@genai-fi/nanogpt 0.10.3 → 0.11.0

package/dist/ops/webgl/fusedSoftmax.js

@@ -1,7 +1,7 @@
-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";
+import { m as b, s as I, r as k } from "../../RealDiv-Ds-jvL09.js";
+import { r as v } from "../../Reshape-Ct266DEk.js";
+import { h as w, af as P } from "../../index-D0RBWjq8.js";
+import { e as S } from "../../axis_util-DofAuy0p.js";
 class T {
   variableNames = ["logits", "maxLogits"];
   outputShape;
@@ -60,11 +60,11 @@ function L(r) {
   o.disposeIntermediateTensorInfo(d);
   const p = I({ inputs: { x: s }, backend: o, attrs: { axis: i, keepDims: !1 } }), a = v({ inputs: { x: p }, backend: o, attrs: { shape: f } });
   if (n !== void 0 && n > 0) {
-    const g = new C(e.shape), h = o.runWebGLProgram(g, [s, a], "float32", [
+    const h = new C(e.shape), g = o.runWebGLProgram(h, [s, a], "float32", [
       [n],
       [c ?? Math.random() * 1e4]
     ]);
-    return o.disposeIntermediateTensorInfo(s), o.disposeIntermediateTensorInfo(p), o.disposeIntermediateTensorInfo(a), h;
+    return o.disposeIntermediateTensorInfo(s), o.disposeIntermediateTensorInfo(p), o.disposeIntermediateTensorInfo(a), g;
   }
   const x = k({ inputs: { a: s, b: a }, backend: o });
   return o.disposeIntermediateTensorInfo(s), o.disposeIntermediateTensorInfo(p), o.disposeIntermediateTensorInfo(a), x;

package/dist/ops/webgl/gatherSub.js

@@ -1,4 +1,4 @@
-import { p as l } from "../../index-DOvlwCh-.js";
+import { h as l } from "../../index-D0RBWjq8.js";
 class u {
   variableNames = ["labels", "logits", "values"];
   outputShape;

package/dist/ops/webgl/gelu.js

@@ -1,5 +1,5 @@
-import { p as a } from "../../index-DOvlwCh-.js";
-import { u as s, C as i } from "../../kernel_funcs_utils-CCzYdUZg.js";
+import { h as a } from "../../index-D0RBWjq8.js";
+import { u as s, C as i } from "../../kernel_funcs_utils-CSaumNDs.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

@@ -1,6 +1,6 @@
-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";
+import { h as e, ao as r } from "../../index-D0RBWjq8.js";
+import { u as s, l as N } from "../../kernel_funcs_utils-CSaumNDs.js";
+import { y as l } from "../../shared-B7USJZgw.js";
 const a = N + `
   return x < 0.0 ? NAN : log(x);
 `, t = `

package/dist/ops/webgl/matMul16.js

@@ -1,27 +1,27 @@
 import { isPackedTensor as k } from "../../utilities/packed.js";
-import { p as g, m as M, b as m } from "../../index-DOvlwCh-.js";
+import { h as g, m as M, b as m } from "../../index-D0RBWjq8.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: d, scaleB: f, forceOutputShape: t, perm: n } = p.attrs, B = !k(r), v = !k(o);
+import { m as G } from "../../mat_mul-DxpNTCRz.js";
+import { r as h } from "../../reshape-CkjKPPqB.js";
+import { t as w } from "../../transpose-JawVKyZy.js";
+function P(l) {
+  const { A: r, B: o } = l.inputs, { transposeA: p, transposeB: c, scale: u, activation: A, scaleA: d, scaleB: f, forceOutputShape: t, perm: n } = l.attrs, B = !k(r), v = !k(o);
   if (B && v) {
     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 = N(a, i, m(u), l, c) : A === "gelu" ? e = U(a, i) : e = G(a, i, l, c), n)
+    if (u !== void 0 ? e = N(a, i, m(u), p, c) : A === "gelu" ? e = U(a, i) : e = G(a, i, p, c), n)
       if (t) {
-        const s = w(e, t);
+        const s = h(e, t);
         e.dispose();
-        const b = h(s, n);
+        const b = w(s, n);
         return s.dispose(), b;
       } else {
-        const s = h(e, n);
+        const s = w(e, n);
         return e.dispose(), s;
       }
     else if (t) {
-      const s = w(e, t);
+      const s = h(e, t);
       return e.dispose(), s;
     } else
       return e;

