@genai-fi/nanogpt 0.19.1 → 0.20.1

package/dist/ops/webgl/adamMoments.js

@@ -1,23 +1,27 @@
-import { c as m } from "../../index-CUXkjxiT.js";
-class i {
-  variableNames = ["moments", "gradient"];
-  outputShape;
-  userCode;
-  customUniforms = [
-    { name: "beta1", type: "float" },
-    { name: "beta2", type: "float" }
-  ];
-  constructor(e) {
-    this.outputShape = e;
-    const t = e.length, n = t === 1 ? "int" : `ivec${Math.min(t, 4)}`, a = t === 1 ? "coords" : `coords[${t - 1}]`, r = t === 1 ? "coords" : e.slice(0, -1).map((s, c) => `coords[${c}]`).join(", ");
-    this.userCode = `
+import { Ii as e } from "../../dist-BewPQWjc.js";
+//#region lib/ops/webgl/adamMoments.ts
+var t = class {
+	variableNames = ["moments", "gradient"];
+	outputShape;
+	userCode;
+	customUniforms = [{
+		name: "beta1",
+		type: "float"
+	}, {
+		name: "beta2",
+		type: "float"
+	}];
+	constructor(e) {
+		this.outputShape = e;
+		let t = e.length, n = t === 1 ? "int" : `ivec${Math.min(t, 4)}`, r = t === 1 ? "coords" : `coords[${t - 1}]`, i = t === 1 ? "coords" : e.slice(0, -1).map((e, t) => `coords[${t}]`).join(", ");
+		this.userCode = `
         void main() {
             float m = getMomentsAtOutCoords();
             ${n} coords = getOutputCoords();
 
             // Add gradient clipping here
-            float g = clamp(getGradient(${r}), -1.0, 1.0);
-            int which = ${a};
+            float g = clamp(getGradient(${i}), -1.0, 1.0);
+            int which = ${r};
 
             float beta = which == 0 ? beta1 : beta2;
             float gg = which == 0 ? g : g * g;
@@ -26,15 +30,15 @@ class i {
             setOutput(newM);
         }
         `;
-  }
-}
-function d(o) {
-  const { moments: e, gradient: t } = o.inputs, { beta1: n, beta2: a } = o.attrs, r = o.backend, s = new i(e.shape);
-  return r.runWebGLProgram(s, [e, t], "float32", [[n], [a]]);
-}
-const g = {
-  kernelName: "AdamMoments",
-  backendName: "webgl",
-  kernelFunc: d
+	}
 };
-m(g);
+function n(e) {
+	let { moments: n, gradient: r } = e.inputs, { beta1: i, beta2: a } = e.attrs, o = e.backend, s = new t(n.shape);
+	return o.runWebGLProgram(s, [n, r], "float32", [[i], [a]]);
+}
+e({
+	kernelName: "AdamMoments",
+	backendName: "webgl",
+	kernelFunc: n
+});
+//#endregion

package/dist/ops/webgl/appendCache.js

@@ -1,12 +1,21 @@
-import { c as p } from "../../index-CUXkjxiT.js";
-class m {
-  variableNames = ["cache", "item"];
-  outputShape;
-  userCode;
-  customUniforms = [{ name: "cacheT", type: "int" }];
-  constructor(t, a, n, o, c) {
-    const s = Math.min(n + 1, c);
-    this.outputShape = [t, a, s, o], this.userCode = `
+import { Ii as e } from "../../dist-BewPQWjc.js";
+//#region lib/ops/webgl/appendCache.ts
+var t = class {
+	variableNames = ["cache", "item"];
+	outputShape;
+	userCode;
+	customUniforms = [{
+		name: "cacheT",
+		type: "int"
+	}];
+	constructor(e, t, n, r, i) {
+		let a = Math.min(n + 1, i);
+		this.outputShape = [
+			e,
+			t,
+			a,
+			r
+		], this.userCode = `
         void main() {
             ivec4 coords = getOutputCoords(); // [b, h, t, d]
             int b = coords.x;
@@ -15,7 +24,7 @@ class m {
             int d = coords.w;
 
             int itemT = 1;
-            int maxSize = ${c};
+            int maxSize = ${i};
             int totalT = cacheT + itemT;
             int start = totalT >= maxSize ? 1 : 0;
 
