@genai-fi/nanogpt 0.23.0 → 1.0.0

package/dist/{kernel_funcs_utils-D-mATnGR.js → kernel_funcs_utils-ql8Y8qPn.js}

@@ -1,8 +1,25 @@
 import { Ci as e, Di as t, In as n, Ms as r, Ya as i, ca as a, ko as o, oc as s, to as c } from "./dist-Da20xy8E.js";
 import { r as l } from "./backend_util-DAV0I34P.js";
-import { a as u, o as d } from "./gpgpu_math-DBYEAAdI.js";
+import { Z as u, a as d } from "./gpgpu_math-20tPK8LM.js";
+//#region node_modules/@tensorflow/tfjs-backend-webgl/dist/binaryop_gpu.js
+var f = "\n  if (isnan(a)) return a;\n  if (isnan(b)) return b;\n", p = class {
+	constructor(e, t, r) {
+		this.variableNames = ["A", "B"], this.outputShape = n(t, r), this.enableShapeUniforms = d(this.outputShape.length), this.userCode = `
+      float binaryOperation(float a, float b) {
+        ${e}
+      }
+
+      void main() {
+        float a = getAAtOutCoords();
+        float b = getBAtOutCoords();
+        setOutput(binaryOperation(a, b));
+      }
+    `;
+	}
+};
+//#endregion
 //#region node_modules/@tensorflow/tfjs-backend-webgl/dist/packing_util.js
-function f(e, t) {
+function m(e, t) {
 	return [
 		"x",
 		"y",
@@ -12,77 +29,33 @@ function f(e, t) {
 		"v"
 	].slice(0, t).map((t) => `${e}.${t}`);
 }
-function p(e, t) {
-	return t === 1 ? [e] : f(e, t);
+function h(e, t) {
+	return t === 1 ? [e] : m(e, t);
 }
-function m(e, t) {
+function g(e, t) {
 	if (e === 1) return "rc";
 	let n = "";
 	for (let r = 0; r < e; r++) n += t[r], r < e - 1 && (n += ",");
 	return n;
 }
 //#endregion
-//#region node_modules/@tensorflow/tfjs-backend-webgl/dist/unaryop_gpu.js
-var h = class {
-	constructor(e, t) {
-		this.variableNames = ["A"], this.outputShape = e, this.enableShapeUniforms = u(this.outputShape.length), this.userCode = `
-      float unaryOperation(float x) {
-        ${t}
-      }
-
-      void main() {
-        float x = getAAtOutCoords();
-        float y = unaryOperation(x);
-
-        setOutput(y);
-      }
-    `;
-	}
-}, g = "if (isnan(x)) return x;", _ = "return x;", v = "return abs(x);", y = "return (x >= 0.0) ? x : (exp(x) - 1.0);", b = g + "\n  return (x < 0.0) ? 0.0 : x;\n", x = g + "\n  return (x < 0.0) ? 0.0 : min(6.0, x);\n", S = "return x;", C = "return 1.0 / (1.0 + exp(-1.0 * x));", w = "return x;", T = "\n  vec4 result;\n\n  result.r = (x.r >= 0.0) ? x.r : (exp(x.r) - 1.0);\n  result.g = (x.g >= 0.0) ? x.g : (exp(x.g) - 1.0);\n  result.b = (x.b >= 0.0) ? x.b : (exp(x.b) - 1.0);\n  result.a = (x.a >= 0.0) ? x.a : (exp(x.a) - 1.0);\n\n  return result;\n", E = "\n  vec4 result = x * vec4(greaterThanEqual(x, vec4(0.0)));\n  bvec4 isNaN = isnan(x);\n\n  result.r = isNaN.r ? x.r : result.r;\n  result.g = isNaN.g ? x.g : result.g;\n  result.b = isNaN.b ? x.b : result.b;\n  result.a = isNaN.a ? x.a : result.a;\n\n  return result;\n", D = "\n  vec4 result = min(x, vec4(6.)) * vec4(greaterThanEqual(x, vec4(0.0)));\n  bvec4 isNaN = isnan(x);\n\n  result.r = isNaN.r ? x.r : result.r;\n  result.g = isNaN.g ? x.g : result.g;\n  result.b = isNaN.b ? x.b : result.b;\n  result.a = isNaN.a ? x.a : result.a;\n\n  return result;\n", O = "return 1.0 / (1.0 + exp(-1.0 * x));", k = class {
-	constructor(e, t) {
-		this.variableNames = ["A"], this.packedInputs = !0, this.packedOutput = !0, this.outputShape = e, this.enableShapeUniforms = u(this.outputShape.length), this.userCode = `
-      vec4 unaryOperation(vec4 x) {
-        ${t}
-      }
-
-      void main() {
-        vec4 x = getAAtOutCoords();
-        vec4 y = unaryOperation(x);
-
-        setOutput(y);
-      }
-    `;
-	}
-}, A = "\n  if (isnan(a)) return a;\n  if (isnan(b)) return b;\n", j = class {
-	constructor(e, t, r) {
