@genai-fi/nanogpt 0.4.3 → 0.4.5

package/dist/layers/CausalSelfAttention.js

@@ -3,15 +3,15 @@ import T from "./BaseLayer.js";
 import { qkv as y } from "../ops/qkv.js";
 import { rope as w } from "../ops/rope.js";
 import { appendCache as E } from "../ops/appendCache.js";
-import { D as z, F as S, t as $, c as L, e as j, H as O } from "../index-C4JCoBvj.js";
+import { D as z, F as S, t as $, c as L, e as j, H as O } from "../index--6vO-cOz.js";
 import { fusedSoftmax as _ } from "../ops/fusedSoftmax.js";
-import { l as W, w as M, d as x } from "../tfjs_backend-Cug-PH75.js";
-import { o as N } from "../ones-Bf3YR48P.js";
-import { v as A } from "../variable-LJT9Ld63.js";
-import { z as q } from "../zeros-dnQxFgAD.js";
-import { r as C, d as I } from "../dropout-DfDdklfL.js";
-import { r as B } from "../reshape-Boe4DuIO.js";
-import { m as F } from "../mat_mul-415y5Qn2.js";
+import { l as W, w as M, d as x } from "../tfjs_backend-DuKis_xG.js";
+import { o as q } from "../ones-D6kB8bdY.js";
+import { v as b } from "../variable-BJTZ3jOy.js";
+import { z as B } from "../zeros-8xl-W2DC.js";
+import { r as C, d as I } from "../dropout-DFEXTPV0.js";
+import { r as F } from "../reshape-z51Eu-re.js";
+import { m as H } from "../mat_mul-BEHRPMh0.js";
 class nt extends T {
   cAttn = null;
   cProj = null;
@@ -23,16 +23,16 @@ class nt extends T {
   units;
   projUnits;
   constructor(t, s) {
-    super(s), this.index = t, this.units = s.gpt.nEmbed * 3, this.projUnits = s.gpt.nEmbed, this.bias = W.bandPart(N([s.gpt.blockSize, s.gpt.blockSize]), -1, 0).cast("bool"), this.divisor = 1 / Math.sqrt(s.gpt.nEmbed / s.gpt.nHead);
-    const o = q([s.gpt.blockSize, s.gpt.blockSize]), e = z([s.gpt.blockSize, s.gpt.blockSize], Number.NEGATIVE_INFINITY);
-    this.maskInf = M(this.bias, o, e);
+    super(s), this.index = t, this.units = s.gpt.nEmbed * 3, this.projUnits = s.gpt.nEmbed, this.bias = W.bandPart(q([s.gpt.blockSize, s.gpt.blockSize]), -1, 0).cast("bool"), this.divisor = 1 / Math.sqrt(s.gpt.nEmbed / s.gpt.nHead);
+    const e = B([s.gpt.blockSize, s.gpt.blockSize]), o = z([s.gpt.blockSize, s.gpt.blockSize], Number.NEGATIVE_INFINITY);
+    this.maskInf = M(this.bias, e, o);
   }
   build() {
-    this.cAttn === null && (this.cAttn = A(
+    this.cAttn === null && (this.cAttn = b(
       C([this.config.gpt.nEmbed, this.units], 0, 0.02),
       !0
       //`block_${this.index}_attn_cAttn_kernel`
-    )), this.cProj === null && (this.cProj = A(
+    )), this.cProj === null && (this.cProj = b(
       C([this.projUnits, this.config.gpt.nEmbed], 0, 0.02),
       !0
       //`block_${this.index}_attn_cProj_kernel`
@@ -53,57 +53,58 @@ class nt extends T {
     t.set(`block_${this.index}_cAttn`, this.cAttn ? [this.cAttn.clone()] : []), t.set(`block_${this.index}_cProj`, this.cProj ? [this.cProj.clone()] : []);
