@genai-fi/nanogpt 0.20.0 → 0.20.1

package/dist/ops/grads/add16.js

@@ -0,0 +1,27 @@
+import { Fi as e, In as t, Rn as n } from "../../dist-BewPQWjc.js";
+import { reshape16 as r } from "../reshape16.js";
+import { sum16 as i } from "../sum16.js";
+//#region lib/ops/grads/add16.ts
+e({
+	kernelName: "Add16",
+	inputsToSave: ["a", "b"],
+	gradFunc: (e, a) => {
+		let [o, s] = a, c = t(o.shape, s.shape);
+		if (Array.isArray(e)) throw Error("Add16 gradFunc expected dy to be a Tensor but got an array");
+		return {
+			a: () => {
+				let t = e, a = n(o.shape, c);
+				a.length > 0 && (t = i(t, a));
+				let s = r(t, o.shape);
+				return t.dispose(), s;
+			},
+			b: () => {
+				let t = e, a = n(s.shape, c);
+				a.length > 0 && (t = i(t, a));
+				let o = r(t, s.shape);
+				return t.dispose(), o;
+			}
+		};
+	}
+});
+//#endregion

package/dist/ops/grads/attentionMask.d.ts

@@ -0,0 +1 @@
+export {};

package/dist/ops/grads/attentionMask.js

@@ -0,0 +1,26 @@
+import { Fi as e } from "../../dist-BewPQWjc.js";
+import { transpose16 as t } from "../transpose16.js";
+import { a as n } from "../../matMul16-BNfZSnNM.js";
+//#region lib/ops/grads/attentionMask.ts
+e({
+	kernelName: "AttentionMask",
+	inputsToSave: ["q", "k"],
+	outputsToSave: [],
+	gradFunc: (e, r, i) => {
+		if (Array.isArray(e)) throw Error("Expected dy to be a single Tensor");
+		let [a, o] = r, { divisor: s } = i;
+		return {
+			q: () => n(e, o, s),
+			k: () => {
+				let r = n(a, e, s, !0, !1), i = t(r, [
+					0,
+					1,
+					3,
+					2
+				]);
+				return r.dispose(), i;
+			}
+		};
+	}
+});
+//#endregion

package/dist/ops/grads/dropout16.d.ts

@@ -0,0 +1 @@
+export {};

package/dist/ops/grads/dropout16.js

@@ -0,0 +1 @@
+import "../dropout16.js";

package/dist/ops/grads/gelu.d.ts

@@ -0,0 +1,2 @@
+import { GradConfig } from '@tensorflow/tfjs-core';
+export declare const geluGradConfig: GradConfig;

package/dist/ops/grads/gelu.js

@@ -0,0 +1,2 @@
+import { r as e } from "../../gelu-Bf1HW1RY.js";
+export { e as geluGradConfig };

package/dist/ops/grads/matMul16.d.ts

@@ -0,0 +1,2 @@
+import { GradConfig } from '@tensorflow/tfjs-core';
+export declare const matMul16GradConfig: GradConfig;

package/dist/ops/grads/matMul16.js

@@ -0,0 +1,2 @@
+import { o as e } from "../../matMul16-BNfZSnNM.js";
+export { e as matMul16GradConfig };

package/dist/ops/grads/matMulGelu.d.ts

@@ -0,0 +1 @@
+export {};

package/dist/ops/grads/matMulGelu.js

@@ -0,0 +1,22 @@
+import { Fi as e, ii as t } from "../../dist-BewPQWjc.js";
+//#region lib/ops/grads/matMulGelu.ts
+function n(e, n, r) {
+	return t().runKernel("MatMulGeluGrad", {
+		dy: e,
+		x: n,
+		kernel: r
+	});
+}
+e({
+	kernelName: "MatMulGelu",
+	inputsToSave: ["x", "kernel"],
+	outputsToSave: [],
+	gradFunc: (e, t) => {
+		let [r, i] = t, [a, o] = n(e, r, i);
+		return {
+			x: () => a,
+			kernel: () => o
+		};
+	}
+});
+//#endregion

