,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAS;IAC9B,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAS;gBAGX,KAAK,EAAE,UAAU,EAClC,MAAM,GAAE,UAAU,GAAG;QAAE,SAAS,CAAC,EAAE,SAAS,CAAA;KAAO;IAWrD,mDAAmD;IACnD,IAAI,SAAS,IAAI,UAAU,CAE1B;IAED,oFAAoF;IACpF,SAAS,IAAI,YAAY;IAIzB;;;;;OAKG;IACH,OAAO,CAAC,WAAW;IAYnB,4EAA4E;IAC5E,GAAG,CAAC,SAAS,EAAE,MAAM,EAAE,EAAE,EAAE,IAAI,GAAE,cAAmB,GAAG,MAAM,EAAE;IA8B/D,QAAQ,CAAC,MAAM,EAAE,MAAM,EAAE,EAAE,IAAI,EAAE,OAAO,sBAAsB,EAAE,iBAAiB,GAAG,MAAM,EAAE;IAU5F,YAAY,CAAC,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,OAAO,sBAAsB,EAAE,SAAS,EAAE,IAAI,EAAE,OAAO,sBAAsB,EAAE,iBAAiB,GAAG,MAAM;IAI7I,iDAAiD;IACjD,gBAAgB,IAAI,WAAW;IAI/B,+EAA+E;IAC/E,MAAM,CAAC,WAAW,CAAC,KAAK,EAAE,UAAU,EAAE,aAAa,EAAE,WAAW,EAAE,SAAS,GAAE,SAAkB,GAAG,cAAc;IAQhH,8EAA8E;IAC9E,KAAK,CAAC,IAAI,GAAE;QAAE,IAAI,CAAC,EAAE,OAAO,CAAA;KAAO,GAAG,WAAW;IAKjD,qFAAqF;IACrF,SAAS,IAAI;QAAE,YAAY,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,CAAC;QAAC,eAAe,EAAE,MAAM,CAAC;QAAC,UAAU,EAAE,MAAM,CAAA;KAAE;CAQtG"}

package/dist/training/lora.js

@@ -0,0 +1,324 @@
+/**
+ * lora.ts — Low-Rank Adaptation (LoRA) + QLoRA for the EvermindLM CPU reference.
+ *
+ * The Meta llama-cookbook's central PEFT idea: freeze the base weights and train
+ * a tiny low-rank delta ΔW = (α/r)·B·A on top. Three payoffs, all realised here:
+ *
+ *   • Cheap — you train r·(rows+cols) params instead of rows·cols. For the tied
+ *     embedding (vocab×dModel) at rank 8 that is orders of magnitude smaller.
+ *   • Composable — the adapter serialises to a few KB (see {@link LoRAAdapter.serialize}),
+ *     so a persona / tenant / project is an MB-scale artifact you swap at load
+ *     time, not a full checkpoint.
+ *   • Forgetting-safe — the base never moves, so an adapter cannot catastrophically
+ *     overwrite the pretrained model (the exact property WSLA approximates with a
+ *     trust region; LoRA gets it structurally).
+ *
+ * QLoRA adds one more: the frozen base is held QUANTIZED (fp16 or int8) and
+ * dequantized on the fly for the merged forward, while the small adapter trains
+ * in fp32. On a single constrained device (our WebGPU target) the frozen base is
+ * where most bytes live, so quantizing it is the biggest memory unlock.
+ *
+ * Pure CPU, exact gradients (finite-difference checked in tests). The WGSL path
+ * is a future acceleration with the same shapes — same contract as EvermindLM.
+ */
+import { AdamW } from "../optim/adamw.js";
+import { SeededRng } from "../utils/rng.js";
+import { quantizeFp16, dequantizeFp16, quantizeInt8, dequantizeInt8, } from "../utils/quantization.js";
+const ADAPTER_MAGIC = 0x4c4f5241; // "LORA"
+/**
+ * A low-rank adapter over one base matrix of shape [rows × cols] (row-major).
+ *
+ *   ΔW[i,j] = (α/r) · Σ_k B[i,k]·A[k,j]          B: [rows×r]  A: [r×cols]
+ *
+ * Standard LoRA init: A ~ N(0, σ²), B = 0, so ΔW starts at exactly zero and the
+ * adapted model equals the base until training moves it. Only A and B are
+ * trainable ({@link parameters}/{@link gradients} expose them to {@link AdamW});
+ * the base is supplied by the caller and never mutated here.
