@genai-fi/nanogpt 0.9.0 → 0.10.0

package/dist/ops/cpu/gelu.js

@@ -1,29 +1,30 @@
-import { f as t, t as d } from "../../index-BzFyqcy-.js";
+import { t as d } from "../../index-ZyQhjEPo.js";
+import { r } from "../../tensor_util-DV-FP5Q3.js";
 const o = 0.7978845608028654, c = 0.044715;
-function m(r) {
-  const { inputs: u } = r, { x: n } = u, e = n;
+function m(t) {
+  const { inputs: u } = t, { x: n } = u, e = n;
   return d(() => {
     const l = e.pow(3), s = e.add(l.mul(c)).mul(o).tanh().add(1).mul(0.5);
     return e.mul(s);
   });
 }
-const p = {
+const N = {
   kernelName: "Gelu",
   backendName: "cpu",
   kernelFunc: m
 };
-t(p);
+r(N);
 const K = {
   kernelName: "Gelu",
   backendName: "tensorflow",
   kernelFunc: m
 };
-t(K);
-function i(r) {
-  const { dy: u, x: n } = r.inputs;
+r(K);
+function i(t) {
+  const { dy: u, x: n } = t.inputs;
   return d(() => {
-    const e = n.square(), l = e.mul(n), a = n.add(l.mul(c)).mul(o).tanh(), f = a.square().neg().add(1), k = e.mul(3 * c).add(1), g = a.add(1).mul(0.5), G = n.mul(f).mul(o).mul(k).mul(0.5), N = g.add(G);
-    return u.mul(N);
+    const e = n.square(), l = e.mul(n), a = n.add(l.mul(c)).mul(o).tanh(), f = a.square().neg().add(1), k = e.mul(3 * c).add(1), g = a.add(1).mul(0.5), G = n.mul(f).mul(o).mul(k).mul(0.5), p = g.add(G);
+    return u.mul(p);
   });
 }
 const x = {
@@ -31,10 +32,10 @@ const x = {
   backendName: "cpu",
   kernelFunc: i
 };
-t(x);
+r(x);
 const h = {
   kernelName: "GeluGrad",
   backendName: "tensorflow",
   kernelFunc: i
 };
-t(h);
+r(h);

package/dist/ops/cpu/matMul16.d.ts

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

package/dist/ops/cpu/matMul16.js

@@ -0,0 +1,16 @@
+import { isPackedTensor as t } from "../../utilities/packed.js";
+import "../../index-ZyQhjEPo.js";
+import { r as p } from "../../tensor_util-DV-FP5Q3.js";
+import { m } from "../../mat_mul-DeAh4uTH.js";
+function l(r) {
+  const { A: e, B: n } = r.inputs, { transposeA: o, transposeB: s } = r.attrs, a = !t(e), c = !t(n);
+  if (a && c)
+    return m(e, n, o, s);
+  throw new Error("MatMul16 CPU kernel only supports packed tensors currently.");
+}
+const u = {
+  kernelName: "MatMul16",
+  backendName: "cpu",
+  kernelFunc: l
+};
+p(u);

package/dist/ops/cpu/matMulGelu.js

@@ -1,10 +1,12 @@
-import { f as e, t as m } from "../../index-BzFyqcy-.js";
-import { g as d, d as M } from "../../gelu-Bp_-935b.js";
+import { t as m } from "../../index-ZyQhjEPo.js";
+import { g as i, d as M } from "../../gelu-CNLFZWea.js";
+import { r as e } from "../../tensor_util-DV-FP5Q3.js";
+import { m as k } from "../../mat_mul-DeAh4uTH.js";
 function c(t) {
-  const { inputs: u } = t, { x: n, kernel: l } = u, a = n, r = l;
+  const { inputs: u } = t, { x: n, kernel: r } = u, a = n, l = r;
   return m(() => {
-    const o = a.matMul(r);
-    return d(o);
+    const o = k(a, l);
+    return i(o);
   });
 }
 const G = {
@@ -13,23 +15,23 @@ const G = {
   kernelFunc: c
 };
 e(G);
-const i = {
+const f = {
   kernelName: "MatMulGelu",
   backendName: "tensorflow",
   kernelFunc: c
 };
-e(i);
-const f = {
+e(f);
+const p = {
   kernelName: "MatMulGelu",
   backendName: "webgpu",
   kernelFunc: c
 };
-e(f);
+e(p);
 function s(t) {
-  const { dy: u, x: n, kernel: l } = t.inputs;
+  const { dy: u, x: n, kernel: r } = t.inputs;
   return m(() => {
-    const a = n.matMul(l), r = M(u, a), o = r.matMul(l.transpose()), k = n.transpose().matMul(r);
-    return [o, k];
+    const a = k(n, r), l = M(u, a), o = l.matMul(r.transpose()), d = n.transpose().matMul(l);
+    return [o, d];
   });
 }
 const g = {
@@ -38,15 +40,15 @@ const g = {
   kernelFunc: s
 };
 e(g);
-const p = {
+const N = {
   kernelName: "MatMulGeluGrad",
   backendName: "tensorflow",
   kernelFunc: s
 };
-e(p);
-const N = {
+e(N);
+const b = {
   kernelName: "MatMulGeluGrad",
   backendName: "webgpu",
   kernelFunc: s
 };
-e(N);
+e(b);