@@ -30,15 +39,15 @@ class m {
             setOutput(val);
         }
         `;
-  }
-}
-function d(e) {
-  const { cache: t, item: a } = e.inputs, { maxSize: n, pastLen: o } = e.attrs, c = e.backend, s = t.shape[0], r = t.shape[2], i = t.shape[1], h = new m(s, i, r, a.shape[3], n);
-  return c.runWebGLProgram(h, [t, a], "float32", [[o]]);
-}
-const l = {
-  kernelName: "AppendCache",
-  backendName: "webgl",
-  kernelFunc: d
+	}
 };
-p(l);
+function n(e) {
+	let { cache: n, item: r } = e.inputs, { maxSize: i, pastLen: a } = e.attrs, o = e.backend, s = n.shape[0], c = n.shape[2], l = n.shape[1], u = new t(s, l, c, r.shape[3], i);
+	return o.runWebGLProgram(u, [n, r], "float32", [[a]]);
+}
+e({
+	kernelName: "AppendCache",
+	backendName: "webgl",
+	kernelFunc: n
+});
+//#endregion

package/dist/ops/webgl/attentionMask.js

@@ -1,15 +1,30 @@
-import { c as m } from "../../index-CUXkjxiT.js";
-class h {
-  variableNames = ["q", "k"];
-  outputShape;
-  userCode;
-  customUniforms = [
-    { name: "divisor", type: "float" },
-    { name: "pastLen", type: "int" },
-    { name: "inf", type: "float" }
-  ];
-  constructor(t, e, s, n, a) {
-    this.outputShape = [t, e, s, n], this.userCode = `
+import { Ii as e } from "../../dist-BewPQWjc.js";
+//#region lib/ops/webgl/attentionMask.ts
+var t = class {
+	variableNames = ["q", "k"];
+	outputShape;
+	userCode;
+	customUniforms = [
+		{
+			name: "divisor",
+			type: "float"
+		},
+		{
+			name: "pastLen",
+			type: "int"
+		},
+		{
+			name: "inf",
+			type: "float"
+		}
+	];
+	constructor(e, t, n, r, i) {
+		this.outputShape = [
+			e,
+			t,
+			n,
+			r
+		], this.userCode = `
         void main() {
             ivec4 coords = getOutputCoords(); // [batch, nh, t1, t2]
             int b = coords.x;
@@ -18,7 +33,7 @@ class h {
             int t2 = coords.w;
 
             float sum = 0.0;
-            for (int i = 0; i < ${a}; ++i) {
+            for (int i = 0; i < ${i}; ++i) {
                 float qv = getQ(b, h, t1, i);
                 float kv = getK(b, h, t2, i);
                 sum += qv * kv;
@@ -31,15 +46,19 @@ class h {
             setOutput((t2 > t1 + pastLen) ? inf : scaled);
         }
         `;
-  }
-}
-function l(o) {
-  const { q: t, k: e } = o.inputs, { divisor: s, pastLen: n } = o.attrs, a = o.backend, i = t.shape[0], r = t.shape[2], c = e.shape[2], u = t.shape[1], p = t.shape[3], d = new h(i, u, r, c, p);
-  return a.runWebGLProgram(d, [t, e], "float32", [[s], [n], [Number.NEGATIVE_INFINITY]]);
-}
-const f = {
-  kernelName: "AttentionMask",
-  backendName: "webgl",
-  kernelFunc: l
+	}
 };
-m(f);
+function n(e) {
+	let { q: n, k: r } = e.inputs, { divisor: i, pastLen: a } = e.attrs, o = e.backend, s = n.shape[0], c = n.shape[2], l = r.shape[2], u = n.shape[1], d = n.shape[3], f = new t(s, u, c, l, d);
+	return o.runWebGLProgram(f, [n, r], "float32", [
+		[i],
+		[a],
+		[-Infinity]
+	]);
+}
+e({
+	kernelName: "AttentionMask",
+	backendName: "webgl",
+	kernelFunc: n
+});
+//#endregion