+ */
+export class LoRAAdapter {
+    rows;
+    cols;
+    rank;
+    alpha;
+    scale;
+    /** [rows × rank], zero-initialised. */
+    B;
+    /** [rank × cols], gaussian-initialised. */
+    A;
+    gB;
+    gA;
+    constructor(rows, cols, config = {}) {
+        if (rows <= 0 || cols <= 0)
+            throw new Error("LoRAAdapter: rows and cols must be > 0");
+        const rank = config.rank ?? 8;
+        if (rank <= 0 || rank > Math.min(rows, cols)) {
+            throw new Error(`LoRAAdapter: rank must be in [1, min(rows,cols)=${Math.min(rows, cols)}]`);
+        }
+        this.rows = rows;
+        this.cols = cols;
+        this.rank = rank;
+        this.alpha = config.alpha ?? rank;
+        this.scale = this.alpha / rank;
+        this.B = new Float32Array(rows * rank); // zero-init ⇒ ΔW starts at 0
+        this.A = new Float32Array(rank * cols);
+        const rng = new SeededRng((config.seed ?? 0x4c6f5241) >>> 0 || 1);
+        // Kaiming-ish small init for A (std ~ 1/sqrt(cols)).
+        const std = 1 / Math.sqrt(cols);
+        for (let i = 0; i < this.A.length; i++) {
+            const u1 = Math.max(rng.next(), 1e-12);
+            const u2 = rng.next();
+            this.A[i] = std * Math.sqrt(-2 * Math.log(u1)) * Math.cos(2 * Math.PI * u2);
+        }
+        this.gB = new Float32Array(this.B.length);
+        this.gA = new Float32Array(this.A.length);
+    }
+    /** Trainable count — the LoRA saving vs a full rows·cols matrix. */
+    numParams() {
+        return this.B.length + this.A.length;
+    }
+    /** Materialise ΔW = scale·B·A as a flat [rows·cols] row-major array. */
+    delta() {
+        const { rows, cols, rank, scale } = this;
+        const out = new Float32Array(rows * cols);
+        for (let i = 0; i < rows; i++) {
+            const bOff = i * rank;
+            const oOff = i * cols;
+            for (let k = 0; k < rank; k++) {
+                const b = this.B[bOff + k] * scale;
+                if (b === 0)
+                    continue;
+                const aOff = k * cols;
+                for (let j = 0; j < cols; j++)
+                    out[oOff + j] = out[oOff + j] + b * this.A[aOff + j];
+            }
+        }
+        return out;
+    }
+    /** Base + ΔW (new array; `base` is not mutated). */
+    applyTo(base) {
+        if (base.length !== this.rows * this.cols) {
+            throw new Error("LoRAAdapter.applyTo: base length mismatch");
+        }
+        const d = this.delta();
+        const out = new Float32Array(base.length);
+        for (let i = 0; i < out.length; i++)
+            out[i] = base[i] + d[i];
+        return out;
+    }
+    /**
+     * Project dL/dW (flat [rows·cols], the gradient the base matrix WOULD receive)
+     * onto the adapter, accumulating dL/dA and dL/dB. Exact chain rule for
+     * ΔW = scale·B·A:  gA = scale·Bᵀ·G,  gB = scale·G·Aᵀ.
+     */
+    accumulateGradient(gW) {
+        const { rows, cols, rank, scale } = this;
+        if (gW.length !== rows * cols)
+            throw new Error("LoRAAdapter.accumulateGradient: gW length mismatch");
+        for (let i = 0; i < rows; i++) {
+            const gOff = i * cols;
+            const bOff = i * rank;
+            for (let k = 0; k < rank; k++) {
+                const aOff = k * cols;
+                let gb = 0;
+                const bScaled = this.B[bOff + k] * scale;
+                for (let j = 0; j < cols; j++) {
+                    const g = gW[gOff + j];
+                    gb += g * this.A[aOff + j]; // gB[i,k] = scale·Σ_j G[i,j]·A[k,j]
+                    this.gA[aOff + j] = this.gA[aOff + j] + bScaled * g; // gA[k,j] += scale·B[i,k]·G[i,j]
+                }
+                this.gB[bOff + k] = this.gB[bOff + k] + scale * gb;
+            }
+        }
+    }
+    // AdamW OptimTarget surface — trains ONLY the adapter.
