@genai-fi/nanogpt 0.10.2 → 0.11.0

package/dist/ops/webgpu/attentionMask32_program.js

@@ -1,5 +1,5 @@
-import { e as r } from "../../webgpu_program-DzaQiqel.js";
-import { f as a, c as u } from "../../webgpu_util-0_ubCEHJ.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

@@ -1,10 +1,8 @@
-import "../../index-D6Q1lPZO.js";
-import { e as x } from "../../webgpu_program-DzaQiqel.js";
-import { f as I, c as D } from "../../webgpu_util-0_ubCEHJ.js";
-import { r as y } from "../../Reshape-CDVLyVfz.js";
-import { r as $ } from "../../tensor_util-DfwaWayG.js";
-import { p as F, s as c } from "../../tensor-CzmOBsdf.js";
-import { a as L, c as d } from "../../concat_util-iBYIyuQe.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 $, c as F } from "../../webgpu_util-BdovYhXr.js";
+import { r as g } from "../../Reshape-Cd6e-Otn.js";
+import { a as L, c as d } from "../../concat_util-CHsJFZJJ.js";
 class T {
   outputShape;
   shaderKey;
@@ -21,7 +19,7 @@ class T {
       t,
       1
       /* axis */
-    ), this.variableNames = t.map((e, a) => `T${a}`), this.dispatchLayout = I(this.outputShape), this.dispatch = D(this.dispatchLayout, this.outputShape, this.workgroupSize, [
+    ), this.variableNames = t.map((e, a) => `T${a}`), this.dispatchLayout = $(this.outputShape), this.dispatch = F(this.dispatchLayout, this.outputShape, this.workgroupSize, [
       this.workPerThread,
       1,
       1
@@ -49,7 +47,7 @@ class T {
         "result[getIndexFromCoords2D(coords, uniforms.outShape)] = T0[getIndexFromCoords2D(vec2<i32>(yR, yC), uniforms.t0Shape)];"
       );
     return `
-      ${x("index")} {
+      ${D("index")} {
         for(var i = 0; i < ${this.workPerThread}; i = i + 1) {
           let flatIndex = index * ${this.workPerThread} + i;
           if(flatIndex < uniforms.size) {
@@ -86,8 +84,8 @@ function m(n, t, e) {
   }
   const l = e.runWebGPUProgram(u, i, i[0].dtype, f);
   i.forEach((o) => e.disposeData(o.dataId));
-  const g = y({ inputs: { x: l }, backend: e, attrs: { shape: s } });
-  return e.disposeData(l.dataId), g.packed = !0, g;
+  const y = g({ inputs: { x: l }, backend: e, attrs: { shape: s } });
+  return e.disposeData(l.dataId), y;
 }
 function P(n, t, e) {
   const a = d(
@@ -95,7 +93,7 @@ function P(n, t, e) {
     t
   );
   return { tensors2D: n.map(
-    (s) => y({
+    (s) => g({
       inputs: { x: s },
       backend: e,
       attrs: {
@@ -105,7 +103,7 @@ function P(n, t, e) {
   ), outShape: a };
 }
 function w(n) {
-  const { inputs: t, backend: e, attrs: a } = n, { axis: i } = a, s = F(i, t[0].shape)[0], h = t.map((r) => r.shape);
+  const { inputs: t, backend: e, attrs: a } = n, { axis: i } = a, s = I(i, t[0].shape)[0], h = t.map((r) => r.shape);
   L(h, s);
   const u = d(
     t.map((r) => r.shape),
@@ -121,7 +119,7 @@ const v = {
   backendName: "webgpu",
   kernelFunc: w
 };
-$(v);
+x(v);
 export {
   T as ConcatProgram,
   v as concatConfig

