@genai-fi/nanogpt 0.2.10 → 0.2.12

package/dist/TeachableLLM.js

@@ -1,6 +1,6 @@
 import { defaultConfig as h } from "./config.js";
-import d from "./NanoGPTModel.js";
-import { saveModel as m } from "./utilities/save.js";
+import m from "./NanoGPTModel.js";
+import { saveModel as d } from "./utilities/save.js";
 import { loadModel as f } from "./utilities/load.js";
 import u from "./Generator.js";
 import _ from "./Trainer.js";
@@ -13,7 +13,9 @@ import "./jszip.min-CjP2V1VV.js";
 import "./ops/scatterSub.js";
 import "./ops/gatherSub.js";
 import "./ops/attentionMask.js";
-import w from "./utilities/profile.js";
+import "./ops/qkv.js";
+import "./ops/rope.js";
+import p from "./utilities/profile.js";
 class a extends c {
   _config;
   _model;
@@ -50,7 +52,7 @@ class a extends c {
   saveModel(t) {
     if (!this._model || !this._tokeniser)
       throw new Error("Model or tokeniser is not initialized.");
-    return m(this._model, this._tokeniser, t);
+    return d(this._model, this._tokeniser, t);
   }
   static loadModel(t, r) {
     const e = new a(t);
@@ -65,7 +67,7 @@ class a extends c {
     }), e;
   }
   static create(t, r = {}) {
-    const e = { ...h, ...r }, o = new g(e.vocabSize), s = new d(t, e), i = new a(t, o, s);
+    const e = { ...h, ...r }, o = new g(e.vocabSize), s = new m(t, e), i = new a(t, o, s);
     return i.setStatus("warmup"), l(s).then(() => {
       i.tokeniser.trained ? i.setStatus("ready") : (i.setStatus("awaitingTokens"), i.tokeniser.once("trainStatus", (n) => {
         n === "trained" && i.setStatus("ready");
@@ -84,7 +86,7 @@ class a extends c {
     if (t) {
       if (!this._model)
         throw new Error("Model is not initialized.");
-      this._model.getProfiler() || this._model.setProfiler(new w());
+      this._model.getProfiler() || this._model.setProfiler(new p());
     } else
       this._model && this._model.setProfiler(void 0);
   }

package/dist/{complex-x7w5HPOS.js → complex-CJ-qCcLB.js}

@@ -1,4 +1,4 @@
-import { o as c, d as s, g as n, E as m, C as r } from "./index-CWQLouWz.js";
+import { o as c, d as s, g as n, E as m, C as r } from "./index-YPKosni4.js";
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.

