@genai-fi/nanogpt 0.10.2 → 0.10.3

package/dist/ops/grads/qkv.js

@@ -1,36 +1,34 @@
-import "../../index-D6Q1lPZO.js";
-import { a as u } from "../../matMul16-fEAJ4smh.js";
-import { concat16 as f } from "../concat16.js";
-import { sum16 as g } from "../sum16.js";
-import { packTensor as k, isPackedTensor as l } from "../../utilities/packed.js";
-import { a as h } from "../../tensor_util-DfwaWayG.js";
-import { s as G } from "../../squeeze-C00Ipm_7.js";
-const m = {
+import { u as c } from "../../index-DOvlwCh-.js";
+import { a as f } from "../../matMul16-BWRSOCWB.js";
+import { concat16 as g } from "../concat16.js";
+import { sum16 as l } from "../sum16.js";
+import { s as k } from "../../squeeze-C7Z2srUo.js";
+const i = {
   kernelName: "QKV",
   inputsToSave: ["x", "kernel"],
   outputsToSave: [],
   gradFunc: (e, s) => {
-    const [o, n, t] = e, [a] = s, p = f([o, n, t], 1);
-    o.dispose(), n.dispose(), t.dispose();
-    const c = [a.shape[0], a.shape[1], 3 * a.shape[2]], i = u.gradFunc(p, s, {
+    const [r, n, t] = e, [a] = s, p = g([r, n, t], 1);
+    r.dispose(), n.dispose(), t.dispose();
+    const m = [a.shape[0], a.shape[1], 3 * a.shape[2]], d = f.gradFunc(p, s, {
       transposeA: !1,
       transposeB: !1,
-      originalShape: c,
+      originalShape: m,
       perm: [0, 2, 1, 3]
     });
     return p.dispose(), {
-      x: () => i.A(),
+      x: () => d.A(),
       kernel: () => {
-        const r = i.B(), d = r.shape[0] === 1 ? G(r, [0]) : g(r, 0);
-        return r.dispose(), l(r) ? k(d) : d;
+        const o = d.B(), u = o.shape[0] === 1 ? k(o, [0]) : l(o, 0);
+        return o.dispose(), u;
       }
     };
   }
 };
-function A(e, s, o) {
-  return m.gradFunc(e, [s, o], {});
+function B(e, s, r) {
+  return i.gradFunc(e, [s, r], {});
 }
-h(m);
+c(i);
 export {
-  A as qkvGrad
+  B as qkvGrad
 };

package/dist/ops/grads/rope.js

@@ -1,7 +1,5 @@
-import "../../utilities/packed.js";
-import "../../index-D6Q1lPZO.js";
-import { a as t } from "../../rope-s4W2XO9B.js";
-import "../../tensor_util-DfwaWayG.js";
+import "../../index-DOvlwCh-.js";
+import { a as p } from "../../rope-Ir4mTyD1.js";
 export {
-  t as ropeGradConfig
+  p as ropeGradConfig
 };

package/dist/ops/grads/softmax16.js

@@ -1,8 +1,7 @@
-import { e as n } from "../../index-D6Q1lPZO.js";
+import { u as n, e as a } from "../../index-DOvlwCh-.js";
 import { isPackedTensor as t } from "../../utilities/packed.js";
-import { a } from "../../tensor_util-DfwaWayG.js";
 function s(r, e) {
-  return n().runKernel("Softmax16Grad", { dy: r, softmaxOutput: e });
+  return a().runKernel("Softmax16Grad", { dy: r, softmaxOutput: e });
 }
 const i = {
   kernelName: "Softmax16",
@@ -20,7 +19,7 @@ const i = {
     };
   }
 };
-a(i);
+n(i);
 export {
   i as softmax16GradConfig
 };

package/dist/ops/grads/unpack16.js

@@ -1,6 +1,5 @@
-import "../../index-D6Q1lPZO.js";
-import { a as i } from "../../pack16-CmVZs6af.js";
-import "../../tensor_util-DfwaWayG.js";
+import "../../index-DOvlwCh-.js";
+import { a as p } from "../../pack16-nQ6JaLo-.js";
 export {
-  i as unpackGradConfig
+  p as unpackGradConfig
 };