-		this.variableNames = ["A", "B"], this.outputShape = n(t, r), this.enableShapeUniforms = u(this.outputShape.length), this.userCode = `
-      float binaryOperation(float a, float b) {
-        ${e}
-      }
-
-      void main() {
-        float a = getAAtOutCoords();
-        float b = getBAtOutCoords();
-        setOutput(binaryOperation(a, b));
-      }
-    `;
-	}
-}, M = "\n  result.r = isNaN.r ? NAN : result.r;\n  result.g = isNaN.g ? NAN : result.g;\n  result.b = isNaN.b ? NAN : result.b;\n  result.a = isNaN.a ? NAN : result.a;\n", N = class {
+//#region node_modules/@tensorflow/tfjs-backend-webgl/dist/binaryop_packed_gpu.js
+var _ = "\n  result.r = isNaN.r ? NAN : result.r;\n  result.g = isNaN.g ? NAN : result.g;\n  result.b = isNaN.b ? NAN : result.b;\n  result.a = isNaN.a ? NAN : result.a;\n", v = class {
 	constructor(e, t, r, i = !1) {
 		this.variableNames = ["A", "B"], this.supportsBroadcasting = !0, this.packedInputs = !0, this.packedOutput = !0, this.outputShape = n(t, r);
 		let a = this.outputShape.length;
-		this.enableShapeUniforms = u(a);
+		this.enableShapeUniforms = d(a);
 		let o = "";
 		if (i) if (a === 0 || s(this.outputShape) === 1) o = "\n          result.y = 0.;\n          result.z = 0.;\n          result.w = 0.;\n        ";
 		else if (o = `
-          ${d(a)} coords = getOutputCoords();
+          ${u(a)} coords = getOutputCoords();
         `, a === 1) this.enableShapeUniforms ? o += "\n            result.y = (coords + 1) >= outShape ? 0. : result.y;\n            result.z = 0.;\n            result.w = 0.;\n          " : o += `
             result.y = (coords + 1) >= ${this.outputShape[0]} ? 0. : result.y;
             result.z = 0.;
             result.w = 0.;
           `;
 		else {
-			let e = p("coords", a);
+			let e = h("coords", a);
 			this.enableShapeUniforms ? o += `
             bool nextRowOutOfBounds =
               (${e[a - 2]} + 1) >= outShape[${a} - 2];
@@ -120,7 +93,7 @@ var h = class {
 };
 //#endregion
 //#region node_modules/@tensorflow/tfjs-backend-webgl/dist/kernels/Identity.js
-function P(e) {
+function y(e) {
 	let { inputs: t, backend: n } = e, { x: r } = t;
 	return n.incRef(r.dataId), {
 		dataId: r.dataId,
@@ -128,48 +101,78 @@ function P(e) {
 		dtype: r.dtype
 	};
 }
-var F = {
+var b = {
 	kernelName: i,
 	backendName: "webgl",
-	kernelFunc: P
+	kernelFunc: y
 };
 //#endregion
 //#region node_modules/@tensorflow/tfjs-backend-webgl/dist/kernels/Complex.js
-function I(e) {
+function x(e) {
 	let { inputs: t, backend: n } = e, { real: r, imag: i } = t, a = n.makeTensorInfo(r.shape, "complex64"), o = n.texData.get(a.dataId);
 	return o.complexTensorInfos = {
-		real: P({
+		real: y({
 			inputs: { x: r },
 			backend: n
 		}),
-		imag: P({
+		imag: y({
 			inputs: { x: i },
 			backend: n
 		})
 	}, a;
 }
-var L = {
+var S = {
 	kernelName: a,
 	backendName: "webgl",
-	kernelFunc: I
-}, R = "return (a < 0.) ? b * a : a;", z = "\n  vec4 aLessThanZero = vec4(lessThan(a, vec4(0.)));\n  return (aLessThanZero * (b * a)) + ((vec4(1.0) - aLessThanZero) * a);\n";
-function B(e) {
-	let { inputs: n, backend: i, attrs: a } = e, { x: o } = n, { alpha: s } = a, c = i.makeTensorInfo([], "float32", t(s, "float32")), l = r().getBool("WEBGL_PACK_BINARY_OPERATIONS") ? new N(z, o.shape, c.shape) : new j(R, o.shape, c.shape), u = i.runWebGLProgram(l, [o, c], "float32");
+	kernelFunc: x
+}, C = "return (a < 0.) ? b * a : a;", w = "\n  vec4 aLessThanZero = vec4(lessThan(a, vec4(0.)));\n  return (aLessThanZero * (b * a)) + ((vec4(1.0) - aLessThanZero) * a);\n";
+function T(e) {
