@genai-fi/nanogpt 0.19.1 → 0.20.1

package/dist/ops/webgl/matMulGelu.js

@@ -1,9 +1,2 @@
-import "../../index-CUXkjxiT.js";
-import "../../Reshape-BW__R4mZ.js";
-import { a as m, b as o, c as p } from "../../matMulGelu-JNLZqKQp.js";
-import "../../mat_mul-DhG0Newp.js";
-export {
-  m as MATMUL_SHARED_DIM_THRESHOLD,
-  o as batchMatMulGeluImpl,
-  p as batchMatMulKernel
-};
+import { n as e, r as t, t as n } from "../../matMulGelu-CPTntosE.js";
+export { n as MATMUL_SHARED_DIM_THRESHOLD, e as batchMatMulGeluImpl, t as batchMatMulKernel };

package/dist/ops/webgl/matMulMul.js

@@ -1,28 +1,24 @@
-import { c as u } from "../../index-CUXkjxiT.js";
-import { b as c } from "../../matMulGelu-JNLZqKQp.js";
-const M = `
-    return a * b;
-`;
-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.");
-  return c({
-    a: t,
-    b: e,
-    transposeA: s,
-    transposeB: i,
-    backend: a,
-    activationSnippet: M,
-    multiplier: l
-  });
+import { Ii as e } from "../../dist-BewPQWjc.js";
+import { n as t } from "../../matMulGelu-CPTntosE.js";
+//#region lib/ops/webgl/matMulMul.ts
+var n = "\n    return a * b;\n";
+function r(e) {
+	let { inputs: r, backend: i, attrs: a } = e, { x: o, kernel: s, y: c } = r, { transposeA: l, transposeB: u } = a;
+	if (o === void 0 || s === void 0) throw Error("BatchMatMul requires two input tensors.");
+	return t({
+		a: o,
+		b: s,
+		transposeA: l,
+		transposeB: u,
+		backend: i,
+		activationSnippet: n,
+		multiplier: c
+	});
 }
-const m = {
-  kernelName: "MatMulMul",
-  backendName: "webgl",
-  kernelFunc: p
-};
-u(m);
-export {
-  p as batchMatMulKernel
-};
+e({
+	kernelName: "MatMulMul",
+	backendName: "webgl",
+	kernelFunc: r
+});
+//#endregion
+export { r as batchMatMulKernel };

package/dist/ops/webgl/mulDropout.js

@@ -1,41 +1,32 @@
-import { c as m } from "../../index-CUXkjxiT.js";
-class f {
-  variableNames = ["a", "b"];
-  outputShape;
-  userCode;
-  customUniforms = [
-    { name: "dropoutRate", type: "float" },
-    { name: "seed", type: "float" }
-  ];
-  constructor(t, r, o) {
-    this.outputShape = [t, r, o, o], 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 a = getA(coords.x, coords.y, coords.z, coords.w);
-            float b = getB(coords.x, coords.y, coords.z, coords.w);
-            
-            float keepProb = 1.0 - dropoutRate;
-            float rand = random(coords);
-            float mask = step(rand, keepProb);
-            setOutput(a * b * mask / keepProb);
-        }
-        `;
-  }
-}
-function b(e) {
-  const { inputs: t, attrs: r } = e, { a: o, b: s } = t, { dropoutRate: a, seed: c } = r, n = e.backend, d = o.shape[0], u = o.shape[2], p = o.shape[1], l = new f(d, p, u);
-  return n.runWebGLProgram(l, [o, s], "float32", [
-    [a ?? 0],
-    [c ?? Math.random() * 1e4]
-  ]);
-}
-const i = {
-  kernelName: "MulDropout",
-  backendName: "webgl",
-  kernelFunc: b
+import { Ii as e } from "../../dist-BewPQWjc.js";
+//#region lib/ops/webgl/mulDropout.ts
+var t = class {
+	variableNames = ["a", "b"];
+	outputShape;
+	userCode;
+	customUniforms = [{
+		name: "dropoutRate",
+		type: "float"
+	}, {
+		name: "seed",
+		type: "float"
+	}];
+	constructor(e, t, n) {
+		this.outputShape = [
+			e,
+			t,
+			n,
+			n
+		], 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 a = getA(coords.x, coords.y, coords.z, coords.w);\n            float b = getB(coords.x, coords.y, coords.z, coords.w);\n            \n            float keepProb = 1.0 - dropoutRate;\n            float rand = random(coords);\n            float mask = step(rand, keepProb);\n            setOutput(a * b * mask / keepProb);\n        }\n        ";
+	}
 };
-m(i);
+function n(e) {
+	let { inputs: n, attrs: r } = e, { a: i, b: a } = n, { dropoutRate: o, seed: s } = r, c = e.backend, l = i.shape[0], u = i.shape[2], d = i.shape[1], f = new t(l, d, u);
+	return c.runWebGLProgram(f, [i, a], "float32", [[o ?? 0], [s ?? Math.random() * 1e4]]);
+}
+e({
+	kernelName: "MulDropout",
+	backendName: "webgl",
+	kernelFunc: n
+});
+//#endregion