package/dist/ops/cpu/matMulMul.js

@@ -1,20 +1,21 @@
-import { f as e, t as i } from "../../index-BzFyqcy-.js";
+import { t as M } from "../../index-ZyQhjEPo.js";
+import { r as e } from "../../tensor_util-DV-FP5Q3.js";
 function n(t) {
-  const { inputs: r, attrs: l } = t, { transposeA: o, transposeB: s } = l, { x: u, kernel: a, y: c } = r, m = u, k = a, M = c;
-  return i(() => m.matMul(k, o, s).mul(M));
+  const { inputs: r, attrs: l } = t, { transposeA: o, transposeB: s } = l, { x: u, kernel: a, y: c } = r, m = u, i = a, k = c;
+  return M(() => m.matMul(i, o, s).mul(k));
 }
-const f = {
+const p = {
   kernelName: "MatMulMul",
   backendName: "cpu",
   kernelFunc: n
 };
-e(f);
-const p = {
+e(p);
+const f = {
   kernelName: "MatMulMul",
   backendName: "tensorflow",
   kernelFunc: n
 };
-e(p);
+e(f);
 const g = {
   kernelName: "MatMulMul",
   backendName: "webgpu",

package/dist/ops/cpu/mulDropout.js

@@ -1,7 +1,8 @@
-import { f as e, a as t } from "../../index-BzFyqcy-.js";
+import { m as u } from "../../index-ZyQhjEPo.js";
+import { r as e } from "../../tensor_util-DV-FP5Q3.js";
 function n(o) {
-  const { inputs: r } = o, { a: l, b: u } = r;
-  return console.warn("Using fallback mulDrop implementation without dropout."), t(l, u);
+  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",

package/dist/ops/cpu/normRMS.js

@@ -1,9 +1,10 @@
-import { f as o, t as d } from "../../index-BzFyqcy-.js";
+import { t as d } from "../../index-ZyQhjEPo.js";
+import { r as a } from "../../tensor_util-DV-FP5Q3.js";
 function i(t) {
-  const { inputs: e } = t, { x: n, gamma: s } = e, r = n, a = s;
+  const { inputs: e } = t, { x: n, gamma: s } = e, r = n, m = s;
   return d(() => {
     const u = r.square().mean(-1, !0).add(1e-8).rsqrt();
-    return r.mul(u).mul(a);
+    return r.mul(u).mul(m);
   });
 }
 const k = {
@@ -11,18 +12,18 @@ const k = {
   backendName: "cpu",
   kernelFunc: i
 };
-o(k);
+a(k);
 const g = {
   kernelName: "RMSNorm",
   backendName: "tensorflow",
   kernelFunc: i
 };
-o(g);
+a(g);
 function N(t) {
   const { dy: e, x: n, gamma: s } = t.inputs;
   return d(() => {
-    const r = n.shape[n.shape.length - 1], a = n.square().mean(-1, !0), m = a.add(1e-8).rsqrt(), u = n.mul(m), l = e.mul(u).sum([0, 1]), c = e.mul(s), f = c.mul(n).sum(-1, !0).div(r);
-    return [c.mul(m).sub(n.mul(f).mul(m).div(a.add(1e-8))), l];
+    const r = n.shape[n.shape.length - 1], m = n.square().mean(-1, !0), o = m.add(1e-8).rsqrt(), u = n.mul(o), l = e.mul(u).sum([0, 1]), c = e.mul(s), f = c.mul(n).sum(-1, !0).div(r);
+    return [c.mul(o).sub(n.mul(f).mul(o).div(m.add(1e-8))), l];
   });
 }
 const S = {
@@ -30,10 +31,10 @@ const S = {
   backendName: "cpu",
   kernelFunc: N
 };
-o(S);
-const R = {
+a(S);
+const p = {
   kernelName: "RMSNormGrad",
   backendName: "tensorflow",
   kernelFunc: N
 };
-o(R);
+a(p);