+	let { inputs: n, backend: i, attrs: a } = e, { x: o } = n, { alpha: s } = a, c = i.makeTensorInfo([], "float32", t(s, "float32")), l = r().getBool("WEBGL_PACK_BINARY_OPERATIONS") ? new v(w, o.shape, c.shape) : new p(C, o.shape, c.shape), u = i.runWebGLProgram(l, [o, c], "float32");
 	return i.disposeIntermediateTensorInfo(c), u;
 }
-var V = {
+var E = {
 	kernelName: c,
 	backendName: "webgl",
-	kernelFunc: B
-}, H = "return (a < 0.) ? b * a : a;", U = "\n  vec4 aLessThanZero = vec4(lessThan(a, vec4(0.)));\n  return (aLessThanZero * (b * a)) + ((vec4(1.0) - aLessThanZero) * a);\n";
-function W(e) {
-	let { inputs: t, backend: n } = e, { x: i, alpha: a } = t, o = r().getBool("WEBGL_PACK_BINARY_OPERATIONS") ? new N(U, i.shape, a.shape) : new j(H, i.shape, a.shape);
+	kernelFunc: T
+}, D = "return (a < 0.) ? b * a : a;", O = "\n  vec4 aLessThanZero = vec4(lessThan(a, vec4(0.)));\n  return (aLessThanZero * (b * a)) + ((vec4(1.0) - aLessThanZero) * a);\n";
+function k(e) {
+	let { inputs: t, backend: n } = e, { x: i, alpha: a } = t, o = r().getBool("WEBGL_PACK_BINARY_OPERATIONS") ? new v(O, i.shape, a.shape) : new p(D, i.shape, a.shape);
 	return n.runWebGLProgram(o, [i, a], "float32");
 }
-var G = {
+var A = {
 	kernelName: o,
 	backendName: "webgl",
-	kernelFunc: W
+	kernelFunc: k
+}, j = class {
+	constructor(e, t) {
+		this.variableNames = ["A"], this.outputShape = e, this.enableShapeUniforms = d(this.outputShape.length), this.userCode = `
+      float unaryOperation(float x) {
+        ${t}
+      }
+
+      void main() {
+        float x = getAAtOutCoords();
+        float y = unaryOperation(x);
+
+        setOutput(y);
+      }
+    `;
+	}
+}, M = "if (isnan(x)) return x;", N = "return x;", P = "return abs(x);", F = "return (x >= 0.0) ? x : (exp(x) - 1.0);", I = M + "\n  return (x < 0.0) ? 0.0 : x;\n", L = M + "\n  return (x < 0.0) ? 0.0 : min(6.0, x);\n", R = "return x;", z = "return 1.0 / (1.0 + exp(-1.0 * x));", B = "return x;", V = "\n  vec4 result;\n\n  result.r = (x.r >= 0.0) ? x.r : (exp(x.r) - 1.0);\n  result.g = (x.g >= 0.0) ? x.g : (exp(x.g) - 1.0);\n  result.b = (x.b >= 0.0) ? x.b : (exp(x.b) - 1.0);\n  result.a = (x.a >= 0.0) ? x.a : (exp(x.a) - 1.0);\n\n  return result;\n", H = "\n  vec4 result = x * vec4(greaterThanEqual(x, vec4(0.0)));\n  bvec4 isNaN = isnan(x);\n\n  result.r = isNaN.r ? x.r : result.r;\n  result.g = isNaN.g ? x.g : result.g;\n  result.b = isNaN.b ? x.b : result.b;\n  result.a = isNaN.a ? x.a : result.a;\n\n  return result;\n", U = "\n  vec4 result = min(x, vec4(6.)) * vec4(greaterThanEqual(x, vec4(0.0)));\n  bvec4 isNaN = isnan(x);\n\n  result.r = isNaN.r ? x.r : result.r;\n  result.g = isNaN.g ? x.g : result.g;\n  result.b = isNaN.b ? x.b : result.b;\n  result.a = isNaN.a ? x.a : result.a;\n\n  return result;\n", W = "return 1.0 / (1.0 + exp(-1.0 * x));", G = class {
+	constructor(e, t) {
+		this.variableNames = ["A"], this.packedInputs = !0, this.packedOutput = !0, this.outputShape = e, this.enableShapeUniforms = d(this.outputShape.length), this.userCode = `
+      vec4 unaryOperation(vec4 x) {
+        ${t}
+      }
+
+      void main() {
+        vec4 x = getAAtOutCoords();
+        vec4 y = unaryOperation(x);
+
+        setOutput(y);
+      }
+    `;
+	}
 }, K = "if (isnan(x)) return x;";
 function q({ opSnippet: e, packedOpSnippet: t, cpuKernelImpl: n, dtype: i }) {
 	return ({ inputs: a, backend: o }) => {
@@ -179,14 +182,14 @@ function q({ opSnippet: e, packedOpSnippet: t, cpuKernelImpl: n, dtype: i }) {
 			return c.makeTensorInfo(s.shape, l, e);
 		}
 		let u = r().getBool("WEBGL_PACK_UNARY_OPERATIONS") && t != null, d;
-		return d = u ? new k(s.shape, t) : new h(s.shape, e), c.runWebGLProgram(d, [s], l);
+		return d = u ? new G(s.shape, t) : new j(s.shape, e), c.runWebGLProgram(d, [s], l);
