@genai-fi/nanogpt 0.3.2 → 0.4.1

package/dist/{sum-C7Mgy9Bw.js → log_sum_exp-DJPkVZZn.js}

@@ -1,7 +1,11 @@
-import { o as c, h as i, E as a, T as f, U as m, V as p } from "./index-pWA4_lUh.js";
+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-XjBAhiFO.js";
+import { e as K } from "./axis_util-DeydwOoC.js";
+import { m as T } from "./max-DJvEiCAJ.js";
+import { r as m } from "./reshape-DFzh97Sc.js";
+import { s as _ } from "./sum-C6u3xMi3.js";
 /**
  * @license
- * Copyright 2017 Google LLC. All Rights Reserved.
+ * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
  * You may obtain a copy of the License at
@@ -15,44 +19,14 @@ import { o as c, h as i, E as a, T as f, U as m, V as p } from "./index-pWA4_lUh
  * limitations under the License.
  * =============================================================================
  */
-function d(e, n) {
-  for (let t = 0; t < e.length; ++t)
-    if (e[e.length - t - 1] !== n - 1 - t)
-      return !1;
-  return !0;
-}
-function x(e, n, t) {
-  const s = e.length + n.length, r = [];
-  let o = 0, l = 0;
-  for (let u = 0; u < s; u++)
-    t.indexOf(u) === -1 ? r.push(e[o++]) : r.push(n[l++]);
-  return r;
-}
-function S(e, n) {
-  const t = [], s = e.length;
-  for (let o = 0; o < s; o++)
-    n.indexOf(o) === -1 && t.push(e[o]);
-  const r = n.map((o) => e[o]);
-  return [t, r];
-}
-function A(e, n) {
-  const t = n.map((s) => 1);
-  return x(e, t, n);
-}
-function E(e, n) {
-  if (d(e, n))
-    return null;
-  const t = [];
-  for (let s = 0; s < n; ++s)
-    e.indexOf(s) === -1 && t.push(s);
-  return e.forEach((s) => t.push(s)), t;
-}
-function O(e) {
-  return e.map((n, t) => [t, n]).sort((n, t) => n[1] - t[1]).map((n) => n[0]);
+function b(s) {
+  const o = { x: p(s, "x", "exp") };
+  return u.runKernel(h, o);
 }
+const N = /* @__PURE__ */ r({ exp_: b });
 /**
  * @license
- * Copyright 2020 Google LLC. All Rights Reserved.
+ * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
  * You may obtain a copy of the License at
@@ -66,14 +40,14 @@ function O(e) {
  * limitations under the License.
  * =============================================================================
  */
-function h(e, n = null, t = !1) {
-  const r = { x: i(e, "x", "max") }, o = { reductionIndices: n, keepDims: t };
-  return a.runKernel(f, r, o);
+function v(s) {
+  const o = { x: p(s, "x", "log", "float32") };
+  return u.runKernel(E, o);
 }
-const T = /* @__PURE__ */ c({ max_: h });
+const w = /* @__PURE__ */ r({ log_: v });
 /**
  * @license
- * Copyright 2018 Google LLC. All Rights Reserved.
+ * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
  * You may obtain a copy of the License at
@@ -87,18 +61,22 @@ const T = /* @__PURE__ */ c({ max_: h });
  * limitations under the License.
  * =============================================================================
  */
-function g(e, n = null, t = !1) {
-  let s = i(e, "x", "sum");
-  s.dtype === "bool" && (s = m(s, "int32"));
-  const r = { x: s }, o = { axis: n, keepDims: t };
-  return a.runKernel(p, r, o);
+function A(s, n = null, o = !1) {
+  const a = p(s, "x", "logSumExp"), t = S(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);
+  if (o) {
+    const g = K(e.shape, t);
+    return m(e, g);
+  }
+  return e;
 }
-const b = /* @__PURE__ */ c({ sum_: g });
+const P = /* @__PURE__ */ r({ logSumExp_: A });
 export {
-  O as a,
-  S as c,
-  A as e,
-  E as g,
-  T as m,
-  b as s
+  w as a,
+  N as e,
+  P as l
 };