package/dist/ops/grads/mul16.d.ts

@@ -0,0 +1 @@
+export {};

package/dist/ops/grads/mul16.js

@@ -0,0 +1 @@
+import "../mul16.js";

package/dist/ops/grads/normRMS.d.ts

@@ -0,0 +1,3 @@
+import { GradConfig } from '@tensorflow/tfjs-core';
+export declare const normRMSGradConfig: GradConfig;
+export declare const normRMSGradConfigNoGamma: GradConfig;

package/dist/ops/grads/normRMS.js

@@ -0,0 +1,37 @@
+import { Fi as e, ii as t } from "../../dist-BewPQWjc.js";
+//#region lib/ops/grads/normRMS.ts
+function n(e, n, r) {
+	return t().runKernel("RMSNormGrad", r ? {
+		dy: e,
+		x: n,
+		gamma: r
+	} : {
+		dy: e,
+		x: n
+	});
+}
+var r = {
+	kernelName: "RMSNorm",
+	inputsToSave: ["x", "gamma"],
+	outputsToSave: [],
+	gradFunc: (e, t) => {
+		let [r, i] = t, [a, o] = n(e, r, i);
+		return {
+			x: () => a,
+			gamma: () => o
+		};
+	}
+};
+e(r);
+var i = {
+	kernelName: "RMSNormNoGamma",
+	inputsToSave: ["x"],
+	outputsToSave: [],
+	gradFunc: (e, t) => {
+		let [r] = t, [i] = n(e, r);
+		return { x: () => i };
+	}
+};
+e(i);
+//#endregion
+export { r as normRMSGradConfig, i as normRMSGradConfigNoGamma };

package/dist/ops/grads/pack16.d.ts

@@ -0,0 +1,2 @@
+import { GradConfig } from '@tensorflow/tfjs-core';
+export declare const packGradConfig: GradConfig;

package/dist/ops/grads/pack16.js

@@ -0,0 +1,2 @@
+import { n as e } from "../../pack16-Ck-spx_F.js";
+export { e as packGradConfig };

package/dist/ops/grads/qkv.d.ts

@@ -0,0 +1,3 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+import { NamedGradientMap } from '@tensorflow/tfjs-core/dist/tape';
+export declare function qkvGrad(dy: Tensor[], x: Tensor, kernel: Tensor): NamedGradientMap;

package/dist/ops/grads/qkv.js

@@ -0,0 +1,46 @@
+import { Fi as e, H as t } from "../../dist-BewPQWjc.js";
+import { sum16 as n } from "../sum16.js";
+import { o as r } from "../../matMul16-BNfZSnNM.js";
+import { concat16 as i } from "../concat16.js";
+//#region lib/ops/grads/qkv.ts
+var a = {
+	kernelName: "QKV",
+	inputsToSave: ["x", "kernel"],
+	outputsToSave: [],
+	gradFunc: (e, a) => {
+		let [o, s, c] = e, [l] = a, u = i([
+			o,
+			s,
+			c
+		], 1);
+		o.dispose(), s.dispose(), c.dispose();
+		let d = [
+			l.shape[0],
+			l.shape[1],
+			3 * l.shape[2]
+		], f = r.gradFunc(u, a, {
+			transposeA: !1,
+			transposeB: !1,
+			originalShape: d,
+			perm: [
+				0,
+				2,
+				1,
+				3
+			]
+		});
+		return u.dispose(), {
+			x: () => f.A(),
+			kernel: () => {
+				let e = f.B(), r = e.shape[0] === 1 ? t(e, [0]) : n(e, 0);
+				return e.dispose(), r;
+			}
+		};
+	}
+};
+function o(e, t, n) {
+	return a.gradFunc(e, [t, n], {});
+}
+e(a);
+//#endregion
+export { o as qkvGrad };

package/dist/ops/grads/rope.d.ts

@@ -0,0 +1,2 @@
+import { GradConfig } from '@tensorflow/tfjs-core';
+export declare const ropeGradConfig: GradConfig;

