@genai-fi/nanogpt 0.18.1 → 0.19.0

package/dist/ops/webgl/matMul16.js

@@ -1,14 +1,14 @@
 import { isPackedTensor as h } from "../../utilities/packed.js";
-import { a as U, m as A, b as c, j as G } from "../../index-DSGwv2Yx.js";
+import { c as b, m as A, a as c, n as G } from "../../index-CUXkjxiT.js";
 import { matMulMul as P } from "../matMulMul.js";
 import { matMulGelu as R } from "../matMulGelu.js";
-import { m as f } from "../../mat_mul-DP86qZtZ.js";
-import { r as B } from "../../relu-DTvZKBsZ.js";
-import { r as v } from "../../reshape-BIN71H3p.js";
-import { t as b } from "../../transpose-Qxz-4os3.js";
-function j(d) {
-  const { A: i, B: m } = d.inputs, { transposeA: o, transposeB: a, scale: M, activation: u, scaleA: k, scaleB: w, forceOutputShape: n, perm: p } = d.attrs, g = !h(i), N = !h(m);
-  if (g && N) {
+import { m as f } from "../../mat_mul-DhG0Newp.js";
+import { r as B } from "../../relu-J_X6MUzx.js";
+import { r as v } from "../../reshape-BDOuCSNW.js";
+import { t as g } from "../../transpose-9kRxIXWR.js";
+function q(d) {
+  const { A: i, B: m } = d.inputs, { transposeA: o, transposeB: a, scale: M, activation: u, scaleA: k, scaleB: w, forceOutputShape: n, perm: p } = d.attrs, N = !h(i), U = !h(m);
+  if (N && U) {
     const t = k !== void 0 ? A(i, c(k)) : i, r = w !== void 0 ? A(m, c(w)) : m;
     let e;
     if (M !== void 0)
@@ -23,10 +23,10 @@ function j(d) {
       if (n) {
         const s = v(e, n);
         e.dispose();
-        const l = b(s, p);
+        const l = g(s, p);
         return s.dispose(), l;
       } else {
-        const s = b(e, p);
+        const s = g(e, p);
         return e.dispose(), s;
       }
     else if (n) {
@@ -37,9 +37,9 @@ function j(d) {
   }
   throw new Error("Not implemented: matMul16 with packed inputs");
 }
-const q = {
+const C = {
   kernelName: "MatMul16",
   backendName: "webgl",
-  kernelFunc: j
+  kernelFunc: q
 };
-U(q);
+b(C);

package/dist/ops/webgl/matMulGelu.js

@@ -1,7 +1,7 @@
-import "../../index-DSGwv2Yx.js";
-import "../../Reshape-BZC-ebeR.js";
-import { a as m, b as o, c as p } from "../../matMulGelu-CsZnh18H.js";
-import "../../mat_mul-DP86qZtZ.js";
+import "../../index-CUXkjxiT.js";
+import "../../Reshape-BW__R4mZ.js";
+import { a as m, b as o, c as p } from "../../matMulGelu-JNLZqKQp.js";
+import "../../mat_mul-DhG0Newp.js";
 export {
   m as MATMUL_SHARED_DIM_THRESHOLD,
   o as batchMatMulGeluImpl,

package/dist/ops/webgl/matMulMul.js

@@ -1,5 +1,5 @@
-import { a as u } from "../../index-DSGwv2Yx.js";
-import { b as c } from "../../matMulGelu-CsZnh18H.js";
+import { c as u } from "../../index-CUXkjxiT.js";
+import { b as c } from "../../matMulGelu-JNLZqKQp.js";
 const M = `
     return a * b;
 `;

package/dist/ops/webgl/mulDropout.js

@@ -1,4 +1,4 @@
-import { a as m } from "../../index-DSGwv2Yx.js";
+import { c as m } from "../../index-CUXkjxiT.js";
 class f {
   variableNames = ["a", "b"];
   outputShape;

package/dist/ops/webgl/normRMS.js

@@ -1,5 +1,5 @@
-import { a as c, e as G } from "../../index-DSGwv2Yx.js";
-import { s as x } from "../../sum-D39FeU5h.js";
+import { c, h as G } from "../../index-CUXkjxiT.js";
+import { s as x } from "../../sum-BpiwSWvg.js";
 class y {
   variableNames = ["x", "meanSquare"];
   outputShape;