package/dist/{index-CWQLouWz.js → index-YPKosni4.js}

@@ -383,7 +383,7 @@ function _t(n, t) {
     return e.set(n, s), e.get(n);
   }
 }
-const Ge = "Abs", ne = "Add", Es = "BatchMatMul", se = "Cast", As = "Complex", ze = "ComplexAbs", We = "RealDiv", Bs = "Elu", vs = "Exp", je = "Fill", Ke = "FloorDiv", Ms = "GatherNd", re = "Identity", Fs = "Imag", $s = "LeakyRelu", Rs = "Log", xs = "Max", Ve = "Maximum", qe = "Multiply", Ns = "Neg", Ds = "Pack", He = "Pow", Cs = "Prelu", _s = "Range", Ps = "Real", Os = "Relu", Ls = "Reshape", Us = "Relu6", Gs = "ScatterNd", zs = "Sigmoid", Je = "Sqrt", Ws = "Sum", js = "Softmax", Xe = "Sub", Ks = "Transpose", Ye = "ZerosLike", Vs = "Step", qs = "_FusedMatMul";
+const Ge = "Abs", ne = "Add", Es = "BatchMatMul", se = "Cast", As = "Complex", ze = "ComplexAbs", Bs = "Concat", We = "RealDiv", vs = "Elu", Ms = "Exp", je = "Fill", Ke = "FloorDiv", Fs = "GatherV2", $s = "GatherNd", re = "Identity", Rs = "Imag", xs = "LeakyRelu", Ns = "Log", Ds = "Max", Ve = "Maximum", qe = "Multiply", Cs = "Neg", _s = "Pack", He = "Pow", Ps = "Prelu", Os = "Range", Ls = "Real", Us = "Relu", Gs = "Reshape", zs = "Relu6", Ws = "ScatterNd", js = "Sigmoid", Je = "Sqrt", Ks = "Sum", Vs = "SplitV", qs = "Softmax", Xe = "Sub", Hs = "Transpose", Ye = "ZerosLike", Js = "Step", Xs = "_FusedMatMul";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -438,11 +438,11 @@ function Wt(n) {
   }
   return e;
 }
-function Hs(n) {
+function Ys(n) {
   const { kernelName: t, backendName: e } = n, s = ie(t, e);
   ht.has(s) && O(`The kernel '${t}' for backend '${e}' is already registered`), ht.set(s, n);
 }
-function Js(n) {
+function Qs(n) {
   const { kernelName: t } = n;
   It.has(t) && S().getBool("DEBUG") && O(`Overriding the gradient for '${t}'`), It.set(t, n);
 }
@@ -1902,7 +1902,7 @@ function I(n, t, e, s = "numeric") {
   const a = r !== "string" ? ae(n, r) : at(n, [], !0);
   return g.makeTensor(a, i, r);
 }
-function Xs(n, t, e, s = "numeric") {
+function Zs(n, t, e, s = "numeric") {
   if (!Array.isArray(n))
     throw new Error(`Argument ${t} passed to ${e} must be a \`Tensor[]\` or \`TensorLike[]\``);
   return n.map((i, o) => I(i, `${t}[${o}]`, e, s));
@@ -2065,10 +2065,10 @@ function Sn(n, t) {
  * limitations under the License.
  * =============================================================================
  */
-function Ys() {
+function tr() {
   return g;
 }
-function Qs() {
+function er() {
   return g.memory();
 }
 function E(n, t) {
@@ -2893,7 +2893,7 @@ function Yn(n, t, e) {
  * limitations under the License.
  * =============================================================================
  */
-function Zs(n, t) {
+function nr(n, t) {
   const e = [];
   for (let s = 0; s < t.length; s++) {
     const r = n[n.length - s - 1], i = t.length - s - 1, o = t[i];
@@ -3061,7 +3061,7 @@ function ss(n, t) {
     a[u] != null && (c[l.name] = a[u]);
   }), s?.forEach((l) => c[l.name] = null), { value: o, grads: c };
 }
-function tr(n) {
+function sr(n) {
   return g.customGrad(n);
 }
 /**
@@ -3841,55 +3841,59 @@ function bs() {
  */
 bs();
 export {
+  Qn as $,
   ds as A,
   Es as B,
   As as C,
-  C as D,
+  w as D,
   g as E,
-  zs as F,
-  Ms as G,
-  Bs as H,
-  Fs as I,
-  $s as J,
-  Cs as K,
-  Rs as L,
-  xs as M,
-  Ns as N,
-  Ps as O,
-  Ds as P,
-  Os as Q,
-  _s as R,
-  Ws as S,
-  Us as T,
-  Vs as U,
-  Ks as V,
-  Zs as W,
-  Qn as X,
-  qs as _,
+  qs as F,
+  $s as G,
+  sr as H,
+  E as I,
+  C as J,
+  js as K,
+  Ns as L,
+  Ds as M,
+  vs as N,
+  Rs as O,
+  _s as P,
+  xs as Q,
+  Gs as R,
+  Ks as S,
+  Cs as T,
+  Ps as U,
+  Ls as V,
+  Us as W,
+  zs as X,
+  Js as Y,
+  Hs as Z,
+  nr as _,
   p as a,
+  Xs as a0,
   Z as b,
-  Js as c,
+  Qs as c,
   I as d,
-  Ys as e,
+  tr as e,
   V as f,
   Is as g,
-  Xs as h,
-  y as i,
-  Ls as j,
-  $t as k,
-  Dt as l,
-  Qs as m,
-  Zt as n,
+  $t as h,
+  Vs as i,
+  Os as j,
+  Zs as k,
+  y as l,
+  er as m,
+  Gn as n,
   F as o,
-  G as p,
-  De as q,
-  Hs as r,
+  Bs as p,
+  Fs as q,
+  Ys as r,
   K as s,
-  Gs as t,
-  vs as u,
-  Ts as v,
-  w,
-  js as x,
-  tr as y,
-  E as z
+  Dt as t,
+  Zt as u,
+  G as v,
+  De as w,
+  Ws as x,
+  Ms as y,
+  Ts as z
 };