package/dist/ops/grads/rope.js

@@ -0,0 +1,2 @@
+import { n as e } from "../../rope-CbeGlsV8.js";
+export { e as ropeGradConfig };

package/dist/ops/grads/softmax16.d.ts

@@ -0,0 +1,2 @@
+import { GradConfig } from '@tensorflow/tfjs-core';
+export declare const softmax16GradConfig: GradConfig;

package/dist/ops/grads/softmax16.js

@@ -0,0 +1,23 @@
+import { Fi as e, ii as t } from "../../dist-BewPQWjc.js";
+import { isPackedTensor as n } from "../../utilities/packed.js";
+//#region lib/ops/grads/softmax16.ts
+function r(e, n) {
+	return t().runKernel("Softmax16Grad", {
+		dy: e,
+		softmaxOutput: n
+	});
+}
+var i = {
+	kernelName: "Softmax16",
+	outputsToSave: [!0],
+	gradFunc: (e, t) => {
+		let [i] = t;
+		if (Array.isArray(e)) throw Error("Expected dy to be a single Tensor");
+		if (!n(i)) throw console.error(i), Error("Softmax16 gradient requires packed y Tensor");
+		if (!n(e)) throw Error("Softmax16 gradient requires packed dy Tensor");
+		return { logits: () => r(e, i) };
+	}
+};
+e(i);
+//#endregion
+export { i as softmax16GradConfig };

package/dist/ops/grads/unpack16.d.ts

@@ -0,0 +1,2 @@
+import { GradConfig } from '@tensorflow/tfjs-core';
+export declare const unpackGradConfig: GradConfig;

package/dist/ops/grads/unpack16.js

@@ -0,0 +1,2 @@
+import { i as e } from "../../pack16-Ck-spx_F.js";
+export { e as unpackGradConfig };

package/dist/ops/grads/utils.d.ts

@@ -0,0 +1,4 @@
+import { TensorInfo } from '@tensorflow/tfjs-core';
+export declare function forceFloat<T extends TensorInfo>(x: T): T;
+export declare function forceInt<T extends TensorInfo>(x: T): T;
+export declare function forcePacked<T extends TensorInfo>(x: T): T;

package/dist/ops/grads/utils.js

@@ -0,0 +1,12 @@
+//#region lib/ops/grads/utils.ts
+function e(e) {
+	return e.dtype = "float32", e;
+}
+function t(e) {
+	return e.dtype = "int32", e;
+}
+function n(e) {
+	return e.dtype = "packedF16", e;
+}
+//#endregion
+export { e as forceFloat, t as forceInt, n as forcePacked };

package/dist/ops/log.js

@@ -0,0 +1 @@
+import "./webgl/log.js";

package/dist/ops/matMul16.d.ts

@@ -0,0 +1,15 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function matMul16(A: Tensor, B: Tensor, transposeA?: boolean, transposeB?: boolean, attrs?: {
+    scale?: number;
+    scaleA?: number;
+    scaleB?: number;
+    activation?: 'gelu' | 'relu2' | 'relu';
+    forceOutputShape?: number[];
+    perm?: number[];
+    causalMask?: boolean;
+    pastLen?: number;
+}): Tensor;
+export declare function matMul16Scaled(A: Tensor, B: Tensor, scale: number, transposeA?: boolean, transposeB?: boolean): Tensor;
+export declare function matMul16ScaleA(A: Tensor, B: Tensor, scale: number, transposeA?: boolean, transposeB?: boolean): Tensor;
+export declare function matMul16ScaleB(A: Tensor, B: Tensor, scale: number, transposeA?: boolean, transposeB?: boolean): Tensor;
+export declare function matMul16Gelu(A: Tensor, B: Tensor, transposeA?: boolean, transposeB?: boolean): Tensor;

package/dist/ops/matMul16.js

@@ -0,0 +1,2 @@
+import { a as e, i as t, n, r, t as i } from "../matMul16-BNfZSnNM.js";
+export { i as matMul16, n as matMul16Gelu, r as matMul16ScaleA, t as matMul16ScaleB, e as matMul16Scaled };