package/dist/ops/webgl/normRMS.js

@@ -1,93 +1,114 @@
-import { c, h as G } from "../../index-CUXkjxiT.js";
-import { s as x } from "../../sum-BpiwSWvg.js";
-class y {
-  variableNames = ["x", "meanSquare"];
-  outputShape;
-  userCode;
-  constructor(a, e, o, t = !0) {
-    t && this.variableNames.push("gamma"), this.outputShape = [a, e, o], this.userCode = `
+import { Ii as e, ii as t, mn as n } from "../../dist-BewPQWjc.js";
+//#region lib/ops/webgl/normRMS.ts
+var r = class {
+	variableNames = ["x", "meanSquare"];
+	outputShape;
+	userCode;
+	constructor(e, t, n, r = !0) {
+		r && this.variableNames.push("gamma"), this.outputShape = [
+			e,
+			t,
+			n
+		], this.userCode = `
         void main() {
             ivec3 coords = getOutputCoords();
             float x = getXAtOutCoords();
             float meanSquare = getMeanSquare(coords.x, coords.y, 0);
-            ${t ? "float gamma = getGammaAtOutCoords();" : ""}
+            ${r ? "float gamma = getGammaAtOutCoords();" : ""}
             float invRms = inversesqrt(meanSquare + 1e-8);
             float normalized = x * invRms;
-            float outVal = normalized ${t ? " * gamma" : ""};
+            float outVal = normalized ${r ? " * gamma" : ""};
             setOutput(outVal);
         }
         `;
-  }
-}
-function S(r) {
-  const { x: a, gamma: e } = r.inputs, o = r.backend, t = a.shape[0], m = a.shape[1], u = a.shape[2], s = a.square().mean(-1, !0), n = new y(t, m, u, e !== void 0);
-  return o.runWebGLProgram(n, e ? [a, s, e] : [a, s], "float32");
+	}
+};
+function i(e) {
+	let { x: t, gamma: n } = e.inputs, i = e.backend, a = t.shape[0], o = t.shape[1], s = t.shape[2], c = t.square().mean(-1, !0), l = new r(a, o, s, n !== void 0);
+	return i.runWebGLProgram(l, n ? [
+		t,
+		c,
+		n
+	] : [t, c], "float32");
 }
-const N = {
-  kernelName: "RMSNorm",
-  backendName: "webgl",
-  kernelFunc: S
-}, b = {
-  kernelName: "RMSNormNoGamma",
-  backendName: "webgl",
-  kernelFunc: S
+var a = {
+	kernelName: "RMSNorm",
+	backendName: "webgl",
+	kernelFunc: i
+}, o = {
+	kernelName: "RMSNormNoGamma",
+	backendName: "webgl",
+	kernelFunc: i
 };
-c(N);
-c(b);
-class C {
-  variableNames = ["x", "meanSquare", "dyGamma", "dyXMean"];
-  outputShape;
-  userCode;
-  constructor(a, e, o) {
-    this.outputShape = [a, e, o], this.userCode = `
+e(a), e(o);
+var s = class {
+	variableNames = [
+		"x",
+		"meanSquare",
+		"dyGamma",
+		"dyXMean"
+	];
+	outputShape;
+	userCode;
+	constructor(e, t, n) {
+		this.outputShape = [
+			e,
+			t,
+			n
+		], this.userCode = `
         void main() {
             ivec3 coords = getOutputCoords();
             float x = getXAtOutCoords();
             float meanSquare = getMeanSquare(coords.x, coords.y, 0) + 1e-8;
             float dyGamma = getDyGammaAtOutCoords();
-            float dyXMean = getDyXMean(coords.x, coords.y, 0) / ${o}.0;