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

@@ -1,3 +1,4 @@
 import { TensorInfo } from '@tensorflow/tfjs-core';
 export declare function forceFloat<T extends TensorInfo>(x: T): T;
 export declare function forceInt<T extends TensorInfo>(x: T): T;
+export declare function forcePacked<T extends TensorInfo>(x: T): T;

package/dist/ops/grads/utils.js

@@ -1,10 +1,14 @@
-function n(t) {
+function e(t) {
   return t.dtype = "float32", t;
 }
-function e(t) {
+function n(t) {
   return t.dtype = "int32", t;
 }
+function r(t) {
+  return t.dtype = "packedF16", t;
+}
 export {
-  n as forceFloat,
-  e as forceInt
+  e as forceFloat,
+  n as forceInt,
+  r as forcePacked
 };

package/dist/ops/matMul16.js

@@ -1,9 +1,9 @@
-import "../index-D6Q1lPZO.js";
-import { b as p, c as u, d as i, e as s, m as M } from "../matMul16-fEAJ4smh.js";
+import "../index-DOvlwCh-.js";
+import { b as p, c as u, d as i, e as s, m as M } from "../matMul16-BWRSOCWB.js";
 import "./webgl/matMul16.js";
 import "./cpu/matMul16.js";
 import "../utilities/packed.js";
-import "../pack16-CmVZs6af.js";
+import "../pack16-nQ6JaLo-.js";
 export {
   p as matMul16,
   u as matMul16Gelu,

package/dist/ops/matMulGelu.js

@@ -1,6 +1,6 @@
-import { e as u } from "../index-D6Q1lPZO.js";
+import { e as u } from "../index-DOvlwCh-.js";
 import "./cpu/matMulGelu.js";
-import "./webgl/matMulGelu.js";
+import "../matMulGelu-CzfgT6Wq.js";
 import "./grads/matMulGelu.js";
 function M(r, e) {
   return u().runKernel("MatMulGelu", { x: r, kernel: e });

package/dist/ops/matMulMul.js

@@ -1,4 +1,4 @@
-import { e as u } from "../index-D6Q1lPZO.js";
+import { e as u } from "../index-DOvlwCh-.js";
 import "./cpu/matMulMul.js";
 import "./webgl/matMulMul.js";
 function m(e, r, t, l = !1, n = !1) {

package/dist/ops/mul16.js

@@ -1,4 +1,4 @@
-import { m as t, e as u } from "../index-D6Q1lPZO.js";
+import { m as t, e as u } from "../index-DOvlwCh-.js";
 import { isPackedTensor as n } from "../utilities/packed.js";
 function i(r, e) {
   return !n(r) && !n(e) ? t(r, e) : u().runKernel("Mul16", { a: r, b: e });

package/dist/ops/mulDrop.js

@@ -1,4 +1,4 @@
-import { e as t } from "../index-D6Q1lPZO.js";
+import { e as t } from "../index-DOvlwCh-.js";
 import "./cpu/mulDropout.js";
 import "./webgl/mulDropout.js";
 function m(r, o, e, n) {

package/dist/ops/normRMS.js

@@ -1,4 +1,4 @@
-import { e as n } from "../index-D6Q1lPZO.js";
+import { e as n } from "../index-DOvlwCh-.js";
 import "./cpu/normRMS.js";
 import "./webgl/normRMS.js";
 import "./grads/normRMS.js";

package/dist/ops/pack16.js

@@ -1,6 +1,5 @@
-import "../utilities/packed.js";
-import "../index-D6Q1lPZO.js";
-import { p as t } from "../pack16-CmVZs6af.js";
+import "../index-DOvlwCh-.js";
+import { p as a } from "../pack16-nQ6JaLo-.js";
 export {
-  t as pack16
+  a as pack16
 };

package/dist/ops/qkv.js

@@ -1,14 +1,10 @@
-import { e as m } from "../index-D6Q1lPZO.js";
+import { e as t } from "../index-DOvlwCh-.js";
 import "./cpu/qkv.js";
 import "./webgl/qkv.js";
 import "./grads/qkv.js";
-import { packTensor as f } from "../utilities/packed.js";
-function l(n, t, e, r = !1) {
-  const o = m().runKernel("QKV", { x: n, kernel: t }, { heads: e, packed: r });
-  return r && o.forEach((i) => {
-    f(i);
-  }), o;
+function u(r, e, n, o = !1) {
+  return t().runKernel("QKV", { x: r, kernel: e }, { heads: n, packed: o });
 }
 export {
-  l as qkv
+  u as qkv
 };