   }
   loadWeights(t) {
-    const s = t.get(`block_${this.index}_cAttn`)?.[0], o = t.get(`block_${this.index}_cProj`)?.[0];
+    const s = t.get(`block_${this.index}_cAttn`)?.[0], e = t.get(`block_${this.index}_cProj`)?.[0];
     if (!s) throw new Error(`Weights for block_${this.index}_cAttn not found`);
-    if (!o) throw new Error(`Weights for block_${this.index}_cProj not found`);
-    this.cAttn ? this.cAttn.assign(s) : this.cAttn = A(s, !0), this.cProj ? this.cProj.assign(o) : this.cProj = A(o, !0);
+    if (!e) throw new Error(`Weights for block_${this.index}_cProj not found`);
+    this.cAttn ? this.cAttn.assign(s) : this.cAttn = b(s, !0), this.cProj ? this.cProj.assign(e) : this.cProj = b(e, !0);
   }
-  getAttentionScores(t, s, o, e) {
+  getAttentionScores(t, s, e, o) {
     const i = P(t, s, this.divisor, this.maskInf);
-    return _(i, o ? this.config.gpt.dropout : 0, e);
+    return _(i, e ? this.config.gpt.dropout : 0, o);
   }
   // Attention with optional past. If pastLen > 0 and T_cur == 1, no mask needed.
-  getAttentionScoresWithPast(t, s, o) {
-    const e = P(t, s, this.divisor, void 0, o);
-    return _(e, 0, 0);
+  getAttentionScoresWithPast(t, s, e) {
+    const o = P(t, s, this.divisor, void 0, e);
+    return _(o, 0, 0);
   }
   getQKV(t) {
     return y(t, this.cAttn, this.config.gpt.nHead);
   }
   getOutputProjection(t) {
-    const s = t.shape[0], o = t.shape[2], e = this.config.gpt.nEmbed, i = t.transpose([0, 2, 1, 3]), n = B(i, [s, o, e]);
+    const s = t.shape[0], e = t.shape[2], o = this.config.gpt.nEmbed, i = t.transpose([0, 2, 1, 3]), n = F(i, [s, e, o]);
     return x(n, this.cProj);
   }
-  updateCache(t, s, o, e) {
-    const i = this.config.gpt.blockSize, n = t.shape[2], r = e?.length || 0, a = o ? t : E(t, i, r, e?.k), p = o ? s : E(s, i, r, e?.v);
-    return {
+  updateCache(t, s, e, o) {
+    const i = this.config.gpt.blockSize, n = t.shape[2], r = o?.length || 0, a = e ? t : E(t, i, r, o?.k);
+    e || (t.dispose(), o?.k.dispose());
+    const p = e ? s : E(s, i, r, o?.v);
+    return e || (s.dispose(), o?.v.dispose()), {
       k: S(a),
       v: S(p),
       length: Math.min(r + n, i),
-      cumulativeLength: e ? e.cumulativeLength + n : n
+      cumulativeLength: o ? o.cumulativeLength + n : n
     };
   }
-  forward(t, s = !1, o, e = !1, i) {
+  forward(t, s = !1, e, o = !1, i) {
     return $(() => {
       this.startMemory();
-      const [n, r, a] = this.getQKV(t), p = i ? i.cumulativeLength : 0, c = this.config.layerConfig.ropeCache, u = c ? w(n, c, p) : n, f = c ? w(r, c, p) : r;
+      const [n, r, a] = this.getQKV(t), p = i ? i.cumulativeLength : 0, c = this.config.layerConfig.ropeCache, u = c ? w(n, c, p) : n, A = c ? w(r, c, p) : r;