package/dist/ops/webgl/dropout16.js

@@ -1,11 +1,12 @@
-import { c as n } from "../../index-CUXkjxiT.js";
-function o(e) {
-  const { x: r } = e.inputs;
-  return r;
+import { Ii as e } from "../../dist-BewPQWjc.js";
+//#region lib/ops/webgl/dropout16.ts
+function t(e) {
+	let { x: t } = e.inputs;
+	return t;
 }
-const t = {
-  kernelName: "Dropout16",
-  backendName: "webgl",
-  kernelFunc: o
-};
-n(t);
+e({
+	kernelName: "Dropout16",
+	backendName: "webgl",
+	kernelFunc: t
+});
+//#endregion

package/dist/ops/webgl/fusedSoftmax.js

@@ -1,80 +1,70 @@
-import { m as b, s as I, r as k } from "../../RealDiv-CGwv0liw.js";
-import { r as v } from "../../Reshape-BW__R4mZ.js";
-import { c as w, ah as P } from "../../index-CUXkjxiT.js";
-import { e as S } from "../../axis_util-GTVlo58H.js";
-class T {
-  variableNames = ["logits", "maxLogits"];
-  outputShape;
-  userCode;
-  constructor(t) {
-    this.outputShape = t, this.userCode = `
-        void main() {
-        ivec4 coords = getOutputCoords(); // [batch, nh, t1, t2]
-            int b = coords.x;
-            int h = coords.y;
-            int t1 = coords.z;
-            int t2 = coords.w;
-            float x = getLogitsAtOutCoords();
-            float maxLogit = getMaxLogits(b, h, t1);
-            setOutput(exp(x - maxLogit));
-        }
-        `;
-  }
-}
-class C {
-  variableNames = ["exp", "sum"];
-  outputShape;
-  userCode;
-  customUniforms = [
-    { name: "dropoutRate", type: "float" },
-    { name: "seed", type: "float" }
-  ];
-  constructor(t) {
-    this.outputShape = t, this.userCode = `
-        float random(ivec4 coords) {
-            float x = float(coords.x * 4096 + coords.y * 256 + coords.z * 16 + coords.w);
-            return fract(sin(seed + x) * 43758.5453123);
-        }
-        void main() {
-            ivec4 coords = getOutputCoords();
-            float numerator = getExp(coords.x, coords.y, coords.z, coords.w);
-            float denominator = getSum(coords.x, coords.y, coords.z, coords.w);
-            float val = numerator / denominator;
-            float keepProb = 1.0 - dropoutRate;
-            float rand = random(coords);
-            float mask = step(rand, keepProb);
-            setOutput(val * mask / keepProb);
-        }
-        `;
-  }
-}
-function L(r) {
-  const { inputs: t, attrs: m } = r, { logits: e } = t, { dim: c, dropoutRate: n, seed: u } = m, o = r.backend;
-  if (!e)
-    throw new Error("Error in softmax: input logits is null");
-  const i = P([c], e.shape), d = b({
-    inputs: { x: e },
-    backend: o,
-    attrs: { reductionIndices: i, keepDims: !1 }
-  }), f = S(d.shape, i), l = new T(e.shape), s = o.runWebGLProgram(l, [e, d], "float32");
-  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 h = new C(e.shape), g = o.runWebGLProgram(h, [s, a], "float32", [
-      [n],
-      [u ?? Math.random() * 1e4]
-    ]);
-    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;
-}
-const E = {
-  kernelName: "FusedSoftmax",
-  backendName: "webgl",
-  kernelFunc: L
-};
-w(E);
-export {
-  L as softmax
+import { Ii as e, Tn as t, nc as n } from "../../dist-BewPQWjc.js";
+import { t as r } from "../../Reshape-CABOPB9d.js";
+import { a as i, r as a, t as o } from "../../RealDiv-DBu0FQqT.js";
+//#region lib/ops/webgl/fusedSoftmax.ts
+var s = class {
+	variableNames = ["logits", "maxLogits"];
+	outputShape;
+	userCode;
+	constructor(e) {
+		this.outputShape = e, this.userCode = "\n        void main() {\n        ivec4 coords = getOutputCoords(); // [batch, nh, t1, t2]\n            int b = coords.x;\n            int h = coords.y;\n            int t1 = coords.z;\n            int t2 = coords.w;\n            float x = getLogitsAtOutCoords();\n            float maxLogit = getMaxLogits(b, h, t1);\n            setOutput(exp(x - maxLogit));\n        }\n        ";
+	}
+}, c = class {
+	variableNames = ["exp", "sum"];
+	outputShape;
+	userCode;
+	customUniforms = [{
+		name: "dropoutRate",
+		type: "float"
+	}, {
+		name: "seed",
+		type: "float"
+	}];
+	constructor(e) {
+		this.outputShape = e, this.userCode = "\n        float random(ivec4 coords) {\n            float x = float(coords.x * 4096 + coords.y * 256 + coords.z * 16 + coords.w);\n            return fract(sin(seed + x) * 43758.5453123);\n        }\n        void main() {\n            ivec4 coords = getOutputCoords();\n            float numerator = getExp(coords.x, coords.y, coords.z, coords.w);\n            float denominator = getSum(coords.x, coords.y, coords.z, coords.w);\n            float val = numerator / denominator;\n            float keepProb = 1.0 - dropoutRate;\n            float rand = random(coords);\n            float mask = step(rand, keepProb);\n            setOutput(val * mask / keepProb);\n        }\n        ";
+	}
 };
+function l(e) {
+	let { inputs: l, attrs: u } = e, { logits: d } = l, { dim: f, dropoutRate: p, seed: m } = u, h = e.backend;
+	if (!d) throw Error("Error in softmax: input logits is null");
+	let g = n([f], d.shape), _ = a({
+		inputs: { x: d },
+		backend: h,
+		attrs: {
+			reductionIndices: g,
+			keepDims: !1
+		}
+	}), v = t(_.shape, g), y = new s(d.shape), b = h.runWebGLProgram(y, [d, _], "float32");
+	h.disposeIntermediateTensorInfo(_);
+	let x = i({
+		inputs: { x: b },
+		backend: h,
+		attrs: {
+			axis: g,
+			keepDims: !1
+		}
+	}), S = r({
+		inputs: { x },
+		backend: h,
+		attrs: { shape: v }
+	});
+	if (p !== void 0 && p > 0) {
+		let e = new c(d.shape), t = h.runWebGLProgram(e, [b, S], "float32", [[p], [m ?? Math.random() * 1e4]]);
+		return h.disposeIntermediateTensorInfo(b), h.disposeIntermediateTensorInfo(x), h.disposeIntermediateTensorInfo(S), t;
+	}
+	let C = o({
+		inputs: {
+			a: b,
+			b: S
+		},
+		backend: h
+	});
+	return h.disposeIntermediateTensorInfo(b), h.disposeIntermediateTensorInfo(x), h.disposeIntermediateTensorInfo(S), C;
+}
+e({
+	kernelName: "FusedSoftmax",
+	backendName: "webgl",
+	kernelFunc: l
+});
+//#endregion
+export { l as softmax };

