@genai-fi/nanogpt 0.10.2 → 0.11.0

package/dist/ops/webgpu/rope.js

@@ -1,10 +1,8 @@
-import { isPackedTensor as w } from "../../utilities/packed.js";
-import { e as x } from "../../webgpu_program-DzaQiqel.js";
-import { f as l, c as m } from "../../webgpu_util-0_ubCEHJ.js";
-import "../../index-D6Q1lPZO.js";
-import { j as b } from "../../tensor-CzmOBsdf.js";
-import { r as v } from "../../tensor_util-DfwaWayG.js";
-class k {
+import { isPackedTensor as y } from "../../utilities/packed.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;
   shaderKey = "Rope";
@@ -13,13 +11,13 @@ class k {
   workgroupSize = [64, 1, 1];
   size = !0;
   uniforms = "pastLen: i32";
-  constructor(e, s, r, t) {
-    this.shaderKey = `Rope_${t}`, this.outputShape = [e, s, r, t], this.dispatchLayout = l(this.outputShape), this.dispatch = m(this.dispatchLayout, this.outputShape, this.workgroupSize);
+  constructor(e, o, a, t) {
+    this.shaderKey = `Rope_${t}`, this.outputShape = [e, o, a, t], this.dispatchLayout = x(this.outputShape), this.dispatch = l(this.dispatchLayout, this.outputShape, this.workgroupSize);
   }
   getUserCode() {
     const e = this.outputShape[3];
     return `
-        ${x("index")} {
+        ${c("index")} {
             if (index < uniforms.size) {
                 let coords = getCoordsFromIndex(index); // [b, h, t, d]
                 let b = coords[0];
@@ -64,7 +62,7 @@ class k {
         `;
   }
 }
-class L {
+class k {
   variableNames = ["x", "sin", "cos"];
   outputShape;
   shaderKey = "Rope";
@@ -73,12 +71,12 @@ class L {
   workgroupSize = [64, 1, 1];
   size = !0;
   uniforms = "pastLen: i32";
-  constructor(e, s, r, t) {
-    this.shaderKey = `Rope_${t}`, this.outputShape = [e, s, r, t / 2], this.dispatchLayout = l(this.outputShape), this.dispatch = m(this.dispatchLayout, this.outputShape, this.workgroupSize);
+  constructor(e, o, a, t) {
+    this.shaderKey = `Rope_${t}`, this.outputShape = [e, o, a, t / 2], this.dispatchLayout = x(this.outputShape), this.dispatch = l(this.dispatchLayout, this.outputShape, this.workgroupSize);
   }
   getUserCode() {
     return `
-        ${x("index")} {
+        ${c("index")} {
             if (index < uniforms.size) {
                 let coords = getCoordsFromIndex(index); // [b, h, t, d]
                 let b = coords[0];
@@ -116,22 +114,22 @@ class L {
         `;
   }
 }
-function P(i) {
-  const { x: e } = i.inputs, { pastLen: s, negSin: r, ropeCache: t } = i.attrs, f = i.backend, a = w(e), p = e.shape[0], h = e.shape[1], n = e.shape[2], d = a ? e.shape[3] * 2 : e.shape[3], o = r ? t.getNegSin() : t.getSin(), u = t.getCos();
-  if (b(o.shape, u.shape, "Error in Rope: "), o.shape[0] < n + s)
+function L(i) {
+  const { x: e } = i.inputs, { pastLen: o, negSin: a, ropeCache: t } = i.attrs, m = i.backend, d = y(e), p = e.shape[0], h = e.shape[1], r = e.shape[2], n = d ? e.shape[3] * 2 : e.shape[3], s = a ? t.getNegSin() : t.getSin(), u = t.getCos();
+  if (b(s.shape, u.shape, "Error in Rope: "), s.shape[0] < r + o)
     throw new Error(
-      `Sin tensor shape ${o.shape} is not compatible with seqLength ${n} and pastLen ${s}.`
+      `Sin tensor shape ${s.shape} is not compatible with seqLength ${r} and pastLen ${o}.`
     );
-  if (o.shape[1] * 2 < d)
-    throw new Error(`Sin tensor shape ${o.shape} is not compatible with feature dimension ${d}.`);
-  if (o.shape.length !== 3)
-    throw new Error(`Sin tensor must be 3-dimensional, but got shape ${o.shape}.`);
-  const S = a ? new L(p, h, n, d) : new k(p, h, n, d), g = [{ type: "int32", data: [s] }], y = a ? "int32" : e.dtype, c = f.runWebGPUProgram(S, [e, o, u], y, g);
-  return c.packed = a, c;
+  if (s.shape[1] * 2 < n)
+    throw new Error(`Sin tensor shape ${s.shape} is not compatible with feature dimension ${n}.`);
+  if (s.shape.length !== 3)
+    throw new Error(`Sin tensor must be 3-dimensional, but got shape ${s.shape}.`);
+  const f = d ? new k(p, h, r, n) : new v(p, h, r, n), S = [{ type: "int32", data: [o] }], g = d ? "packedF16" : e.dtype;
+  return m.runWebGPUProgram(f, [e, s, u], g, S);
 }
-const $ = {
+const P = {
   kernelName: "Rope",
   backendName: "webgpu",
-  kernelFunc: P
+  kernelFunc: L
 };
-v($);
+w(P);