       c && (n.dispose(), r.dispose());
-      const g = i ? i.length : 0, d = this.updateCache(f, a, s, i), l = d.k, m = d.v;
-      i && (f.dispose(), a.dispose());
+      const f = i ? i.length : 0, d = this.updateCache(A, a, s, i), l = d.k, g = d.v;
       let h;
-      g > 0 ? h = this.getAttentionScoresWithPast(u, l, g) : h = this.getAttentionScores(u, l, s, o), u.dispose(), s && l.dispose();
-      const b = F(h, m);
-      e || h.dispose(), s && m.dispose();
-      const k = this.getOutputProjection(b);
-      b.dispose();
-      const v = e ? h.mean(1) : void 0;
+      f > 0 ? h = this.getAttentionScoresWithPast(u, l, f) : h = this.getAttentionScores(u, l, s, e), u.dispose(), s && l.dispose();
+      const m = H(h, g);
+      o || h.dispose(), s && g.dispose();
+      const k = this.getOutputProjection(m);
+      m.dispose();
+      const v = o ? h.mean(1) : void 0;
       return this.endMemory("CausalSelfAttention"), { output: k, attention: v, presentKV: s ? void 0 : d };
     });
   }
-  call(t, s = !1, o = !1, e) {
-    if (e && !this.config.gpt.useRope)
+  call(t, s = !1, e = !1, o) {
+    if (o && !this.config.gpt.useRope)
       throw new Error("Cannot use pastKV without RoPE enabled");
-    if (s && e)
+    if (s && o)
       throw new Error("Cannot use pastKV during training");
     if (t.shape.length !== 3)
       throw new Error(`Input tensor must be rank 3 [B, T, C], got shape ${t.shape}`);
@@ -115,15 +116,15 @@ class nt extends T {
       const r = L(
         // @ts-expect-error Invalid params
         (a, p, c, u) => {
-          const f = this.forward(a, !0, i);
+          const A = this.forward(a, !0, i);
           u([a]);
-          const g = (d, l) => {
-            const [m] = l, h = j().state.activeTape;
+          const f = (d, l) => {
+            const [g] = l, h = j().state.activeTape;
             j().state.activeTape = [];
-            const b = O((k, v, R) => this.forward(k, !0, i).output)([m, p, c], d);
-            return j().state.activeTape = h, b;
+            const m = O((k, v, R) => this.forward(k, !0, i).output)([g, p, c], d);
+            return j().state.activeTape = h, m;
           };
-          return { value: f.output, gradFunc: g };
+          return { value: A.output, gradFunc: f };
         }
       )(t, this.cAttn, this.cProj);
       if (this.config.gpt.dropout > 0) {
@@ -132,7 +133,7 @@ class nt extends T {
       } else
         return { output: r };
     } else {
-      const n = this.forward(t, s, i, o, e);
+      const n = this.forward(t, s, i, e, o);
       if (this.config.gpt.dropout > 0) {
         const r = I(n.output, this.config.gpt.dropout);
         return n.output.dispose(), { output: r, attention: n.attention, presentKV: n.presentKV };