+    parameters() {
+        return [{ data: this.B }, { data: this.A }];
+    }
+    gradients() {
+        return [{ data: this.gB }, { data: this.gA }];
+    }
+    zeroGrad() {
+        this.gB.fill(0);
+        this.gA.fill(0);
+    }
+    /** Compact self-describing adapter blob (magic, rows, cols, rank, alpha, B, A). */
+    serialize() {
+        const headerEls = 5;
+        const headerBytes = headerEls * 4;
+        const buf = new ArrayBuffer(headerBytes + (this.B.length + this.A.length) * 4);
+        const head = new Uint32Array(buf, 0, headerEls);
+        head[0] = ADAPTER_MAGIC;
+        head[1] = this.rows;
+        head[2] = this.cols;
+        head[3] = this.rank;
+        new Float32Array(buf, 16, 1)[0] = this.alpha;
+        const body = new Float32Array(buf, headerBytes);
+        body.set(this.B, 0);
+        body.set(this.A, this.B.length);
+        return buf;
+    }
+    static deserialize(buffer) {
+        const head = new Uint32Array(buffer, 0, 5);
+        if (head[0] !== ADAPTER_MAGIC)
+            throw new Error("LoRAAdapter.deserialize: bad magic");
+        const rows = head[1];
+        const cols = head[2];
+        const rank = head[3];
+        const alpha = new Float32Array(buffer, 16, 1)[0];
+        const adapter = new LoRAAdapter(rows, cols, { rank, alpha });
+        const body = new Float32Array(buffer, 20);
+        adapter.B.set(body.subarray(0, adapter.B.length));
+        adapter.A.set(body.subarray(adapter.B.length, adapter.B.length + adapter.A.length));
+        return adapter;
+    }
+}
+/** Quantize a base matrix for QLoRA storage; returns a dequantized view + byte cost. */
+export function quantizeBase(base, mode) {
+    switch (mode) {
+        case "none":
+            return { view: base, bytes: base.length * 4 };
+        case "fp16": {
+            const q = quantizeFp16(base);
+            return { view: dequantizeFp16(q), bytes: q.length * 2 };
+        }
+        case "int8": {
+            const q = quantizeInt8(base);
+            return { view: dequantizeInt8(q.data, q.scale), bytes: q.data.length + 4 };
+        }
+    }
+}
+/**
+ * LoRA / QLoRA fine-tuning of an {@link EvermindLM} through its tied token
+ * embedding — the dominant parameter (vocab×dModel) and the natural adapter
+ * target (input lookup and output head share it). The base model is frozen; only
+ * the {@link LoRAAdapter} trains.
+ *
+ * QLoRA: pass `baseQuant: "fp16" | "int8"` and the frozen base is stored
+ * quantized and dequantized on the fly for the merged forward, so the resident
+ * base costs half (fp16) or a quarter (int8) of the bytes while the adapter
+ * trains full-precision.
+ *
+ * The adapter is the shippable artifact: {@link serializeAdapter} emits a few KB
+ * you swap per persona/tenant/project, versus rewriting the whole checkpoint.
+ */
+export class EvermindLMLoRA {
+    model;
+    adapter;
+    baseQuant;
+    /** Frozen base embedding as used in the merged forward (dequantized under QLoRA). */
+    frozenBase;
+    baseBytesStored;
+    rows;
+    cols;
+    constructor(model, config = {}) {
+        this.model = model;
+        this.rows = model.config.vocabSize;
+        this.cols = model.config.dModel;
+        this.baseQuant = config.baseQuant ?? "none";
+        const { view, bytes } = quantizeBase(Float32Array.from(model.emb), this.baseQuant);
+        this.frozenBase = view;
+        this.baseBytesStored = bytes;
+        this.adapter = new LoRAAdapter(this.rows, this.cols, config);
+    }
+    /** The frozen base model this adapter rides on. */
+    get baseModel() {
+        return this.model;
+    }
+    /** Effective embedding used for training/generation: frozenBase + adapter delta. */
+    mergedEmb() {
+        return this.adapter.applyTo(this.frozenBase);
+    }
+    /**
+     * One forward+backward on the merged weights, projecting the base-embedding
+     * gradient onto the adapter. The underlying model's own weights are left
+     * exactly as they were (base frozen); adapter gradients ACCUMULATE (caller
+     * zeroes between optimiser windows).