package/dist/main.js

@@ -1,11 +1,11 @@
-import { default as P } from "./NanoGPTModel.js";
-import { default as h } from "./TeachableLLM.js";
-import { default as y } from "./tokeniser/CharTokeniser.js";
-import { default as U } from "./tokeniser/bpe.js";
-import { default as v } from "./utilities/waitForModel.js";
-import { default as B } from "./data/textLoader.js";
-import { estimateMemoryUsage as E, estimateParameterCount as F, estimateResources as G, estimateTrainingMemoryUsage as N, validateConfig as R } from "./utilities/parameters.js";
-import "./index-pWA4_lUh.js";
+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-XjBAhiFO.js";
 import "./ops/cpu/scatterSub.js";
 import "./ops/webgl/scatterSub.js";
 import "./ops/cpu/gatherSub.js";
@@ -22,16 +22,19 @@ import "./ops/webgl/rope.js";
 import "./ops/grads/rope.js";
 import "./ops/cpu/appendCache.js";
 import "./ops/webgl/appendCache.js";
+import "./ops/cpu/fusedSoftmax.js";
+import "./ops/webgl/fusedSoftmax.js";
+import "./ops/grads/fusedSoftmax.js";
 export {
-  U as BPETokeniser,
-  y as CharTokeniser,
-  P as NanoGPT,
-  h as TeachableLLM,
-  E as estimateMemoryUsage,
-  F as estimateParameterCount,
-  G as estimateResources,
-  N as estimateTrainingMemoryUsage,
-  B as loadTextData,
-  R as validateConfig,
-  v as waitForModel
+  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
 };

package/dist/{mat_mul-D7_a4KJn.js → mat_mul-CKwFEV1Q.js}

@@ -1,4 +1,4 @@
-import { o as m, h as s, p as c, E as M, B as p } from "./index-pWA4_lUh.js";
+import { o as m, h as s, p as c, E as M, B as p } from "./index-XjBAhiFO.js";
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.

package/dist/max-DJvEiCAJ.js

@@ -0,0 +1,25 @@
+import { o as r, h as e, E as x, M as c } from "./index-XjBAhiFO.js";
+/**
+ * @license
+ * Copyright 2020 Google LLC. All Rights Reserved.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ * =============================================================================
+ */
+function m(n, o = null, s = !1) {
+  const t = { x: e(n, "x", "max") }, a = { reductionIndices: o, keepDims: s };
+  return x.runKernel(c, t, a);
+}
+const l = /* @__PURE__ */ r({ max_: m });
+export {
+  l as m
+};

package/dist/moments-CrWRPcR3.js

@@ -0,0 +1,53 @@
+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-XjBAhiFO.js";
+import { e as v } from "./axis_util-DeydwOoC.js";
+import { r as E } from "./reshape-DFzh97Sc.js";
+/**
+ * @license
+ * Copyright 2020 Google Inc. All Rights Reserved.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ * =============================================================================
+ */
+function S(a, t = null, e = !1) {
+  const s = { x: c(a, "x", "mean") }, o = { axis: t, keepDims: e };
+  return f.runKernel(i, s, o);
+}
+const r = /* @__PURE__ */ m({ mean_: S });
+/**
+ * @license
+ * Copyright 2020 Google LLC. All Rights Reserved.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ * =============================================================================
+ */
+function T(a, t = null, e = !1) {
+  a = c(a, "x", "moments");
+  const n = l(t, a.shape), s = r(a, n, e);
+  let o = s.shape;
+  e || (o = v(s.shape, n));
+  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 });
+export {
+  N as a,
+  r as m
+};