package/dist/ops/webgpu/scatterSub.js

@@ -1,8 +1,6 @@
-import { e as p } from "../../webgpu_program-DzaQiqel.js";
-import { f as u, c as d } from "../../webgpu_util-0_ubCEHJ.js";
-import "../../index-D6Q1lPZO.js";
-import { j as s } from "../../tensor-CzmOBsdf.js";
-import { r as h } from "../../tensor_util-DfwaWayG.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;
@@ -28,11 +26,11 @@ class b {
             `;
   }
 }
-function f(o) {
-  const { logits: t, labels: e, dy: a } = o.inputs, c = o.backend, r = e.shape[0], i = t.shape[1];
-  s(a.shape, [r], "Error in EfficientScatterSub dy: "), s(t.shape, [r, i], "Error in EfficientScatterSub logits: "), s(e.shape, [r], "Error in EfficientScatterSub labels: ");
+function f(a) {
+  const { logits: t, labels: e, dy: s } = a.inputs, c = a.backend, r = e.shape[0], i = t.shape[1];
+  o(s.shape, [r], "Error in EfficientScatterSub dy: "), o(t.shape, [r, i], "Error in EfficientScatterSub logits: "), o(e.shape, [r], "Error in EfficientScatterSub labels: ");
   const n = new b(r, i);
-  return c.runWebGPUProgram(n, [e, t, a], "float32");
+  return c.runWebGPUProgram(n, [e, t, s], "float32");
 }
 const l = {
   kernelName: "EfficientScatterSub",