 	};
 }
 function J({ opSnippet: t, packedOpSnippet: n, checkOutOfBounds: i = !1, supportsComplex: a = !1, cpuKernelImpl: o, dtype: s }) {
 	return ({ inputs: c, backend: u }) => {
-		let { a: d, b: f } = c, p = u;
+		let { a: d, b: f } = c, m = u;
 		if (a && d.dtype === "complex64") {
-			let n = p.texData.get(d.dataId), r = p.texData.get(f.dataId), [i, a] = [[n.complexTensorInfos.real, r.complexTensorInfos.real], [n.complexTensorInfos.imag, r.complexTensorInfos.imag]].map((n) => {
+			let n = m.texData.get(d.dataId), r = m.texData.get(f.dataId), [i, a] = [[n.complexTensorInfos.real, r.complexTensorInfos.real], [n.complexTensorInfos.imag, r.complexTensorInfos.imag]].map((n) => {
 				let [r, i] = n, a = {
 					dataId: r.dataId,
 					dtype: r.dtype,
@@ -195,35 +198,35 @@ function J({ opSnippet: t, packedOpSnippet: n, checkOutOfBounds: i = !1, support
 					dataId: i.dataId,
 					dtype: i.dtype,
 					shape: f.shape
-				}, s = new j(t, d.shape, f.shape);
-				return p.runWebGLProgram(s, [a, o], e(r.dtype, i.dtype));
-			}), o = I({
+				}, s = new p(t, d.shape, f.shape);
+				return m.runWebGLProgram(s, [a, o], e(r.dtype, i.dtype));
+			}), o = x({
 				inputs: {
 					real: i,
 					imag: a
 				},
-				backend: p
+				backend: m
 			});
-			return p.disposeIntermediateTensorInfo(i), p.disposeIntermediateTensorInfo(a), o;
+			return m.disposeIntermediateTensorInfo(i), m.disposeIntermediateTensorInfo(a), o;
 		}
-		let m = s || e(d.dtype, f.dtype);
-		if ((d.dtype === "string" || f.dtype === "string" || p.shouldExecuteOnCPU([d, f])) && o != null) {
-			let e = p.texData.get(d.dataId).values, t = p.texData.get(f.dataId).values, n = d.dtype === "string" ? l(e) : e, r = d.dtype === "string" ? l(t) : t, [i, a] = o(d.shape, f.shape, n, r, m), s = p.makeTensorInfo(a, m), c = p.texData.get(s.dataId);
+		let h = s || e(d.dtype, f.dtype);
+		if ((d.dtype === "string" || f.dtype === "string" || m.shouldExecuteOnCPU([d, f])) && o != null) {
+			let e = m.texData.get(d.dataId).values, t = m.texData.get(f.dataId).values, n = d.dtype === "string" ? l(e) : e, r = d.dtype === "string" ? l(t) : t, [i, a] = o(d.shape, f.shape, n, r, h), s = m.makeTensorInfo(a, h), c = m.texData.get(s.dataId);
 			return c.values = i, s;
 		}
-		let h = r().getBool("WEBGL_PACK_BINARY_OPERATIONS") && n != null, g;
-		return g = h ? new N(n, d.shape, f.shape, i) : new j(t, d.shape, f.shape), p.runWebGLProgram(g, [d, f], m);
+		let g = r().getBool("WEBGL_PACK_BINARY_OPERATIONS") && n != null, _;
+		return _ = g ? new v(n, d.shape, f.shape, i) : new p(t, d.shape, f.shape), m.runWebGLProgram(_, [d, f], h);
 	};
 }
 function Y(e, t = !1) {
-	if (e === "linear") return t ? w : _;
-	if (e === "relu") return t ? E : b;
-	if (e === "elu") return t ? T : y;
-	if (e === "relu6") return t ? D : x;
-	if (e === "prelu") return t ? U : H;
-	if (e === "leakyrelu") return t ? z : R;
-	if (e === "sigmoid") return t ? O : C;
+	if (e === "linear") return t ? B : N;
+	if (e === "relu") return t ? H : I;
+	if (e === "elu") return t ? V : F;
+	if (e === "relu6") return t ? U : L;
+	if (e === "prelu") return t ? O : D;
+	if (e === "leakyrelu") return t ? w : C;
+	if (e === "sigmoid") return t ? W : z;
 	throw Error(`Activation ${e} has not been implemented for the WebGL backend.`);
 }
 //#endregion
-export { f as S, g as _, G as a, p as b, L as c, N as d, M as f, v as g, k as h, q as i, P as l, A as m, J as n, V as o, j as p, Y as r, I as s, K as t, F as u, S as v, m as x, h as y };
+export { f as S, _, G as a, m as b, R as c, E as d, x as f, v as g, b as h, q as i, j as l, y as m, J as n, P as o, S as p, Y as r, M as s, K as t, A as u, h as v, p as x, g as y };