+     */
+    _accumulate(seq) {
+        const merged = this.mergedEmb();
+        const saved = this.model.emb;
+        this.model.emb = merged; // swap in effective weights (tied lookup + head)
+        this.model.zeroGrad();
+        const loss = this.model.lossAndBackward(seq);
+        const gW = this.model.gradients()[0].data; // dL/dW_eff == dL/dΔW, shape [rows×cols]
+        this.adapter.accumulateGradient(gW);
+        this.model.emb = saved; // restore the frozen base
+        return loss;
+    }
+    /** Train the adapter (only) with AdamW + optional gradient accumulation. */
+    fit(sequences, opts = {}) {
+        const epochs = opts.epochs ?? 1;
+        const accum = Math.max(1, opts.accumSteps ?? 1);
+        const adam = new AdamW(this.adapter, opts);
+        const grads = this.adapter.gradients();
+        const history = [];
+        for (let e = 0; e < epochs; e++) {
+            let total = 0;
+            let n = 0;
+            let pending = 0;
+            this.adapter.zeroGrad();
+            const flush = () => {
+                if (pending === 0)
+                    return;
+                if (pending !== 1)
+                    for (const g of grads)
+                        for (let i = 0; i < g.data.length; i++)
+                            g.data[i] = g.data[i] / pending;
+                adam.step();
+                this.adapter.zeroGrad();
+                pending = 0;
+            };
+            for (const seq of sequences) {
+                if (seq.length < 2)
+                    continue;
+                total += this._accumulate(seq);
+                n++;
+                if (++pending >= accum)
+                    flush();
+            }
+            flush();
+            history.push(n > 0 ? total / n : 0);
+        }
+        return history;
+    }
+    generate(prompt, opts) {
+        const saved = this.model.emb;
+        this.model.emb = this.mergedEmb();
+        try {
+            return this.model.generate(prompt, opts);
+        }
+        finally {
+            this.model.emb = saved;
+        }
+    }
+    generateText(prompt, codec, opts) {
+        return codec.decode(this.generate(codec.encode(prompt), opts));
+    }
+    /** The shippable adapter artifact (a few KB). */
+    serializeAdapter() {
+        return this.adapter.serialize();
+    }
+    /** Reconstruct a fine-tuned model from a base model + a serialized adapter. */
+    static loadAdapter(model, adapterBuffer, baseQuant = "none") {
+        const loaded = LoRAAdapter.deserialize(adapterBuffer);
+        const wrap = new EvermindLMLoRA(model, { rank: loaded.rank, alpha: loaded.alpha, baseQuant });
+        wrap.adapter.B.set(loaded.B);
+        wrap.adapter.A.set(loaded.A);
+        return wrap;
+    }
+    /** Bake the adapter into the base and return a standalone EVL0 checkpoint. */
+    merge(opts = {}) {
+        this.model.emb.set(this.mergedEmb());
+        return this.model.exportWeights(opts);
+    }
+    /** Byte cost of the trainable adapter vs the frozen base — the LoRA/QLoRA saving. */