package/dist/ops/matMulGelu.d.ts

@@ -0,0 +1,3 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function matMulGelu(x: Tensor, kernel: Tensor): Tensor;
+export declare function dMatMulGelu(dy: Tensor, x: Tensor, kernel: Tensor): Tensor;

package/dist/ops/matMulGelu.js

@@ -0,0 +1,20 @@
+import { ii as e } from "../dist-BewPQWjc.js";
+import "../matMulGelu-CPTntosE.js";
+import "./cpu/matMulGelu.js";
+import "./grads/matMulGelu.js";
+//#region lib/ops/matMulGelu.ts
+function t(t, n) {
+	return e().runKernel("MatMulGelu", {
+		x: t,
+		kernel: n
+	});
+}
+function n(t, n, r) {
+	return e().runKernel("MatMulGeluGrad", {
+		dy: t,
+		x: n,
+		kernel: r
+	});
+}
+//#endregion
+export { n as dMatMulGelu, t as matMulGelu };

package/dist/ops/matMulMul.d.ts

@@ -0,0 +1,2 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function matMulMul(x: Tensor, kernel: Tensor, y: Tensor, transposeA?: boolean, transposeB?: boolean): Tensor;

package/dist/ops/matMulMul.js

@@ -0,0 +1,16 @@
+import { ii as e } from "../dist-BewPQWjc.js";
+import "./cpu/matMulMul.js";
+import "./webgl/matMulMul.js";
+//#region lib/ops/matMulMul.ts
+function t(t, n, r, i = !1, a = !1) {
+	return e().runKernel("MatMulMul", {
+		x: t,
+		kernel: n,
+		y: r
+	}, {
+		transposeA: i,
+		transposeB: a
+	});
+}
+//#endregion
+export { t as matMulMul };

package/dist/ops/mul16.d.ts

@@ -0,0 +1,2 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function mul16(a: Tensor, b: Tensor): Tensor;

package/dist/ops/mul16.js

@@ -0,0 +1,43 @@
+import { Fi as e, In as t, Rn as n, Wr as r, ii as i } from "../dist-BewPQWjc.js";
+import { isPackedTensor as a } from "../utilities/packed.js";
+import { reshape16 as o } from "./reshape16.js";
+import { sum16 as s } from "./sum16.js";
+//#region lib/ops/grads/mul16.ts
+e({
+	kernelName: "Mul16",
+	inputsToSave: ["a", "b"],
+	gradFunc: (e, r) => {
+		let [i, a] = r, l = t(i.shape, a.shape);
+		if (Array.isArray(e)) throw Error("Mul16 gradFunc expected dy to be a Tensor but got an array");
+		return {
+			a: () => {
+				let t = c(e, a), r = n(i.shape, l);
+				if (r.length > 0) {
+					let e = s(t, r);
+					t.dispose(), t = e;
+				}
+				let u = o(t, i.shape);
+				return t.dispose(), u;
+			},
+			b: () => {
+				let t = c(e, i), r = n(a.shape, l);
+				if (r.length > 0) {
+					let e = s(t, r);
+					t.dispose(), t = e;
+				}
+				let u = o(t, a.shape);
+				return t.dispose(), u;
+			}
+		};
+	}
+});
+//#endregion
+//#region lib/ops/mul16.ts
+function c(e, t) {
+	return !a(e) && !a(t) ? r(e, t) : i().runKernel("Mul16", {
+		a: e,
+		b: t
+	});
+}
+//#endregion
+export { c as mul16 };

package/dist/ops/mulDrop.d.ts

@@ -0,0 +1,2 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function mulDrop(a: Tensor, b: Tensor, dropoutRate: number, seed: number): Tensor;