package/dist/ops/webgpu/gatherSub.js

@@ -1,8 +1,6 @@
-import { e as u } from "../../webgpu_program-DzaQiqel.js";
-import { f as p, c as h } from "../../webgpu_util-0_ubCEHJ.js";
-import "../../index-D6Q1lPZO.js";
-import { j as s } from "../../tensor-CzmOBsdf.js";
-import { r as c } from "../../tensor_util-DfwaWayG.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;
@@ -11,8 +9,8 @@ class l {
   dispatch;
   workgroupSize = [64, 1, 1];
   size = !0;
-  constructor(e) {
-    this.outputShape = [e], this.dispatchLayout = p(this.outputShape), this.dispatch = h(this.dispatchLayout, this.outputShape, this.workgroupSize);
+  constructor(t) {
+    this.outputShape = [t], this.dispatchLayout = h(this.outputShape), this.dispatch = p(this.dispatchLayout, this.outputShape, this.workgroupSize);
   }
   getUserCode() {
     return `
@@ -27,11 +25,11 @@ class l {
     `;
   }
 }
-function d(t) {
-  const { logits: e, labels: a, values: r } = t.inputs, o = t.backend, i = a.shape[0];
-  s(r.shape, [i], "Error in EfficientGatherSub: "), s(a.shape, [i], "Error in EfficientGatherSub: ");
+function d(e) {
+  const { logits: t, labels: a, values: s } = e.inputs, o = e.backend, i = a.shape[0];
+  r(s.shape, [i], "Error in EfficientGatherSub: "), r(a.shape, [i], "Error in EfficientGatherSub: ");
   const n = new l(i);
-  return o.runWebGPUProgram(n, [a, e, r], "float32");
+  return o.runWebGPUProgram(n, [a, t, s], "float32");
 }
 const f = {
   kernelName: "EfficientGatherSub",

package/dist/ops/webgpu/gelu.js

@@ -1,10 +1,9 @@
-import "../../index-D6Q1lPZO.js";
-import { e as s } from "../../webgpu_program-DzaQiqel.js";
-import { f as o, c as p } from "../../webgpu_util-0_ubCEHJ.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";
-import { r as h } from "../../tensor_util-DfwaWayG.js";
-const r = 0.7978845608028654, u = 0.044715;
-class x {
+const u = 0.7978845608028654, r = 0.044715;
+class c {
   outputShape;
   shaderKey;
   dispatchLayout;
@@ -13,7 +12,7 @@ class x {
   workgroupSize;
   size = !0;
   constructor(e) {
-    this.workgroupSize = [128, 1, 1], this.outputShape = e, this.dispatchLayout = o(this.outputShape), this.dispatch = p(this.dispatchLayout, this.outputShape, this.workgroupSize), this.shaderKey = "unary_gelu";
+    this.workgroupSize = [128, 1, 1], this.outputShape = e, this.dispatchLayout = n(this.outputShape), this.dispatch = o(this.dispatchLayout, this.outputShape, this.workgroupSize), this.shaderKey = "unary_gelu";
   }
   getUserCode() {
     return `
@@ -23,8 +22,8 @@ class x {
             }
             fn unaryOperation(x : f32) -> f32 {
                 let x3 = x * x * x;
-                var inner = fma(${u}, x3, x);
-                inner = ${r} * inner;
+                var inner = fma(${r}, x3, x);
+                inner = ${u} * inner;
                 inner = tanhComplete(inner);
                 inner = 0.5 * (1.0 + inner);
                 return x * inner;
@@ -38,17 +37,17 @@ class x {
       `;
   }
 }
-function g(t) {
-  const { x: e } = t.inputs, a = t.backend, i = new x(e.shape);
+function x(t) {
+  const { x: e } = t.inputs, a = t.backend, i = new c(e.shape);
   return a.runWebGPUProgram(i, [e], "float32");
 }
-const f = {
+const g = {
   kernelName: "Gelu",
   backendName: "webgpu",
-  kernelFunc: g
+  kernelFunc: x
 };
-h(f);
-class m {
+d(g);
+class f {
   // Inputs: dy, x
   variableNames = ["dy", "x"];
   outputShape;
@@ -58,7 +57,7 @@ class m {
   workgroupSize = [128, 1, 1];
   size = !0;
   constructor(e) {
-    this.outputShape = e, this.dispatchLayout = o(this.outputShape), this.dispatch = p(this.dispatchLayout, this.outputShape, this.workgroupSize);
+    this.outputShape = e, this.dispatchLayout = n(this.outputShape), this.dispatch = o(this.dispatchLayout, this.outputShape, this.workgroupSize);
   }
   getUserCode() {
     return `
@@ -69,10 +68,10 @@ class m {
             fn activationGrad(dy: f32, X: f32) -> f32 {
                 let x2 = X * X;
                 let x3 = x2 * X;
-                let u  = ${r} * (X + ${u} * x3);
+                let u  = ${u} * (X + ${r} * x3);
                 let t  = tanhComplete(u);
                 let sech2 = 1.0 - t * t;
-                let du_dx = ${r} * (1.0 + 3.0 * ${u} * x2);
+                let du_dx = ${u} * (1.0 + 3.0 * ${r} * x2);
                 let dgelu = 0.5 * (1.0 + t) + 0.5 * X * sech2 * du_dx;
                 return dy *dgelu;
             }
@@ -89,7 +88,7 @@ class m {
             }`;
   }
 }
-class y {
+class m {
   // Inputs: dy, x
   variableNames = ["dy", "x"];
   outputShape;
@@ -99,7 +98,7 @@ class y {
   workgroupSize = [128, 1, 1];
   size = !0;
   constructor(e) {
-    this.outputShape = e, this.dispatchLayout = o(this.outputShape), this.dispatch = p(this.dispatchLayout, this.outputShape, this.workgroupSize);
+    this.outputShape = e, this.dispatchLayout = n(this.outputShape), this.dispatch = o(this.dispatchLayout, this.outputShape, this.workgroupSize);
   }
   getUserCode() {
     return `
@@ -110,10 +109,10 @@ class y {
             fn activationGrad(dy: f32, X: f32) -> f32 {
                 let x2 = X * X;
                 let x3 = x2 * X;
-                let u  = ${r} * (X + ${u} * x3);
+                let u  = ${u} * (X + ${r} * x3);
                 let t  = tanhComplete(u);
                 let sech2 = 1.0 - t * t;
-                let du_dx = ${r} * (1.0 + 3.0 * ${u} * x2);
+                let du_dx = ${u} * (1.0 + 3.0 * ${r} * x2);
                 let dgelu = 0.5 * (1.0 + t) + 0.5 * X * sech2 * du_dx;
                 return dy *dgelu;
             }
@@ -127,16 +126,16 @@ class y {
             }`;
   }
 }
-function b(t) {
-  const { dy: e, x: a } = t.inputs, i = t.backend, n = l(e), c = n ? new m(a.shape) : new y(a.shape), d = i.runWebGPUProgram(c, [e, a], n ? "int32" : "float32");
-  return d.packed = n, d;
+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);
+  return i.runWebGPUProgram(h, [e, a], p ? "packedF16" : "float32");
 }
-const k = {
+const b = {
   kernelName: "GeluGrad",
   backendName: "webgpu",
-  kernelFunc: b
+  kernelFunc: y
 };
-h(k);
+d(b);
 export {
-  x as GeluProgram
+  c as GeluProgram
 };