package/dist/layers/MLP.js

@@ -1,11 +1,11 @@
-import { t as _, c as M, e as d, H as v } from "../index-C4JCoBvj.js";
-import x from "./BaseLayer.js";
-import { g as L } from "../gelu-CnCt17Lk.js";
-import { v as n } from "../variable-LJT9Ld63.js";
-import { r as p, d as u } from "../dropout-DfDdklfL.js";
-import { r as l } from "../reshape-Boe4DuIO.js";
-import { m as f } from "../mat_mul-415y5Qn2.js";
-class V extends x {
+import { t as F, c as _, e as h, H as M } from "../index--6vO-cOz.js";
+import v from "./BaseLayer.js";
+import { matMulGelu as x } from "../ops/matMulGelu.js";
+import { v as c } from "../variable-BJTZ3jOy.js";
+import { r as d, d as u } from "../dropout-DFEXTPV0.js";
+import { r as p } from "../reshape-z51Eu-re.js";
+import { m as L } from "../mat_mul-BEHRPMh0.js";
+class G extends v {
   cFc = null;
   cProj = null;
   index;
@@ -15,12 +15,12 @@ class V extends x {
     super(s), this.index = t, this.hiddenUnits = s.gpt.mlpFactor * s.gpt.nEmbed;
   }
   build() {
-    this.cFc === null && (this.cFc = n(
-      p([this.config.gpt.nEmbed, this.hiddenUnits], 0, 0.02),
+    this.cFc === null && (this.cFc = c(
+      d([this.config.gpt.nEmbed, this.hiddenUnits], 0, 0.02),
       !0
       //`block_${this.index}_attn_cAttn_kernel`
-    )), this.cProj === null && (this.cProj = n(
-      p(
+    )), this.cProj === null && (this.cProj = c(
+      d(
         [this.hiddenUnits, this.config.gpt.nEmbed],
         0,
         0.02 / Math.sqrt(2 * this.config.gpt.nLayer)
@@ -45,43 +45,41 @@ class V extends x {
     const s = t.get(`block_${this.index}_mlpOut`)?.[0], i = t.get(`block_${this.index}_mlpHidden`)?.[0];
     if (!s || !i)
       throw new Error(`Weights for block ${this.index} not found`);
-    this.cFc ? this.cFc.assign(i) : this.cFc = n(i, !0), this.cProj ? this.cProj.assign(s) : this.cProj = n(s, !0);
+    this.cFc ? this.cFc.assign(i) : this.cFc = c(i, !0), this.cProj ? this.cProj.assign(s) : this.cProj = c(s, !0);
   }
   forward(t) {
-    return _(() => {
+    return F(() => {
       this.startMemory();
-      const [s, i, o] = t.shape, r = l(t, [s * i, o]), e = f(r, this.cFc), c = L(e);
+      const [s, i, r] = t.shape, o = p(t, [s * i, r]), e = x(o, this.cFc), n = L(e, this.cProj);
       e.dispose();
-      const a = f(c, this.cProj);
-      c.dispose();
-      const h = l(a, [s, i, o]);
-      return this.endMemory("MLP"), h;
+      const a = p(n, [s, i, r]);
+      return this.endMemory("MLP"), a;
     });
   }
   call(t, s = !1) {
     if (this.build(), s && this.config.layerConfig.checkpointMLP) {
-      const o = M(
+      const r = _(
         // @ts-expect-error Invalid params
-        (r, e, c, a) => {
-          const h = this.forward(r);
-          return a([r]), { value: h, gradFunc: (g, m) => {
-            const [b] = m, P = d().state.activeTape;
-            d().state.activeTape = [];
-            const j = v((F, w, T) => this.forward(F))([b, e, c], g);
-            return d().state.activeTape = P, j;
+        (o, e, n, a) => {
+          const l = this.forward(o);
+          return a([o]), { value: l, gradFunc: (f, g) => {
+            const [m] = g, b = h().state.activeTape;
+            h().state.activeTape = [];
+            const P = M((j, w, T) => this.forward(j))([m, e, n], f);
+            return h().state.activeTape = b, P;
           } };
         }
       )(t, this.cFc, this.cProj);
       if (this.config.gpt.dropout > 0) {
-        const r = u(o, this.config.gpt.dropout);
-        return o.dispose(), r;
+        const o = u(r, this.config.gpt.dropout);
+        return r.dispose(), o;
       }
-      return o;
+      return r;
     } else {
       const i = this.forward(t);
       if (s && this.config.gpt.dropout > 0) {
-        const o = u(i, this.config.gpt.dropout);
-        return i.dispose(), o;
+        const r = u(i, this.config.gpt.dropout);
+        return i.dispose(), r;
       }
       return i;
     }
@@ -91,5 +89,5 @@ class V extends x {
   }
 }
 export {
-  V as default
+  G as default
 };

package/dist/layers/RMSNorm.d.ts

@@ -2,8 +2,7 @@ import { Tensor, Variable } from '@tensorflow/tfjs-core';
 import { default as BaseLayer, GPTLayerConfig } from './BaseLayer';
 export default class RMSNorm extends BaseLayer {
     private gamma;
-    private epsilon;
-    constructor(config: GPTLayerConfig, epsilon?: number, name?: string);
+    constructor(config: GPTLayerConfig, name?: string);
     get trainableWeights(): Variable[];
     set trainable(value: boolean);
     getWeights(): Tensor[];