package/dist/ops/mulDrop.js

@@ -0,0 +1,15 @@
+import { ii as e } from "../dist-BewPQWjc.js";
+import "./cpu/mulDropout.js";
+import "./webgl/mulDropout.js";
+//#region lib/ops/mulDrop.ts
+function t(t, n, r, i) {
+	return e().runKernel("MulDropout", {
+		a: t,
+		b: n
+	}, {
+		dropoutRate: r,
+		seed: i
+	});
+}
+//#endregion
+export { t as mulDrop };

package/dist/ops/normRMS.d.ts

@@ -0,0 +1,2 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function normRMS(x: Tensor, gamma?: Tensor): Tensor;

package/dist/ops/normRMS.js

@@ -0,0 +1,22 @@
+import { ii as e } from "../dist-BewPQWjc.js";
+import "./cpu/normRMS.js";
+import "./webgl/normRMS.js";
+import "./grads/normRMS.js";
+//#region lib/ops/normRMS.ts
+function t(t, n) {
+	if (t.rank === 4) {
+		let [r, i, ...a] = t.shape, o = [r * i, ...a], s = t.reshape(o), c = e().runKernel(n ? "RMSNorm" : "RMSNormNoGamma", n ? {
+			x: s,
+			gamma: n
+		} : { x: s });
+		s.dispose();
+		let l = c.reshape(t.shape);
+		return c.dispose(), l;
+	}
+	return e().runKernel(n ? "RMSNorm" : "RMSNormNoGamma", n ? {
+		x: t,
+		gamma: n
+	} : { x: t });
+}
+//#endregion
+export { t as normRMS };

package/dist/ops/pack16.d.ts

@@ -0,0 +1,2 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function pack16(x: Tensor, scaling?: number, padding?: number): Tensor;

package/dist/ops/pack16.js

@@ -0,0 +1,2 @@
+import { t as e } from "../pack16-Ck-spx_F.js";
+export { e as pack16 };

package/dist/ops/qkv.d.ts

@@ -0,0 +1,2 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function qkv(x: Tensor, kernel: Tensor, heads: number, packed?: boolean): Tensor[];

package/dist/ops/qkv.js

@@ -0,0 +1,16 @@
+import { ii as e } from "../dist-BewPQWjc.js";
+import "./cpu/qkv.js";
+import "./webgl/qkv.js";
+import "./grads/qkv.js";
+//#region lib/ops/qkv.ts
+function t(t, n, r, i = !1) {
+	return e().runKernel("QKV", {
+		x: t,
+		kernel: n
+	}, {
+		heads: r,
+		packed: i
+	});
+}
+//#endregion
+export { t as qkv };

package/dist/ops/reshape16.d.ts

@@ -0,0 +1,2 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function reshape16(x: Tensor, shape: number[]): Tensor;

package/dist/ops/reshape16.js

@@ -0,0 +1,33 @@
+import { Fi as e, Ii as t, br as n, ii as r } from "../dist-BewPQWjc.js";
+//#region lib/ops/reshape16.ts
+e({
+	kernelName: "Reshape16",
+	inputsToSave: ["x"],
+	gradFunc: (e, t) => {
+		let [n] = t;
+		if (Array.isArray(e)) throw Error("Reshape16 gradient does not support multiple outputs.");
+		return { x: () => a(e, n.shape) };
+	}
+});
+function i(e) {
+	let { inputs: t, attrs: r } = e, { x: i } = t, { shape: a } = r;
+	return n(i, a);
+}
+t({
+	kernelName: "Reshape16",
+	backendName: "webgpu",
+	kernelFunc: i
+}), t({
+	kernelName: "Reshape16",
+	backendName: "webgl",
+	kernelFunc: i
+}), t({
+	kernelName: "Reshape16",
+	backendName: "cpu",
+	kernelFunc: i
+});
+function a(e, t) {
+	return r().runKernel("Reshape16", { x: e }, { shape: t });
+}
+//#endregion
+export { a as reshape16 };

package/dist/ops/rope.d.ts

@@ -0,0 +1,3 @@
+import { default as RoPECache } from '../../layers/RoPECache';
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function rope(x: Tensor, cache: RoPECache, pastLength: number, negSin?: boolean): Tensor;