package/dist/ops/webgl/qkv.js

@@ -1,4 +1,4 @@
-import { a as i } from "../../index-DSGwv2Yx.js";
+import { c as i } from "../../index-CUXkjxiT.js";
 class l {
   variableNames = ["x", "kernel"];
   outputShape;

package/dist/ops/webgl/rope.js

@@ -1,4 +1,4 @@
-import { a as h } from "../../index-DSGwv2Yx.js";
+import { c as h } from "../../index-CUXkjxiT.js";
 class g {
   variableNames = ["x", "sin", "cos"];
   outputShape;

package/dist/ops/webgl/scatterSub.js

@@ -1,4 +1,4 @@
-import { a as i } from "../../index-DSGwv2Yx.js";
+import { c as i } from "../../index-CUXkjxiT.js";
 class u {
   variableNames = ["labels", "softmaxProbs", "dy"];
   outputShape;
@@ -16,7 +16,7 @@ class u {
   }
 }
 function d(o) {
-  const { logits: e, labels: t, dy: r } = o.inputs, s = o.backend, a = t.shape[0], n = e.shape[1], c = new u(a, n);
+  const { logits: e, labels: t, dy: r } = o.inputs, s = o.backend, n = t.shape[0], a = e.shape[1], c = new u(n, a);
   return s.runWebGLProgram(c, [t, e, r], "float32");
 }
 const b = {

package/dist/ops/webgpu/adamAdjust.js

@@ -1,6 +1,6 @@
-import { e as d } from "../../webgpu_program-CbjdYLYk.js";
-import { f as l, c as f } from "../../webgpu_util-DuofJBMo.js";
-import { a as h, a6 as c } from "../../index-DSGwv2Yx.js";
+import { e as d } from "../../webgpu_program-B4HmApL1.js";
+import { f as l, c as f } from "../../webgpu_util-DYlGSwOJ.js";
+import { c as h, a6 as c } from "../../index-CUXkjxiT.js";
 class v {
   variableNames = ["moments", "value"];
   outputShape;

package/dist/ops/webgpu/adamMoments.js

@@ -1,6 +1,6 @@
-import { e as u } from "../../webgpu_program-CbjdYLYk.js";
-import { f as p, c } from "../../webgpu_util-DuofJBMo.js";
-import { a as d, a6 as f } from "../../index-DSGwv2Yx.js";
+import { e as u } from "../../webgpu_program-B4HmApL1.js";
+import { f as p, c } from "../../webgpu_util-DYlGSwOJ.js";
+import { c as d, a6 as f } from "../../index-CUXkjxiT.js";
 class h {
   variableNames = ["moments", "gradient", "scaling"];
   outputShape;

package/dist/ops/webgpu/add16.js

@@ -1,13 +1,13 @@
-import { a as t } from "../../index-DSGwv2Yx.js";
-import { BinaryOpProgram as p } from "./utils/binary_op.js";
-import { B as s } from "../../binary_op_util-pKXltfxI.js";
-function c(e) {
-  const { a: r, b: n } = e.inputs, a = e.backend, o = new p(s.ADD, r.shape, n.shape);
+import { c as t } from "../../index-CUXkjxiT.js";
+import { BinaryOpProgram as c } from "./utils/binary_op.js";
+import { B as p } from "../../binary_op_util-pKXltfxI.js";
+function s(e) {
+  const { a: r, b: n } = e.inputs, a = e.backend, o = new c(p.ADD, r.shape, n.shape);
   return a.runWebGPUProgram(o, [r, n], "packedF16");
 }
 const m = {
   kernelName: "Add16",
   backendName: "webgpu",
-  kernelFunc: c
+  kernelFunc: s
 };
 t(m);

package/dist/ops/webgpu/appendCache.js

@@ -1,7 +1,7 @@
 import { isPackedTensor as T } from "../../utilities/packed.js";
-import { e as p } from "../../webgpu_program-CbjdYLYk.js";
-import { f as d, c as u } from "../../webgpu_util-DuofJBMo.js";
-import { a as S, a6 as g } from "../../index-DSGwv2Yx.js";
+import { e as p } from "../../webgpu_program-B4HmApL1.js";
+import { f as d, c as u } from "../../webgpu_util-DYlGSwOJ.js";
+import { c as S, a6 as g } from "../../index-CUXkjxiT.js";
 class x {
   variableNames = ["cache", "item"];
   outputShape;