+            float dyXMean = getDyXMean(coords.x, coords.y, 0) / ${n}.0;
             float invRms = inversesqrt(meanSquare);
             float dx = dyGamma * invRms - x * dyXMean * invRms / meanSquare;
             setOutput(dx);
         }
         `;
-  }
-}
-class k {
-  variableNames = ["x", "meanSquare", "dy"];
-  outputShape;
-  userCode;
-  constructor(a, e, o) {
-    this.outputShape = [a, e, o], this.userCode = `
-        void main() {
-            ivec3 coords = getOutputCoords();
-            float x = getXAtOutCoords();
-            float meanSquare = getMeanSquare(coords.x, coords.y, 0) + 1e-8;
-            float dy = getDyAtOutCoords();
-            float invRms = inversesqrt(meanSquare);
-            float dGamma = dy * (x * invRms);
-            setOutput(dGamma);
-        }
-        `;
-  }
-}
-function M(r) {
-  const { dy: a, x: e, gamma: o } = r.inputs, t = r.backend, m = e.shape[0], u = e.shape[1], s = e.shape[2], n = o ? a.mul(o) : a, i = n.mul(e), l = i.sum(-1, !0);
-  i.dispose();
-  const g = e.square(), d = g.mean(-1, !0);
-  g.dispose();
-  const h = new C(m, u, s), p = t.runWebGLProgram(h, [e, d, n, l], "float32");
-  if (o && n.dispose(), l.dispose(), o) {
-    const q = new k(m, u, s), f = t.runWebGLProgram(q, [e, d, a], "float32");
-    d.dispose();
-    const v = x(G().makeTensorFromTensorInfo(f), [0, 1]);
-    return t.disposeIntermediateTensorInfo(f), [p, v];
-  } else
-    return d.dispose(), [p];
-}
-const R = {
-  kernelName: "RMSNormGrad",
-  backendName: "webgl",
-  kernelFunc: M
+	}
+}, c = class {
+	variableNames = [
+		"x",
+		"meanSquare",
+		"dy"
+	];
+	outputShape;
+	userCode;
+	constructor(e, t, n) {
+		this.outputShape = [
+			e,
+			t,
+			n
+		], this.userCode = "\n        void main() {\n            ivec3 coords = getOutputCoords();\n            float x = getXAtOutCoords();\n            float meanSquare = getMeanSquare(coords.x, coords.y, 0) + 1e-8;\n            float dy = getDyAtOutCoords();\n            float invRms = inversesqrt(meanSquare);\n            float dGamma = dy * (x * invRms);\n            setOutput(dGamma);\n        }\n        ";
+	}
 };
-c(R);
+function l(e) {
+	let { dy: r, x: i, gamma: a } = e.inputs, o = e.backend, l = i.shape[0], u = i.shape[1], d = i.shape[2], f = a ? r.mul(a) : r, p = f.mul(i), m = p.sum(-1, !0);
+	p.dispose();
+	let h = i.square(), g = h.mean(-1, !0);
+	h.dispose();
+	let _ = new s(l, u, d), v = o.runWebGLProgram(_, [
+		i,
+		g,
+		f,
+		m
+	], "float32");
+	if (a && f.dispose(), m.dispose(), a) {
+		let e = new c(l, u, d), a = o.runWebGLProgram(e, [
+			i,
+			g,
+			r
+		], "float32");
+		g.dispose();
+		let s = n(t().makeTensorFromTensorInfo(a), [0, 1]);
+		return o.disposeIntermediateTensorInfo(a), [v, s];
+	} else return g.dispose(), [v];
+}
+e({
+	kernelName: "RMSNormGrad",
+	backendName: "webgl",
+	kernelFunc: l
+});
+//#endregion