package/dist/layers/MLP.d.ts

@@ -1,6 +1,6 @@
 import { Tensor } from '@tensorflow/tfjs-core';
 import { default as BaseLayer, ForwardAttributes } from './BaseLayer';
-import { GPTConfig } from '../../main';
+import { GPTConfig } from '../../models/config';
 export interface MLPConfig {
     activation?: 'gelu' | 'relu2';
     hiddenFactor?: number;

package/dist/layers/PositionEmbedding.d.ts

@@ -1,6 +1,7 @@
 import { Tensor } from '@tensorflow/tfjs-core';
 import { default as BaseLayer } from './BaseLayer';
-import { GPTConfig, ModelForwardAttributes } from '../../main';
+import { GPTConfig } from '../../models/config';
+import { ModelForwardAttributes } from '../../models/model';
 export default class PositionEmbedding extends BaseLayer {
     private wpe?;
     constructor(config: GPTConfig, name?: string, parent?: BaseLayer);

package/dist/layers/PositionEmbedding.js

@@ -1,5 +1,5 @@
 import { Tt as e, _n as t, di as n, dt as r, vi as i } from "../dist-Da20xy8E.js";
-import { n as a, t as o } from "../dist-BqAU9-yi.js";
+import { n as a, t as o } from "../dist-CwK5S7Ls.js";
 import s from "./BaseLayer.js";
 //#region lib/layers/PositionEmbedding.ts
 var c = class extends s {