package/dist/ops/webgpu/attentionMask.js

@@ -1,6 +1,6 @@
-import { a as d, a6 as b } from "../../index-DSGwv2Yx.js";
+import { c as d, a6 as b } from "../../index-CUXkjxiT.js";
 import { isPackedTensor as p } from "../../utilities/packed.js";
-import { b as l } from "../../matMul16-BIT70Vya.js";
+import { b as l } from "../../matMul16-BcVC_E62.js";
 import M from "./attentionMask32_program.js";
 function w(n) {
   const { q: t, k: e } = n.inputs, { divisor: a, pastLen: o } = n.attrs, m = n.backend;

package/dist/ops/webgpu/attentionMask32_program.js

@@ -1,5 +1,5 @@
-import { e as r } from "../../webgpu_program-CbjdYLYk.js";
-import { f as a, c as u } from "../../webgpu_util-DuofJBMo.js";
+import { e as r } from "../../webgpu_program-B4HmApL1.js";
+import { f as a, c as u } from "../../webgpu_util-DYlGSwOJ.js";
 class p {
   variableNames = ["q", "k"];
   outputShape;

package/dist/ops/webgpu/clipScale.js

@@ -1,5 +1,5 @@
 import { reduce as d, ReduceProgram as g } from "./utils/reductions.js";
-import { a as S } from "../../index-DSGwv2Yx.js";
+import { c as S } from "../../index-CUXkjxiT.js";
 import h from "./utils/deviceInfo.js";
 class k extends g {
   shaderKey = "clipscale";
@@ -32,19 +32,19 @@ class k extends g {
   }
 }
 function w(o) {
-  const { x: e } = o.inputs, { invLossScaling: i, clipNorm: n } = o.attrs, s = o.backend, a = [];
+  const { x: e } = o.inputs, { invLossScaling: i, clipNorm: n } = o.attrs, c = o.backend, a = [];
   if (e.shape.length !== 1)
     throw new Error(`ClipScaleProgram requires 1D input, but got shape ${e.shape}`);
-  const l = h(s);
-  let r = 128, c = 1;
+  const l = h(c);
+  let r = 128, s = 1;
   const t = e.shape[0];
-  t <= 16 ? r = 16 : t <= 32 ? r = 32 : t <= 64 && (r = 64), t > 128 && (c = Math.ceil(t / 128));
+  t <= 16 ? r = 16 : t <= 32 ? r = 32 : t <= 64 && (r = 64), t > 128 && (s = Math.ceil(t / 128));
   const p = {
-    inSize: r * c,
+    inSize: r * s,
     outSize: 2,
     batchSize: 1,
     windowSize: r
-  }, u = new k(l, p, r), m = d(u, [e], s, [
+  }, u = new k(l, p, r), m = d(u, [e], c, [
     { type: "float32", data: [i] },
     { type: "float32", data: [n] }
   ]);

package/dist/ops/webgpu/concat16.js

@@ -1,8 +1,8 @@
-import { a as x, ah as I, N as c } from "../../index-DSGwv2Yx.js";
-import { e as D } from "../../webgpu_program-CbjdYLYk.js";
-import { f as $, c as F } from "../../webgpu_util-DuofJBMo.js";
-import { r as g } from "../../Reshape-pwprEaej.js";
-import { a as L, c as d } from "../../concat_util-C1Mxe27t.js";
+import { c as x, ah as I, U as c } from "../../index-CUXkjxiT.js";
+import { e as D } from "../../webgpu_program-B4HmApL1.js";
+import { f as $, c as F } from "../../webgpu_util-DYlGSwOJ.js";
+import { r as g } from "../../Reshape-CPBkTIH2.js";
+import { a as L, c as d } from "../../concat_util-D18dJ4fD.js";
 class T {
   outputShape;
   shaderKey;