package/dist/ops/webgl/matMulGelu.js

@@ -1,7 +1,7 @@
-import "../../index-DOvlwCh-.js";
-import "../../Reshape-Zt6eb7yh.js";
-import { a as m, b as o, c as p } from "../../matMulGelu-CzfgT6Wq.js";
-import "../../mat_mul-SjpJRLyL.js";
+import "../../index-D0RBWjq8.js";
+import "../../Reshape-Ct266DEk.js";
+import { a as m, b as o, c as p } from "../../matMulGelu-B2s_80-H.js";
+import "../../mat_mul-DxpNTCRz.js";
 export {
   m as MATMUL_SHARED_DIM_THRESHOLD,
   o as batchMatMulGeluImpl,

package/dist/ops/webgl/matMulMul.js

@@ -1,9 +1,9 @@
-import { p as u } from "../../index-DOvlwCh-.js";
-import { b as c } from "../../matMulGelu-CzfgT6Wq.js";
-const p = `
+import { h as u } from "../../index-D0RBWjq8.js";
+import { b as c } from "../../matMulGelu-B2s_80-H.js";
+const M = `
     return a * b;
 `;
-function M(r) {
+function p(r) {
   const { inputs: n, backend: a, attrs: o } = r, { x: t, kernel: e, y: l } = n, { transposeA: s, transposeB: i } = o;
   if (t === void 0 || e === void 0)
     throw new Error("BatchMatMul requires two input tensors.");
@@ -13,16 +13,16 @@ function M(r) {
     transposeA: s,
     transposeB: i,
     backend: a,
-    activationSnippet: p,
+    activationSnippet: M,
     multiplier: l
   });
 }
 const m = {
   kernelName: "MatMulMul",
   backendName: "webgl",
-  kernelFunc: M
+  kernelFunc: p
 };
 u(m);
 export {
-  M as batchMatMulKernel
+  p as batchMatMulKernel
 };

package/dist/ops/webgl/mulDropout.js

@@ -1,4 +1,4 @@
-import { p as m } from "../../index-DOvlwCh-.js";
+import { h as m } from "../../index-D0RBWjq8.js";
 class f {
   variableNames = ["a", "b"];
   outputShape;

package/dist/ops/webgl/normRMS.js

@@ -1,5 +1,5 @@
-import { p as g, e as G } from "../../index-DOvlwCh-.js";
-import { s as x } from "../../sum-DWAtNGez.js";
+import { h as p, e as G } from "../../index-D0RBWjq8.js";
+import { s as x } from "../../sum-BdplSvq_.js";
 class y {
   variableNames = ["x", "meanSquare", "gamma"];
   outputShape;
@@ -28,7 +28,7 @@ const C = {
   backendName: "webgl",
   kernelFunc: v
 };
-g(C);
+p(C);
 class b {
   variableNames = ["x", "meanSquare", "dyGamma", "dyXMean"];
   outputShape;
@@ -73,14 +73,14 @@ function M(t) {
   l.dispose();
   const f = new b(n, m, u), S = r.runWebGLProgram(f, [e, d, s, i], "float32");
   s.dispose(), i.dispose();
-  const h = new N(n, m, u), p = r.runWebGLProgram(h, [e, d, a], "float32");
+  const h = new N(n, m, u), g = r.runWebGLProgram(h, [e, d, a], "float32");
   d.dispose();
-  const q = x(G().makeTensorFromTensorInfo(p), [0, 1]);
-  return r.disposeIntermediateTensorInfo(p), [S, q];
+  const q = x(G().makeTensorFromTensorInfo(g), [0, 1]);
+  return r.disposeIntermediateTensorInfo(g), [S, q];
 }
 const k = {
   kernelName: "RMSNormGrad",
   backendName: "webgl",
   kernelFunc: M
 };
-g(k);
+p(k);

package/dist/ops/webgl/qkv.js

@@ -1,4 +1,4 @@
-import { p as i } from "../../index-DOvlwCh-.js";
+import { h as i } from "../../index-D0RBWjq8.js";
 class l {
   variableNames = ["x", "kernel"];
   outputShape;