package/dist/ops/webgpu/matMul16.js

@@ -1,34 +1,32 @@
-import { m as y, b as B, j as Q } from "../../index-D6Q1lPZO.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 V } from "../matMulMul.js";
+import { matMulMul as Q } from "../matMulMul.js";
 import { matMulGelu as X } from "../matMulGelu.js";
 import Y from "./matMul16_program.js";
-import { r as Z } from "../../tensor_util-DfwaWayG.js";
-import { m as _ } from "../../mat_mul-C59XWcJd.js";
-import { r as x } from "../../reshape-CaPQzFvz.js";
-import { t as C } from "../../transpose-DH4gmHvu.js";
-import { s as E } from "../../tensor-CzmOBsdf.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: m, causalMask: g, pastLen: W } = p.attrs, z = p.backend, S = !R(e), M = !R(s);
+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 ? y(e, B(c)) : e, b = u !== void 0 ? y(s, B(u)) : s;
+    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 = V(A, b, B(i), d, f) : k === "gelu" ? a = X(A, b) : a = _(A, b, d, f), m)
+    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 n = x(a, o);
+        const r = x(a, o);
         a.dispose();
-        const J = C(n, m);
-        return n.dispose(), J;
+        const q = C(r, h);
+        return r.dispose(), q;
       } else {
-        const n = C(a, m);
-        return a.dispose(), n;
+        const r = C(a, h);
+        return a.dispose(), r;
       }
     else if (o) {
-      const n = x(a, o);
-      return a.dispose(), n;
+      const r = x(a, o);
+      return a.dispose(), r;
     } else
       return a;
   }
@@ -36,23 +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 h = e.shape.length, l = s.shape.length, F = e.shape.slice(0, -2), I = s.shape.slice(0, -2), v = E(F), w = E(I), N = Q(e.shape.slice(0, -2), s.shape.slice(0, -2)), j = Math.max(v, w), K = e.shape[h - 2], L = s.shape[l - 2], T = e.shape[h - 1] * 2, q = s.shape[l - 1] * 2, D = U(e, [v, e.shape[h - 2], e.shape[h - 1]]), G = U(s, [w, s.shape[l - 2], s.shape[l - 1]]), t = new Y(j, K, L, T, q, d, f), r = [];
-  i !== void 0 && (t.useScale(), r.push({ type: "float32", data: [i] })), c !== void 0 && (t.useScaleA(), r.push({ type: "float32", data: [c] })), u !== void 0 && (t.useScaleB(), r.push({ type: "float32", data: [u] })), k !== void 0 && t.useActivation(k), g && (t.useCausalMask(), r.push({ type: "int32", data: [W || 0] }));
+  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 H = o ?? N.concat([t.outputShape[O - 2], t.outputShape[O - 1]]);
-  t.setOutputShape(H, m);
-  const P = z.runWebGPUProgram(
+  const $ = o ?? I.concat([t.outputShape[O - 2], t.outputShape[O - 1]]);
+  t.setOutputShape($, h);
+  const j = F.runWebGPUProgram(
     t,
     [D, G],
-    "int32",
-    r.length > 0 ? r : void 0
+    "packedF16",
+    n.length > 0 ? n : void 0
   );
-  return P.packed = !0, D.dispose(), G.dispose(), P;
+  return D.dispose(), G.dispose(), j;
 }
 const ee = {
   kernelName: "MatMul16",
   backendName: "webgpu",
-  kernelFunc: $
+  kernelFunc: _
 };
-Z(ee);
+H(ee);