package/dist/ops/webgpu/dropout16.js

@@ -1,7 +1,7 @@
-import { e as u } from "../../webgpu_program-CbjdYLYk.js";
-import { f as d, c as i } from "../../webgpu_util-DuofJBMo.js";
-import { a as p } from "../../index-DSGwv2Yx.js";
-class c {
+import { e as u } from "../../webgpu_program-B4HmApL1.js";
+import { f as d, c as i } from "../../webgpu_util-DYlGSwOJ.js";
+import { c } from "../../index-CUXkjxiT.js";
+class p {
   variableNames = ["x"];
   outputShape;
   shaderKey = "Dropout16";
@@ -37,7 +37,7 @@ class c {
   }
 }
 function f(t) {
-  const { x: e } = t.inputs, { dropout: o, seed: r } = t.attrs, s = t.backend, a = new c(e.shape), n = [
+  const { x: e } = t.inputs, { dropout: o, seed: r } = t.attrs, s = t.backend, a = new p(e.shape), n = [
     { type: "float32", data: [o] },
     { type: "float32", data: [r] }
   ];
@@ -48,4 +48,4 @@ const l = {
   backendName: "webgpu",
   kernelFunc: f
 };
-p(l);
+c(l);

package/dist/ops/webgpu/gatherSub.js

@@ -1,6 +1,6 @@
-import { e as u } from "../../webgpu_program-CbjdYLYk.js";
-import { f as h, c as p } from "../../webgpu_util-DuofJBMo.js";
-import { a as c, a6 as r } from "../../index-DSGwv2Yx.js";
+import { e as u } from "../../webgpu_program-B4HmApL1.js";
+import { f as h, c as p } from "../../webgpu_util-DYlGSwOJ.js";
+import { c, a6 as r } from "../../index-CUXkjxiT.js";
 class l {
   variableNames = ["labels", "logits", "values"];
   outputShape;

package/dist/ops/webgpu/gelu.js

@@ -1,9 +1,9 @@
-import { a as d } from "../../index-DSGwv2Yx.js";
-import { e as s } from "../../webgpu_program-CbjdYLYk.js";
-import { f as n, c as o } from "../../webgpu_util-DuofJBMo.js";
-import { isPackedTensor as l } from "../../utilities/packed.js";
+import { c as d } from "../../index-CUXkjxiT.js";
+import { e as s } from "../../webgpu_program-B4HmApL1.js";
+import { f as n, c as o } from "../../webgpu_util-DYlGSwOJ.js";
+import { isPackedTensor as c } from "../../utilities/packed.js";
 const u = 0.7978845608028654, r = 0.044715;
-class c {
+class l {
   outputShape;
   shaderKey;
   dispatchLayout;
@@ -38,7 +38,7 @@ class c {
   }
 }
 function x(t) {
-  const { x: e } = t.inputs, a = t.backend, i = new c(e.shape);
+  const { x: e } = t.inputs, a = t.backend, i = new l(e.shape);
   return a.runWebGPUProgram(i, [e], "float32");
 }
 const g = {
@@ -127,7 +127,7 @@ class m {
   }
 }
 function y(t) {
-  const { dy: e, x: a } = t.inputs, i = t.backend, p = l(e), h = p ? new f(a.shape) : new m(a.shape);
+  const { dy: e, x: a } = t.inputs, i = t.backend, p = c(e), h = p ? new f(a.shape) : new m(a.shape);
   return i.runWebGPUProgram(h, [e, a], p ? "packedF16" : "float32");
 }
 const b = {
@@ -137,5 +137,5 @@ const b = {
 };
 d(b);
 export {
-  c as GeluProgram
+  l as GeluProgram
 };