package/dist/ops/webgl/gatherSub.js

@@ -1,27 +1,28 @@
-import { c as l } from "../../index-CUXkjxiT.js";
-class u {
-  variableNames = ["labels", "logits", "values"];
-  outputShape;
-  userCode;
-  constructor(e) {
-    this.outputShape = [e], this.userCode = `
-          void main() {
-              int coords = getOutputCoords();
-              int index = int(getLabelsAtOutCoords());
-              float val = getValuesAtOutCoords();
-              float logit = getLogits(coords, index);
-              setOutput(val - logit);
-          }
-    `;
-  }
-}
-function i(t) {
-  const { logits: e, labels: o, values: s } = t.inputs, a = t.backend, r = o.shape[0], n = new u(r);
-  return a.runWebGLProgram(n, [o, e, s], "float32");
-}
-const c = {
-  kernelName: "EfficientGatherSub",
-  backendName: "webgl",
-  kernelFunc: i
+import { Ii as e } from "../../dist-BewPQWjc.js";
+//#region lib/ops/webgl/gatherSub.ts
+var t = class {
+	variableNames = [
+		"labels",
+		"logits",
+		"values"
+	];
+	outputShape;
+	userCode;
+	constructor(e) {
+		this.outputShape = [e], this.userCode = "\n          void main() {\n              int coords = getOutputCoords();\n              int index = int(getLabelsAtOutCoords());\n              float val = getValuesAtOutCoords();\n              float logit = getLogits(coords, index);\n              setOutput(val - logit);\n          }\n    ";
+	}
 };
-l(c);
+function n(e) {
+	let { logits: n, labels: r, values: i } = e.inputs, a = e.backend, o = r.shape[0], s = new t(o);
+	return a.runWebGLProgram(s, [
+		r,
+		n,
+		i
+	], "float32");
+}
+e({
+	kernelName: "EfficientGatherSub",
+	backendName: "webgl",
+	kernelFunc: n
+});
+//#endregion