package/dist/ops/webgpu/matMul16_program.js

@@ -1,7 +1,6 @@
-import "../../index-D6Q1lPZO.js";
-import { e as h } from "../../webgpu_program-DzaQiqel.js";
-import { s as f } from "../../tensor-CzmOBsdf.js";
-class A {
+import { V as f } from "../../index-D0RBWjq8.js";
+import { e as h } from "../../webgpu_program-CAE4RICo.js";
+class B {
   variableNames = ["A", "B"];
   outputShape;
   shaderKey = "MatMul16TB";
@@ -332,5 +331,5 @@ class A {
   }
 }
 export {
-  A as default
+  B as default
 };

package/dist/ops/webgpu/mul16.js

@@ -1,14 +1,13 @@
-import "../../index-D6Q1lPZO.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";
-import { r as c } from "../../tensor_util-DfwaWayG.js";
-function i(r) {
-  const { a: e, b: n } = r.inputs, t = r.backend, a = new m(p.MUL, e.shape, n.shape), o = t.runWebGPUProgram(a, [e, n], "int32");
-  return o.packed = !0, o;
+function s(e) {
+  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 s = {
+const c = {
   kernelName: "Mul16",
   backendName: "webgpu",
-  kernelFunc: i
+  kernelFunc: s
 };
-c(s);
+t(c);

package/dist/ops/webgpu/normRMS.js

@@ -1,30 +1,28 @@
-import "../../index-D6Q1lPZO.js";
-import { createReduceInfo as g, reduce as l } from "./utils/reductions.js";
-import { j as w } from "../../tensor-CzmOBsdf.js";
-import { isPackedTensor as f } from "../../utilities/packed.js";
-import { p as k } from "../../pack16-CmVZs6af.js";
-import S from "./normRMS16_program.js";
-import z from "./normRMS32_program.js";
-import N from "./utils/deviceInfo.js";
-import { r as b } from "../../tensor_util-DfwaWayG.js";
-function P(m) {
-  const { x: e, gamma: n } = m.inputs, c = m.backend, i = N(c), s = f(e), a = f(n), o = s || a, r = !o || s ? e : k(e), p = !o || a ? n : k(n), h = [r, p], t = g(h, -1), u = o ? new S(i, t) : new z(i, t);
-  if (w(p.shape, [r.shape[r.shape.length - 1]], "Error in RMSNorm: "), e.shape.length !== 3)
+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-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 = 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 (t.inSize !== r.shape[r.shape.length - 1])
+  if (o.inSize !== r.shape[r.shape.length - 1])
     throw new Error(
-      `rmsNormGPU: reduction size ${t.inSize} does not match expected size ${r.shape[r.shape.length - 1]}`
+      `rmsNormGPU: reduction size ${o.inSize} does not match expected size ${r.shape[r.shape.length - 1]}`
     );
-  if (t.batchSize !== e.shape[0] * e.shape[1])
+  if (o.batchSize !== e.shape[0] * e.shape[1])
     throw new Error(
-      `rmsNormGPU: batch size ${t.batchSize} does not match expected size ${e.shape[0] * e.shape[1]}`
+      `rmsNormGPU: batch size ${o.batchSize} does not match expected size ${e.shape[0] * e.shape[1]}`
     );
-  const d = l(u, h, c);
-  return d.packed = o, o && !s && r.dispose(), o && !a && p.dispose(), d;
+  const k = S(u, h, m);
+  return n && !t && r.dispose(), n && !a && p.dispose(), k;
 }
 const G = {
   kernelName: "RMSNorm",
   backendName: "webgpu",
   kernelFunc: P
 };
-b(G);
+g(G);