package/dist/ops/webgpu/matMul16.js

@@ -1,17 +1,17 @@
-import { a as V, m as y, b as g, j as X, N as U, _ as Y } from "../../index-DSGwv2Yx.js";
+import { c as V, m as U, a as g, n as X, U as y, _ as Y } from "../../index-CUXkjxiT.js";
 import { isPackedTensor as x } from "../../utilities/packed.js";
 import { reshape16 as C } from "../reshape16.js";
 import { matMulMul as Z } from "../matMulMul.js";
 import { matMulGelu as $ } from "../matMulGelu.js";
 import ee from "./matMul16_program.js";
-import { m as v } from "../../mat_mul-DP86qZtZ.js";
-import { r as E } from "../../relu-DTvZKBsZ.js";
-import { r as F } from "../../reshape-BIN71H3p.js";
-import { t as N } from "../../transpose-Qxz-4os3.js";
+import { m as v } from "../../mat_mul-DhG0Newp.js";
+import { r as E } from "../../relu-J_X6MUzx.js";
+import { r as F } from "../../reshape-BDOuCSNW.js";
+import { t as W } from "../../transpose-9kRxIXWR.js";
 function se(m) {
-  const { A: s, B: t } = m.inputs, { transposeA: p, transposeB: i, scale: h, activation: c, scaleA: d, scaleB: f, forceOutputShape: r, perm: k, causalMask: w, pastLen: W } = m.attrs, z = m.backend, b = !x(s), B = !x(t);
-  if (b && B) {
-    const u = d !== void 0 ? y(s, g(d)) : s, l = f !== void 0 ? y(t, g(f)) : t;
+  const { A: s, B: t } = m.inputs, { transposeA: p, transposeB: i, scale: h, activation: c, scaleA: d, scaleB: f, forceOutputShape: r, perm: k, causalMask: w, pastLen: z } = m.attrs, I = m.backend, B = !x(s), b = !x(t);
+  if (B && b) {
+    const u = d !== void 0 ? U(s, g(d)) : s, l = f !== void 0 ? U(t, g(f)) : t;
     if (w)
       throw new Error("Causal mask is not supported for unpacked MatMul16.");
     let e;
@@ -27,10 +27,10 @@ function se(m) {
       if (r) {
         const o = F(e, r);
         e.dispose();
-        const A = N(o, k);
+        const A = W(o, k);
         return o.dispose(), A;
       } else {
-        const o = N(e, k);
+        const o = W(e, k);
         return e.dispose(), o;
       }
     else if (r) {
@@ -39,17 +39,17 @@ function se(m) {
     } else
       return e;
   }
-  if (b && !B)
+  if (B && !b)
     throw new Error("When using mixed precision, A must be packed if B is packed.");
-  if (!b && B)
+  if (!B && b)
     throw new Error("When using mixed precision, B must be packed if A is packed.");
-  const M = s.shape.length, S = t.shape.length, I = s.shape.slice(0, -2), j = t.shape.slice(0, -2), R = U(I), D = U(j), q = Y(s.shape.slice(0, -2), t.shape.slice(0, -2)), K = Math.max(R, D), L = s.shape[M - 2], T = t.shape[S - 2], _ = s.shape[M - 1] * 2, H = t.shape[S - 1] * 2, G = C(s, [R, s.shape[M - 2], s.shape[M - 1]]), O = C(t, [D, t.shape[S - 2], t.shape[S - 1]]), a = new ee(K, L, T, _, H, p, i), n = [];
-  h !== void 0 && (a.useScale(), n.push({ type: "float32", data: [h] })), d !== void 0 && (a.useScaleA(), n.push({ type: "float32", data: [d] })), f !== void 0 && (a.useScaleB(), n.push({ type: "float32", data: [f] })), c !== void 0 && a.useActivation(c), w && (a.useCausalMask(), n.push({ type: "int32", data: [W || 0] }));
+  const M = s.shape.length, S = t.shape.length, N = s.shape.slice(0, -2), q = t.shape.slice(0, -2), R = y(N), D = y(q), K = Y(s.shape.slice(0, -2), t.shape.slice(0, -2)), L = Math.max(R, D), T = s.shape[M - 2], _ = t.shape[S - 2], j = s.shape[M - 1] * 2, H = t.shape[S - 1] * 2, G = C(s, [R, s.shape[M - 2], s.shape[M - 1]]), O = C(t, [D, t.shape[S - 2], t.shape[S - 1]]), a = new ee(L, T, _, j, H, p, i), n = [];
+  h !== void 0 && (a.useScale(), n.push({ type: "float32", data: [h] })), d !== void 0 && (a.useScaleA(), n.push({ type: "float32", data: [d] })), f !== void 0 && (a.useScaleB(), n.push({ type: "float32", data: [f] })), c !== void 0 && a.useActivation(c), w && (a.useCausalMask(), n.push({ type: "int32", data: [z || 0] }));
   const P = a.outputShape.length;
   r && (m.attrs.originalShape = a.outputShape);
-  const J = r ?? q.concat([a.outputShape[P - 2], a.outputShape[P - 1]]);
+  const J = r ?? K.concat([a.outputShape[P - 2], a.outputShape[P - 1]]);
   a.setOutputShape(J, k);
-  const Q = z.runWebGPUProgram(
+  const Q = I.runWebGPUProgram(
     a,
     [G, O],
     "packedF16",