package/dist/layers/RMSNorm.js

@@ -1,12 +1,12 @@
-import { t as r } from "../index-C4JCoBvj.js";
+import { t as r } from "../index--6vO-cOz.js";
 import m from "./BaseLayer.js";
-import { v as i } from "../variable-LJT9Ld63.js";
-import { o } from "../ones-Bf3YR48P.js";
-class d extends m {
+import { normRMS as s } from "../ops/normRMS.js";
+import { v as e } from "../variable-BJTZ3jOy.js";
+import { o as i } from "../ones-D6kB8bdY.js";
+class u extends m {
   gamma;
-  epsilon;
-  constructor(t, s = 1e-8, a = "") {
-    super(t), this.epsilon = s, this.gamma = i(o([t.gpt.nEmbed]), !0, `${a}_gamma`, "float32");
+  constructor(t, a = "") {
+    super(t), this.gamma = e(i([t.gpt.nEmbed]), !0, `${a}_gamma`, "float32");
   }
   get trainableWeights() {
     return [this.gamma];
@@ -23,8 +23,8 @@ class d extends m {
   apply(t) {
     return r(() => {
       this.startMemory();
-      const a = t.square().mean(-1, !0).add(this.epsilon).rsqrt(), e = t.mul(a).mul(this.gamma);
-      return this.endMemory("RMSNorm"), e;
+      const a = s(t, this.gamma);
+      return this.endMemory("RMSNorm"), a;
     });
   }
   dispose() {
@@ -32,5 +32,5 @@ class d extends m {
   }
 }
 export {
-  d as default
+  u as default
 };

package/dist/layers/RoPECache.js

@@ -1,6 +1,6 @@
-import { o as h, h as c, E as f, N as l, f as n, O as m, t as u, F as p } from "../index-C4JCoBvj.js";
-import { c as d, s as C } from "../sin-KmhiDuMa.js";
-import { r as a } from "../range-9AzeApCc.js";
+import { o as h, h as c, E as f, T as l, f as n, U as m, t as u, F as p } from "../index--6vO-cOz.js";
+import { c as d, s as C } from "../sin-H567uayl.js";
+import { r as a } from "../range-C_vpUjBu.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.

package/dist/layers/TiedEmbedding.js

@@ -1,8 +1,8 @@
-import { T as a } from "../TiedEmbedding-CnJ1bx4q.js";
-import "../index-C4JCoBvj.js";
-import "../tfjs_backend-Cug-PH75.js";
-import "../variable-LJT9Ld63.js";
-import "../gather-ZYRWhmXR.js";
+import { T as a } from "../TiedEmbedding-DznFwzcB.js";
+import "../index--6vO-cOz.js";
+import "../tfjs_backend-DuKis_xG.js";
+import "../variable-BJTZ3jOy.js";
+import "../gather-C5D8PxwA.js";
 export {
   a as default
 };

package/dist/layers/TransformerBlock.js

@@ -2,7 +2,7 @@ import h from "./CausalSelfAttention.js";
 import o from "./MLP.js";
 import a from "./RMSNorm.js";
 import p from "./BaseLayer.js";
-import { t as d } from "../index-C4JCoBvj.js";
+import { t as d } from "../index--6vO-cOz.js";
 class W extends p {
   ln1;
   attn;
@@ -12,7 +12,7 @@ class W extends p {
   _trainable = !0;
   skipped = !1;
   constructor(t, s) {
-    super(s), this.index = t, this.ln1 = new a(s, 1e-8, `block_${this.index}_rms1`), this.attn = new h(this.index, s), this.ln2 = new a(s, 1e-8, `block_${this.index}_rms2`), this.mlp = new o(this.index, s);
+    super(s), this.index = t, this.ln1 = new a(s, `block_${this.index}_rms1`), this.attn = new h(this.index, s), this.ln2 = new a(s, `block_${this.index}_rms2`), this.mlp = new o(this.index, s);
   }
   get variables() {
     return [

package/dist/{log_sum_exp-BswFnwOb.js → log_sum_exp-CiEy1aUe.js}

@@ -1,8 +1,8 @@
-import { o as r, h as p, E as u, a3 as h, a4 as E, Y as S, s as $, a5 as d } from "./index-C4JCoBvj.js";
-import { e as K } from "./axis_util-BgTGy5w8.js";
-import { m as T } from "./max-CP_9O2Yd.js";
-import { r as m } from "./reshape-Boe4DuIO.js";
-import { s as _ } from "./sum-R28pucR5.js";
+import { o as r, h as p, E as u, a6 as h, a7 as E, $, s as S, a8 as d } from "./index--6vO-cOz.js";
+import { e as K } from "./axis_util-QP0LdI1v.js";
+import { m as T } from "./max-BUShNgfh.js";
+import { r as m } from "./reshape-z51Eu-re.js";
+import { s as _ } from "./sum-DdkDf2MG.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -62,12 +62,12 @@ const w = /* @__PURE__ */ r({ log_: v });
  * =============================================================================
  */
 function A(s, n = null, o = !1) {
-  const a = p(s, "x", "logSumExp"), t = S(n, a.shape), x = T(
+  const a = p(s, "x", "logSumExp"), t = $(n, a.shape), x = T(
     a,
     t,
     !0
     /* keepDims */
-  ), i = $(a, x), l = N(i), f = _(l, t), c = w(f), e = d(m(x, c.shape), c);
+  ), i = S(a, x), l = N(i), f = _(l, t), c = w(f), e = d(m(x, c.shape), c);
   if (o) {
     const g = K(e.shape, t);
     return m(e, g);