package/dist/ops/reshape16.js

@@ -1,43 +1,41 @@
-import { e as c } from "../index-D6Q1lPZO.js";
-import { isPackedTensor as u, packTensor as i } from "../utilities/packed.js";
-import { r as p } from "../reshape-CaPQzFvz.js";
-import { a as l, r as t } from "../tensor_util-DfwaWayG.js";
-const m = {
+import { u as p, p as s, e as u } from "../index-DOvlwCh-.js";
+import { r as c } from "../reshape-ByE68wS9.js";
+const i = {
   kernelName: "Reshape16",
   inputsToSave: ["x"],
   gradFunc: (e, r) => {
     const [n] = r;
     if (Array.isArray(e))
       throw new Error("Reshape16 gradient does not support multiple outputs.");
-    return { x: () => f(e, n.shape) };
+    return { x: () => m(e, n.shape) };
   }
 };
-l(m);
+p(i);
 function a(e) {
-  const { inputs: r, attrs: n } = e, { x: s } = r, { shape: o } = n;
-  return u(s) ? i(p(s, o)) : p(s, o);
+  const { inputs: r, attrs: n } = e, { x: t } = r, { shape: o } = n;
+  return c(t, o);
 }
-const k = {
+const l = {
   kernelName: "Reshape16",
   backendName: "webgpu",
   kernelFunc: a
 };
-t(k);
+s(l);
 const g = {
   kernelName: "Reshape16",
   backendName: "webgl",
   kernelFunc: a
 };
-t(g);
+s(g);
 const h = {
   kernelName: "Reshape16",
   backendName: "cpu",
   kernelFunc: a
 };
-t(h);
-function f(e, r) {
-  return c().runKernel("Reshape16", { x: e }, { shape: r });
+s(h);
+function m(e, r) {
+  return u().runKernel("Reshape16", { x: e }, { shape: r });
 }
 export {
-  f as reshape16
+  m as reshape16
 };

package/dist/ops/rope.d.ts

@@ -1,3 +1,3 @@
 import { default as RoPECache } from '../layers/RoPECache';
-import { Tensor } from '@tensorflow/tfjs';
+import { Tensor } from '@tensorflow/tfjs-core';
 export declare function rope(x: Tensor, cache: RoPECache, pastLength: number, negSin?: boolean): Tensor;

package/dist/ops/rope.js

@@ -1,12 +1,7 @@
-import "../index-D6Q1lPZO.js";
-import "../random_width-BVV9HveY.js";
-import "../register_all_kernels-nvj2k7OC.js";
-import "../index-Cp39cXWe.js";
-import "../dataset-D2P7rHAw.js";
+import "../index-DOvlwCh-.js";
 import "./cpu/rope.js";
 import "./webgl/rope.js";
-import { r as x } from "../rope-s4W2XO9B.js";
-import "../utilities/packed.js";
+import { r as i } from "../rope-Ir4mTyD1.js";
 export {
-  x as rope
+  i as rope
 };

package/dist/ops/scatterSub.js

@@ -1,4 +1,4 @@
-import { e as i } from "../index-D6Q1lPZO.js";
+import { e as i } from "../index-DOvlwCh-.js";
 import "./cpu/scatterSub.js";
 import "./webgl/scatterSub.js";
 function c(t, r, e) {

package/dist/ops/slice16.js

@@ -1,6 +1,6 @@
 import { isPackedTensor as n } from "../utilities/packed.js";
-import { e as c } from "../index-D6Q1lPZO.js";
-import { s as i } from "../slice-DvovR5wq.js";
+import { e as c } from "../index-DOvlwCh-.js";
+import { s as i } from "../slice-BfEGSH82.js";
 function a(r, e, o) {
   return n(r) ? c().runKernel("Slice16", { x: r }, { begin: e, size: o }) : i(r, e, o);
 }

package/dist/ops/softmax16.js

@@ -1,12 +1,9 @@
-import { e } from "../index-D6Q1lPZO.js";
+import { e as n } from "../index-DOvlwCh-.js";
 import "./grads/softmax16.js";
-import { isPackedTensor as m, packTensor as a } from "../utilities/packed.js";
-function p(r) {
-  if (!m(r))
-    return e().runKernel("Softmax", { logits: r }, { dim: r.rank - 1 });
-  const n = e().runKernel("Softmax16", { logits: r }, { dim: r.rank - 1 });
-  return m(r) ? a(n) : n;
+import { isPackedTensor as e } from "../utilities/packed.js";
+function t(r) {
+  return e(r) ? n().runKernel("Softmax16", { logits: r }, { dim: r.rank - 1 }) : n().runKernel("Softmax", { logits: r }, { dim: r.rank - 1 });
 }
 export {
-  p as softmax16
+  t as softmax16
 };