package/dist/layers/CausalSelfAttention.d.ts

@@ -21,7 +21,9 @@ export default class CausalSelfAttention extends BaseLayer {
     private divisor;
     private index;
     private _trainable;
+    private units;
     constructor(tf: typeof TF, index: number, config: GPTConfig, ropeCache?: RoPECache | undefined);
+    private build;
     get variables(): TF.Variable[];
     get trainable(): boolean;
     set trainable(value: boolean);

package/dist/layers/CausalSelfAttention.js

@@ -1,17 +1,10 @@
-import { attentionMask as z } from "../ops/attentionMask.js";
-import S from "./BaseLayer.js";
-class C extends S {
+import { attentionMask as x } from "../ops/attentionMask.js";
+import j from "./BaseLayer.js";
+import { qkv as w } from "../ops/qkv.js";
+import { rope as y } from "../ops/rope.js";
+class N extends j {
   constructor(t, i, s, e) {
-    super(), this.ropeCache = e, this.config = s, this.tf = t, this.index = i, this.cAttn = this.tf.layers.dense({
-      units: 3 * s.nEmbed,
-      useBias: s.biasInLinear,
-      name: `block_${i}_attn_cAttn`,
-      kernelInitializer: this.tf.initializers.randomNormal({
-        mean: 0,
-        stddev: 0.02
-      }),
-      biasInitializer: "zeros"
-    }), this.cProj = this.tf.layers.dense({
+    super(), this.ropeCache = e, this.config = s, this.tf = t, this.index = i, this.units = s.nEmbed * 3, this.cProj = this.tf.layers.dense({
       units: s.nEmbed,
       useBias: s.biasInLinear,
       name: `block_${i}_attn_cProj`,
@@ -21,11 +14,11 @@ class C extends S {
       }),
       biasInitializer: "zeros"
     }), this.attnDropout = this.tf.layers.dropout({ rate: s.dropout }), this.residDropout = this.tf.layers.dropout({ rate: s.dropout }), this.bias = this.tf.linalg.bandPart(this.tf.ones([s.blockSize, s.blockSize]), -1, 0).cast("bool"), this.divisor = 1 / Math.sqrt(s.nEmbed / s.nHead);
-    const o = this.tf.zeros([s.blockSize, s.blockSize]), c = this.tf.fill([s.blockSize, s.blockSize], Number.NEGATIVE_INFINITY);
-    this.maskInf = this.tf.where(this.bias, o, c);
+    const o = this.tf.zeros([s.blockSize, s.blockSize]), a = this.tf.fill([s.blockSize, s.blockSize], Number.NEGATIVE_INFINITY);
+    this.maskInf = this.tf.where(this.bias, o, a);
   }
   config;
-  cAttn;
+  cAttn = null;
   cProj;
   attnDropout;
   residDropout;
@@ -35,26 +28,35 @@ class C extends S {
   divisor;
   index;
   _trainable = !0;
+  units;
+  build() {
+    this.cAttn === null && (this.cAttn = this.tf.variable(
+      this.tf.randomNormal([this.config.nEmbed, this.units], 0, 0.02),
+      !0
+      //`block_${this.index}_attn_cAttn_kernel`
+    ));
+  }
   get variables() {
-    return [
-      ...this.cAttn.trainableWeights.map((t) => t.read()),
-      ...this.cProj.trainableWeights.map((t) => t.read())
-    ];
+    if (this.cAttn === null)