package/dist/ops/webgpu/matMul16_program.js

@@ -1,5 +1,5 @@
-import { N as f } from "../../index-DSGwv2Yx.js";
-import { e as p } from "../../webgpu_program-CbjdYLYk.js";
+import { U as f } from "../../index-CUXkjxiT.js";
+import { e as p } from "../../webgpu_program-B4HmApL1.js";
 class B {
   variableNames = ["A", "B"];
   outputShape;

package/dist/ops/webgpu/mul16.js

@@ -1,13 +1,13 @@
-import { a as t } from "../../index-DSGwv2Yx.js";
-import { BinaryOpScalarProgram as s, BinaryOpProgram as m } from "./utils/binary_op.js";
+import { c as t } from "../../index-CUXkjxiT.js";
+import { BinaryOpScalarProgram as s, BinaryOpProgram as c } from "./utils/binary_op.js";
 import { B as a } from "../../binary_op_util-pKXltfxI.js";
-function c(n) {
-  const { a: e, b: r } = n.inputs, o = n.backend, p = r.shape.length === 0 ? new s(a.MUL, e.shape) : new m(a.MUL, e.shape, r.shape);
+function m(n) {
+  const { a: e, b: r } = n.inputs, o = n.backend, p = r.shape.length === 0 ? new s(a.MUL, e.shape) : new c(a.MUL, e.shape, r.shape);
   return o.runWebGPUProgram(p, [e, r], "packedF16");
 }
 const i = {
   kernelName: "Mul16",
   backendName: "webgpu",
-  kernelFunc: c
+  kernelFunc: m
 };
 t(i);

package/dist/ops/webgpu/norm2.js

@@ -1,5 +1,5 @@
 import { reduce as g, ReduceProgram as h } from "./utils/reductions.js";
-import { a as w, N as S } from "../../index-DSGwv2Yx.js";
+import { c as w, U as S } from "../../index-CUXkjxiT.js";
 import k from "./utils/deviceInfo.js";
 class v extends h {
   shaderKey = "norm2";

package/dist/ops/webgpu/normRMS.js

@@ -1,7 +1,7 @@
-import { a as k, a6 as N } from "../../index-DSGwv2Yx.js";
+import { c as k, a6 as N } from "../../index-CUXkjxiT.js";
 import { createReduceInfo as w, reduce as S } from "./utils/reductions.js";
 import { isPackedTensor as d } from "../../utilities/packed.js";
-import { p as f } from "../../pack16-WlOSOuZA.js";
+import { p as f } from "../../pack16-bqltoUlR.js";
 import b from "./normRMS16_program.js";
 import z from "./normRMS32_program.js";
 import G from "./utils/deviceInfo.js";

package/dist/ops/webgpu/normRMSGrad.js

@@ -1,8 +1,8 @@
-import { a as $, a6 as y, e as L } from "../../index-DSGwv2Yx.js";
+import { c as $, a6 as y, h as L } from "../../index-CUXkjxiT.js";
 import { createReduceInfo as M } from "./utils/reductions.js";
-import { f as w } from "../../webgpu_util-DuofJBMo.js";
-import { e as _ } from "../../webgpu_program-CbjdYLYk.js";
-import { p as x, u as R } from "../../pack16-WlOSOuZA.js";
+import { f as w } from "../../webgpu_util-DYlGSwOJ.js";
+import { e as _ } from "../../webgpu_program-B4HmApL1.js";
+import { p as x, u as R } from "../../pack16-bqltoUlR.js";
 import { isPackedTensor as h } from "../../utilities/packed.js";
 import { reshape16 as z } from "../reshape16.js";
 import { sum16 as N } from "../sum16.js";

package/dist/ops/webgpu/pack16.js

@@ -1,14 +1,14 @@
-import { a as i } from "../../index-DSGwv2Yx.js";
+import { c as i } from "../../index-CUXkjxiT.js";
 import p from "./pack16_program.js";
 function m(e) {
-  const { x: n } = e.inputs, { scaling: a, padding: r } = e.attrs, s = e.backend;
+  const { x: n } = e.inputs, { scaling: a, padding: r } = e.attrs, c = e.backend;
   if (n.shape[n.shape.length - 1] % 2 !== 0)
     throw new Error("Last dimension of input tensor must be even to use Pack16.");
   e.attrs && (e.attrs.originalShape = n.shape);
   const t = new p(n.shape, r), o = a !== 1;
   o && t.useScaling();
-  const c = [{ type: "float32", data: [a] }];
-  return s.runWebGPUProgram(t, [n], "packedF16", o ? c : void 0);
+  const s = [{ type: "float32", data: [a] }];
+  return c.runWebGPUProgram(t, [n], "packedF16", o ? s : void 0);
 }
 const k = {
   kernelName: "Pack16",