+    footprint() {
+        return {
+            adapterBytes: this.adapter.serialize().byteLength,
+            baseBytes: this.baseBytesStored,
+            trainableParams: this.adapter.numParams(),
+            baseParams: this.rows * this.cols,
+        };
+    }
+}
+//# sourceMappingURL=lora.js.map

package/dist/training/lora.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"lora.js","sourceRoot":"","sources":["../../src/training/lora.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAEH,OAAO,EAAE,KAAK,EAAqB,MAAM,mBAAmB,CAAC;AAC7D,OAAO,EAAE,SAAS,EAAE,MAAM,iBAAiB,CAAC;AAC5C,OAAO,EACL,YAAY,EACZ,cAAc,EACd,YAAY,EACZ,cAAc,GACf,MAAM,0BAA0B,CAAC;AAclC,MAAM,aAAa,GAAG,UAAU,CAAC,CAAC,SAAS;AAE3C;;;;;;;;;GASG;AACH,MAAM,OAAO,WAAW;IACb,IAAI,CAAS;IACb,IAAI,CAAS;IACb,IAAI,CAAS;IACb,KAAK,CAAS;IACd,KAAK,CAAS;IACvB,uCAAuC;IAC9B,CAAC,CAAe;IACzB,2CAA2C;IAClC,CAAC,CAAe;IACR,EAAE,CAAe;IACjB,EAAE,CAAe;IAElC,YAAY,IAAY,EAAE,IAAY,EAAE,SAAqB,EAAE;QAC7D,IAAI,IAAI,IAAI,CAAC,IAAI,IAAI,IAAI,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,wCAAwC,CAAC,CAAC;QACtF,MAAM,IAAI,GAAG,MAAM,CAAC,IAAI,IAAI,CAAC,CAAC;QAC9B,IAAI,IAAI,IAAI,CAAC,IAAI,IAAI,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,EAAE,IAAI,CAAC,EAAE,CAAC;YAC7C,MAAM,IAAI,KAAK,CAAC,mDAAmD,IAAI,CAAC,GAAG,CAAC,IAAI,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC;QAC9F,CAAC;QACD,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;QACjB,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;QACjB,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;QACjB,IAAI,CAAC,KAAK,GAAG,MAAM,CAAC,KAAK,IAAI,IAAI,CAAC;QAClC,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC;QAE/B,IAAI,CAAC,CAAC,GAAG,IAAI,YAAY,CAAC,IAAI,GAAG,IAAI,CAAC,CAAC,CAAC,6BAA6B;QACrE,IAAI,CAAC,CAAC,GAAG,IAAI,YAAY,CAAC,IAAI,GAAG,IAAI,CAAC,CAAC;QACvC,MAAM,GAAG,GAAG,IAAI,SAAS,CAAC,CAAC,MAAM,CAAC,IAAI,IAAI,UAAU,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC;QAClE,qDAAqD;QACrD,MAAM,GAAG,GAAG,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QAChC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACvC,MAAM,EAAE,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,IAAI,EAAE,EAAE,KAAK,CAAC,CAAC;YACvC,MAAM,EAAE,GAAG,GAAG,CAAC,IAAI,EAAE,CAAC;YACtB,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,GAAG,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,EAAE,GAAG,EAAE,CAAC,CAAC;QAC9E,CAAC;QACD,IAAI,CAAC,EAAE,GAAG,IAAI,YAAY,CAAC,IAAI,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC;QAC1C,IAAI,CAAC,EAAE,GAAG,IAAI,YAAY,CAAC,IAAI,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC;IAC5C,CAAC;IAED,oEAAoE;IACpE,SAAS;QACP,OAAO,IAAI,CAAC,CAAC,CAAC,MAAM,GAAG,IAAI,CAAC,CAAC,CAAC,MAAM,CAAC;IACvC,CAAC;IAED,wEAAwE;IACxE,KAAK;QACH,MAAM,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,GAAG,IAAI,CAAC;QACzC,MAAM,GAAG,GAAG,IAAI,YAAY,CAAC,IAAI,GAAG,IAAI,CAAC,CAAC;QAC1C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,EAAE,CAAC,EAAE,EAAE,CAAC;YAC9B,MAAM,IAAI,GAAG,CAAC,GAAG,IAAI,CAAC;YACtB,MAAM,IAAI,GAAG,CAAC,GAAG,IAAI,CAAC;YACtB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,EAAE