package/dist/ops/webgl/qkv.js

@@ -1,13 +1,21 @@
-import { c as i } from "../../index-CUXkjxiT.js";
-class l {
-  variableNames = ["x", "kernel"];
-  outputShape;
-  userCode;
-  // enableShapeUniforms = true;
-  customUniforms = [{ name: "mode", type: "int" }];
-  constructor(e, t, s, o) {
-    const n = o / t;
-    this.outputShape = [e, t, s, n], this.userCode = `
+import { Ii as e } from "../../dist-BewPQWjc.js";
+//#region lib/ops/webgl/qkv.ts
+var t = class {
+	variableNames = ["x", "kernel"];
+	outputShape;
+	userCode;
+	customUniforms = [{
+		name: "mode",
+		type: "int"
+	}];
+	constructor(e, t, n, r) {
+		let i = r / t;
+		this.outputShape = [
+			e,
+			t,
+			n,
+			i
+		], this.userCode = `
         void main() {
             ivec4 coords = getOutputCoords(); // [b, h, t, d]
             int b = coords.x;
@@ -16,10 +24,10 @@ class l {
             int d = coords.w;
 
             // Compute output channel index in fused kernel
-            int out_offset = mode * ${t} * ${n} + h * ${n} + d;
+            int out_offset = mode * ${t} * ${i} + h * ${i} + d;
 
             float sum = 0.0;
-            for (int c = 0; c < ${o}; ++c) {
+            for (int c = 0; c < ${r}; ++c) {
                 float xval = getX(b, t, c); // fetch from x
                 float kval = getKernel(c, out_offset); // fetch from kernel
                 sum += xval * kval;
@@ -28,19 +36,19 @@ class l {
             setOutput(sum);
         }
         `;
-  }
-}
-function m(r) {
-  const { x: e, kernel: t } = r.inputs, { heads: s } = r.attrs, o = r.backend, n = e.shape[0], c = e.shape[1], u = e.shape[2], a = new l(n, s, c, u);
-  return [
-    o.runWebGLProgram(a, [e, t], "float32", [[0]]),
-    o.runWebGLProgram(a, [e, t], "float32", [[1]]),
-    o.runWebGLProgram(a, [e, t], "float32", [[2]])
-  ];
-}
-const d = {
-  kernelName: "QKV",
-  backendName: "webgl",
-  kernelFunc: m
+	}
 };
-i(d);
+function n(e) {
+	let { x: n, kernel: r } = e.inputs, { heads: i } = e.attrs, a = e.backend, o = n.shape[0], s = n.shape[1], c = n.shape[2], l = new t(o, i, s, c);
+	return [
+		a.runWebGLProgram(l, [n, r], "float32", [[0]]),
+		a.runWebGLProgram(l, [n, r], "float32", [[1]]),
+		a.runWebGLProgram(l, [n, r], "float32", [[2]])
+	];
+}
+e({
+	kernelName: "QKV",
+	backendName: "webgl",
+	kernelFunc: n
+});
+//#endregion