package/dist/ops/webgpu/pack16_program.js

@@ -1,5 +1,5 @@
-import { f as o, c as a } from "../../webgpu_util-DuofJBMo.js";
-import { e as s } from "../../webgpu_program-CbjdYLYk.js";
+import { f as o, c as a } from "../../webgpu_util-DYlGSwOJ.js";
+import { e as s } from "../../webgpu_program-B4HmApL1.js";
 class h {
   outputShape;
   shaderKey = "Pack16";

package/dist/ops/webgpu/qkv.js

@@ -1,5 +1,5 @@
-import { a as h, a6 as l } from "../../index-DSGwv2Yx.js";
-import { b as f } from "../../matMul16-BIT70Vya.js";
+import { c as h, a6 as l } from "../../index-CUXkjxiT.js";
+import { b as f } from "../../matMul16-BcVC_E62.js";
 import { slice16 as a } from "../slice16.js";
 import { isPackedTensor as u } from "../../utilities/packed.js";
 function k(i) {

package/dist/ops/webgpu/rope.js

@@ -1,7 +1,7 @@
 import { isPackedTensor as y } from "../../utilities/packed.js";
-import { e as c } from "../../webgpu_program-CbjdYLYk.js";
-import { f as x, c as l } from "../../webgpu_util-DuofJBMo.js";
-import { a as w, a6 as b } from "../../index-DSGwv2Yx.js";
+import { e as c } from "../../webgpu_program-B4HmApL1.js";
+import { f as x, c as l } from "../../webgpu_util-DYlGSwOJ.js";
+import { c as w, a6 as b } from "../../index-CUXkjxiT.js";
 class v {
   variableNames = ["x", "sin", "cos"];
   outputShape;

package/dist/ops/webgpu/scatterSub.js

@@ -1,6 +1,6 @@
-import { e as p } from "../../webgpu_program-CbjdYLYk.js";
-import { f as u, c as d } from "../../webgpu_util-DuofJBMo.js";
-import { a as h, a6 as o } from "../../index-DSGwv2Yx.js";
+import { e as p } from "../../webgpu_program-B4HmApL1.js";
+import { f as u, c as d } from "../../webgpu_util-DYlGSwOJ.js";
+import { c as h, a6 as o } from "../../index-CUXkjxiT.js";
 class b {
   variableNames = ["labels", "softmaxProbs", "dy"];
   outputShape;

package/dist/ops/webgpu/slice16.js

@@ -1,7 +1,7 @@
-import { b as u, c as m, e as l } from "../../webgpu_program-CbjdYLYk.js";
-import { f, c as g } from "../../webgpu_util-DuofJBMo.js";
-import { a as S, N as k } from "../../index-DSGwv2Yx.js";
-import { p as y, a as $ } from "../../slice_util-DPY56GzQ.js";
+import { b as u, c as m, e as l } from "../../webgpu_program-B4HmApL1.js";
+import { f, c as g } from "../../webgpu_util-DYlGSwOJ.js";
+import { c as S, U as k } from "../../index-CUXkjxiT.js";
+import { p as y, a as $ } from "../../slice_util-CC35pLmT.js";
 function x(o) {
   switch (o) {
     case 1:

package/dist/ops/webgpu/softmax16.js

@@ -1,12 +1,12 @@
-import { a as S, N as h, e as b } from "../../index-DSGwv2Yx.js";
+import { c as h, U as S, h as b } from "../../index-CUXkjxiT.js";
 import { reshape16 as d } from "../reshape16.js";
 import x from "./softmax16_program.js";
 import k from "./softmax16_subgroup_program.js";
 import l from "./utils/deviceInfo.js";
-import { r as z } from "../../reshape-BIN71H3p.js";
+import { r as z } from "../../reshape-BDOuCSNW.js";
 function F(a) {
   const { inputs: t, backend: o, attrs: p } = a, { logits: e } = t, { dim: r } = p, m = o.subgroupMinSize, i = o.subgroupMaxSize, c = l(o).subgroupsSupported, s = z(e, [
-    h(e.shape) / e.shape[r],
+    S(e.shape) / e.shape[r],
     e.shape[r]
   ]), u = c ? new k(s.shape, m, i) : new x(s.shape), f = o.runWebGPUProgram(u, [s], "packedF16");
   s.dispose();
@@ -18,4 +18,4 @@ const I = {
   backendName: "webgpu",
   kernelFunc: F
 };
-S(I);
+h(I);