package/dist/ops/webgl/rope.js

@@ -1,4 +1,4 @@
-import { p as h } from "../../index-DOvlwCh-.js";
+import { h as l } from "../../index-D0RBWjq8.js";
 class g {
   variableNames = ["x", "sin", "cos"];
   outputShape;
@@ -45,12 +45,12 @@ class g {
   }
 }
 function f(o) {
-  const { x: t } = o.inputs, { pastLen: s, ropeCache: e, negSin: n } = o.attrs, a = n ? e.getNegSin() : e.getSin(), r = e.getCos(), d = o.backend, i = t.shape[0], c = t.shape[1], p = t.shape[2], u = t.shape[3], l = new g(i, c, p, u);
-  return d.runWebGLProgram(l, [t, a, r], "float32", [[s]]);
+  const { x: t } = o.inputs, { pastLen: s, ropeCache: e, negSin: n } = o.attrs, a = n ? e.getNegSin() : e.getSin(), r = e.getCos(), d = o.backend, i = t.shape[0], c = t.shape[1], p = t.shape[2], u = t.shape[3], h = new g(i, c, p, u);
+  return d.runWebGLProgram(h, [t, a, r], "float32", [[s]]);
 }
 const m = {
   kernelName: "Rope",
   backendName: "webgl",
   kernelFunc: f
 };
-h(m);
+l(m);

package/dist/ops/webgl/scatterSub.js

@@ -1,4 +1,4 @@
-import { p as i } from "../../index-DOvlwCh-.js";
+import { h as i } from "../../index-D0RBWjq8.js";
 class u {
   variableNames = ["labels", "softmaxProbs", "dy"];
   outputShape;

package/dist/ops/webgpu/adamAdjust.js

@@ -1,6 +1,6 @@
-import { e as p } from "../../webgpu_program-DuOXPQol.js";
-import { f as d, c as l } from "../../webgpu_util-RxEF33Rj.js";
-import { p as f, ab as c } from "../../index-DOvlwCh-.js";
+import { e as p } from "../../webgpu_program-CAE4RICo.js";
+import { f as d, c as l } from "../../webgpu_util-BdovYhXr.js";
+import { h as f, a7 as c } from "../../index-D0RBWjq8.js";
 class h {
   variableNames = ["moments", "value"];
   outputShape;

package/dist/ops/webgpu/adamMoments.js

@@ -1,6 +1,6 @@
-import { e as u } from "../../webgpu_program-DuOXPQol.js";
-import { f as p, c as d } from "../../webgpu_util-RxEF33Rj.js";
-import { p as c, ab as f } from "../../index-DOvlwCh-.js";
+import { e as u } from "../../webgpu_program-CAE4RICo.js";
+import { f as p, c as d } from "../../webgpu_util-BdovYhXr.js";
+import { h as c, a7 as f } from "../../index-D0RBWjq8.js";
 class l {
   variableNames = ["moments", "gradient"];
   outputShape;

package/dist/ops/webgpu/add16.js

@@ -1,4 +1,4 @@
-import { p as t } from "../../index-DOvlwCh-.js";
+import { h as t } from "../../index-D0RBWjq8.js";
 import { BinaryOpProgram as p } from "./utils/binary_op.js";
 import { B as s } from "../../binary_op_util-pKXltfxI.js";
 function c(e) {

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-DuOXPQol.js";
-import { f as d, c as u } from "../../webgpu_util-RxEF33Rj.js";
-import { p as S, ab as g } from "../../index-DOvlwCh-.js";
+import { e as p } from "../../webgpu_program-CAE4RICo.js";
+import { f as d, c as u } from "../../webgpu_util-BdovYhXr.js";
+import { h as S, a7 as g } from "../../index-D0RBWjq8.js";
 class x {
   variableNames = ["cache", "item"];
   outputShape;

package/dist/ops/webgpu/attentionMask.js

@@ -1,6 +1,6 @@
-import { p as d, ab as b } from "../../index-DOvlwCh-.js";
+import { h as d, a7 as b } from "../../index-D0RBWjq8.js";
 import { isPackedTensor as p } from "../../utilities/packed.js";
-import { b as l } from "../../matMul16-BWRSOCWB.js";
+import { b as l } from "../../matMul16-cDxwemKj.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;
@@ -11,12 +11,12 @@ function w(n) {
     throw new Error("Divisor must be non-zero in AttentionMask");
   if (o < 0)
     throw new Error("pastLen must be non-negative in AttentionMask");
-  const u = new M(s, i, k, r, c), f = [
+  const u = new M(s, i, k, r, c), h = [
     { type: "float32", data: [a] },
     { type: "int32", data: [o] },
     { type: "float32", data: [Number.NEGATIVE_INFINITY] }
-  ], h = t.dtype;
-  return m.runWebGPUProgram(u, [t, e], h, f);
+  ], f = t.dtype;
+  return m.runWebGPUProgram(u, [t, e], f, h);
 }
 const A = {
   kernelName: "AttentionMask",