,CAAC,EAAE,EAAE,CAAC;gBAC9B,MAAM,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,CAAE,GAAG,KAAK,CAAC;gBACpC,IAAI,CAAC,KAAK,CAAC;oBAAE,SAAS;gBACtB,MAAM,IAAI,GAAG,CAAC,GAAG,IAAI,CAAC;gBACtB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,EAAE,CAAC,EAAE;oBAAE,GAAG,CAAC,IAAI,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC,IAAI,GAAG,CAAC,CAAE,GAAG,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,CAAE,CAAC;YACxF,CAAC;QACH,CAAC;QACD,OAAO,GAAG,CAAC;IACb,CAAC;IAED,oDAAoD;IACpD,OAAO,CAAC,IAAkB;QACxB,IAAI,IAAI,CAAC,MAAM,KAAK,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;YAC1C,MAAM,IAAI,KAAK,CAAC,2CAA2C,CAAC,CAAC;QAC/D,CAAC;QACD,MAAM,CAAC,GAAG,IAAI,CAAC,KAAK,EAAE,CAAC;QACvB,MAAM,GAAG,GAAG,IAAI,YAAY,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QAC1C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,GAAG,CAAC,MAAM,EAAE,CAAC,EAAE;YAAE,GAAG,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC;QAC/D,OAAO,GAAG,CAAC;IACb,CAAC;IAED;;;;OAIG;IACH,kBAAkB,CAAC,EAAgB;QACjC,MAAM,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,GAAG,IAAI,CAAC;QACzC,IAAI,EAAE,CAAC,MAAM,KAAK,IAAI,GAAG,IAAI;YAAE,MAAM,IAAI,KAAK,CAAC,oDAAoD,CAAC,CAAC;QACrG,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,EAAE,CAAC,EAAE,EAAE,CAAC;YAC9B,MAAM,IAAI,GAAG,CAAC,GAAG,IAAI,CAAC;YACtB,MAAM,IAAI,GAAG,CAAC,GAAG,IAAI,CAAC;YACtB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,EAAE,CAAC,EAAE,EAAE,CAAC;gBAC9B,MAAM,IAAI,GAAG,CAAC,GAAG,IAAI,CAAC;gBACtB,IAAI,EAAE,GAAG,CAAC,CAAC;gBACX,MAAM,OAAO,GAAG,IAAI,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,CAAE,GAAG,KAAK,CAAC;gBAC1C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,EAAE,CAAC,EAAE,EAAE,CAAC;oBAC9B,MAAM,CAAC,GAAG,EAAE,CAAC,IAAI,GAAG,CAAC,CAAE,CAAC;oBACxB,EAAE,IAAI,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,CAAE,CAAC,CAAC,oCAAoC;oBACjE,IAAI,CAAC,EAAE,CAAC,IAAI,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,EAAE,CAAC,IAAI,GAAG,CAAC,CAAE,GAAG,OAAO,GAAG,CAAC,CAAC,CAAC,iCAAiC;gBACzF,CAAC;gBACD,IAAI,CAAC,EAAE,CAAC,IAAI,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,EAAE,CAAC,IAAI,GAAG,CAAC,CAAE,GAAG,KAAK,GAAG,EAAE,CAAC;YACtD,CAAC;QACH,CAAC;IACH,CAAC;IAED,uDAAuD;IACvD,UAAU;QACR,OAAO,CAAC,EAAE,IAAI,EAAE,IAAI,CAAC,CAAC,EAAE,EAAE,EAAE,IAAI,EAAE,IAAI,CAAC,CAAC,EAAE,CAAC,CAAC;IAC9C,CAAC;IACD,SAAS;QACP,OAAO,CAAC,EAAE,IAAI,EAAE,IAAI,CAAC,EAAE,EAAE,EAAE,EAAE,IAAI,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;IAChD,CAAC;IACD,QAAQ;QACN,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QAChB,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IAClB,CAAC;IAED,mFAAmF;IACnF,SAAS;QACP,MAAM,SAAS,GAAG,CAAC,CAAC;QACpB,MAAM,WAAW,GAAG,SAAS,GAAG,CAAC,CAAC;QAClC,MAAM,GAAG,GAAG,IAAI,WAAW,CAAC,WAAW,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,MAAM,GAAG,IAAI,CAAC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC;QAC/E,MAAM,IAAI,GAAG,IAAI,WAAW,CAAC,GAAG,EAAE,CAAC,EAAE,SAAS,CAAC,CAAC;QAChD,IAAI,CAAC,CAAC,CAAC,GAAG,aAAa,CAAC;QACxB,IAAI,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC;QACpB,IAAI,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC;QACpB,IAAI,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC;QACpB,IAAI,YAAY,CAAC,GAAG,EAAE,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC;QAC7C,MAAM,IAAI,GAAG,IAAI,YAAY,CAAC,GAAG,EAAE,WAAW,CAAC,CAAC;QAChD,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;QACpB,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC;QAChC,OAAO,GAAG,CAAC;IACb,CAAC;IAED,MAAM,CAAC,WAAW,CAAC,MAAmB;QACpC,MAAM,IAAI,GAAG,IAAI,WAAW,CAAC,MAAM,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC;QAC3C,IAAI,IAAI,CAAC,CAAC,CAAC,KAAK,aAAa;YAAE,MAAM,IAAI,KAAK,CAAC,oCAAoC,CAAC,CAAC;QACrF,MAAM,IAAI,GAAG,IAAI,CAAC,CAAC,CAAE,CAAC;QACtB,MAAM,IAAI,GAAG,IAAI,CAAC,CAAC,CAAE,CAAC;QACtB,MAAM,IAAI,GAAG,IAAI,CAAC,CAAC,CAAE,CAAC;QACtB,MAAM,KAAK,GAAG,IAAI,YAAY,CAAC,MAAM,EAAE,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC;QAClD,MAAM,OAAO,GAAG,IAAI,WAAW,CAAC,IAAI,EAAE,IAAI,EAAE,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,CAAC;QAC7D,MAAM,IAAI,GAAG,IAAI,YAAY,CAAC,MAAM,EAAE,EAAE,CAAC,CAAC;QAC1C,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC,EAAE,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC;QAClD,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,MAAM,GAAG,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC;QACpF,OAAO,OAAO,CAAC;IACjB,CAAC;CACF;AAED,wFAAwF;AACxF,MAAM,UAAU,YAAY,CAAC,IAAkB,EAAE,IAAe;IAC9D,QAAQ,IAAI,EAAE,CAAC;QACb,KAAK,MAAM;YACT,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAChD,KAAK,MAAM,CAAC,CAAC,CAAC;YACZ,MAAM,CAAC,GAAG,YAAY,CAAC,IAAI,CAAC,CAAC;YAC7B,OAAO,EAAE,IAAI,EAAE,cAAc,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC1D,CAAC;QACD,KAAK,MAAM,CAAC,CAAC,CAAC;YACZ,MAAM,CAAC,GAAG,YAAY,CAAC,IAAI,CAAC,CAAC;YAC7B,OAAO,EAAE,IAAI,EAAE,cAAc,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,KAAK,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC,IAAI,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC7E,CAAC;IACH,CAAC;AACH,CAAC;AAcD;;;;;;;;;;;;;GAaG;AACH,MAAM,OAAO,cAAc;IAUN;IATV,OAAO,CAAc;IACrB,SAAS,CAAY;IAC9B,qFAAqF;IACpE,UAAU,CAAe;IACzB,eAAe,CAAS;IACxB,IAAI,CAAS;IACb,IAAI,CAAS;IAE9B,YACmB,KAAiB,EAClC,SAAiD,EAAE;QADlC,UAAK,GAAL,KAAK,CAAY;QAGlC,IAAI,CAAC,IAAI,GAAG,KAAK,CAAC,MAAM,CAAC,SAAS,CAAC;QACnC,IAAI,CAAC,IAAI,GAAG,KAAK,CAAC,MAAM,CAAC,MAAM,CAAC;QAChC,IAAI,CAAC,SAAS,GAAG,MAAM,CAAC,SAAS,IAAI,MAAM,CAAC;QAC5C,MAAM,EAAE,IAAI,EAAE,KAAK,EAAE,GAAG,YAAY,CAAC,YAAY,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,SAAS,CAAC,CAAC;QACnF,IAAI,CAAC,UAAU,GAAG,IAAI,CAAC;QACvB,IAAI,CAAC,eAAe,GAAG,KAAK,CAAC;QAC7B,IAAI,CAAC,OAAO,GAAG,IAAI,WAAW,CAAC,IAAI,CAAC,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;IAC/D,CAAC;IAED,mDAAmD;IACnD,IAAI,SAAS;QACX,OAAO,IAAI,CAAC,KAAK,CAAC;IACpB,CAAC;IAED,oFAAoF;IACpF,SAAS;QACP,OAAO,IAAI,CAAC,OAAO,CAAC,OAAO,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC;IAC/C,CAAC;IAED;;;;;OAKG;IACK,WAAW,CAAC,GAAa;QAC/B,MAAM,MAAM,GAAG,IAAI,CAAC,SAAS,EAAE,CAAC;QAChC,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC;QAC7B,IAAI,CAAC,KAAK,CAAC,GAAG,GAAG,MAAM,CAAC,CAAC,iDAAiD;QAC1E,IAAI,CAAC,KAAK,CAAC,QAAQ,EAAE,CAAC;QACtB,