package/dist/ops/webgpu/softmax16_program.js

@@ -1,5 +1,5 @@
-import { e } from "../../webgpu_program-CbjdYLYk.js";
-import { f as o } from "../../webgpu_util-DuofJBMo.js";
+import { e } from "../../webgpu_program-B4HmApL1.js";
+import { f as o } from "../../webgpu_util-DYlGSwOJ.js";
 class i {
   variableNames = ["logits"];
   outputShape;

package/dist/ops/webgpu/softmax16_subgroup_program.js

@@ -1,5 +1,5 @@
-import { e as o } from "../../webgpu_program-CbjdYLYk.js";
-import { f as u } from "../../webgpu_util-DuofJBMo.js";
+import { e as o } from "../../webgpu_program-B4HmApL1.js";
+import { f as u } from "../../webgpu_util-DYlGSwOJ.js";
 class i {
   variableNames = ["logits"];
   outputShape;

package/dist/ops/webgpu/softmax16grad.js

@@ -1,4 +1,4 @@
-import { a as p } from "../../index-DSGwv2Yx.js";
+import { c as p } from "../../index-CUXkjxiT.js";
 import { createReduceInfo as m, reduce as l, ReduceProgram as i } from "./utils/reductions.js";
 import { isPackedTensor as n } from "../../utilities/packed.js";
 import k from "./utils/deviceInfo.js";
@@ -23,11 +23,11 @@ class x extends i {
   }
 }
 function b(o) {
-  const { dy: e, softmaxOutput: t } = o.inputs, a = o.backend, s = k(a), u = n(e), c = n(t);
+  const { dy: e, softmaxOutput: t } = o.inputs, r = o.backend, s = k(r), u = n(e), c = n(t);
   if (!(u && c))
     throw new Error("softmaxGradGPU: dy and softmaxOutput must be packed tensors");
-  const r = [e, t], f = m(r, -1), d = new x(s, f);
-  return l(d, r, a);
+  const a = [e, t], f = m(a, -1), d = new x(s, f);
+  return l(d, a, r);
 }
 const v = {
   kernelName: "Softmax16Grad",

package/dist/ops/webgpu/sub16.js

@@ -1,13 +1,13 @@
-import { a as t } from "../../index-DSGwv2Yx.js";
+import { c as t } from "../../index-CUXkjxiT.js";
 import { BinaryOpProgram as s } from "./utils/binary_op.js";
-import { B as p } from "../../binary_op_util-pKXltfxI.js";
-function c(e) {
-  const { a: r, b: n } = e.inputs, a = e.backend, o = new s(p.SUB, r.shape, n.shape);
-  return a.runWebGPUProgram(o, [r, n], "packedF16");
+import { B as c } from "../../binary_op_util-pKXltfxI.js";
+function p(e) {
+  const { a: r, b: n } = e.inputs, o = e.backend, a = new s(c.SUB, r.shape, n.shape);
+  return o.runWebGPUProgram(a, [r, n], "packedF16");
 }
 const m = {
   kernelName: "Sub16",
   backendName: "webgpu",
-  kernelFunc: c
+  kernelFunc: p
 };
 t(m);

package/dist/ops/webgpu/sum16.js

@@ -1,10 +1,10 @@
 import { createReduceInfo as h, reduce as g, ReduceProgram as x } from "./utils/reductions.js";
-import { a as k, ah as A } from "../../index-DSGwv2Yx.js";
+import { c as k, ah as A } from "../../index-CUXkjxiT.js";
 import { isPackedTensor as P } from "../../utilities/packed.js";
 import { transpose16 as b } from "../transpose16.js";
 import I from "./utils/deviceInfo.js";
-import { s as w } from "../../sum-D39FeU5h.js";
-import { a as D, b as K } from "../../axis_util-QWWgLjut.js";
+import { s as w } from "../../sum-BpiwSWvg.js";
+import { a as D, b as K } from "../../axis_util-GTVlo58H.js";
 class v extends x {
   shaderKey = "sum16";
   constructor(e, t, o) {