package/dist/ops/webgpu/attentionMask32_program.js

@@ -1,5 +1,5 @@
-import { e as r } from "../../webgpu_program-DuOXPQol.js";
-import { f as a, c as u } from "../../webgpu_util-RxEF33Rj.js";
+import { e as r } from "../../webgpu_program-CAE4RICo.js";
+import { f as a, c as u } from "../../webgpu_util-BdovYhXr.js";
 class p {
   variableNames = ["q", "k"];
   outputShape;

package/dist/ops/webgpu/concat16.js

@@ -1,8 +1,8 @@
-import { p as x, af as I, h as c } from "../../index-DOvlwCh-.js";
-import { e as D } from "../../webgpu_program-DuOXPQol.js";
-import { f as $, c as F } from "../../webgpu_util-RxEF33Rj.js";
-import { r as g } from "../../Reshape-BYkmUnAv.js";
-import { a as L, c as d } from "../../concat_util-DpW8mL_l.js";
+import { h as x, af as I, V as c } from "../../index-D0RBWjq8.js";
+import { e as D } from "../../webgpu_program-CAE4RICo.js";
+import { f as $, c as F } from "../../webgpu_util-BdovYhXr.js";
+import { r as g } from "../../Reshape-Cd6e-Otn.js";
+import { a as L, c as d } from "../../concat_util-CHsJFZJJ.js";
 class T {
   outputShape;
   shaderKey;

package/dist/ops/webgpu/gatherSub.js

@@ -1,6 +1,6 @@
-import { e as u } from "../../webgpu_program-DuOXPQol.js";
-import { f as h, c as p } from "../../webgpu_util-RxEF33Rj.js";
-import { p as c, ab as r } from "../../index-DOvlwCh-.js";
+import { e as u } from "../../webgpu_program-CAE4RICo.js";
+import { f as h, c as p } from "../../webgpu_util-BdovYhXr.js";
+import { h as c, a7 as r } from "../../index-D0RBWjq8.js";
 class l {
   variableNames = ["labels", "logits", "values"];
   outputShape;
@@ -31,9 +31,9 @@ function d(e) {
   const n = new l(i);
   return o.runWebGPUProgram(n, [a, t, s], "float32");
 }
-const b = {
+const f = {
   kernelName: "EfficientGatherSub",
   backendName: "webgpu",
   kernelFunc: d
 };
-c(b);
+c(f);

package/dist/ops/webgpu/gelu.js

@@ -1,6 +1,6 @@
-import { p as d } from "../../index-DOvlwCh-.js";
-import { e as s } from "../../webgpu_program-DuOXPQol.js";
-import { f as n, c as o } from "../../webgpu_util-RxEF33Rj.js";
+import { h as d } from "../../index-D0RBWjq8.js";
+import { e as s } from "../../webgpu_program-CAE4RICo.js";
+import { f as n, c as o } from "../../webgpu_util-BdovYhXr.js";
 import { isPackedTensor as l } from "../../utilities/packed.js";
 const u = 0.7978845608028654, r = 0.044715;
 class c {