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,eAAe,CAAC,GAAG,CAAC,CAAC;QAC7C,MAAM,EAAE,GAAG,IAAI,CAAC,KAAK,CAAC,SAAS,EAAE,CAAC,CAAC,CAAE,CAAC,IAAI,CAAC,CAAC,yCAAyC;QACrF,IAAI,CAAC,OAAO,CAAC,kBAAkB,CAAC,EAAE,CAAC,CAAC;QACpC,IAAI,CAAC,KAAK,CAAC,GAAG,GAAG,KAAK,CAAC,CAAC,0BAA0B;QAClD,OAAO,IAAI,CAAC;IACd,CAAC;IAED,4EAA4E;IAC5E,GAAG,CAAC,SAAqB,EAAE,OAAuB,EAAE;QAClD,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,IAAI,CAAC,CAAC;QAChC,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,UAAU,IAAI,CAAC,CAAC,CAAC;QAChD,MAAM,IAAI,GAAG,IAAI,KAAK,CAAC,IAAI,CAAC,OAAO,EAAE,IAAI,CAAC,CAAC;QAC3C,MAAM,KAAK,GAAG,IAAI,CAAC,OAAO,CAAC,SAAS,EAAE,CAAC;QACvC,MAAM,OAAO,GAAa,EAAE,CAAC;QAC7B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YAChC,IAAI,KAAK,GAAG,CAAC,CAAC;YACd,IAAI,CAAC,GAAG,CAAC,CAAC;YACV,IAAI,OAAO,GAAG,CAAC,CAAC;YAChB,IAAI,CAAC,OAAO,CAAC,QAAQ,EAAE,CAAC;YACxB,MAAM,KAAK,GAAG,GAAG,EAAE;gBACjB,IAAI,OAAO,KAAK,CAAC;oBAAE,OAAO;gBAC1B,IAAI,OAAO,KAAK,CAAC;oBAAE,KAAK,MAAM,CAAC,IAAI,KAAK;wBAAE,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE;4BAAE,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC,CAAE,GAAG,OAAO,CAAC;gBACnH,IAAI,CAAC,IAAI,EAAE,CAAC;gBACZ,IAAI,CAAC,OAAO,CAAC,QAAQ,EAAE,CAAC;gBACxB,OAAO,GAAG,CAAC,CAAC;YACd,CAAC,CAAC;YACF,KAAK,MAAM,GAAG,IAAI,SAAS,EAAE,CAAC;gBAC5B,IAAI,GAAG,CAAC,MAAM,GAAG,CAAC;oBAAE,SAAS;gBAC7B,KAAK,IAAI,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC;gBAC/B,CAAC,EAAE,CAAC;gBACJ,IAAI,EAAE,OAAO,IAAI,KAAK;oBAAE,KAAK,EAAE,CAAC;YAClC,CAAC;YACD,KAAK,EAAE,CAAC;YACR,OAAO,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QACtC,CAAC;QACD,OAAO,OAAO,CAAC;IACjB,CAAC;IAED,QAAQ,CAAC,MAAgB,EAAE,IAAsD;QAC/E,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC;QAC7B,IAAI,CAAC,KAAK,CAAC,GAAG,GAAG,IAAI,CAAC,SAAS,EAAE,CAAC;QAClC,IAAI,CAAC;YACH,OAAO,IAAI,CAAC,KAAK,CAAC,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAC,CAAC;QAC3C,CAAC;gBAAS,CAAC;YACT,IAAI,CAAC,KAAK,CAAC,GAAG,GAAG,KAAK,CAAC;QACzB,CAAC;IACH,CAAC;IAED,YAAY,CAAC,MAAc,EAAE,KAA+C,EAAE,IAAsD;QAClI,OAAO,KAAK,CAAC,MAAM,CAAC,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,IAAI,CAAC,CAAC,CAAC;IACjE,CAAC;IAED,iDAAiD;IACjD,gBAAgB;QACd,OAAO,IAAI,CAAC,OAAO,CAAC,SAAS,EAAE,CAAC;IAClC,CAAC;IAED,+EAA+E;IAC/E,MAAM,CAAC,WAAW,CAAC,KAAiB,EAAE,aAA0B,EAAE,YAAuB,MAAM;QAC7F,MAAM,MAAM,GAAG,WAAW,CAAC,WAAW,CAAC,aAAa,CAAC,CAAC;QACtD,MAAM,IAAI,GAAG,IAAI,cAAc,CAAC,KAAK,EAAE,EAAE,IAAI,EAAE,MAAM,CAAC,IAAI,EAAE,KAAK,EAAE,MAAM,CAAC,KAAK,EAAE,SAAS,EAAE,CAAC,CAAC;QAC9F,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;QAC7B,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;QAC7B,OAAO,IAAI,CAAC;IACd,CAAC;IAED,8EAA8E;IAC9E,KAAK,CAAC,OAA2B,EAAE;QACjC,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,GAAG,CAAC,IAAI,CAAC,SAAS,EAAE,CAAC,CAAC;QACrC,OAAO,IAAI,CAAC,KAAK,CAAC,aAAa,CAAC,IAAI,CAAC,CAAC;IACxC,CAAC;IAED,qFAAqF;IACrF,SAAS;QACP,OAAO;YACL,YAAY,EAAE,IAAI,CAAC,OAAO,CAAC,SAAS,EAAE,CAAC,UAAU;YACjD,SAAS,EAAE,IAAI,CAAC,eAAe;YAC/B,eAAe,EAAE,IAAI,CAAC,OAAO,CAAC,SAAS,EAAE;YACzC,UAAU,EAAE,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC,IAAI;SAClC,CAAC;IACJ,CAAC;CACF"}