package/dist/ops/webgpu/slice16.js

@@ -1,10 +1,8 @@
-import { b as c, c as f, e as l } from "../../webgpu_program-DzaQiqel.js";
-import { f as g, c as S } from "../../webgpu_util-0_ubCEHJ.js";
-import "../../index-D6Q1lPZO.js";
-import { r as k } from "../../tensor_util-DfwaWayG.js";
-import { p as y, a as $ } from "../../slice_util-DyjSAD0u.js";
-import { s as x } from "../../tensor-CzmOBsdf.js";
-function b(o) {
+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:
       return "1D";
@@ -17,7 +15,7 @@ function b(o) {
   }
   return "";
 }
-class w {
+class b {
   variableNames = ["source"];
   uniforms;
   outputShape;
@@ -30,42 +28,42 @@ class w {
   start;
   size = !0;
   constructor(e, t) {
-    this.outputShape = t, this.rank = t.length, this.dispatchLayout = g(this.outputShape), this.dispatch = S(this.dispatchLayout, this.outputShape, this.workgroupSize, [
+    this.outputShape = t, this.rank = t.length, this.dispatchLayout = f(this.outputShape), this.dispatch = g(this.dispatchLayout, this.outputShape, this.workgroupSize, [
       this.workPerThread,
       1,
       1
-    ]), this.start = e, this.uniforms = `start : ${c(e.length)}, `, this.shaderKey = "slice";
+    ]), this.start = e, this.uniforms = `start : ${u(e.length)}, `, this.shaderKey = "slice";
   }
   getUserCode() {
-    const e = c(this.rank);
+    const e = u(this.rank);
     let t;
-    return this.start.length === 1 ? t = this.outputShape.map(() => "sourceLoc = uniforms.start + coords;") : t = this.outputShape.map((r, s) => `sourceLoc.${p[s]} = uniforms.start.${f(s)} + coords.${p[s]};`), `
+    return this.start.length === 1 ? t = this.outputShape.map(() => "sourceLoc = uniforms.start + coords;") : t = this.outputShape.map((r, s) => `sourceLoc.${c[s]} = uniforms.start.${m(s)} + coords.${c[s]};`), `
       ${l("index")} {
         if (index < uniforms.size) {
           var sourceLoc : ${e};
           let coords = getCoordsFromIndex(index);
           ${t.join(`
 `)}
-          result[index] = source[getIndexFromCoords${b(this.rank)}(sourceLoc, uniforms.sourceShape)];
+          result[index] = source[getIndexFromCoords${x(this.rank)}(sourceLoc, uniforms.sourceShape)];
         }
       }
     `;
   }
 }
-const p = ["x", "y", "z", "w", "u", "v"];
-function C(o) {
-  const { inputs: e, backend: t, attrs: n } = o, { x: r } = e, { begin: s, size: h } = n, [i, a] = y(r, s, h);
-  if ($(r, i, a), x(a) === 0)
+const c = ["x", "y", "z", "w", "u", "v"];
+function w(o) {
+  const { inputs: e, backend: t, attrs: n } = o, { x: r } = e, { begin: s, size: p } = n, [i, a] = y(r, s, p);
+  if ($(r, i, a), k(a) === 0)
     return t.makeTensorInfo(a, r.dtype, []);
-  const d = new w(i, a), m = [{ type: "int32", data: i }], u = t.runWebGPUProgram(d, [r], r.dtype, m);
-  return u.packed = !0, u;
+  const h = new b(i, a), d = [{ type: "int32", data: i }];
+  return t.runWebGPUProgram(h, [r], r.dtype, d);
 }
-const D = {
+const C = {
   kernelName: "Slice16",
   backendName: "webgpu",
-  kernelFunc: C
+  kernelFunc: w
 };
-k(D);
+S(C);
 export {
-  C as slice
+  w as slice
 };

package/dist/ops/webgpu/softmax16.js

@@ -1,23 +1,21 @@
-import { e as S } from "../../index-D6Q1lPZO.js";
-import { reshape16 as h } from "../reshape16.js";
-import b from "./softmax16_program.js";
-import d from "./softmax16_subgroup_program.js";
-import x from "./utils/deviceInfo.js";
-import { r as k } from "../../tensor_util-DfwaWayG.js";
-import { r as l } from "../../reshape-CaPQzFvz.js";
-import { s as z } from "../../tensor-CzmOBsdf.js";
-function I(a) {
-  const { inputs: m, backend: e, attrs: p } = a, { logits: o } = m, { dim: s } = p, i = e.subgroupMinSize, c = e.subgroupMaxSize, u = x(e).subgroupsSupported, r = l(o, [
-    z(o.shape) / o.shape[s],
-    o.shape[s]
-  ]), f = u ? new d(r.shape, i, c) : new b(r.shape), n = e.runWebGPUProgram(f, [r], "int32");
-  n.packed = !0, r.dispose();
-  const t = S().makeTensorFromTensorInfo(n), g = h(t, o.shape);
-  return t.dispose(), g;
+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-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],
+    e.shape[r]
+  ]), u = c ? new k(s.shape, m, i) : new x(s.shape), f = o.runWebGPUProgram(u, [s], "packedF16");
+  s.dispose();
+  const n = b().makeTensorFromTensorInfo(f), g = d(n, e.shape);
+  return n.dispose(), g;
 }
-const P = {
+const I = {
   kernelName: "Softmax16",
   backendName: "webgpu",
-  kernelFunc: I
+  kernelFunc: F
 };
-k(P);
+h(I);