package/dist/main.js

@@ -1,11 +1,11 @@
-import { default as k } from "./NanoGPTModel.js";
-import { default as L } from "./TeachableLLM.js";
-import { default as b } from "./tokeniser/CharTokeniser.js";
-import { default as w } from "./tokeniser/bpe.js";
-import { default as D } from "./utilities/waitForModel.js";
-import { default as F } from "./data/textLoader.js";
-import { estimateMemoryUsage as N, estimateParameterCount as R, estimateResources as j, estimateTrainingMemoryUsage as q, validateConfig as z } from "./utilities/parameters.js";
-import "./index-C4JCoBvj.js";
+import { default as E } from "./NanoGPTModel.js";
+import { default as G } from "./TeachableLLM.js";
+import { default as R } from "./tokeniser/CharTokeniser.js";
+import { default as q } from "./tokeniser/bpe.js";
+import { default as A } from "./utilities/waitForModel.js";
+import { default as I } from "./data/textLoader.js";
+import { estimateMemoryUsage as K, estimateParameterCount as O, estimateResources as Q, estimateTrainingMemoryUsage as S, validateConfig as V } from "./utilities/parameters.js";
+import "./index--6vO-cOz.js";
 import "./ops/cpu/scatterSub.js";
 import "./ops/webgl/scatterSub.js";
 import "./ops/cpu/gatherSub.js";
@@ -25,16 +25,25 @@ import "./ops/webgl/appendCache.js";
 import "./ops/cpu/fusedSoftmax.js";
 import "./ops/webgl/fusedSoftmax.js";
 import "./ops/grads/fusedSoftmax.js";
+import "./ops/cpu/matMulGelu.js";
+import "./ops/webgl/matMulGelu.js";
+import "./ops/grads/matMulGelu.js";
+import "./ops/cpu/gelu.js";
+import "./ops/webgl/gelu.js";
+import "./ops/grads/gelu.js";
+import "./ops/cpu/normRMS.js";
+import "./ops/webgl/normRMS.js";
+import "./ops/grads/normRMS.js";
 export {
-  w as BPETokeniser,
-  b as CharTokeniser,
-  k as NanoGPT,
-  L as TeachableLLM,
-  N as estimateMemoryUsage,
-  R as estimateParameterCount,
-  j as estimateResources,
-  q as estimateTrainingMemoryUsage,
-  F as loadTextData,
-  z as validateConfig,
-  D as waitForModel
+  q as BPETokeniser,
+  R as CharTokeniser,
+  E as NanoGPT,
+  G as TeachableLLM,
+  K as estimateMemoryUsage,
+  O as estimateParameterCount,
+  Q as estimateResources,
+  S as estimateTrainingMemoryUsage,
+  I as loadTextData,
+  V as validateConfig,
+  A as waitForModel
 };

package/dist/{mat_mul-415y5Qn2.js → mat_mul-BEHRPMh0.js}

@@ -1,4 +1,4 @@
-import { o as m, h as s, p as c, E as M, B as p } from "./index-C4JCoBvj.js";
+import { o as m, h as s, p as c, E as M, B as p } from "./index--6vO-cOz.js";
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.

package/dist/{max-CP_9O2Yd.js → max-BUShNgfh.js}

@@ -1,4 +1,4 @@
-import { o as r, h as e, E as x, M as c } from "./index-C4JCoBvj.js";
+import { o as r, h as e, E as x, M as c } from "./index--6vO-cOz.js";
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.