package/dist/ops/webgpu/normRMSGrad.js

@@ -1,14 +1,12 @@
-import { e as _ } from "../../index-D6Q1lPZO.js";
-import { createReduceInfo as D } from "./utils/reductions.js";
-import { f as X } from "../../webgpu_util-0_ubCEHJ.js";
-import { e as $ } from "../../webgpu_program-DzaQiqel.js";
-import { j as z } from "../../tensor-CzmOBsdf.js";
-import { p as k, u as M } from "../../pack16-CmVZs6af.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-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 R } from "../reshape16.js";
-import { sum16 as L } from "../sum16.js";
-import { slice16 as w } from "../slice16.js";
-import { r as P } from "../../tensor_util-DfwaWayG.js";
+import { reshape16 as L } from "../reshape16.js";
+import { sum16 as P } from "../sum16.js";
+import { slice16 as z } from "../slice16.js";
 class N {
   outputShape;
   shaderKey = "RMSNormGrad";
@@ -23,7 +21,7 @@ class N {
   packed = !1;
   outputComponent;
   constructor(a, e = 4, o = !1) {
-    if (this.packed = o, this.shaderKey = `RMSNormGrad_${e}`, this.rowsPerWorkgroup = e, this.inputShape = [a.batchSize, a.inSize], this.outputShape = [a.batchSize + a.batchSize / this.rowsPerWorkgroup, a.inSize], this.dispatchLayout = X(this.outputShape), this.dispatch = [a.batchSize / this.rowsPerWorkgroup, 1, 1], a.batchSize % this.rowsPerWorkgroup !== 0)
+    if (this.packed = o, this.shaderKey = `RMSNormGrad_${e}`, this.rowsPerWorkgroup = e, this.inputShape = [a.batchSize, a.inSize], this.outputShape = [a.batchSize + a.batchSize / this.rowsPerWorkgroup, a.inSize], this.dispatchLayout = $(this.outputShape), this.dispatch = [a.batchSize / this.rowsPerWorkgroup, 1, 1], a.batchSize % this.rowsPerWorkgroup !== 0)
       throw new Error(
         `RMSNormGradProgram: batch size ${a.batchSize} must be divisible by rowsPerWorkgroup ${this.rowsPerWorkgroup}`
       );
@@ -87,7 +85,7 @@ class N {
 
           ${o}
 
-          ${$("index")} {
+          ${M("index")} {
             // One workgroup per row (batch).
             let Length = uniforms.reduceSize;
             let BatchSize = uniforms.batchSize;
@@ -145,10 +143,10 @@ class N {
 }
 function W(p) {
   const { dy: a, x: e, gamma: o } = p.inputs, n = 4;
-  z(e.shape, a.shape, "Error in RMSNormGrad dy: ");
+  y(e.shape, a.shape, "Error in RMSNormGrad dy: ");
   const s = h(e), i = h(o), u = h(a), r = s || i || u, m = !r || s ? e : k(e), c = !r || i ? o : k(o), d = !r || u ? a : k(a);
-  z(c.shape, [m.shape[m.shape.length - 1]], "Error in RMSNormGrad gamma: ");
-  const G = p.backend, t = D([m, c, d], -1), f = new N(t, n, r), v = [
+  y(c.shape, [m.shape[m.shape.length - 1]], "Error in RMSNormGrad gamma: ");
+  const w = p.backend, t = X([m, c, d], -1), f = new N(t, n, r), G = [
     { type: "int32", data: [f.inputShape[1]] },
     // Reduce size
     { type: "int32", data: [f.inputShape[0]] }
@@ -156,27 +154,22 @@ function W(p) {
   ];
   if (t.inSize > 1024)
     throw new Error(`rmsNormGradGPU: inSize ${t.inSize} exceeds max of 1024`);
-  const x = G.runWebGPUProgram(
-    f,
-    [m, c, d],
-    r ? "int32" : "float32",
-    v
-  );
-  x.packed = r, r && !s && m.dispose(), r && !i && c.dispose(), r && !u && d.dispose();
-  const l = _().makeTensorFromTensorInfo(x), S = w(l, [0, 0], [t.batchSize, t.inSize]), g = w(
+  const v = w.runWebGPUProgram(f, [m, c, d], r ? "packedF16" : "float32", G);
+  r && !s && m.dispose(), r && !i && c.dispose(), r && !u && d.dispose();
+  const l = D().makeTensorFromTensorInfo(v), x = z(l, [0, 0], [t.batchSize, t.inSize]), S = z(
     l,
     [t.batchSize, 0],
     [t.batchSize / n, t.inSize]
   );
   l.dispose();
-  const b = R(S, e.shape);
-  S.dispose();
-  const y = L(g, [0]);
-  return g.dispose(), [b, !r || i ? y : M(y)];
+  const b = L(x, e.shape);
+  x.dispose();
+  const g = P(S, [0]);
+  return S.dispose(), [b, !r || i ? g : R(g)];
 }
 const Y = {
   kernelName: "RMSNormGrad",
   backendName: "webgpu",
   kernelFunc: W
 };
-P(Y);
+_(Y);