package/dist/ops/webgl/gelu.js

@@ -1,50 +1,48 @@
-import { c as a } from "../../index-CUXkjxiT.js";
-import { u as s, C as i } from "../../kernel_funcs_utils-pq0CK9co.js";
-const t = 0.7978845608028654, r = 0.044715, c = i + `
+import { Ii as e } from "../../dist-BewPQWjc.js";
+import { _ as t, i as n } from "../../kernel_funcs_utils-HiXOOx3f.js";
+//#region lib/ops/webgl/gelu.ts
+var r = .7978845608028654, i = .044715, a = n({ opSnippet: t + `
     float x3 = x * x * x;
-    float inner = x + ${r} * x3;
-    inner = ${t} * inner;
+    float inner = x + ${i} * x3;
+    inner = ${r} * inner;
     inner = abs(inner) > 15.0 ? sign(inner) : tanh(inner);
     inner = 0.5 * (1.0 + inner);
     inner = x * inner;
     return inner;
-`, d = s({ opSnippet: c }), x = {
-  kernelName: "Gelu",
-  backendName: "webgl",
-  kernelFunc: d
-};
-a(x);
-class f {
-  // Inputs: dy, x
-  variableNames = ["dy", "x"];
-  outputShape;
-  userCode;
-  constructor(n) {
-    this.outputShape = n, this.userCode = `
+` });
+e({
+	kernelName: "Gelu",
+	backendName: "webgl",
+	kernelFunc: a
+});
+var o = class {
+	variableNames = ["dy", "x"];
+	outputShape;
+	userCode;
+	constructor(e) {
+		this.outputShape = e, this.userCode = `
             void main() {
                 float dy = getDyAtOutCoords();
                 float x  = getXAtOutCoords();
                 float x2 = x * x;
                 float x3 = x2 * x;
-                float u  = ${t} * (x + ${r} * x3);
+                float u  = ${r} * (x + ${i} * x3);
                 float t = abs(u) > 15.0 ? sign(u) : tanh(u);
                 float sech2 = 1.0 - t * t;
-                float du_dx = ${t} * (1.0 + 3.0 * ${r} * x2);
+                float du_dx = ${r} * (1.0 + 3.0 * ${i} * x2);
                 float dgelu = 0.5 * (1.0 + t) + 0.5 * x * sech2 * du_dx;
                 setOutput(dy * dgelu);
             }`;
-  }
-}
-function g(e) {
-  const { dy: n, x: o } = e.inputs, u = e.backend, l = new f(o.shape);
-  return u.runWebGLProgram(l, [n, o], "float32");
-}
-const p = {
-  kernelName: "GeluGrad",
-  backendName: "webgl",
-  kernelFunc: g
-};
-a(p);
-export {
-  d as gelu
+	}
 };
+function s(e) {
+	let { dy: t, x: n } = e.inputs, r = e.backend, i = new o(n.shape);
+	return r.runWebGLProgram(i, [t, n], "float32");
+}
+e({
+	kernelName: "GeluGrad",
+	backendName: "webgl",
+	kernelFunc: s
+});
+//#endregion
+export { a as gelu };