package/dist/ops/webgpu/softmax16_program.js

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

@@ -1,8 +1,7 @@
-import "../../index-D6Q1lPZO.js";
-import { createReduceInfo as p, reduce as m, ReduceProgram as i } from "./utils/reductions.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 l from "./utils/deviceInfo.js";
-import { r as k } from "../../tensor_util-DfwaWayG.js";
+import k from "./utils/deviceInfo.js";
 class x extends i {
   constructor(e, t) {
     super(e, t, { reductionOp: "sum", elementwise: !0 }, !0), this.shaderKey = "SoftmaxGrad16", this.variableNames = ["dy", "softmaxOutput"], this.variableComponents = [1, 1];
@@ -24,15 +23,15 @@ class x extends i {
   }
 }
 function b(o) {
-  const { dy: e, softmaxOutput: t } = o.inputs, r = o.backend, s = l(r), 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 a = [e, t], f = p(a, -1), d = new x(s, f);
-  return m(d, a, r);
+  const a = [e, t], f = m(a, -1), d = new x(s, f);
+  return l(d, a, r);
 }
 const v = {
   kernelName: "Softmax16Grad",
   backendName: "webgpu",
   kernelFunc: b
 };
-k(v);
+p(v);

package/dist/ops/webgpu/sub16.js

@@ -1,14 +1,13 @@
-import "../../index-D6Q1lPZO.js";
-import { BinaryOpProgram as p } from "./utils/binary_op.js";
-import { B as m } from "../../binary_op_util-pKXltfxI.js";
-import { r as s } from "../../tensor_util-DfwaWayG.js";
-function c(r) {
-  const { a: e, b: n } = r.inputs, t = r.backend, a = new p(m.SUB, e.shape, n.shape), o = t.runWebGPUProgram(a, [e, n], "int32");
-  return o.packed = !0, o;
+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 s(p.SUB, r.shape, n.shape);
+  return o.runWebGPUProgram(a, [r, n], "packedF16");
 }
-const i = {
+const m = {
   kernelName: "Sub16",
   backendName: "webgpu",
   kernelFunc: c
 };
-s(i);
+t(m);

package/dist/ops/webgpu/sum16.js

@@ -1,40 +1,38 @@
-import { createReduceInfo as g, reduce as h, ReduceProgram as x } from "./utils/reductions.js";
-import "../../index-D6Q1lPZO.js";
-import { isPackedTensor as k } from "../../utilities/packed.js";
-import { transpose16 as A } from "../transpose16.js";
-import P from "./utils/deviceInfo.js";
-import { r as b } from "../../tensor_util-DfwaWayG.js";
-import { s as I } from "../../sum-ywRJj3Zr.js";
-import { p as w } from "../../tensor-CzmOBsdf.js";
-import { a as D, b as K } from "../../axis_util-BsIr9ZNu.js";
+import { createReduceInfo as h, reduce as g, ReduceProgram as x } from "./utils/reductions.js";
+import { h as k, af as A } from "../../index-D0RBWjq8.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-BdplSvq_.js";
+import { a as D, b as K } from "../../axis_util-DofAuy0p.js";
 class v extends x {
   shaderKey = "sum16";
-  constructor(e, o, t) {
+  constructor(e, t, o) {
     super(
       e,
-      o,
+      t,
       {
         reductionOp: "sum",
         elementwise: !1
       },
-      t
-    ), t && (this.shaderKey += "_packed");
+      o
+    ), o && (this.shaderKey += "_packed");
   }
 }
 function y(r) {
-  const { x: e } = r.inputs, { axis: o, keepDims: t } = r.attrs, m = r.backend, a = [], p = P(m), c = k(e);
-  if (!c)
-    return I(e, o, t);
-  let n = w(o ?? -1, e.shape);
+  const { x: e } = r.inputs, { axis: t, keepDims: o } = r.attrs, a = r.backend, c = [], u = I(a), m = P(e);
+  if (!m)
+    return w(e, t, o);
+  let n = A(t ?? -1, e.shape);
   const i = D(n, e.shape.length);
   let s = e;
-  i != null && (s = A(e, i), n = K(n.length, s.shape.length), a.push(s));
-  const u = g([s], -1), f = new v(p, u, c), d = h(f, [s], m);
-  return a.forEach((l) => l.dispose()), d;
+  i != null && (s = b(e, i), n = K(n.length, s.shape.length), c.push(s));
+  const p = h([s], -1), f = new v(u, p, m), d = g(f, [s], a);
+  return c.forEach((l) => l.dispose()), d;
 }
 const N = {
   kernelName: "Sum16",
   backendName: "webgpu",
   kernelFunc: y
 };
-b(N);
+k(N);