package/dist/norm-BzY929B_.js

@@ -0,0 +1,86 @@
+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-XjBAhiFO.js";
+import { e as A } from "./axis_util-DeydwOoC.js";
+import { m as f } from "./max-DJvEiCAJ.js";
+import { r as h } from "./reshape-DFzh97Sc.js";
+import { s as t } from "./sum-C6u3xMi3.js";
+/**
+ * @license
+ * Copyright 2020 Google Inc. All Rights Reserved.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ * =============================================================================
+ */
+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 s = /* @__PURE__ */ l({ min_: k });
+/**
+ * @license
+ * Copyright 2018 Google LLC. All Rights Reserved.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ * =============================================================================
+ */
+function T(n, e = "euclidean", r = null, m = !1) {
+  n = c(n, "x", "norm");
+  const i = d(n, e, r);
+  let a = i.shape;
+  if (m) {
+    const p = w(r, n.shape);
+    a = A(i.shape, p);
+  }
+  return h(i, a);
+}
+function d(n, e, r = null) {
+  if (n.rank === 0)
+    return o(n);
+  if (n.rank !== 1 && r === null)
+    return d(h(n, [-1]), e, r);
+  if (n.rank === 1 || typeof r == "number" || Array.isArray(r) && r.length === 1) {
+    if (e === 1)
+      return t(o(n), r);
+    if (e === 1 / 0)
+      return f(o(n), r);
+    if (e === -1 / 0)
+      return s(o(n), r);
+    if (e === "euclidean" || e === 2)
+      return u(t(v(o(n), I(2, "int32")), r));
+    throw new Error(`Error in norm: invalid ord value: ${e}`);
+  }
+  if (Array.isArray(r) && r.length === 2) {
+    if (e === 1)
+      return f(t(o(n), r[0]), r[1] - 1);
+    if (e === 1 / 0)
+      return f(t(o(n), r[1]), r[0]);
+    if (e === -1 / 0)
+      return s(t(o(n), r[1]), r[0]);
+    if (e === "fro" || e === "euclidean")
+      return u(t($(n), r));
+    throw new Error(`Error in norm: invalid ord value: ${e}`);
+  }
+  throw new Error(`Error in norm: invalid axis: ${r}`);
+}
+const K = /* @__PURE__ */ l({ norm_: T });
+export {
+  s as m,
+  K as n
+};

package/dist/{ones-Cog-G2ag.js → ones-BO01zpJG.js}

@@ -1,5 +1,5 @@
-import { k as n, l as t, n as m, E as i } from "./index-pWA4_lUh.js";
-import { z as l, c } from "./zeros-CCy9C3uU.js";
+import { k as n, l as t, n as m, E as i } from "./index-XjBAhiFO.js";
+import { z as l, c } from "./zeros-CmJFiC84.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.

package/dist/ops/appendCache.js

@@ -1,4 +1,4 @@
-import { e as p } from "../index-pWA4_lUh.js";
+import { e as p } from "../index-XjBAhiFO.js";
 import "./cpu/appendCache.js";
 import "./webgl/appendCache.js";
 function a(e, n, r) {

package/dist/ops/attentionMask.js

@@ -1,4 +1,4 @@
-import { e as o } from "../index-pWA4_lUh.js";
+import { e as o } from "../index-XjBAhiFO.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 a, e as m } from "../../index-pWA4_lUh.js";
-import { c as d } from "../../concat-BIZS_td9.js";
+import { r as a, e as m } from "../../index-XjBAhiFO.js";
+import { c as d } from "../../concat-DS_qH7MI.js";
 function r(n) {
   const { cache: c, item: t } = n.inputs, { maxSize: o } = n.attrs, e = d([c, t], 2), s = e.shape[2];
   if (s > o) {

package/dist/ops/cpu/attentionMask.js

@@ -1,5 +1,5 @@
-import { r as s, f as i } from "../../index-pWA4_lUh.js";
-import { m as k } from "../../mat_mul-D7_a4KJn.js";
+import { r as s, f as i } from "../../index-XjBAhiFO.js";
+import { m as k } from "../../mat_mul-CKwFEV1Q.js";
 function a(t) {
   const { q: e, k: o, mask: r } = t.inputs, { divisor: c } = t.attrs, n = e.shape[2], m = k(e, o, !1, !0).mul(i(c)), l = r.slice([0, 0], [n, n]).expandDims(0).expandDims(0);
   return m.add(l);