package/dist/{moments-CjeIaVdp.js → moments-DYOHXoRV.js}

@@ -1,6 +1,6 @@
-import { o as m, h as c, E as f, X as i, Y as l, Z as h, s as x, x as d } from "./index-C4JCoBvj.js";
-import { e as v } from "./axis_util-BgTGy5w8.js";
-import { r as E } from "./reshape-Boe4DuIO.js";
+import { o as m, h as c, E as f, _ as i, $ as l, a0 as h, s as x, x as d } from "./index--6vO-cOz.js";
+import { e as v } from "./axis_util-QP0LdI1v.js";
+import { r as E } from "./reshape-z51Eu-re.js";
 /**
  * @license
  * Copyright 2020 Google Inc. All Rights Reserved.
@@ -46,8 +46,8 @@ function T(a, t = null, e = !1) {
   const p = h(x(d(a, "float32"), E(s, o))), u = r(p, n, e);
   return { mean: s, variance: u };
 }
-const N = /* @__PURE__ */ m({ moments_: T });
+const K = /* @__PURE__ */ m({ moments_: T });
 export {
-  N as a,
+  K as a,
   r as m
 };

package/dist/{norm-CZM380I3.js → norm-DSva3hI3.js}

@@ -1,8 +1,8 @@
-import { o as l, h as c, E as y, _ as E, Y as w, $ as o, a0 as u, O as v, f as I, Z as $ } from "./index-C4JCoBvj.js";
-import { e as A } from "./axis_util-BgTGy5w8.js";
-import { m as f } from "./max-CP_9O2Yd.js";
-import { r as h } from "./reshape-Boe4DuIO.js";
-import { s as t } from "./sum-R28pucR5.js";
+import { o as l, h as c, E as y, a1 as E, $ as w, a2 as o, a3 as u, U as v, f as I, a0 as $ } from "./index--6vO-cOz.js";
+import { e as A } from "./axis_util-QP0LdI1v.js";
+import { m as f } from "./max-BUShNgfh.js";
+import { r as h } from "./reshape-z51Eu-re.js";
+import { s as t } from "./sum-DdkDf2MG.js";
 /**
  * @license
  * Copyright 2020 Google Inc. All Rights Reserved.
@@ -20,8 +20,8 @@ import { s as t } from "./sum-R28pucR5.js";
  * =============================================================================
  */
 function k(n, e = null, r = !1) {
-  const i = { x: c(n, "x", "min") }, a = { axis: e, keepDims: r };
-  return y.runKernel(E, i, a);
+  const a = { x: c(n, "x", "min") }, i = { axis: e, keepDims: r };
+  return y.runKernel(E, a, i);
 }
 const s = /* @__PURE__ */ l({ min_: k });
 /**
@@ -42,13 +42,13 @@ const s = /* @__PURE__ */ l({ min_: k });
  */
 function T(n, e = "euclidean", r = null, m = !1) {
   n = c(n, "x", "norm");
-  const i = d(n, e, r);
-  let a = i.shape;
+  const a = d(n, e, r);
+  let i = a.shape;
   if (m) {
     const p = w(r, n.shape);
-    a = A(i.shape, p);
+    i = A(a.shape, p);
   }
-  return h(i, a);
+  return h(a, i);
 }
 function d(n, e, r = null) {
   if (n.rank === 0)
@@ -79,8 +79,8 @@ function d(n, e, r = null) {
   }
   throw new Error(`Error in norm: invalid axis: ${r}`);
 }
-const K = /* @__PURE__ */ l({ norm_: T });
+const N = /* @__PURE__ */ l({ norm_: T });
 export {
   s as m,
-  K as n
+  N as n
 };

package/dist/{ones-Bf3YR48P.js → ones-D6kB8bdY.js}

@@ -1,5 +1,5 @@
-import { k as n, l as t, n as m, E as i } from "./index-C4JCoBvj.js";
-import { z as l, c } from "./zeros-dnQxFgAD.js";
+import { k as n, l as t, n as m, E as i } from "./index--6vO-cOz.js";
+import { z as l, c } from "./zeros-8xl-W2DC.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.