+      throw new Error("Layer not built yet");
+    return [this.cAttn, ...this.cProj.trainableWeights.map((t) => t.read())];
   }
   get trainable() {
     return this._trainable;
   }
   set trainable(t) {
-    this._trainable = t, this.cAttn.trainable = t, this.cProj.trainable = t;
+    this._trainable = t, this.cAttn && (this.cAttn.trainable = t), this.cProj.trainable = t;
   }
   saveWeights(t) {
-    t.set(`block_${this.index}_cAttn`, this.cAttn.getWeights()), t.set(`block_${this.index}_cProj`, this.cProj.getWeights());
+    t.set(`block_${this.index}_cAttn`, this.cAttn ? [this.cAttn.clone()] : []), t.set(`block_${this.index}_cProj`, this.cProj.getWeights());
   }
   loadWeights(t) {
-    this.cAttn.setWeights(t.get(`block_${this.index}_cAttn`) || []), this.cProj.setWeights(t.get(`block_${this.index}_cProj`) || []);
+    const i = t.get(`block_${this.index}_cAttn`)?.[0];
+    if (!i) throw new Error(`Weights for block_${this.index}_cAttn not found`);
+    this.cAttn ? this.cAttn.assign(i) : this.cAttn = this.tf.variable(i, !0), this.cProj.setWeights(t.get(`block_${this.index}_cProj`) || []);
   }
   getAttentionScores(t, i, s) {
-    const e = z(t, i, this.maskInf, this.divisor), o = this.tf.softmax(e, -1);
+    const e = x(t, i, this.maskInf, this.divisor), o = this.tf.softmax(e, -1);
     return this.attnDropout.apply(o, { training: s });
   }
   // Attention with optional past. If pastLen > 0 and T_cur == 1, no mask needed.
@@ -64,61 +66,49 @@ class C extends S {
     if (o > 1 && e > 0)
       throw new Error("Cannot use past with T_cur > 1");
     if (o > 1) {
-      const a = this.maskInf.slice([0, 0], [o, o]).expandDims(0).expandDims(0);
-      r = r.add(a);
+      const c = this.maskInf.slice([0, 0], [o, o]).expandDims(0).expandDims(0);
+      r = r.add(c);
     }
     const h = this.tf.softmax(r, -1);
     return this.attnDropout.apply(h, { training: s });
   }
   getQKV(t) {
-    const [i, s, e] = t.shape, o = this.cAttn.apply(t), [c, r, h] = this.tf.split(o, 3, -1);
-    o.dispose();
-    const a = e / this.config.nHead, u = this.tf.reshape(c, [i, s, this.config.nHead, a]);
-    c.dispose();
-    const f = u.transpose([0, 2, 1, 3]);
-    u.dispose();
-    const d = this.tf.reshape(r, [i, s, this.config.nHead, a]);
-    r.dispose();
-    const n = d.transpose([0, 2, 1, 3]);
-    d.dispose();
-    const l = this.tf.reshape(h, [i, s, this.config.nHead, a]);
-    h.dispose();
-    const p = l.transpose([0, 2, 1, 3]);
-    return l.dispose(), [f, n, p];
+    return w(t, this.cAttn, this.config.nHead);
   }
   getOutputProjection(t, i) {
-    const s = t.shape[0], e = t.shape[2], o = this.config.nEmbed, c = t.transpose([0, 2, 1, 3]), r = this.tf.reshape(c, [s, e, o]), h = this.cProj.apply(r);