package/dist/ops/cpu/qkv.js

@@ -1,25 +1,29 @@
-import { f as q } from "../../index-BzFyqcy-.js";
-import { r as o } from "../../reshape-CnIwVG1c.js";
-import { s as x } from "../../split-DK2k5eHf.js";
+import "../../index-ZyQhjEPo.js";
+import { r as q } from "../../tensor_util-DV-FP5Q3.js";
+import { r as o } from "../../reshape-DevtBWtf.js";
+import { s as x } from "../../split-DrfihRpZ.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);
+  const { x: c, kernel: K } = p.inputs, { heads: n, packed: C } = p.attrs;
+  if (C)
+    throw new Error("QKV CPU implementation does not support packed tensors.");
+  const [s, e, t] = c.shape, a = o(c, [s * e, t]), i = a.dot(K);
   a.dispose();
   const d = o(i, [s, e, 3 * t]);
   i.dispose();
-  const [k, l, m] = x(d, 3, -1);
+  const [k, m, l] = x(d, 3, -1);
   d.dispose();
   const r = t / n, f = o(k, [s, e, n, r]);
   k.dispose();
-  const C = f.transpose([0, 2, 1, 3]);
+  const w = f.transpose([0, 2, 1, 3]);
   f.dispose();
-  const h = o(l, [s, e, n, r]);
-  l.dispose();
+  const h = o(m, [s, e, n, r]);
+  m.dispose();
   const N = h.transpose([0, 2, 1, 3]);
   h.dispose();
-  const u = o(m, [s, e, n, r]);
-  m.dispose();
+  const u = o(l, [s, e, n, r]);
+  l.dispose();
   const T = u.transpose([0, 2, 1, 3]);
-  return u.dispose(), [C, N, T];
+  return u.dispose(), [w, N, T];
 }
 const F = {
   kernelName: "QKV",
@@ -27,12 +31,12 @@ const F = {
   kernelFunc: v
 };
 q(F);
-const R = {
+const Q = {
   kernelName: "QKV",
   backendName: "tensorflow",
   kernelFunc: v
 };
-q(R);
+q(Q);
 export {
   v as qkvCPU
 };