package/dist/ops/webgpu/matMul16.js

@@ -1,25 +1,25 @@
-import { p as Q, m as P, b as B, h as y, a3 as V } from "../../index-DOvlwCh-.js";
+import { h as H, m as P, b as B, V as y, $ as J } from "../../index-D0RBWjq8.js";
 import { isPackedTensor as R } from "../../utilities/packed.js";
 import { reshape16 as U } from "../reshape16.js";
-import { matMulMul as X } from "../matMulMul.js";
-import { matMulGelu as Y } from "../matMulGelu.js";
-import Z from "./matMul16_program.js";
-import { m as _ } from "../../mat_mul-SjpJRLyL.js";
-import { r as x } from "../../reshape-ByE68wS9.js";
-import { t as C } from "../../transpose-ClWiBS_b.js";
-function $(p) {
+import { matMulMul as Q } from "../matMulMul.js";
+import { matMulGelu as X } from "../matMulGelu.js";
+import Y from "./matMul16_program.js";
+import { m as Z } from "../../mat_mul-DxpNTCRz.js";
+import { r as x } from "../../reshape-CkjKPPqB.js";
+import { t as C } from "../../transpose-JawVKyZy.js";
+function _(p) {
   const { A: e, B: s } = p.inputs, { transposeA: d, transposeB: f, scale: i, activation: k, scaleA: c, scaleB: u, forceOutputShape: o, perm: h, causalMask: g, pastLen: E } = p.attrs, F = p.backend, S = !R(e), M = !R(s);
   if (S && M) {
     const A = c !== void 0 ? P(e, B(c)) : e, b = u !== void 0 ? P(s, B(u)) : s;
     if (g)
       throw new Error("Causal mask is not supported for unpacked MatMul16.");
     let a;
-    if (i !== void 0 ? a = X(A, b, B(i), d, f) : k === "gelu" ? a = Y(A, b) : a = _(A, b, d, f), h)
+    if (i !== void 0 ? a = Q(A, b, B(i), d, f) : k === "gelu" ? a = X(A, b) : a = Z(A, b, d, f), h)
       if (o) {
         const r = x(a, o);
         a.dispose();
-        const J = C(r, h);
-        return r.dispose(), J;
+        const q = C(r, h);
+        return r.dispose(), q;
       } else {
         const r = C(a, h);
         return a.dispose(), r;
@@ -34,23 +34,23 @@ function $(p) {
     throw new Error("When using mixed precision, A must be packed if B is packed.");
   if (!S && M)
     throw new Error("When using mixed precision, B must be packed if A is packed.");
-  const l = e.shape.length, m = s.shape.length, W = e.shape.slice(0, -2), z = s.shape.slice(0, -2), v = y(W), w = y(z), I = V(e.shape.slice(0, -2), s.shape.slice(0, -2)), N = Math.max(v, w), K = e.shape[l - 2], L = s.shape[m - 2], T = e.shape[l - 1] * 2, j = s.shape[m - 1] * 2, D = U(e, [v, e.shape[l - 2], e.shape[l - 1]]), G = U(s, [w, s.shape[m - 2], s.shape[m - 1]]), t = new Z(N, K, L, T, j, d, f), n = [];
+  const l = e.shape.length, m = s.shape.length, W = e.shape.slice(0, -2), z = s.shape.slice(0, -2), v = y(W), w = y(z), I = J(e.shape.slice(0, -2), s.shape.slice(0, -2)), N = Math.max(v, w), K = e.shape[l - 2], L = s.shape[m - 2], T = e.shape[l - 1] * 2, V = s.shape[m - 1] * 2, D = U(e, [v, e.shape[l - 2], e.shape[l - 1]]), G = U(s, [w, s.shape[m - 2], s.shape[m - 1]]), t = new Y(N, K, L, T, V, d, f), n = [];
   i !== void 0 && (t.useScale(), n.push({ type: "float32", data: [i] })), c !== void 0 && (t.useScaleA(), n.push({ type: "float32", data: [c] })), u !== void 0 && (t.useScaleB(), n.push({ type: "float32", data: [u] })), k !== void 0 && t.useActivation(k), g && (t.useCausalMask(), n.push({ type: "int32", data: [E || 0] }));
   const O = t.outputShape.length;
   o && (p.attrs.originalShape = t.outputShape);
-  const q = o ?? I.concat([t.outputShape[O - 2], t.outputShape[O - 1]]);
-  t.setOutputShape(q, h);
-  const H = F.runWebGPUProgram(
+  const $ = o ?? I.concat([t.outputShape[O - 2], t.outputShape[O - 1]]);
+  t.setOutputShape($, h);
+  const j = F.runWebGPUProgram(
     t,
     [D, G],
     "packedF16",
     n.length > 0 ? n : void 0
   );
-  return D.dispose(), G.dispose(), H;
+  return D.dispose(), G.dispose(), j;
 }
 const ee = {
   kernelName: "MatMul16",
   backendName: "webgpu",
-  kernelFunc: $
+  kernelFunc: _
 };
-Q(ee);
+H(ee);

package/dist/ops/webgpu/matMul16_program.js

@@ -1,5 +1,5 @@
-import { h as f } from "../../index-DOvlwCh-.js";
-import { e as h } from "../../webgpu_program-DuOXPQol.js";
+import { V as f } from "../../index-D0RBWjq8.js";
+import { e as h } from "../../webgpu_program-CAE4RICo.js";
 class B {
   variableNames = ["A", "B"];
   outputShape;