package/dist/ops/sub16.js

@@ -1,4 +1,4 @@
-import { c as s, e as t } from "../index-D6Q1lPZO.js";
+import { c as s, e as t } from "../index-DOvlwCh-.js";
 import { isPackedTensor as n } from "../utilities/packed.js";
 function c(r, e) {
   return !n(r) && !n(e) ? s(r, e) : t().runKernel("Sub16", { a: r, b: e });

package/dist/ops/sum16.js

@@ -1,6 +1,6 @@
-import { e as t } from "../index-D6Q1lPZO.js";
+import { e as t } from "../index-DOvlwCh-.js";
 import { isPackedTensor as s } from "../utilities/packed.js";
-import { s as n } from "../sum-ywRJj3Zr.js";
+import { s as n } from "../sum-DWAtNGez.js";
 function p(r, o, e = !1) {
   if (!s(r))
     return n(r, o, e);

package/dist/ops/transpose16.js

@@ -1,41 +1,40 @@
-import { e as i } from "../index-D6Q1lPZO.js";
-import { forceInt as u, forceFloat as l } from "./grads/utils.js";
-import { g as m } from "../axis_util-BsIr9ZNu.js";
-import { isPackedTensor as f, packTensor as g } from "../utilities/packed.js";
-import { t as a } from "../transpose-DH4gmHvu.js";
-import { a as d, r as p } from "../tensor_util-DfwaWayG.js";
-const k = {
+import { u as i, p, e as u } from "../index-DOvlwCh-.js";
+import { forcePacked as l, forceFloat as m } from "./grads/utils.js";
+import { g } from "../axis_util-BaG7mf5A.js";
+import { isPackedTensor as f } from "../utilities/packed.js";
+import { t as a } from "../transpose-ClWiBS_b.js";
+const d = {
   kernelName: "Transpose16",
-  gradFunc: (e, s, o) => {
-    if (Array.isArray(e))
+  gradFunc: (r, s, t) => {
+    if (Array.isArray(r))
       throw new Error("Transpose16 gradient does not support multiple outputs.");
-    const n = o, { perm: r } = n, t = m(r);
-    return { x: () => w(e, t) };
+    const n = t, { perm: e } = n, o = g(e);
+    return { x: () => T(r, o) };
   }
 };
-d(k);
-function c(e) {
-  const { inputs: s, attrs: o } = e, { x: n } = s, { perm: r } = o, t = f(n);
-  if (t && r[r.length - 1] !== n.shape.length - 1)
+i(d);
+function c(r) {
+  const { inputs: s, attrs: t } = r, { x: n } = s, { perm: e } = t, o = f(n);
+  if (o && e[e.length - 1] !== n.shape.length - 1)
     throw new Error("Transpose16 currently only supports the last axis being unchanged.");
-  return t ? g(u(a(l(n), r))) : a(n, r);
+  return o ? l(a(m(n), e)) : a(n, e);
 }
-const h = {
+const k = {
   kernelName: "Transpose16",
   backendName: "webgl",
   kernelFunc: c
 };
-p(h);
-const T = {
+p(k);
+const h = {
   kernelName: "Transpose16",
   backendName: "cpu",
   kernelFunc: c
 };
-p(T);
-function w(e, s) {
-  return s == null && (s = e.shape.map((n, r) => r).reverse()), i().runKernel("Transpose16", { x: e }, { perm: s });
+p(h);
+function T(r, s) {
+  return s == null && (s = r.shape.map((n, e) => e).reverse()), u().runKernel("Transpose16", { x: r }, { perm: s });
 }
 export {
-  w as transpose16,
-  k as transpose16GradConfig
+  T as transpose16,
+  d as transpose16GradConfig
 };

package/dist/ops/unpack16.js

@@ -1,5 +1,5 @@
-import "../index-D6Q1lPZO.js";
-import { u as t } from "../pack16-CmVZs6af.js";
+import "../index-DOvlwCh-.js";
+import { u as t } from "../pack16-nQ6JaLo-.js";
 import "../utilities/packed.js";
 export {
   t as unpack16