package/dist/ops/rope.js

@@ -0,0 +1,2 @@
+import { t as e } from "../rope-CbeGlsV8.js";
+export { e as rope };

package/dist/ops/scatterSub.d.ts

@@ -0,0 +1,2 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function scatterSub(probabilities: Tensor, labels: Tensor, scale: Tensor): Tensor;

package/dist/ops/scatterSub.js

@@ -0,0 +1,13 @@
+import { ii as e } from "../dist-BewPQWjc.js";
+import "./cpu/scatterSub.js";
+import "./webgl/scatterSub.js";
+//#region lib/ops/scatterSub.ts
+function t(t, n, r) {
+	return e().runKernel("EfficientScatterSub", {
+		logits: t,
+		labels: n,
+		dy: r
+	}, {});
+}
+//#endregion
+export { t as scatterSub };

package/dist/ops/slice16.d.ts

@@ -0,0 +1,2 @@
+import { Rank, Tensor } from '@tensorflow/tfjs-core';
+export declare function slice16<R extends Rank = Rank>(x: Tensor<R>, begin: number | number[], size: number | number[]): Tensor<R>;

package/dist/ops/slice16.js

@@ -0,0 +1,11 @@
+import { ii as e, mr as t } from "../dist-BewPQWjc.js";
+import { isPackedTensor as n } from "../utilities/packed.js";
+//#region lib/ops/slice16.ts
+function r(r, i, a) {
+	return n(r) ? e().runKernel("Slice16", { x: r }, {
+		begin: i,
+		size: a
+	}) : t(r, i, a);
+}
+//#endregion
+export { r as slice16 };

package/dist/ops/softmax16.d.ts

@@ -0,0 +1,2 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function softmax16(logits: Tensor): Tensor;

package/dist/ops/softmax16.js

@@ -0,0 +1,9 @@
+import { ii as e } from "../dist-BewPQWjc.js";
+import { isPackedTensor as t } from "../utilities/packed.js";
+import "./grads/softmax16.js";
+//#region lib/ops/softmax16.ts
+function n(n) {
+	return t(n) ? e().runKernel("Softmax16", { logits: n }, { dim: n.rank - 1 }) : e().runKernel("Softmax", { logits: n }, { dim: n.rank - 1 });
+}
+//#endregion
+export { n as softmax16 };

package/dist/ops/sub16.d.ts

@@ -0,0 +1,2 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function sub16(a: Tensor, b: Tensor): Tensor;

package/dist/ops/sub16.js

@@ -0,0 +1,11 @@
+import { Bt as e, ii as t } from "../dist-BewPQWjc.js";
+import { isPackedTensor as n } from "../utilities/packed.js";
+//#region lib/ops/sub16.ts
+function r(r, i) {
+	return !n(r) && !n(i) ? e(r, i) : t().runKernel("Sub16", {
+		a: r,
+		b: i
+	});
+}
+//#endregion
+export { r as sub16 };

package/dist/ops/sum16.d.ts

@@ -0,0 +1,2 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function sum16(x: Tensor, axis?: number | number[], keepDims?: boolean): Tensor;

package/dist/ops/sum16.js

@@ -0,0 +1,13 @@
+import { ii as e, mn as t } from "../dist-BewPQWjc.js";
+import { isPackedTensor as n } from "../utilities/packed.js";
+//#region lib/ops/sum16.ts
+function r(r, i, a = !1) {
+	if (!n(r)) return t(r, i, a);
+	if (a) throw Error("sum16 with keepDims=true not supported for packed tensors");
+	return e().runKernel("Sum16", { x: r }, {
+		axis: i ?? -1,
+		keepDims: a
+	});
+}
+//#endregion
+export { r as sum16 };

package/dist/ops/transpose16.d.ts

@@ -0,0 +1,3 @@
+import { GradConfig, Tensor } from '@tensorflow/tfjs-core';
+export declare const transpose16GradConfig: GradConfig;
+export declare function transpose16(x: Tensor, perm?: number[]): Tensor;