package/dist/ops/webgpu/mul16.js

@@ -1,8 +1,8 @@
-import { p as t } from "../../index-DOvlwCh-.js";
-import { BinaryOpProgram as p } from "./utils/binary_op.js";
-import { B as m } from "../../binary_op_util-pKXltfxI.js";
+import { h as t } from "../../index-D0RBWjq8.js";
+import { BinaryOpProgram as m } 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, o = e.backend, a = new p(m.MUL, r.shape, n.shape);
+  const { a: r, b: n } = e.inputs, o = e.backend, a = new m(p.MUL, r.shape, n.shape);
   return o.runWebGPUProgram(a, [r, n], "packedF16");
 }
 const c = {

package/dist/ops/webgpu/normRMS.js

@@ -1,12 +1,12 @@
-import { p as g, ab as l } from "../../index-DOvlwCh-.js";
+import { h as g, a7 as l } from "../../index-D0RBWjq8.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-nQ6JaLo-.js";
-import b from "./normRMS16_program.js";
-import z from "./normRMS32_program.js";
-import N from "./utils/deviceInfo.js";
+import { p as f } from "../../pack16-k4jq6aMX.js";
+import z from "./normRMS16_program.js";
+import N from "./normRMS32_program.js";
+import b from "./utils/deviceInfo.js";
 function P(c) {
-  const { x: e, gamma: s } = c.inputs, m = c.backend, i = N(m), t = d(e), a = d(s), n = t || a, r = !n || t ? e : f(e), p = !n || a ? s : f(s), h = [r, p], o = w(h, -1), u = n ? new b(i, o) : new z(i, o);
+  const { x: e, gamma: s } = c.inputs, m = c.backend, i = b(m), t = d(e), a = d(s), n = t || a, r = !n || t ? e : f(e), p = !n || a ? s : f(s), h = [r, p], o = w(h, -1), u = n ? new z(i, o) : new N(i, o);
   if (l(p.shape, [r.shape[r.shape.length - 1]], "Error in RMSNorm: "), e.shape.length !== 3)
     throw new Error(`rmsNormGPU: input rank ${e.shape.length} not supported, only rank 3 is supported`);
   if (o.inSize !== r.shape[r.shape.length - 1])

package/dist/ops/webgpu/normRMSGrad.js

@@ -1,8 +1,8 @@
-import { p as _, ab as y, e as D } from "../../index-DOvlwCh-.js";
+import { h as _, a7 as y, e as D } from "../../index-D0RBWjq8.js";
 import { createReduceInfo as X } from "./utils/reductions.js";
-import { f as $ } from "../../webgpu_util-RxEF33Rj.js";
-import { e as M } from "../../webgpu_program-DuOXPQol.js";
-import { p as k, u as R } from "../../pack16-nQ6JaLo-.js";
+import { f as $ } from "../../webgpu_util-BdovYhXr.js";
+import { e as M } from "../../webgpu_program-CAE4RICo.js";
+import { p as k, u as R } from "../../pack16-k4jq6aMX.js";
 import { isPackedTensor as h } from "../../utilities/packed.js";
 import { reshape16 as L } from "../reshape16.js";
 import { sum16 as P } from "../sum16.js";

package/dist/ops/webgpu/pack16.js

@@ -1,4 +1,4 @@
-import { p as i } from "../../index-DOvlwCh-.js";
+import { h as i } from "../../index-D0RBWjq8.js";
 import p from "./pack16_program.js";
 function m(e) {
   const { x: n } = e.inputs, { scaling: a, padding: r } = e.attrs, s = e.backend;

package/dist/ops/webgpu/pack16_program.js

@@ -1,5 +1,5 @@
-import { f as o, c as a } from "../../webgpu_util-RxEF33Rj.js";
-import { e as s } from "../../webgpu_program-DuOXPQol.js";
+import { f as o, c as a } from "../../webgpu_util-BdovYhXr.js";
+import { e as s } from "../../webgpu_program-CAE4RICo.js";
 class h {
   outputShape;
   shaderKey = "Pack16";