package/dist/ops/cpu/rope.js

@@ -1,37 +1,38 @@
-import { f as S } from "../../index-BzFyqcy-.js";
-import { r as F } from "../../range-CWcz7xFA.js";
-import { g as I } from "../../gather-Dnpgw-YQ.js";
-import { s as E } from "../../stack-DFatutCx.js";
-import { c as T } from "../../concat-B912vBbo.js";
-function U(t, c, p, o, r) {
-  const n = o.shape[3], s = p;
-  if (s > n) return o;
-  const e = o.shape[2], i = s / 2, a = c.slice([r, 0, 0], [e, i, 1]).reshape([1, 1, e, i]), d = t.slice([r, 0, 0], [e, i, 1]).reshape([1, 1, e, i]), l = o.shape[0], m = o.shape[1], h = F(0, s, 2, "int32"), k = F(1, s, 2, "int32"), D = ((R) => {
-    const g = R.slice([0, 0, 0, 0], [l, m, e, s]), v = s < n ? R.slice([0, 0, 0, s], [l, m, e, n - s]) : null, u = I(g, h, 3), f = I(g, k, 3), C = u.mul(a), N = f.mul(d), P = C.sub(N), b = f.mul(a), x = u.mul(d), K = b.add(x);
-    u.dispose(), f.dispose(), a.dispose(), d.dispose(), C.dispose(), N.dispose(), b.dispose(), x.dispose();
-    const w = E([P, K], -1);
-    P.dispose(), K.dispose();
-    const y = w.reshape([l, m, e, s]);
-    return w.dispose(), v ? T([y, v], 3) : y;
-  })(o);
-  return h.dispose(), k.dispose(), D;
+import "../../index-ZyQhjEPo.js";
+import { r as I } from "../../tensor_util-DV-FP5Q3.js";
+import { r as y } from "../../range-BMS52eQi.js";
+import { g as F } from "../../gather-DykLGqmW.js";
+import { s as E } from "../../stack-yOIAalTq.js";
+import { c as T } from "../../concat-BHlIJeyT.js";
+function U(c, r, p, e, n) {
+  const t = e.shape[3], s = p;
+  if (s > t) return e;
+  const o = e.shape[2], i = s / 2, a = r.slice([n, 0, 0], [o, i, 1]).reshape([1, 1, o, i]), d = c.slice([n, 0, 0], [o, i, 1]).reshape([1, 1, o, i]), l = e.shape[0], m = e.shape[1], h = y(0, s, 2, "int32"), g = y(1, s, 2, "int32"), D = ((k) => {
+    const C = k.slice([0, 0, 0, 0], [l, m, o, s]), R = s < t ? k.slice([0, 0, 0, s], [l, m, o, t - s]) : null, u = F(C, h, 3), f = F(C, g, 3), v = u.mul(a), N = f.mul(d), S = v.sub(N), P = f.mul(a), b = u.mul(d), x = P.add(b);
+    u.dispose(), f.dispose(), a.dispose(), d.dispose(), v.dispose(), N.dispose(), P.dispose(), b.dispose();
+    const K = E([S, x], -1);
+    S.dispose(), x.dispose();
+    const w = K.reshape([l, m, o, s]);
+    return K.dispose(), R ? T([w, R], 3) : w;
+  })(e);
+  return h.dispose(), g.dispose(), D;
 }
-function B(t) {
-  const { x: c, sin: p, cos: o } = t.inputs, { pastLen: r } = t.attrs, n = c.shape[3];
-  return U(p, o, n, c, r);
+function B(c) {
+  const { x: r } = c.inputs, { pastLen: p, negSin: e, ropeCache: n } = c.attrs, t = r.shape[3], s = e ? n.getNegSin() : n.getSin(), o = n.getCos();
+  return U(s, o, t, r, p);
 }
 const j = {
   kernelName: "Rope",
   backendName: "cpu",
   kernelFunc: B
 };
-S(j);
+I(j);
 const z = {
   kernelName: "Rope",
   backendName: "tensorflow",
   kernelFunc: B
 };
-S(z);
+I(z);
 export {
   U as applyRoPE,
   B as ropeCPU

package/dist/ops/cpu/scatterSub.js

@@ -1,39 +1,25 @@
-import { E as f, M as g, F as r, G as l, a7 as N, f as b, c as S, a as h } from "../../index-BzFyqcy-.js";
-import { v as D } from "../../scatter_nd_util-BQdz--Gn.js";
-import { r as k } from "../../range-CWcz7xFA.js";
-import { s as v } from "../../stack-DFatutCx.js";
-import { o as E } from "../../ones-tIJeHlq-.js";
-/**
- * @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 I(a, e, s) {
-  g(s);
-  const n = r(a, "indices", "scatterND", "int32"), t = r(e, "updates", "scatterND");
+import { A as f, B as c, E as g, c as l, m as N } from "../../index-ZyQhjEPo.js";
+import { j as b, r as S } from "../../tensor_util-DV-FP5Q3.js";
+import { d as h } from "../../tensor-DdQUJZlz.js";
+import { v as D } from "../../scatter_nd_util-5EL-8VAQ.js";
+import { r as k } from "../../range-BMS52eQi.js";
+import { s as v } from "../../stack-yOIAalTq.js";
+import { o as E } from "../../ones-CAMiP4I2.js";
+function I(r, e, s) {
+  h(s);
+  const n = c(r, "indices", "scatterND", "int32"), t = c(e, "updates", "scatterND");
   D(t, n, s);
-  const c = { indices: n, updates: t }, o = { shape: s };
-  return l.runKernel(N, c, o);
+  const o = { indices: n, updates: t }, a = { shape: s };
+  return g.runKernel(b, o, a);
 }
 const K = /* @__PURE__ */ f({ scatterND_: I });
-function L(a) {
-  const { logits: e, labels: s, dy: n } = a.inputs, t = s.shape[0], c = e.shape[1], o = k(0, t, 1, "int32"), i = v([o, s], 1), d = E([t]), u = K(i, d, [t, c]), p = S(e, u), m = n.reshape([t, 1]);
-  return h(p, m);
+function L(r) {
+  const { logits: e, labels: s, dy: n } = r.inputs, t = s.shape[0], o = e.shape[1], a = k(0, t, 1, "int32"), i = v([a, s], 1), d = E([t]), p = K(i, d, [t, o]), u = l(e, p), m = n.reshape([t, 1]);
+  return N(u, m);
 }
 const T = {
   kernelName: "EfficientScatterSub",
   backendName: "cpu",
   kernelFunc: L
 };
-b(T);
+S(T);