package/dist/ops/webgl/rope.js

@@ -1,12 +1,24 @@
-import { c as h } from "../../index-CUXkjxiT.js";
-class g {
-  variableNames = ["x", "sin", "cos"];
-  outputShape;
-  userCode;
-  // enableShapeUniforms = true;
-  customUniforms = [{ name: "pastLen", type: "int" }];
-  constructor(t, s, e, n) {
-    this.outputShape = [t, s, e, n], this.userCode = `
+import { Ii as e } from "../../dist-BewPQWjc.js";
+//#region lib/ops/webgl/rope.ts
+var t = class {
+	variableNames = [
+		"x",
+		"sin",
+		"cos"
+	];
+	outputShape;
+	userCode;
+	customUniforms = [{
+		name: "pastLen",
+		type: "int"
+	}];
+	constructor(e, t, n, r) {
+		this.outputShape = [
+			e,
+			t,
+			n,
+			r
+		], this.userCode = `
         void main() {
             ivec4 coords = getOutputCoords(); // [b, h, t, d]
             int b = coords.x;
@@ -14,7 +26,7 @@ class g {
             int t = coords.z;
             int d = coords.w;
 
-            int rotaryDim = ${n};
+            int rotaryDim = ${r};
 
             float outVal = 0.0;
 
@@ -42,15 +54,19 @@ class g {
             setOutput(outVal);
         }
         `;
-  }
-}
-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 m = {
-  kernelName: "Rope",
-  backendName: "webgl",
-  kernelFunc: f
+	}
 };
-h(m);
+function n(e) {
+	let { x: n } = e.inputs, { pastLen: r, ropeCache: i, negSin: a } = e.attrs, o = a ? i.getNegSin() : i.getSin(), s = i.getCos(), c = e.backend, l = n.shape[0], u = n.shape[1], d = n.shape[2], f = n.shape[3], p = new t(l, u, d, f);
+	return c.runWebGLProgram(p, [
+		n,
+		o,
+		s
+	], "float32", [[r]]);
+}
+e({
+	kernelName: "Rope",
+	backendName: "webgl",
+	kernelFunc: n
+});
+//#endregion

package/dist/ops/webgl/scatterSub.js

@@ -1,27 +1,28 @@
-import { c as i } from "../../index-CUXkjxiT.js";
-class u {
-  variableNames = ["labels", "softmaxProbs", "dy"];
-  outputShape;
-  userCode;
-  constructor(e, t) {
-    this.outputShape = [e, t], this.userCode = `
-            void main() {
-                ivec2 coords = getOutputCoords();
-                int index = int(getLabels(coords.x));
-                float prob = getSoftmaxProbsAtOutCoords();
-                float dy = getDy(coords.x);
-                setOutput((index == coords.y ? prob - 1.0 : prob) * dy);
-            }
-            `;
-  }
-}
-function d(o) {
-  const { logits: e, labels: t, dy: r } = o.inputs, s = o.backend, n = t.shape[0], a = e.shape[1], c = new u(n, a);
-  return s.runWebGLProgram(c, [t, e, r], "float32");
-}
-const b = {
-  kernelName: "EfficientScatterSub",
-  backendName: "webgl",
-  kernelFunc: d
+import { Ii as e } from "../../dist-BewPQWjc.js";
+//#region lib/ops/webgl/scatterSub.ts
+var t = class {
+	variableNames = [
+		"labels",
+		"softmaxProbs",
+		"dy"
+	];
+	outputShape;
+	userCode;
+	constructor(e, t) {
+		this.outputShape = [e, t], this.userCode = "\n            void main() {\n                ivec2 coords = getOutputCoords();\n                int index = int(getLabels(coords.x));\n                float prob = getSoftmaxProbsAtOutCoords();\n                float dy = getDy(coords.x);\n                setOutput((index == coords.y ? prob - 1.0 : prob) * dy);\n            }\n            ";
+	}
 };
-i(b);
+function n(e) {
+	let { logits: n, labels: r, dy: i } = e.inputs, a = e.backend, o = r.shape[0], s = n.shape[1], c = new t(o, s);
+	return a.runWebGLProgram(c, [
+		r,
+		n,
+		i
+	], "float32");
+}
+e({
+	kernelName: "EfficientScatterSub",
+	backendName: "webgl",
+	kernelFunc: n
+});
+//#endregion