package/dist/ops/webgl/adamAdjust.js

@@ -1,6 +1,5 @@
-import { r as n } from "../../Reshape-CHdUjC72.js";
-import "../../index-D6Q1lPZO.js";
-import { r as f } from "../../tensor_util-DfwaWayG.js";
+import { r as n } from "../../Reshape-Zt6eb7yh.js";
+import { p as f } from "../../index-DOvlwCh-.js";
 class v {
   variableNames = ["moments", "value"];
   outputShape;

package/dist/ops/webgl/adamMoments.js

@@ -1,5 +1,4 @@
-import "../../index-D6Q1lPZO.js";
-import { r as m } from "../../tensor_util-DfwaWayG.js";
+import { p as m } from "../../index-DOvlwCh-.js";
 class i {
   variableNames = ["moments", "gradient"];
   outputShape;

package/dist/ops/webgl/appendCache.js

@@ -1,5 +1,4 @@
-import "../../index-D6Q1lPZO.js";
-import { r as p } from "../../tensor_util-DfwaWayG.js";
+import { p } from "../../index-DOvlwCh-.js";
 class m {
   variableNames = ["cache", "item"];
   outputShape;

package/dist/ops/webgl/attentionMask.js

@@ -1,5 +1,4 @@
-import "../../index-D6Q1lPZO.js";
-import { r as d } from "../../tensor_util-DfwaWayG.js";
+import { p as m } from "../../index-DOvlwCh-.js";
 class h {
   variableNames = ["q", "k"];
   outputShape;
@@ -35,12 +34,12 @@ class h {
   }
 }
 function l(o) {
-  const { q: t, k: e } = o.inputs, { divisor: s, pastLen: n } = o.attrs, a = o.backend, i = t.shape[0], r = t.shape[2], c = e.shape[2], u = t.shape[1], p = t.shape[3], m = new h(i, u, r, c, p);
-  return a.runWebGLProgram(m, [t, e], "float32", [[s], [n], [Number.NEGATIVE_INFINITY]]);
+  const { q: t, k: e } = o.inputs, { divisor: s, pastLen: n } = o.attrs, a = o.backend, i = t.shape[0], r = t.shape[2], c = e.shape[2], u = t.shape[1], p = t.shape[3], d = new h(i, u, r, c, p);
+  return a.runWebGLProgram(d, [t, e], "float32", [[s], [n], [Number.NEGATIVE_INFINITY]]);
 }
 const f = {
   kernelName: "AttentionMask",
   backendName: "webgl",
   kernelFunc: l
 };
-d(f);
+m(f);

package/dist/ops/webgl/fusedSoftmax.js

@@ -1,9 +1,7 @@
-import { m as b, s as I, r as k } from "../../RealDiv-zz7FpkKX.js";
-import { r as v } from "../../Reshape-CHdUjC72.js";
-import "../../index-D6Q1lPZO.js";
-import { r as w } from "../../tensor_util-DfwaWayG.js";
-import { p as P } from "../../tensor-CzmOBsdf.js";
-import { e as S } from "../../axis_util-BsIr9ZNu.js";
+import { m as b, s as I, r as k } from "../../RealDiv-KAPDe8zB.js";
+import { r as v } from "../../Reshape-Zt6eb7yh.js";
+import { p as w, af as P } from "../../index-DOvlwCh-.js";
+import { e as S } from "../../axis_util-BaG7mf5A.js";
 class T {
   variableNames = ["logits", "maxLogits"];
   outputShape;

package/dist/ops/webgl/gatherSub.js

@@ -1,6 +1,5 @@
-import "../../index-D6Q1lPZO.js";
-import { r as i } from "../../tensor_util-DfwaWayG.js";
-class l {
+import { p as l } from "../../index-DOvlwCh-.js";
+class u {
   variableNames = ["labels", "logits", "values"];
   outputShape;
   userCode;
@@ -16,13 +15,13 @@ class l {
     `;
   }
 }
-function u(t) {
-  const { logits: e, labels: o, values: s } = t.inputs, a = t.backend, r = o.shape[0], n = new l(r);
+function i(t) {
+  const { logits: e, labels: o, values: s } = t.inputs, a = t.backend, r = o.shape[0], n = new u(r);
   return a.runWebGLProgram(n, [o, e, s], "float32");
 }
 const c = {
   kernelName: "EfficientGatherSub",
   backendName: "webgl",
-  kernelFunc: u
+  kernelFunc: i
 };
-i(c);
+l(c);