package/dist/ops/webgpu/transpose16.js

@@ -1,29 +1,28 @@
 import { isPackedTensor as d } from "../../utilities/packed.js";
-import { e as k } from "../../index-D6Q1lPZO.js";
-import { reshape16 as i } from "../reshape16.js";
-import l from "./transpose16_shared_program.js";
-import P from "./transpose16_program.js";
-import { r as b } from "../../tensor_util-DfwaWayG.js";
-import { t as T } from "../../transpose-DH4gmHvu.js";
-function w(a) {
-  const { inputs: u, attrs: h } = a, { x: e } = u, { perm: r } = h, m = a.backend, c = d(e);
-  if (c && r[r.length - 1] !== e.shape.length - 1) {
-    const n = e.shape.length, t = n === 4 ? r.map((s) => s - 1).filter((s) => s >= 0) : r, p = n === 4 ? i(e, [e.shape[0] * e.shape[1], e.shape[2], e.shape[3]]) : e, f = new l(p.shape, t), o = m.runWebGPUProgram(f, [p], "int32");
-    if (o.packed = !0, n === 4) {
-      p.dispose();
-      const s = k().makeTensorFromTensorInfo(o), g = i(s, [
+import { h as k, e as l } from "../../index-D0RBWjq8.js";
+import { reshape16 as u } from "../reshape16.js";
+import P from "./transpose16_shared_program.js";
+import b from "./transpose16_program.js";
+import { t as T } from "../../transpose-JawVKyZy.js";
+function w(p) {
+  const { inputs: h, attrs: i } = p, { x: e } = h, { perm: r } = i, a = p.backend, m = d(e);
+  if (m && r[r.length - 1] !== e.shape.length - 1) {
+    const o = e.shape.length, c = o === 4 ? r.map((s) => s - 1).filter((s) => s >= 0) : r, t = o === 4 ? u(e, [e.shape[0] * e.shape[1], e.shape[2], e.shape[3]]) : e, f = new P(t.shape, c), n = a.runWebGPUProgram(f, [t], "packedF16");
+    if (o === 4) {
+      t.dispose();
+      const s = l().makeTensorFromTensorInfo(n), g = u(s, [
         e.shape[0],
         e.shape[1],
-        o.shape[1],
-        o.shape[2]
+        n.shape[1],
+        n.shape[2]
       ]);
       return s.dispose(), g;
     }
-    return o;
+    return n;
   }
-  if (c) {
-    const n = new P(e.shape, r), t = m.runWebGPUProgram(n, [e], "int32");
-    return t.packed = !0, t;
+  if (m) {
+    const o = new b(e.shape, r);
+    return a.runWebGPUProgram(o, [e], "packedF16");
   } else
     return T(e, r);
 }
@@ -32,4 +31,4 @@ const F = {
   backendName: "webgpu",
   kernelFunc: w
 };
-b(F);
+k(F);

package/dist/ops/webgpu/transpose16_program.js

@@ -1,5 +1,5 @@
-import { f as a, c as i } from "../../webgpu_util-0_ubCEHJ.js";
-import { b as h, e as d, c as n } from "../../webgpu_program-DzaQiqel.js";
+import { f as a, c as i } from "../../webgpu_util-BdovYhXr.js";
+import { b as h, e as d, c as n } from "../../webgpu_program-CAE4RICo.js";
 function p(r) {
   const e = r.length;
   if (e > 6)