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

@@ -0,0 +1,9 @@
+import { SoftmaxAttrs, SoftmaxInputs, TensorInfo } from '@tensorflow/tfjs-core';
+interface FusedSoftmaxAttrs extends SoftmaxAttrs {
+    dropoutRate?: number;
+}
+export declare function softmaxCPU(args: {
+    inputs: SoftmaxInputs;
+    attrs: FusedSoftmaxAttrs;
+}): TensorInfo;
+export {};

package/dist/ops/cpu/fusedSoftmax.js

@@ -0,0 +1,23 @@
+import { r as n } from "../../index-XjBAhiFO.js";
+import { s as f } from "../../softmax-4DOn6cPq.js";
+function r(t) {
+  const { inputs: s, attrs: i } = t, { logits: o } = s, { dim: a, dropoutRate: e } = i;
+  if (!o)
+    throw new Error("Error in softmax: input logits is null");
+  return e !== void 0 && e > 0 && console.warn("Dropout in fusedSoftmax not implemented for CPU backend, skipping dropout."), f(o, a);
+}
+const m = {
+  kernelName: "FusedSoftmax",
+  backendName: "cpu",
+  kernelFunc: r
+};
+n(m);
+const l = {
+  kernelName: "FusedSoftmax",
+  backendName: "tensorflow",
+  kernelFunc: r
+};
+n(l);
+export {
+  r as softmaxCPU
+};

package/dist/ops/cpu/gatherSub.js

@@ -1,6 +1,6 @@
-import { o as u, h as c, E as g, H as h, r as m, s as p } from "../../index-pWA4_lUh.js";
-import { r as l } from "../../range-CcDl05lo.js";
-import { s as N } from "../../stack-BMm-efee.js";
+import { o as u, h as c, E as g, L as h, r as m, s as p } from "../../index-XjBAhiFO.js";
+import { r as l } from "../../range-DQMNzBWs.js";
+import { s as N } from "../../stack-DaIMO5iX.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.

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

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

package/dist/ops/cpu/mulDropout.js

@@ -0,0 +1,17 @@
+import { r as e, b as u } from "../../index-XjBAhiFO.js";
+function n(o) {
+  const { inputs: r } = o, { a: l, b: t } = r;
+  return console.warn("Using fallback mulDrop implementation without dropout."), u(l, t);
+}
+const a = {
+  kernelName: "MulDropout",
+  backendName: "cpu",
+  kernelFunc: n
+};
+e(a);
+const c = {
+  kernelName: "MulDropout",
+  backendName: "tensorflow",
+  kernelFunc: n
+};
+e(c);

package/dist/ops/cpu/qkv.js

@@ -1,6 +1,6 @@
-import { r as q } from "../../index-pWA4_lUh.js";
-import { r as o } from "../../reshape-C8CR_Bad.js";
-import { s as x } from "../../split-DZbvruEP.js";
+import { r as q } from "../../index-XjBAhiFO.js";
+import { r as o } from "../../reshape-DFzh97Sc.js";
+import { s as x } from "../../split-CkbeVdF8.js";
 function v(p) {
   const { x: c, kernel: K } = p.inputs, { heads: n } = p.attrs, [s, e, t] = c.shape, a = o(c, [s * e, t]), i = a.dot(K);
   a.dispose();

package/dist/ops/cpu/rope.js

@@ -1,8 +1,8 @@
-import { r as S } from "../../index-pWA4_lUh.js";
-import { r as F } from "../../range-CcDl05lo.js";
-import { g as I } from "../../gather-BPGW8RsB.js";
-import { s as E } from "../../stack-BMm-efee.js";
-import { c as T } from "../../concat-BIZS_td9.js";
+import { r as S } from "../../index-XjBAhiFO.js";
+import { r as F } from "../../range-DQMNzBWs.js";
+import { g as I } from "../../gather-BUmJIS8n.js";
+import { s as E } from "../../stack-DaIMO5iX.js";
+import { c as T } from "../../concat-DS_qH7MI.js";
 function U(t, c, p, o, r) {
   const n = o.shape[3], s = p;
   if (s > n) return o;