+    const s = t.shape[0], e = t.shape[2], o = this.config.nEmbed, a = t.transpose([0, 2, 1, 3]), r = this.tf.reshape(a, [s, e, o]), h = this.cProj.apply(r);
     return this.residDropout.apply(h, { training: i });
   }
   // Added optional KV cache support (pastKV). Returns presentKV for chaining.
   call(t, i = !1, s = !1, e) {
     if (e && !this.config.useRope)
       throw new Error("Cannot use pastKV without RoPE enabled");
-    return this.tf.tidy(() => {
+    return this.build(), this.tf.tidy(() => {
       this.startMemory();
-      const [o, c, r] = this.getQKV(t), h = o.shape[2], a = this.config.blockSize, u = e ? e.cumulativeLength : 0, [f, d] = this.ropeCache ? this.ropeCache.applyRoPE(o, c, u) : [o, c];
-      let n = d, l = r, p = 0;
-      e && (p = e.length, n = this.tf.concat([e.k, d], 2), l = this.tf.concat([e.v, r], 2));
+      const [o, a, r] = this.getQKV(t), h = o.shape[2], c = this.config.blockSize, d = e ? e.cumulativeLength : 0, f = this.ropeCache ? y(o, this.ropeCache, d) : o, m = this.ropeCache ? y(a, this.ropeCache, d) : a;
+      this.ropeCache && (o.dispose(), a.dispose());
+      let n = m, l = r, u = 0;
+      e && (u = e.length, n = this.tf.concat([e.k, m], 2), l = this.tf.concat([e.v, r], 2));
       const b = n.shape[2];
-      if (b > a) {
-        const k = b - a, g = n.shape[0], A = n.shape[1], I = n.shape[3];
-        n = n.slice([0, 0, k, 0], [g, A, a, I]), l = l.slice([0, 0, k, 0], [g, A, a, I]), p = a - h;
+      if (b > c) {
+        const k = b - c, A = n.shape[0], g = n.shape[1], _ = n.shape[3];
+        n = n.slice([0, 0, k, 0], [A, g, c, _]), l = l.slice([0, 0, k, 0], [A, g, c, _]), u = c - h;
       }
-      let m;
-      p > 0 ? m = this.getAttentionScoresWithPast(f, n, i, p) : m = this.getAttentionScores(f, n, i);
-      const _ = this.tf.matMul(m, l), v = this.getOutputProjection(_, i), y = {
+      let p;
+      u > 0 ? p = this.getAttentionScoresWithPast(f, n, i, u) : p = this.getAttentionScores(f, n, i);
+      const P = this.tf.matMul(p, l), S = this.getOutputProjection(P, i), v = {
         k: this.tf.keep(n),
         v: this.tf.keep(l),
-        length: p + h,
+        length: u + h,
         cumulativeLength: e ? e.cumulativeLength + h : h
-      }, P = s ? m.mean(1) : void 0;
-      return this.endMemory("CausalSelfAttention"), { output: v, attention: P, presentKV: y };
+      }, I = s ? p.mean(1) : void 0;
+      return this.endMemory("CausalSelfAttention"), { output: S, attention: I, presentKV: v };
     });
   }
   dispose() {
-    this.cAttn.dispose(), this.cProj.dispose(), this.attnDropout.dispose(), this.residDropout.dispose(), this.bias.dispose(), this.maskInf.dispose();
+    this.cAttn?.dispose(), this.cProj.dispose(), this.attnDropout.dispose(), this.residDropout.dispose(), this.bias.dispose(), this.maskInf.dispose();
   }
 }
 export {
-  C as default
+  N as default
 };