package/dist/ops/webgpu/adamAdjust.js

@@ -1,24 +1,29 @@
-import { e as d } from "../../webgpu_program-B4HmApL1.js";
-import { f as l, c as f } from "../../webgpu_util-DYlGSwOJ.js";
-import { c as h, a6 as c } from "../../index-CUXkjxiT.js";
-class v {
-  variableNames = ["moments", "value"];
-  outputShape;
-  shaderKey = "AdamAdjust";
-  dispatchLayout;
-  dispatch;
-  workgroupSize = [64, 1, 1];
-  size = !0;
-  uniforms = "invbeta1: f32, invbeta2: f32, learningRate: f32, epsilon: f32";
-  outputComponent = 1;
-  variableComponents = [2, 1];
-  useWeightDecay;
-  constructor(a, t) {
-    this.outputShape = a, this.dispatchLayout = l(this.outputShape), this.dispatch = f(this.dispatchLayout, this.outputShape, this.workgroupSize), this.useWeightDecay = t, t && (this.uniforms += ", weightDecay: f32");
-  }
-  getUserCode() {
-    return `
-        ${d("index")} {
+import { Fs as e, Ii as t } from "../../dist-BewPQWjc.js";
+import { s as n } from "../../webgpu_program-WOyIVMlZ.js";
+import { c as r, i } from "../../webgpu_util-B_F3SShA.js";
+//#region lib/ops/webgpu/adamAdjust.ts
+var a = class {
+	variableNames = ["moments", "value"];
+	outputShape;
+	shaderKey = "AdamAdjust";
+	dispatchLayout;
+	dispatch;
+	workgroupSize = [
+		64,
+		1,
+		1
+	];
+	size = !0;
+	uniforms = "invbeta1: f32, invbeta2: f32, learningRate: f32, epsilon: f32";
+	outputComponent = 1;
+	variableComponents = [2, 1];
+	useWeightDecay;
+	constructor(e, t) {
+		this.outputShape = e, this.dispatchLayout = r(this.outputShape), this.dispatch = i(this.dispatchLayout, this.outputShape, this.workgroupSize), this.useWeightDecay = t, t && (this.uniforms += ", weightDecay: f32");
+	}
+	getUserCode() {
+		return `
+        ${n("index")} {
             if (index < uniforms.size) {
                 let moments: vec2<f32> = moments[index];
                 let value: f32 = value[index];
@@ -36,22 +41,37 @@ class v {
             }
         }
     `;
-  }
-}
-function g(e) {
-  const { moments: a, value: t } = e.inputs, { beta1: i, beta2: o, learningRate: u, epsilon: r, weightDecay: s } = e.attrs, m = e.backend;
-  c(a.shape, [...t.shape, 2], "Error in AdamAdjust: ");
-  const p = new v(t.shape, s > 0), n = [
-    { type: "float32", data: [1 / i] },
-    { type: "float32", data: [1 / o] },
-    { type: "float32", data: [u] },
-    { type: "float32", data: [r] }
-  ];
-  return s > 0 && n.push({ type: "float32", data: [s] }), m.runWebGPUProgram(p, [a, t], "float32", n);
-}
-const y = {
-  kernelName: "AdamAdjust",
-  backendName: "webgpu",
-  kernelFunc: g
+	}
 };
-h(y);
+function o(t) {
+	let { moments: n, value: r } = t.inputs, { beta1: i, beta2: o, learningRate: s, epsilon: c, weightDecay: l } = t.attrs, u = t.backend;
+	e(n.shape, [...r.shape, 2], "Error in AdamAdjust: ");
+	let d = new a(r.shape, l > 0), f = [
+		{
+			type: "float32",
+			data: [1 / i]
+		},
+		{
+			type: "float32",
+			data: [1 / o]
+		},
+		{
+			type: "float32",
+			data: [s]
+		},
+		{
+			type: "float32",
+			data: [c]
+		}
+	];
+	return l > 0 && f.push({
+		type: "float32",
+		data: [l]
+	}), u.runWebGPUProgram(d, [n, r], "float32", f);
+}
+t({
+	kernelName: "AdamAdjust",
+	backendName: "webgpu",
+	kernelFunc: o
+});
+//#endregion