package/dist/ops/cpu/scatterSub.js

@@ -1,27 +1,27 @@
-import { o as l, k, h, E as g, Q as w, r as $, s as d, b as m } from "../../index-pWA4_lUh.js";
-import { r as b } from "../../range-CcDl05lo.js";
-import { s as E } from "../../stack-BMm-efee.js";
-import { o as D } from "../../ones-Cog-G2ag.js";
-function N(a, r, t) {
-  const s = r.rank > 1 ? r.shape[r.rank - 1] : 1, e = r.rank > 1 ? r.rank - 1 : 1, o = `Must have updates.shape = indices.shape[:batchDim] + shape[sliceDim:], got updates.shape: ${t.shape}, indices.shape: ${r.shape}, shape: ${a}, sliceDim: ${s}, and batchDim: ${e}.`;
+import { o as l, k, h, E as g, a2 as w, r as $, s as d, b as m } from "../../index-XjBAhiFO.js";
+import { r as b } from "../../range-DQMNzBWs.js";
+import { s as E } from "../../stack-DaIMO5iX.js";
+import { o as D } from "../../ones-BO01zpJG.js";
+function N(n, r, t) {
+  const s = r.rank > 1 ? r.shape[r.rank - 1] : 1, e = r.rank > 1 ? r.rank - 1 : 1, o = `Must have updates.shape = indices.shape[:batchDim] + shape[sliceDim:], got updates.shape: ${t.shape}, indices.shape: ${r.shape}, shape: ${n}, sliceDim: ${s}, and batchDim: ${e}.`;
   if (t.rank < e)
     throw new Error(o + ` update.rank < ${e}. `);
-  if (a.length < s + (t.rank - e))
+  if (n.length < s + (t.rank - e))
     throw new Error(o + ` Output shape length < ${s + (t.rank - e)}`);
-  if (t.rank !== e + a.length - s)
-    throw new Error(o + ` update.rank != ${e + a.length - s}`);
-  for (let n = 0; n < e; ++n)
-    if (t.shape[n] !== r.shape[n])
-      throw new Error(o + ` updates.shape[${n}] (${t.shape[n]}) != indices.shape[${n}] (${r.shape[n]}).`);
-  for (let n = 0; n < t.rank - e; ++n)
-    if (t.shape[n + e] !== a[n + s])
-      throw new Error(o + ` updates.shape[${n + e}] (${t.shape[n + e]}) != shape[${n + e}] (${a[n + e]})`);
+  if (t.rank !== e + n.length - s)
+    throw new Error(o + ` update.rank != ${e + n.length - s}`);
+  for (let a = 0; a < e; ++a)
+    if (t.shape[a] !== r.shape[a])
+      throw new Error(o + ` updates.shape[${a}] (${t.shape[a]}) != indices.shape[${a}] (${r.shape[a]}).`);
+  for (let a = 0; a < t.rank - e; ++a)
+    if (t.shape[a + e] !== n[a + s])
+      throw new Error(o + ` updates.shape[${a + e}] (${t.shape[a + e]}) != shape[${a + e}] (${n[a + e]})`);
 }
-function S(a, r, t) {
+function S(n, r, t) {
   if (r.rank < 1)
     throw new Error(`tf.scatterND() expects the indices to be rank 1 or higher, but the rank was ${r.rank}.`);
-  if (a.rank < 1)
-    throw new Error(`tf.scatterND() expects the updates to be rank 1 or higher, but the rank was ${a.rank}.`);
+  if (n.rank < 1)
+    throw new Error(`tf.scatterND() expects the updates to be rank 1 or higher, but the rank was ${n.rank}.`);
   if (r.dtype !== "int32")
     throw new Error(`The dtype of 'indices' should be int32, but got dtype: ${r.dtype}`);
   if (t.length < 1)
@@ -29,10 +29,10 @@ function S(a, r, t) {
   if (t.length === 0) {
     if (r.size === 0)
       throw new Error(`Indices specified for empty output. indices shape: ${r.shape}`);
-    if (a.size === 0)
-      throw new Error(`Updates specified for empty output. updates shape: ${a.shape}`);
+    if (n.size === 0)
+      throw new Error(`Updates specified for empty output. updates shape: ${n.shape}`);
   }
-  N(t, r, a);
+  N(t, r, n);
 }
 /**
  * @license
@@ -50,16 +50,16 @@ function S(a, r, t) {
  * limitations under the License.
  * =============================================================================
  */
-function y(a, r, t) {
+function y(n, r, t) {
   k(t);
-  const s = h(a, "indices", "scatterND", "int32"), e = h(r, "updates", "scatterND");
+  const s = h(n, "indices", "scatterND", "int32"), e = h(r, "updates", "scatterND");
   S(e, s, t);
-  const o = { indices: s, updates: e }, n = { shape: t };
-  return g.runKernel(w, o, n);
+  const o = { indices: s, updates: e }, a = { shape: t };
+  return g.runKernel(w, o, a);
 }
 const v = /* @__PURE__ */ l({ scatterND_: y });
-function I(a) {
-  const { logits: r, labels: t, dy: s } = a.inputs, e = t.shape[0], o = r.shape[1], n = b(0, e, 1, "int32"), i = E([n, t], 1), c = D([e]), p = v(i, c, [e, o]), f = d(r, p), u = s.reshape([e, 1]);
+function I(n) {
+  const { logits: r, labels: t, dy: s } = n.inputs, e = t.shape[0], o = r.shape[1], a = b(0, e, 1, "int32"), i = E([a, t], 1), c = D([e]), p = v(i, c, [e, o]), f = d(r, p), u = s.reshape([e, 1]);
   return m(f, u);
 }
 const T = {