package/dist/ops/webgpu/qkv.js

@@ -1,8 +1,8 @@
-import { p as h, ab as l } from "../../index-DOvlwCh-.js";
-import { b as f } from "../../matMul16-BWRSOCWB.js";
+import { h, a7 as l } from "../../index-D0RBWjq8.js";
+import { b as f } from "../../matMul16-cDxwemKj.js";
 import { slice16 as a } from "../slice16.js";
 import { isPackedTensor as u } from "../../utilities/packed.js";
-function b(i) {
+function k(i) {
   const { x: n, kernel: c } = i.inputs, { heads: e } = i.attrs, r = n.shape[0], t = n.shape[1], s = n.shape[2], p = u(n);
   if (l(c.shape, [p ? s * 2 : s, 3 * s], "Error in QKV: "), s % e !== 0)
     throw new Error(`Channel dimension ${s} must be divisible by number of heads ${e} in QKV.`);
@@ -16,9 +16,9 @@ function b(i) {
   ];
   return o.dispose(), m;
 }
-const k = {
+const b = {
   kernelName: "QKV",
   backendName: "webgpu",
-  kernelFunc: b
+  kernelFunc: k
 };
-h(k);
+h(b);

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-DuOXPQol.js";
-import { f as x, c as l } from "../../webgpu_util-RxEF33Rj.js";
-import { p as w, ab as b } from "../../index-DOvlwCh-.js";
+import { e as c } from "../../webgpu_program-CAE4RICo.js";
+import { f as x, c as l } from "../../webgpu_util-BdovYhXr.js";
+import { h as w, a7 as b } from "../../index-D0RBWjq8.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-DuOXPQol.js";
-import { f as u, c as d } from "../../webgpu_util-RxEF33Rj.js";
-import { p as h, ab as o } from "../../index-DOvlwCh-.js";
+import { e as p } from "../../webgpu_program-CAE4RICo.js";
+import { f as u, c as d } from "../../webgpu_util-BdovYhXr.js";
+import { h, a7 as o } from "../../index-D0RBWjq8.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-DuOXPQol.js";
-import { f, c as g } from "../../webgpu_util-RxEF33Rj.js";
-import { p as S, h as k } from "../../index-DOvlwCh-.js";
-import { p as y, a as $ } from "../../slice_util-uTKwiEpW.js";
+import { b as u, c as m, e as l } from "../../webgpu_program-CAE4RICo.js";
+import { f, c as g } from "../../webgpu_util-BdovYhXr.js";
+import { h as S, V as k } from "../../index-D0RBWjq8.js";
+import { p as y, a as $ } from "../../slice_util-D8CQRenR.js";
 function x(o) {
   switch (o) {
     case 1:

package/dist/ops/webgpu/softmax16.js

@@ -1,9 +1,9 @@
-import { p as h, h as S, e as b } from "../../index-DOvlwCh-.js";
+import { h, V as S, e as b } from "../../index-D0RBWjq8.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-ByE68wS9.js";
+import { r as z } from "../../reshape-CkjKPPqB.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, [
     S(e.shape) / e.shape[r],

package/dist/ops/webgpu/softmax16_program.js

@@ -1,5 +1,5 @@
-import { e } from "../../webgpu_program-DuOXPQol.js";
-import { f as o } from "../../webgpu_util-RxEF33Rj.js";
+import { e } from "../../webgpu_program-CAE4RICo.js";
+import { f as o } from "../../webgpu_util-BdovYhXr.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-DuOXPQol.js";
-import { f as u } from "../../webgpu_util-RxEF33Rj.js";
+import { e as o } from "../../webgpu_program-CAE4RICo.js";
+import { f as u } from "../../webgpu_util-BdovYhXr.js";
 class i {
   variableNames = ["logits"];
   outputShape;

package/dist/ops/webgpu/softmax16grad.js

@@ -1,4 +1,4 @@
-import { p } from "../../index-DOvlwCh-.js";
+import { h as p } from "../../index-D0RBWjq8.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";

package/dist/ops/webgpu/sub16.js

@@ -1,8 +1,8 @@
-import { p as t } from "../../index-DOvlwCh-.js";
-import { BinaryOpProgram as p } from "./utils/binary_op.js";
-import { B as s } from "../../binary_op_util-pKXltfxI.js";
+import { h as t } from "../../index-D0RBWjq8.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, o = e.backend, a = new p(s.SUB, r.shape, n.shape);
+  const { a: r, b: n } = e.inputs, o = e.backend, a = new s(p.SUB, r.shape, n.shape);
   return o.runWebGPUProgram(a, [r, n], "packedF16");
 }
 const m = {