package/dist/layers/RMSNorm.d.ts

@@ -1,6 +1,6 @@
 import { Tensor } from '@tensorflow/tfjs-core';
 import { default as BaseLayer, ForwardAttributes } from './BaseLayer';
-import { GPTConfig } from '../../main';
+import { GPTConfig } from '../../models/config';
 export interface RMSNormConfig {
     useGamma?: boolean;
 }

package/dist/layers/TiedEmbedding.js

@@ -1,5 +1,5 @@
 import { T as e, an as t } from "../dist-Da20xy8E.js";
-import { n } from "../dist-BqAU9-yi.js";
+import { n } from "../dist-CwK5S7Ls.js";
 import { isPackedTensor as r } from "../utilities/packed.js";
 import { transpose16 as i } from "../ops/transpose16.js";
 import { t as a } from "../pack16-BhuXNUS7.js";

package/dist/layers.d.ts

@@ -0,0 +1,4 @@
+export { default as CausalSelfAttention } from './layers/CausalSelfAttention';
+export { default as MLP } from './layers/MLP';
+export { default as TransformerBlock } from './layers/TransformerBlock';
+export { default as RoPECache } from './layers/RoPECache';

package/dist/layers.js

@@ -0,0 +1,14 @@
+import { n as e } from "./chunk-CWhphoD1.js";
+import t from "./layers/CausalSelfAttention.js";
+import n from "./layers/MLP.js";
+import r from "./layers/TransformerBlock.js";
+import i from "./layers/RoPECache.js";
+//#region lib/layers.ts
+var a = /* @__PURE__ */ e({
+	CausalSelfAttention: () => t,
+	MLP: () => n,
+	RoPECache: () => i,
+	TransformerBlock: () => r
+});
+//#endregion
+export { t as CausalSelfAttention, n as MLP, i as RoPECache, r as TransformerBlock, a as t };

package/dist/loader/load.js

@@ -1,2 +1,58 @@
-import { c as e, l as t } from "../main-BSaDGH7I.js";
-export { e as VERSION, t as loadModel };
+import { i as e } from "../chunk-CWhphoD1.js";
+import { t } from "../jszip.min-DKa1Rjyn.js";
+import { load_safetensors as n } from "../utilities/safetensors.js";
+import r from "./oldZipLoad.js";
+import i from "./newZipLoad.js";
+import a from "./loadHF.js";
+import { loadZipMeta as o } from "./loadZipMeta.js";
+//#region lib/loader/load.ts
+var s = /* @__PURE__ */ e(t(), 1), c = 2;
+async function l(e) {
+	let t = await fetch(e);
+	if (!t.ok) throw Error(`Failed to fetch ${e}: ${t.statusText}`);
+	return t.arrayBuffer();
+}
+async function u(e, t, r) {
+	let i = e.file("model.safetensors");
+	if (!i) return;
+	let a = await n(await i.async("arraybuffer")), o = /* @__PURE__ */ new Map();
+	for (let [e, t] of Object.entries(a)) o.set(e, [t]);
+	t.weightStore.loadWeights(o, r);
+}
+async function d(e, t) {
+	let n = e.file("config.json");
+	if (!n) return;
+	let r = await n.async("string"), i = JSON.parse(r);
+	if (i.loraName) {
+		if (t.hasLoRA()) throw Error("Model already has LoRA attached");
+		t.attachLoRA(i.loraName);
+	}
+}
+async function f(e, t) {
+	if (t.reference) {
+		let n = await p(t.reference);
+		return await u(e, n.model, !!t.url), await d(e, n.model), n.model.config.loraName && n.model.attachLoRA(n.model.config.loraName), {
+			...n,
+			metaData: {
+				...n.metaData,
+				...t
+			}
+		};
+	} else if (e.file("manifest.json")) return r(e, t);
+	else {
+		let n = await i(e, t);
+		return n.model.config.loraName && n.model.attachLoRA(n.model.config.loraName), n;
+	}
+}
+async function p(e, t) {
+	if (typeof e == "string") if (e.startsWith("http://") || e.startsWith("https://")) {
+		let t = await l(e), n = await s.default.loadAsync(t), r = await o(n);
+		return r.url = e, f(n, r);
+	} else return a(e);
+	else {
+		let n = await s.default.loadAsync(e), r = await o(n);
+		return r.url = t?.sourceURL || void 0, f(n, r);
+	}
+}
+//#endregion
+export { c as VERSION, p as loadModel };