package/dist/ops/dot16.d.ts

@@ -0,0 +1,2 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function dot16(a: Tensor, b: Tensor, transposeA?: boolean, transposeB?: boolean): Tensor;

package/dist/ops/dot16.js

@@ -0,0 +1,42 @@
+import { b as d } from "../matMul16--R5hOwDG.js";
+import { transpose16 as w } from "./transpose16.js";
+import { reshape16 as n } from "./reshape16.js";
+import { isPackedTensor as p } from "../utilities/packed.js";
+import { d as x } from "../tfjs_backend-kNyO5L2d.js";
+function E(e, s, h = !1, c = !1) {
+  if (!p(e) && !p(s))
+    return x(e, s);
+  if (e.rank < 2 || s.rank < 2)
+    throw new Error(
+      `dot requires both inputs to be rank >= 2 but got x shape = ${e.shape} and y shape = ${s.shape}`
+    );
+  if (s.rank >= 3) {
+    const r = e.shape.slice(-1)[0], i = s.shape.slice(-2)[0];
+    if (r !== i)
+      throw new Error(
+        `If rank y >= 3, then the second last dim of y must equal the last dim of x but got x shape = ${e.shape} and  y shape = ${s.shape}`
+      );
+  }
+  if (!p(e) || !p(s))
+    throw new Error("dot16 requires both inputs to be packed Tensors.");
+  if (e.rank === 2 && s.rank === 2)
+    return d(e, s, h, c);
+  {
+    const r = e.shape.slice(), i = r.pop();
+    e = n(e, [-1, i]);
+    const a = s.shape.slice(), l = a.pop(), m = a.pop(), k = [...a, l], f = Array.from({ length: s.rank }, (o, t) => t === 0 ? s.rank - 2 : t <= s.rank - 2 ? t - 1 : t);
+    if (f.every((o, t) => o === t))
+      s = n(s, [m, -1]);
+    else {
+      const o = w(s, f);
+      s = n(o, [m, -1]), o.dispose();
+    }
+    const y = [...r, ...k], u = d(e, s, h, c);
+    e.dispose(), s.dispose();
+    const D = n(u, y);
+    return u.dispose(), D;
+  }
+}
+export {
+  E as dot16
+};

package/dist/ops/gatherSub.js

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

package/dist/ops/gelu.js

@@ -1,7 +1,7 @@
-import "../index-BzFyqcy-.js";
+import "../index-ZyQhjEPo.js";
 import "./cpu/gelu.js";
 import "./webgl/gelu.js";
-import { d as e, g as i } from "../gelu-Bp_-935b.js";
+import { d as e, g as i } from "../gelu-CNLFZWea.js";
 export {
   e as dGelu,
   i as gelu

package/dist/ops/grads/add16.d.ts

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

package/dist/ops/grads/add16.js

@@ -0,0 +1,27 @@
+import { j as u, q as d } from "../../index-ZyQhjEPo.js";
+import { sum16 as p } from "../sum16.js";
+import { reshape16 as c } from "../reshape16.js";
+import { a as h } from "../../tensor_util-DV-FP5Q3.js";
+const m = {
+  kernelName: "Add16",
+  inputsToSave: ["a", "b"],
+  gradFunc: (s, i) => {
+    const [t, a] = i, n = u(t.shape, a.shape);
+    if (Array.isArray(s))
+      throw new Error("Add16 gradFunc expected dy to be a Tensor but got an array");
+    return { a: () => {
+      let e = s;
+      const r = d(t.shape, n);
+      r.length > 0 && (e = p(e, r));
+      const o = c(e, t.shape);
+      return e.dispose(), o;
+    }, b: () => {
+      let e = s;
+      const r = d(a.shape, n);
+      r.length > 0 && (e = p(e, r));
+      const o = c(e, a.shape);
+      return e.dispose(), o;
+    } };
+  }
+};
+h(m);