package/dist/ops/fusedSoftmax.d.ts

@@ -0,0 +1,2 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function fusedSoftmax(logits: Tensor, dropoutRate: number, seed: number): Tensor;

package/dist/ops/fusedSoftmax.js

@@ -0,0 +1,10 @@
+import { e as t } from "../index-XjBAhiFO.js";
+import "./cpu/fusedSoftmax.js";
+import "./webgl/fusedSoftmax.js";
+import "./grads/fusedSoftmax.js";
+function p(r, e, o) {
+  return t().runKernel("FusedSoftmax", { logits: r }, { dim: -1, dropoutRate: e, seed: o });
+}
+export {
+  p as fusedSoftmax
+};

package/dist/ops/gatherSub.js

@@ -1,4 +1,4 @@
-import { e as n } from "../index-pWA4_lUh.js";
+import { e as n } from "../index-XjBAhiFO.js";
 import "./cpu/gatherSub.js";
 import "./webgl/gatherSub.js";
 function f(r, e, t) {

package/dist/ops/grads/attentionMask.js

@@ -1,4 +1,4 @@
-import { g as i } from "../../index-pWA4_lUh.js";
+import { g as i } from "../../index-XjBAhiFO.js";
 const u = {
   kernelName: "AttentionMask",
   inputsToSave: ["q", "k"],

package/dist/ops/grads/fusedSoftmax.d.ts

@@ -0,0 +1,2 @@
+import { GradConfig } from '@tensorflow/tfjs-core';
+export declare const softmaxGradConfig: GradConfig;

package/dist/ops/grads/fusedSoftmax.js

@@ -0,0 +1,17 @@
+import { g as p, b as m, s as d } from "../../index-XjBAhiFO.js";
+import { mulDrop as c } from "../mulDrop.js";
+import { s as f } from "../../sum-C6u3xMi3.js";
+const g = {
+  kernelName: "FusedSoftmax",
+  outputsToSave: [!0],
+  gradFunc: (s, a, u) => {
+    const [o] = a, { dim: i, dropoutRate: t, seed: r } = u, n = !0, e = t && r ? c(s, o, t, r) : m(s, o);
+    return {
+      logits: () => d(e, m(f(e, [i], n), o))
+    };
+  }
+};
+p(g);
+export {
+  g as softmaxGradConfig
+};

package/dist/ops/grads/qkv.js

@@ -1,4 +1,4 @@
-import { g as v } from "../../index-pWA4_lUh.js";
+import { g as v } from "../../index-XjBAhiFO.js";
 const g = {
   kernelName: "QKV",
   inputsToSave: ["x", "kernel"],

package/dist/ops/grads/rope.js

@@ -1,4 +1,4 @@
-import { g as a, e as i } from "../../index-pWA4_lUh.js";
+import { g as a, e as i } from "../../index-XjBAhiFO.js";
 function p(n, e, s, o) {
   return i().runKernel("Rope", { x: n, sin: e, cos: s }, { pastLen: o });
 }

package/dist/ops/mulDrop.d.ts

@@ -0,0 +1,2 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function mulDrop(a: Tensor, b: Tensor, dropoutRate: number, seed: number): Tensor;

package/dist/ops/mulDrop.js

@@ -0,0 +1,9 @@
+import { e as t } from "../index-XjBAhiFO.js";
+import "./cpu/mulDropout.js";
+import "./webgl/mulDropout.js";
+function m(r, o, e, n) {
+  return t().runKernel("MulDropout", { a: r, b: o }, { dropoutRate: e, seed: n });
+}
+export {
+  m as mulDrop
+};