package/dist/layers/RoPECache.d.ts

@@ -3,14 +3,15 @@ import { GPTConfig } from '../config';
 export default class RoPECache {
     private readonly tf;
     private readonly config;
-    private rotaryDim;
+    readonly rotaryDim: number;
     private ropeBase;
     private ropeInvFreq;
     private ropeCos;
     private ropeSin;
     private ropeCacheLen;
     constructor(tf: typeof TF, config: GPTConfig);
-    private ensureRopeCache;
-    applyRoPE(q: TF.Tensor, k: TF.Tensor, pastLen: number): [TF.Tensor, TF.Tensor];
+    ensureRopeCache(needed: number): void;
+    getCos(): TF.Tensor | null;
+    getSin(): TF.Tensor | null;
     dispose(): void;
 }

package/dist/layers/RoPECache.js

@@ -1,12 +1,12 @@
-class b {
-  constructor(s, r) {
-    this.tf = s, this.config = r;
-    const o = this.config.nEmbed / this.config.nHead;
-    if (this.rotaryDim = o, this.rotaryDim % 2 !== 0)
+class n {
+  constructor(i, e) {
+    this.tf = i, this.config = e;
+    const t = this.config.nEmbed / this.config.nHead;
+    if (this.rotaryDim = t, this.rotaryDim % 2 !== 0)
       throw new Error("rotaryDim must be even");
     this.ropeBase = 1e4;
-    const i = this.tf.range(0, this.rotaryDim, 2, "float32"), t = i.div(this.tf.scalar(this.rotaryDim, "float32")), e = this.tf.pow(this.tf.scalar(this.ropeBase, "float32"), t);
-    this.ropeInvFreq = this.tf.reciprocal(e), t.dispose(), e.dispose(), i.dispose(), this.config.useRope === !1 ? (this.ropeCos = null, this.ropeSin = null, this.ropeCacheLen = 0) : this.tf.tidy(() => {
+    const s = this.tf.range(0, this.rotaryDim, 2, "float32"), o = s.div(this.tf.scalar(this.rotaryDim, "float32")), r = this.tf.pow(this.tf.scalar(this.ropeBase, "float32"), o);
+    this.ropeInvFreq = this.tf.reciprocal(r), o.dispose(), r.dispose(), s.dispose(), this.config.useRope === !1 ? (this.ropeCos = null, this.ropeSin = null, this.ropeCacheLen = 0) : this.tf.tidy(() => {
       this.ensureRopeCache(this.config.blockSize * 4);
     });
   }
@@ -18,27 +18,22 @@ class b {
   ropeSin = null;
   // [cacheLen, rotaryDim/2]
   ropeCacheLen = 0;
-  ensureRopeCache(s) {
-    if (s <= this.ropeCacheLen) return;
+  ensureRopeCache(i) {
+    if (i <= this.ropeCacheLen) return;
     this.ropeCos && this.ropeCos.dispose(), this.ropeSin && this.ropeSin.dispose();
-    const o = this.tf.range(0, s, 1, "float32").expandDims(1).mul(this.ropeInvFreq.expandDims(0));
-    this.ropeCos = this.tf.keep(this.tf.cos(o).expandDims(-1)), this.ropeSin = this.tf.keep(this.tf.sin(o).expandDims(-1)), this.ropeCacheLen = s;
+    const e = Math.max(i, this.ropeCacheLen + this.config.blockSize * 4), s = this.tf.range(0, e, 1, "float32").expandDims(1).mul(this.ropeInvFreq.expandDims(0));
+    this.ropeCos = this.tf.keep(this.tf.cos(s).expandDims(-1)), this.ropeSin = this.tf.keep(this.tf.sin(s).expandDims(-1)), this.ropeCacheLen = e;
   }
-  applyRoPE(s, r, o) {
-    const i = s.shape[3], t = this.rotaryDim;
-    if (t > i) return [s, r];
-    const e = s.shape[2], v = o + e;
-    this.ensureRopeCache(v);
-    const n = t / 2, p = this.ropeCos.slice([o, 0, 0], [e, n, 1]).reshape([1, 1, e, n]), a = this.ropeSin.slice([o, 0, 0], [e, n, 1]).reshape([1, 1, e, n]), h = s.shape[0], c = s.shape[1], f = this.tf.range(0, t, 2, "int32"), l = this.tf.range(1, t, 2, "int32"), d = (u) => {
-      const m = u.slice([0, 0, 0, 0], [h, c, e, t]), C = t < i ? u.slice([0, 0, 0, t], [h, c, e, i - t]) : null, D = this.tf.gather(m, f, 3), g = this.tf.gather(m, l, 3), x = D.mul(p).sub(g.mul(a)), k = g.mul(p).add(D.mul(a)), R = this.tf.stack([x, k], -1).reshape([h, c, e, t]);
-      return C ? this.tf.concat([R, C], 3) : R;
-    }, y = d(s), S = d(r);
-    return f.dispose(), l.dispose(), [y, S];
+  getCos() {
+    return this.ropeCos;
+  }
+  getSin() {
+    return this.ropeSin;
   }
   dispose() {
     this.ropeCos && this.ropeCos.dispose(), this.ropeSin && this.ropeSin.dispose(), this.ropeInvFreq.dispose();
   }
 }
 export {
-  b as default
+  n as default
 };