package/dist/ops/grads/attentionMask.js

@@ -1,29 +1,22 @@
-import { i as m, b as i } from "../../index-BzFyqcy-.js";
-import { matMulMul as u } from "../matMulMul.js";
-const p = {
+import "../../index-ZyQhjEPo.js";
+import { m as o } from "../../matMul16--R5hOwDG.js";
+import { transpose16 as m } from "../transpose16.js";
+import { a as c } from "../../tensor_util-DV-FP5Q3.js";
+const l = {
   kernelName: "AttentionMask",
   inputsToSave: ["q", "k"],
   outputsToSave: [],
-  gradFunc: (t, c, l) => {
-    if (Array.isArray(t))
+  gradFunc: (r, s, n) => {
+    if (Array.isArray(r))
       throw new Error("Expected dy to be a single Tensor");
-    const [e, n] = c, { divisor: a } = l;
+    const [a, i] = s, { divisor: e } = n;
     return {
-      q: () => u(t, n, i(a)),
+      q: () => o(r, i, e),
       k: () => {
-        const s = e.transpose([0, 1, 3, 2]), r = u(s, t, i(a));
-        s.dispose();
-        const o = r.transpose([0, 1, 3, 2]);
-        return r.dispose(), o;
-      },
-      mask: () => t,
-      divisor: () => {
-        const s = e.matMul(n, !1, !0), r = t.mul(s);
-        s.dispose();
-        const o = r.sum();
-        return r.dispose(), o;
+        const t = o(a, r, e, !0, !1), u = m(t, [0, 1, 3, 2]);
+        return t.dispose(), u;
       }
     };
   }
 };
-m(p);
+c(l);

package/dist/ops/grads/gelu.js

@@ -1,5 +1,6 @@
-import "../../index-BzFyqcy-.js";
-import { a as e } from "../../gelu-Bp_-935b.js";
+import "../../index-ZyQhjEPo.js";
+import { a as m } from "../../gelu-CNLFZWea.js";
+import "../../tensor_util-DV-FP5Q3.js";
 export {
-  e as geluGradConfig
+  m as geluGradConfig
 };

package/dist/ops/grads/matMul16.d.ts

@@ -0,0 +1,2 @@
+import { GradConfig } from '@tensorflow/tfjs-core';
+export declare const matMul16GradConfig: GradConfig;

package/dist/ops/grads/matMul16.js

@@ -0,0 +1,9 @@
+import "../../index-ZyQhjEPo.js";
+import { a as f } from "../../matMul16--R5hOwDG.js";
+import "../../gelu-CNLFZWea.js";
+import "../transpose16.js";
+import "../reshape16.js";
+import "../../tensor_util-DV-FP5Q3.js";
+export {
+  f as matMul16GradConfig
+};

package/dist/ops/grads/matMulGelu.js

@@ -1,17 +1,18 @@
-import { i as a, e as o } from "../../index-BzFyqcy-.js";
-function i(e, n, r) {
-  return o().runKernel("MatMulGeluGrad", { dy: e, x: n, kernel: r });
+import { e as l } from "../../index-ZyQhjEPo.js";
+import { a as o } from "../../tensor_util-DV-FP5Q3.js";
+function i(e, r, n) {
+  return l().runKernel("MatMulGeluGrad", { dy: e, x: r, kernel: n });
 }
 const s = {
   kernelName: "MatMulGelu",
   inputsToSave: ["x", "kernel"],
   outputsToSave: [],
-  gradFunc: (e, n) => {
-    const [r, t] = n, [u, l] = i(e, r, t);
+  gradFunc: (e, r) => {
+    const [n, t] = r, [u, a] = i(e, n, t);
     return {
       x: () => u,
-      kernel: () => l
+      kernel: () => a
     };
   }
 };
-a(s);
+o(s);

package/dist/ops/grads/normRMS.js

@@ -1,20 +1,21 @@
-import { i as t, e as g } from "../../index-BzFyqcy-.js";
-function i(r, a, n) {
-  return g().runKernel("RMSNormGrad", { dy: r, x: a, gamma: n });
+import { e as t } from "../../index-ZyQhjEPo.js";
+import { a as g } from "../../tensor_util-DV-FP5Q3.js";
+function i(r, a, m) {
+  return t().runKernel("RMSNormGrad", { dy: r, x: a, gamma: m });
 }
 const s = {
   kernelName: "RMSNorm",
   inputsToSave: ["x", "gamma"],
   outputsToSave: [],
   gradFunc: (r, a) => {
-    const [n, e] = a, [m, o] = i(r, n, e);
+    const [m, n] = a, [o, e] = i(r, m, n);
     return {
-      x: () => m,
-      gamma: () => o
+      x: () => o,
+      gamma: () => e
     };
   }
 };
-t(s);
+g(s);
 export {
   s as normRMSGradConfig
 };

package/dist/ops/grads/{fusedSoftmax.d.ts → pack16.d.ts}

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