package/dist/ops/webgl/gelu.js

@@ -1,6 +1,5 @@
-import "../../index-D6Q1lPZO.js";
-import { u as s, C as i } from "../../kernel_funcs_utils-CWfOAPGO.js";
-import { r as a } from "../../tensor_util-DfwaWayG.js";
+import { p as a } from "../../index-DOvlwCh-.js";
+import { u as s, C as i } from "../../kernel_funcs_utils-CCzYdUZg.js";
 const t = 0.7978845608028654, r = 0.044715, c = i + `
     float x3 = x * x * x;
     float inner = x + ${r} * x3;

package/dist/ops/webgl/log.js

@@ -1,10 +1,9 @@
-import "../../index-D6Q1lPZO.js";
-import { u as r, l as e } from "../../kernel_funcs_utils-CWfOAPGO.js";
-import { y as s } from "../../shared-CHhxz-O5.js";
-import { r as N, L as l } from "../../tensor_util-DfwaWayG.js";
-const t = e + `
+import { p as e, ao as r } from "../../index-DOvlwCh-.js";
+import { u as s, l as N } from "../../kernel_funcs_utils-CCzYdUZg.js";
+import { y as l } from "../../shared-DT1TkE6w.js";
+const a = N + `
   return x < 0.0 ? NAN : log(x);
-`, a = `
+`, t = `
   vec4 result = log(x);
   bvec4 isNaN = isnan(x);
   result.r = isNaN.r ? x.r : (x.r < 0.0 ? NAN : result.r);
@@ -12,13 +11,13 @@ const t = e + `
   result.b = isNaN.b ? x.b : (x.b < 0.0 ? NAN : result.b);
   result.a = isNaN.a ? x.a : (x.a < 0.0 ? NAN : result.a);
   return result;
-`, n = r({
-  opSnippet: t,
-  packedOpSnippet: a,
-  cpuKernelImpl: s
+`, n = s({
+  opSnippet: a,
+  packedOpSnippet: t,
+  cpuKernelImpl: l
 }), o = {
-  kernelName: l,
+  kernelName: r,
   backendName: "webgl",
   kernelFunc: n
 };
-N(o);
+e(o);

package/dist/ops/webgl/matMul16.js

@@ -1,17 +1,16 @@
 import { isPackedTensor as k } from "../../utilities/packed.js";
-import { m as M, b as m } from "../../index-D6Q1lPZO.js";
-import { matMulMul as g } from "../matMulMul.js";
-import { matMulGelu as N } from "../matMulGelu.js";
-import { r as U } from "../../tensor_util-DfwaWayG.js";
-import { m as G } from "../../mat_mul-C59XWcJd.js";
-import { r as w } from "../../reshape-CaPQzFvz.js";
-import { t as h } from "../../transpose-DH4gmHvu.js";
+import { p as g, m as M, b as m } from "../../index-DOvlwCh-.js";
+import { matMulMul as N } from "../matMulMul.js";
+import { matMulGelu as U } from "../matMulGelu.js";
+import { m as G } from "../../mat_mul-SjpJRLyL.js";
+import { r as w } from "../../reshape-ByE68wS9.js";
+import { t as h } from "../../transpose-ClWiBS_b.js";
 function P(p) {
-  const { A: r, B: o } = p.inputs, { transposeA: l, transposeB: c, scale: u, activation: A, scaleA: f, scaleB: d, forceOutputShape: t, perm: n } = p.attrs, B = !k(r), v = !k(o);
+  const { A: r, B: o } = p.inputs, { transposeA: l, transposeB: c, scale: u, activation: A, scaleA: d, scaleB: f, forceOutputShape: t, perm: n } = p.attrs, B = !k(r), v = !k(o);
   if (B && v) {
-    const a = f !== void 0 ? M(r, m(f)) : r, i = d !== void 0 ? M(o, m(d)) : o;
+    const a = d !== void 0 ? M(r, m(d)) : r, i = f !== void 0 ? M(o, m(f)) : o;
     let e;
-    if (u !== void 0 ? e = g(a, i, m(u), l, c) : A === "gelu" ? e = N(a, i) : e = G(a, i, l, c), n)
+    if (u !== void 0 ? e = N(a, i, m(u), l, c) : A === "gelu" ? e = U(a, i) : e = G(a, i, l, c), n)
       if (t) {
         const s = w(e, t);
         e.dispose();
@@ -34,4 +33,4 @@ const C = {
   backendName: "webgl",
   kernelFunc: P
 };
-U(C);
+g(C);