package/dist/layers/TiedEmbedding.js

@@ -1,7 +1,8 @@
-import { o as h, d as i, E as o, F as V, H as X, I as Y, J as Z, N as ee, K as te, O as se, Q as ne, T as re, U as ue, i as L, z as ae, V as A, a as ie, W as oe, w as le, f as q, p as C, X as P, y as U, _ as H } from "../index-CWQLouWz.js";
-import { s as ce, r as f } from "../sum-CnIf1YOh.js";
-import { m } from "../mat_mul-4v7St11W.js";
-import { c as pe } from "../complex-x7w5HPOS.js";
+import { o as h, d as i, E as o, K as X, N as Y, O as Z, Q as J, T as ee, U as te, V as se, W as ne, X as re, Y as ue, l as L, I as ae, Z as A, a as ie, _ as oe, D as le, f as q, v as C, $ as P, H as U, a0 as H } from "../index-YPKosni4.js";
+import { r as f } from "../reshape-DmnmKT6r.js";
+import { s as ce } from "../sum-D7fu15XL.js";
+import { m } from "../mat_mul-Bu7bhLms.js";
+import { c as pe } from "../complex-CJ-qCcLB.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -20,7 +21,7 @@ import { c as pe } from "../complex-x7w5HPOS.js";
  */
 function he(t) {
   const s = { x: i(t, "x", "sigmoid", "float32") };
-  return o.runKernel(V, s);
+  return o.runKernel(X, s);
 }
 const fe = /* @__PURE__ */ h({ sigmoid_: he });
 /**
@@ -41,7 +42,7 @@ const fe = /* @__PURE__ */ h({ sigmoid_: he });
  */
 function de(t) {
   const s = { x: i(t, "x", "elu", "float32") };
-  return o.runKernel(X, s);
+  return o.runKernel(Y, s);
 }
 const me = /* @__PURE__ */ h({ elu_: de });
 /**
@@ -62,7 +63,7 @@ const me = /* @__PURE__ */ h({ elu_: de });
  */
 function ge(t) {
   const s = { input: i(t, "input", "imag") };
-  return o.runKernel(Y, s);
+  return o.runKernel(Z, s);
 }
 const $e = /* @__PURE__ */ h({ imag_: ge });
 /**
@@ -83,7 +84,7 @@ const $e = /* @__PURE__ */ h({ imag_: ge });
  */
 function xe(t, e = 0.2) {
   const n = { x: i(t, "x", "leakyRelu") }, r = { alpha: e };
-  return o.runKernel(Z, n, r);
+  return o.runKernel(J, n, r);
 }
 const ke = /* @__PURE__ */ h({ leakyRelu_: xe });
 /**
@@ -169,7 +170,7 @@ function Me(t) {
   const s = { x: i(t, "x", "relu") };
   return o.runKernel(ne, s);
 }
-const we = /* @__PURE__ */ h({ relu_: Me });
+const We = /* @__PURE__ */ h({ relu_: Me });
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
@@ -186,11 +187,11 @@ const we = /* @__PURE__ */ h({ relu_: Me });
  * limitations under the License.
  * =============================================================================
  */
-function We(t) {
+function we(t) {
   const s = { x: i(t, "x", "relu6") };
   return o.runKernel(re, s);
 }
-const ze = /* @__PURE__ */ h({ relu6_: We });
+const ze = /* @__PURE__ */ h({ relu6_: we });
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -273,7 +274,7 @@ function Te(t, e, s, n) {
   if (e === "linear")
     return t;
   if (e === "relu")
-    return we(t);
+    return We(t);
   if (e === "elu")
     return me(t);
   if (e === "relu6")
@@ -310,42 +311,42 @@ function Ne({ a: t, b: e, transposeA: s = !1, transposeB: n = !1, bias: r, activ
   }
   let u = i(t, "a", "fused matMul"), a = i(e, "b", "fused matMul");
   [u, a] = q(u, a);
-  const D = s ? u.shape[u.rank - 2] : u.shape[u.rank - 1], b = n ? a.shape[a.rank - 1] : a.shape[a.rank - 2], w = s ? u.shape[u.rank - 1] : u.shape[u.rank - 2], W = n ? a.shape[a.rank - 2] : a.shape[a.rank - 1], T = u.shape.slice(0, -2), y = a.shape.slice(0, -2), B = C(T), N = C(y);
+  const D = s ? u.shape[u.rank - 2] : u.shape[u.rank - 1], b = n ? a.shape[a.rank - 1] : a.shape[a.rank - 2], W = s ? u.shape[u.rank - 1] : u.shape[u.rank - 2], w = n ? a.shape[a.rank - 2] : a.shape[a.rank - 1], T = u.shape.slice(0, -2), y = a.shape.slice(0, -2), B = C(T), N = C(y);