package/dist/ops/webgl/log.js

@@ -1,23 +1,14 @@
-import { c as e, aO as r } from "../../index-CUXkjxiT.js";
-import { u as s, l as N } from "../../kernel_funcs_utils-pq0CK9co.js";
-import { y as l } from "../../shared-3agzAqQ_.js";
-const a = N + `
-  return x < 0.0 ? NAN : log(x);
-`, t = `
-  vec4 result = log(x);
-  bvec4 isNaN = isnan(x);
-  result.r = isNaN.r ? x.r : (x.r < 0.0 ? NAN : result.r);
-  result.g = isNaN.g ? x.g : (x.g < 0.0 ? NAN : result.g);
-  result.b = isNaN.b ? x.b : (x.b < 0.0 ? NAN : result.b);
-  result.a = isNaN.a ? x.a : (x.a < 0.0 ? NAN : result.a);
-  return result;
-`, n = s({
-  opSnippet: a,
-  packedOpSnippet: t,
-  cpuKernelImpl: l
-}), o = {
-  kernelName: r,
-  backendName: "webgl",
-  kernelFunc: n
-};
-e(o);
+import { Ii as e } from "../../dist-BewPQWjc.js";
+import { y as t } from "../../shared-DOgWaqvL.js";
+import { i as n, t as r } from "../../kernel_funcs_utils-HiXOOx3f.js";
+//#region lib/ops/webgl/log.ts
+e({
+	kernelName: "Log",
+	backendName: "webgl",
+	kernelFunc: n({
+		opSnippet: r + "\n  return x < 0.0 ? NAN : log(x);\n",
+		packedOpSnippet: "\n  vec4 result = log(x);\n  bvec4 isNaN = isnan(x);\n  result.r = isNaN.r ? x.r : (x.r < 0.0 ? NAN : result.r);\n  result.g = isNaN.g ? x.g : (x.g < 0.0 ? NAN : result.g);\n  result.b = isNaN.b ? x.b : (x.b < 0.0 ? NAN : result.b);\n  result.a = isNaN.a ? x.a : (x.a < 0.0 ? NAN : result.a);\n  return result;\n",
+		cpuKernelImpl: t
+	})
+});
+//#endregion

package/dist/ops/webgl/matMul16.js

@@ -1,45 +1,37 @@
-import { isPackedTensor as h } from "../../utilities/packed.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-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)
-      e = P(t, r, c(M), o, a);
-    else if (u === "gelu")
-      e = R(t, r);
-    else if (u === "relu2") {
-      const s = f(t, r, o, a), l = B(s);
-      s.dispose(), e = G(l), l.dispose();
-    } else u === "relu" ? e = B(f(t, r, o, a)) : e = f(t, r, o, a);
-    if (p)
-      if (n) {
-        const s = v(e, n);
-        e.dispose();
-        const l = g(s, p);
-        return s.dispose(), l;
-      } else {
-        const s = g(e, p);
-        return e.dispose(), s;
-      }
-    else if (n) {
-      const s = v(e, n);
-      return e.dispose(), s;
-    } else
-      return e;
-  }
-  throw new Error("Not implemented: matMul16 with packed inputs");
+import { Ii as e, Wr as t, _n as n, br as r, gr as i, hn as a, lt as o, w as s } from "../../dist-BewPQWjc.js";
+import { isPackedTensor as c } from "../../utilities/packed.js";
+import { matMulMul as l } from "../matMulMul.js";
+import { matMulGelu as u } from "../matMulGelu.js";
+//#region lib/ops/webgl/matMul16.ts
+function d(e) {
+	let { A: d, B: f } = e.inputs, { transposeA: p, transposeB: m, scale: h, activation: g, scaleA: _, scaleB: v, forceOutputShape: y, perm: b } = e.attrs, x = !c(d), S = !c(f);
+	if (x && S) {
+		let e = _ === void 0 ? d : t(d, n(_)), c = v === void 0 ? f : t(f, n(v)), x;
+		if (h !== void 0) x = l(e, c, n(h), p, m);
+		else if (g === "gelu") x = u(e, c);
+		else if (g === "relu2") {
+			let t = i(e, c, p, m), n = o(t);
+			t.dispose(), x = a(n), n.dispose();
+		} else x = g === "relu" ? o(i(e, c, p, m)) : i(e, c, p, m);
+		if (b) if (y) {
+			let e = r(x, y);
+			x.dispose();
+			let t = s(e, b);
+			return e.dispose(), t;
+		} else {
+			let e = s(x, b);
+			return x.dispose(), e;
+		}
+		else if (y) {
+			let e = r(x, y);
+			return x.dispose(), e;
+		} else return x;
+	}
+	throw Error("Not implemented: matMul16 with packed inputs");
 }
-const C = {
-  kernelName: "MatMul16",
-  backendName: "webgl",
-  kernelFunc: q
-};
-b(C);
+e({
+	kernelName: "MatMul16",
+	backendName: "webgl",
+	kernelFunc: d
+});
+//#endregion