package/dist/ops/webgpu/transpose16_shared_program.js

@@ -1,8 +1,7 @@
-import { c as a } from "../../webgpu_util-0_ubCEHJ.js";
-import { e as p } from "../../webgpu_program-DzaQiqel.js";
-import "../../index-D6Q1lPZO.js";
-import { a as l } from "../../tensor-CzmOBsdf.js";
-class y {
+import { c as a } from "../../webgpu_util-BdovYhXr.js";
+import { e as l } from "../../webgpu_program-CAE4RICo.js";
+import { y as p } from "../../index-D0RBWjq8.js";
+class n {
   variableNames = ["A"];
   outputShape;
   shaderKey;
@@ -12,22 +11,22 @@ class y {
   // Nick: Reduce to 8x8
   workgroupSize = [8, 8, 1];
   constructor(t, o) {
-    const i = t.length, e = new Array(i), u = t.slice();
-    u[u.length - 1] *= 2;
-    for (let r = 0; r < e.length; r++)
-      e[r] = u[o[r]];
+    const i = t.length, e = new Array(i), r = t.slice();
+    r[r.length - 1] *= 2;
+    for (let u = 0; u < e.length; u++)
+      e[u] = r[o[u]];
     e[e.length - 1] /= 2, this.outputShape = e, this.dispatchLayout = i === 2 ? { x: [0], y: [1] } : { x: [1], y: [2], z: [0] }, this.dispatch = a(this.dispatchLayout, this.outputShape, this.workgroupSize, [2, 1, 1]), this.shaderKey = `transposeShared16_${i}`;
   }
   getUserCode() {
     const t = this.outputShape.length;
-    l(
+    p(
       this.workgroupSize[0] === this.workgroupSize[1],
       () => `Must be a square tile, current tile shape is ${this.workgroupSize[0]} x ${this.workgroupSize[1]}`
     );
     const o = this.workgroupSize[0] * 2;
     return `
       var<workgroup> tile : array<array<f32, ${o + 1}>, ${o}>;
-      ${p()} {
+      ${l()} {
         var x = i32(workgroupId.x) * ${o / 2} + i32(localId.x);
         var y = i32(workgroupId.y) * ${o} + i32(localId.y);
         let batch = ${t === 3 ? "i32(workgroupId.z)" : "0"};
@@ -67,5 +66,5 @@ class y {
   }
 }
 export {
-  y as default
+  n as default
 };

package/dist/ops/webgpu/unpack16.js

@@ -1,7 +1,6 @@
-import { f as c, c as r } from "../../webgpu_util-0_ubCEHJ.js";
-import { e as u } from "../../webgpu_program-DzaQiqel.js";
-import "../../index-D6Q1lPZO.js";
-import { r as p } from "../../tensor_util-DfwaWayG.js";
+import { f as c, c as r } from "../../webgpu_util-BdovYhXr.js";
+import { e as u } from "../../webgpu_program-CAE4RICo.js";
+import { h as p } from "../../index-D0RBWjq8.js";
 class l {
   outputShape;
   shaderKey = "Unpack16";

package/dist/ops/webgpu/utils/binary_op.js

@@ -1,10 +1,9 @@
-import { f as s, c as n } from "../../../webgpu_util-0_ubCEHJ.js";
+import { f as s, c as n } from "../../../webgpu_util-BdovYhXr.js";
 import { g as p } from "../../../binary_op_util-pKXltfxI.js";
-import { B as g } from "../../../binary_op_util-pKXltfxI.js";
-import { j as u } from "../../../index-D6Q1lPZO.js";
-import { e as l } from "../../../webgpu_program-DzaQiqel.js";
-import { l as r } from "../../../tensor-CzmOBsdf.js";
-class y {
+import { B as b } from "../../../binary_op_util-pKXltfxI.js";
+import { $ as u, ag as r } from "../../../index-D0RBWjq8.js";
+import { e as l } from "../../../webgpu_program-CAE4RICo.js";
+class x {
   dispatch;
   dispatchLayout;
   outputComponent;
@@ -74,6 +73,6 @@ class y {
   }
 }
 export {
-  y as BinaryOpProgram,
-  g as BinaryOpType
+  x as BinaryOpProgram,
+  b as BinaryOpType
 };