   L(D === b, () => `Error in fused matMul: inner shapes (${D}) and (${b}) of Tensors with shapes ${u.shape} and ${a.shape} and transposeA=${s} and transposeB=${n} must match.`);
-  const O = P(u.shape.slice(0, -2), a.shape.slice(0, -2)).concat([w, W]), F = s ? f(u, [B, D, w]) : f(u, [B, w, D]), R = n ? f(a, [N, W, b]) : f(a, [N, b, W]);
+  const O = P(u.shape.slice(0, -2), a.shape.slice(0, -2)).concat([W, w]), F = s ? f(u, [B, D, W]) : f(u, [B, W, D]), R = n ? f(a, [N, w, b]) : f(a, [N, b, w]);
   let S;
   r != null && (S = i(r, "bias", "fused matMul"), [S] = q(S, u), P(O, S.shape));
-  let G;
-  l != null && (G = i(l, "prelu weights", "fused matMul"));
-  const I = (x, M) => {
+  let v;
+  l != null && (v = i(l, "prelu weights", "fused matMul"));
+  const G = (x, M) => {
     const [g, $, k, z] = M, d = Ae(f(x, k.shape), k, c);
     let K, _;
     if (!s && !n ? (K = m(d, $, !1, !0), _ = m(g, d, !0, !1)) : !s && n ? (K = m(d, $, !1, !1), _ = m(d, g, !0, !1)) : s && !n ? (K = m($, d, !1, !0), _ = m(g, d, !1, !1)) : (K = m($, d, !0, !0), _ = m(d, g, !0, !0)), r != null) {
-      const Q = Le(z, d);
-      return [K, _, Q];
+      const V = Le(z, d);
+      return [K, _, V];
     } else
       return [K, _];
-  }, v = {
+  }, I = {
     a: F,
     b: R,
     bias: S,
-    preluActivationWeights: G
+    preluActivationWeights: v
   }, j = { transposeA: s, transposeB: n, activation: c, leakyreluAlpha: p };
   return r == null ? U((M, g, $) => {
     const k = (
       // tslint:disable-next-line: no-unnecessary-type-assertion
-      o.runKernel(H, v, j)
+      o.runKernel(H, I, j)
     );
-    return $([M, g, k]), { value: f(k, O), gradFunc: I };
+    return $([M, g, k]), { value: f(k, O), gradFunc: G };
   })(F, R) : U((M, g, $, k) => {
     const z = (
       // tslint:disable-next-line: no-unnecessary-type-assertion
-      o.runKernel(H, v, j)
+      o.runKernel(H, I, j)
     );
-    return k([M, g, z, $]), { value: f(z, O), gradFunc: I };
+    return k([M, g, z, $]), { value: f(z, O), gradFunc: G };
   })(F, R, S);
 }
-const J = /* @__PURE__ */ h({ fusedMatMul_: Ne });
+const Q = /* @__PURE__ */ h({ fusedMatMul_: Ne });
 /**
  * @license
  * Copyright 2018 Google LLC
@@ -369,7 +370,7 @@ class E extends Error {
  * https://opensource.org/licenses/MIT.
  * =============================================================================
  */
-function Ge(t, e, s, n) {
+function ve(t, e, s, n) {
   if (t.rank < 2 || e.rank < 2)
     throw new E(`dot requires both inputs to be rank >= 2 but got x shape = ${t.shape} and y shape = ${e.shape}`);
   if (e.rank >= 3) {
@@ -378,7 +379,7 @@ function Ge(t, e, s, n) {
       throw new E(`If rank y >= 3, then the second last dim of y must equal the last dim of x but got x shape = ${t.shape} and  y shape = ${e.shape}`);
   }
   if (t.rank === 2 && e.rank === 2)
-    return J({
+    return Q({
       a: t,
       b: e,
       transposeA: !1,
@@ -392,7 +393,7 @@ function Ge(t, e, s, n) {
     const l = e.shape.slice(), p = l.pop(), u = l.pop(), a = [...l, p], D = Array.from({ length: e.rank }, (T, y) => y === 0 ? e.rank - 2 : y <= e.rank - 2 ? y - 1 : y);
     e = f(Re(e, D), [u, -1]);
     const b = [...r, ...a];
-    return f(J({
+    return f(Q({
       a: t,
       b: e,
       transposeA: !1,
@@ -402,7 +403,7 @@ function Ge(t, e, s, n) {
     }), b);
   }
 }
-class Pe {
+class Ue {
   vocabSize;
   embedDim;
   tf;
@@ -425,7 +426,7 @@ class Pe {
     return this.tf.gather(this.tiedWeights, e, 0);
   }
   project(e) {
-    return Ge(e, this.tiedWeights.transpose());
+    return ve(e, this.tiedWeights.transpose());
   }
   getWeights() {
     return [this.tiedWeights];
@@ -444,5 +445,5 @@ class Pe {
   }
 }
 export {
-  Pe as default
+  Ue as default
 };