package/dist/ops/webgpu/pack16.js

@@ -1,19 +1,18 @@
-import "../../index-D6Q1lPZO.js";
-import c from "./pack16_program.js";
-import { r as p } from "../../tensor_util-DfwaWayG.js";
-function m(n) {
-  const { x: e } = n.inputs, { scaling: t, padding: r } = n.attrs, i = n.backend;
-  if (e.shape[e.shape.length - 1] % 2 !== 0)
+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;
+  if (n.shape[n.shape.length - 1] % 2 !== 0)
     throw new Error("Last dimension of input tensor must be even to use Pack16.");
-  n.attrs && (n.attrs.originalShape = e.shape);
-  const a = new c(e.shape, r), o = t !== 1;
-  o && a.useScaling();
-  const s = [{ type: "float32", data: [t] }];
-  return i.runWebGPUProgram(a, [e], "int32", o ? s : void 0);
+  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 u = {
+const k = {
   kernelName: "Pack16",
   backendName: "webgpu",
   kernelFunc: m
 };
-p(u);
+i(k);

package/dist/ops/webgpu/pack16_program.js

@@ -1,5 +1,5 @@
-import { f as o, c as a } from "../../webgpu_util-0_ubCEHJ.js";
-import { e as s } from "../../webgpu_program-DzaQiqel.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

@@ -1,26 +1,24 @@
-import "../../index-D6Q1lPZO.js";
-import { j as h } from "../../tensor-CzmOBsdf.js";
-import { b as f } from "../../matMul16-fEAJ4smh.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 l } from "../../utilities/packed.js";
-import { r as u } from "../../tensor_util-DfwaWayG.js";
+import { isPackedTensor as u } from "../../utilities/packed.js";
 function k(i) {
-  const { x: r, kernel: c } = i.inputs, { heads: e } = i.attrs, t = r.shape[0], n = r.shape[1], s = r.shape[2], m = l(r);
-  if (h(c.shape, [m ? s * 2 : s, 3 * s], "Error in QKV: "), s % e !== 0)
+  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.`);
-  const o = f(r, c, !1, !1, {
-    forceOutputShape: [t, n, 3 * e, s / e],
+  const o = f(n, c, !1, !1, {
+    forceOutputShape: [r, t, 3 * e, s / e],
     perm: [0, 2, 1, 3]
-  }), p = [
-    a(o, [0, 0, 0, 0], [t, e, n, s / e]),
-    a(o, [0, e, 0, 0], [t, e, n, s / e]),
-    a(o, [0, 2 * e, 0, 0], [t, e, n, s / e])
+  }), m = [
+    a(o, [0, 0, 0, 0], [r, e, t, s / e]),
+    a(o, [0, e, 0, 0], [r, e, t, s / e]),
+    a(o, [0, 2 * e, 0, 0], [r, e, t, s / e])
   ];
-  return o.dispose(), p;
+  return o.dispose(), m;
 }
 const b = {
   kernelName: "QKV",
   backendName: "webgpu",
   kernelFunc: k
 };
-u(b);
+h(b);