package/dist/ops/webgpu/utils/reductions.js

@@ -1,9 +1,8 @@
-import { e as l } from "../../../index-D6Q1lPZO.js";
-import { e as a } from "../../../webgpu_program-DzaQiqel.js";
-import { reshape16 as S } from "../../reshape16.js";
-import { f as b } from "../../../webgpu_util-0_ubCEHJ.js";
-import { p as f, s as c } from "../../../tensor-CzmOBsdf.js";
-import { c as h } from "../../../axis_util-BsIr9ZNu.js";
+import { af as l, V as c, e as S } from "../../../index-D0RBWjq8.js";
+import { e as a } from "../../../webgpu_program-CAE4RICo.js";
+import { reshape16 as b } from "../../reshape16.js";
+import { f } from "../../../webgpu_util-BdovYhXr.js";
+import { c as h } from "../../../axis_util-DofAuy0p.js";
 function d(e, u, t, i) {
   return e && !u ? `
             bestValue = subgroupAdd(bestValue);
@@ -173,11 +172,11 @@ function x(e) {
            }
          `;
 }
-function C(e, u) {
-  const t = e[0], n = f(u, t.shape), [, r] = h(t.shape, n), o = c(r), p = c(t.shape) / o;
+function L(e, u) {
+  const t = e[0], n = l(u, t.shape), [, r] = h(t.shape, n), o = c(r), p = c(t.shape) / o;
   return { windowSize: o, inSize: o, batchSize: p, outSize: 1 };
 }
-class y {
+class C {
   outputShape;
   shaderKey = "reduce16";
   dispatchLayout;
@@ -198,7 +197,7 @@ class y {
   constructor(u, t, i, n) {
     this.params = i, this.inputShape = [t.batchSize, t.inSize], this.deviceInfo = u, this.packed = n;
     const r = t.inSize % 64 === 0 ? 64 : 32;
-    u.subgroupsSupported ? (this.workgroupSize = [Math.min(r, u.subgroupMaxSize), 1, 1], this.subgroups = !0, u.variableSubgroups && (this.subgroupBuiltins = !0)) : this.workgroupSize[0] = r, this.outputShape = i.elementwise ? [t.batchSize, t.inSize] : [t.batchSize / 2], this.dispatchLayout = b(this.outputShape), this.dispatch = [i.elementwise ? t.batchSize : t.batchSize / 2, 1, 1], this.outputComponent = 1, this.variableComponents = [1], this.elementwise = i.elementwise === !0;
+    u.subgroupsSupported ? (this.workgroupSize = [Math.min(r, u.subgroupMaxSize), 1, 1], this.subgroups = !0, u.variableSubgroups && (this.subgroupBuiltins = !0)) : this.workgroupSize[0] = r, this.outputShape = i.elementwise ? [t.batchSize, t.inSize] : [t.batchSize / 2], this.dispatchLayout = f(this.outputShape), this.dispatch = [i.elementwise ? t.batchSize : t.batchSize / 2, 1, 1], this.outputComponent = 1, this.variableComponents = [1], this.elementwise = i.elementwise === !0;
   }
   getWriteSnippet() {
     return this.packed ? "result[outputIndex] = i32(pack2x16float(bestValue));" : "result[outputIndex] = bestValue;";
@@ -238,22 +237,15 @@ class y {
     });
   }
 }
-function X(e, u, t) {
-  const i = u[0], r = [{ type: "int32", data: [e.inputShape[e.inputShape.length - 1]] }], o = t.runWebGPUProgram(
-    e,
-    u,
-    e.packed ? "int32" : "float32",
-    r
-  );
-  o.packed = e.packed ?? !1;
-  const s = l().makeTensorFromTensorInfo(o), p = S(
+function y(e, u, t) {
+  const i = u[0], r = [{ type: "int32", data: [e.inputShape[e.inputShape.length - 1]] }], o = t.runWebGPUProgram(e, u, e.packed ? "packedF16" : "float32", r), s = S().makeTensorFromTensorInfo(o), p = b(
     s,
     e.elementwise ? i.shape : e.packed ? [...i.shape.slice(0, -2), i.shape[i.shape.length - 2] / 2] : [...i.shape.slice(0, -2), i.shape[i.shape.length - 2]]
   );
   return s.dispose(), p;
 }
 export {
-  y as ReduceProgram,
-  C as createReduceInfo,
-  X as reduce
+  C as ReduceProgram,
+  L as createReduceInfo,
+  y as reduce
 };