package/dist/ops/appendCache.js

@@ -1,8 +1,8 @@
-import { e as a } from "../index-C4JCoBvj.js";
+import { e as a } from "../index--6vO-cOz.js";
 import "./cpu/appendCache.js";
 import "./webgl/appendCache.js";
-import { z as s } from "../zeros-dnQxFgAD.js";
-import { c } from "../concat-CuRsVY-K.js";
+import { z as s } from "../zeros-8xl-W2DC.js";
+import { c } from "../concat-DvWM7HGZ.js";
 function i(r, p, n, o) {
   if (!o) {
     const e = r.shape[2];

package/dist/ops/attentionMask.js

@@ -1,4 +1,4 @@
-import { e as i } from "../index-C4JCoBvj.js";
+import { e as i } from "../index--6vO-cOz.js";
 import "./cpu/attentionMask.js";
 import "./webgl/attentionMask.js";
 import "./grads/attentionMask.js";

package/dist/ops/cpu/appendCache.js

@@ -1,5 +1,5 @@
-import { r as d } from "../../index-C4JCoBvj.js";
-import { c as h } from "../../concat-CuRsVY-K.js";
+import { r as d } from "../../index--6vO-cOz.js";
+import { c as h } from "../../concat-DvWM7HGZ.js";
 function u(p) {
   const { cache: n, item: s } = p.inputs, { maxSize: r, pastLen: c } = p.attrs, t = n.shape[0], o = n.shape[1], a = n.shape[3], e = s.shape[2];
   if (c + e <= r) {

package/dist/ops/cpu/attentionMask.js

@@ -1,5 +1,5 @@
-import { r as o, f as k } from "../../index-C4JCoBvj.js";
-import { m as d } from "../../mat_mul-415y5Qn2.js";
+import { r as o, f as k } from "../../index--6vO-cOz.js";
+import { m as d } from "../../mat_mul-BEHRPMh0.js";
 function r(t) {
   const { q: e, k: n, mask: s } = t.inputs, { divisor: c } = t.attrs, m = e.shape[2], i = n.shape[2], a = d(e, n, !1, !0).mul(k(c));
   if (s) {

package/dist/ops/cpu/fusedSoftmax.js

@@ -1,5 +1,5 @@
-import { r as n } from "../../index-C4JCoBvj.js";
-import { s as f } from "../../softmax-Cujsg4ay.js";
+import { r as n } from "../../index--6vO-cOz.js";
+import { s as f } from "../../softmax-Dsxflvdl.js";
 function r(t) {
   const { inputs: s, attrs: i } = t, { logits: o } = s, { dim: a, dropoutRate: e } = i;
   if (!o)

package/dist/ops/cpu/gatherSub.js

@@ -1,6 +1,6 @@
-import { o as u, h as c, E as g, L as h, r as m, s as p } from "../../index-C4JCoBvj.js";
-import { r as l } from "../../range-9AzeApCc.js";
-import { s as N } from "../../stack-D1YjmgKN.js";
+import { o as u, h as c, E as g, N as h, r as m, s as p } from "../../index--6vO-cOz.js";
+import { r as N } from "../../range-C_vpUjBu.js";
+import { s as l } from "../../stack-CmqSdsfs.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -23,7 +23,7 @@ function f(e, s) {
 }
 const b = /* @__PURE__ */ u({ gatherND_: f });
 function d(e) {
-  const { values: s, labels: n, logits: r } = e.inputs, t = n.shape[0], o = l(0, t, 1, "int32"), a = N([o, n], 1), i = b(r, a);
+  const { values: s, labels: n, logits: r } = e.inputs, t = n.shape[0], o = N(0, t, 1, "int32"), a = l([o, n], 1), i = b(r, a);
   return p(s, i);
 }
 const k = {

package/dist/ops/cpu/gelu.js

@@ -1,4 +1,4 @@
-import { r as t, t as d } from "../../index-C4JCoBvj.js";
+import { r as t, t as d } from "../../index--6vO-cOz.js";
 const o = 0.7978845608028654, c = 0.044715;
 function m(u) {
   const { inputs: l } = u, { x: e } = l, n = e;

package/dist/ops/cpu/matMulGelu.d.ts

@@ -0,0 +1 @@
+export {};