package/dist/loader/loadHF.d.ts

@@ -1,4 +1,4 @@
-import { ITokeniser } from '../../main';
+import { ITokeniser } from '../../tokeniser/type';
 import { default as Model, ModelForwardAttributes } from '../../models/model';
 import { TransformersMetadata } from './types';
 export default function loadHuggingFace(name: string): Promise<{

package/dist/loader/loadHF.js

@@ -1,2 +1,17 @@
-import { u as e } from "../main-BSaDGH7I.js";
-export { e as default };
+import e from "./loadTransformers.js";
+//#region lib/loader/loadHF.ts
+async function t(t) {
+	let n = `https://huggingface.co/${t}/resolve/main/config.json`, r = `https://huggingface.co/${t}/resolve/main/tokeniser.json`, i = `https://huggingface.co/${t}/resolve/main/meta.json`, a = `https://huggingface.co/${t}/resolve/main/model.safetensors`, [o, s, c, l] = await Promise.all([
+		fetch(n),
+		fetch(r),
+		fetch(i),
+		fetch(a)
+	]);
+	if (!o.ok) throw Error(`Failed to fetch config from ${n}: ${o.statusText}`);
+	if (!s.ok) throw Error(`Failed to fetch tokeniser from ${r}: ${s.statusText}`);
+	if (!c.ok) throw Error(`Failed to fetch meta from ${i}: ${c.statusText}`);
+	if (!l.ok) throw Error(`Failed to fetch weights from ${a}: ${l.statusText}`);
+	return e(await o.json(), await s.json(), await c.json(), await l.arrayBuffer());
+}
+//#endregion
+export { t as default };

package/dist/loader/loadTransformers.js

@@ -1,2 +1,46 @@
-import { f as e, p as t } from "../main-BSaDGH7I.js";
-export { e as default, t as mapTransformersConfigToGPTConfig };
+import e from "../tokeniser/CharTokeniser.js";
+import t from "../tokeniser/bpe.js";
+import { ri as n } from "../dist-Da20xy8E.js";
+import { load_safetensors as r } from "../utilities/safetensors.js";
+import { dummyPassAsync as i } from "../utilities/dummy.js";
+import a from "../models/factory.js";
+//#region lib/loader/loadTransformers.ts
+function o(e) {
+	let t;
+	return t = e.model_type === "GenAI_NanoGPT_v1" ? {
+		modelType: "GenAI_NanoGPT_v1",
+		vocabSize: e.vocab_size,
+		blockSize: e.block_size,
+		nLayer: e.num_hidden_layers,
+		nHead: e.num_attention_heads,
+		nEmbed: e.hidden_size,
+		mlpFactor: e.mlpFactor,
+		useRope: e.useRope
+	} : {
+		modelType: "GenAI_NanoGPT_v2",
+		vocabSize: e.vocab_size,
+		blockSize: e.block_size,
+		nLayer: e.num_hidden_layers,
+		nHead: e.num_attention_heads,
+		nEmbed: e.hidden_size,
+		mlpFactor: e.mlpFactor,
+		loraConfig: e.loraConfig ? new Map(Object.entries(e.loraConfig)) : void 0,
+		loraName: e.loraName,
+		windowSize: e.windowSize
+	}, t;
+}
+async function s(s, c, l, u) {
+	let d = o(s), f = (c.type ?? "char") === "char" ? new e(c.vocab) : new t(c.vocab, c.merges);
+	c.datasetID && (f.datasetID = c.datasetID), c.id ? f.id = c.id : f.generateID();
+	let p = await r(u), m = /* @__PURE__ */ new Map();
+	for (let [e, t] of Object.entries(p)) m.set(e, [t]);
+	n();
+	let h = a(d);
+	return h.metaData = l, await i(h), h.weightStore.loadWeights(m, !!l.url), {
+		model: h,
+		tokeniser: f,
+		metaData: l
+	};
+}
+//#endregion
+export { s as default, o as mapTransformersConfigToGPTConfig };

package/dist/loader/newZipLoad.js

@@ -1,2 +1,25 @@
-import { d as e } from "../main-BSaDGH7I.js";
-export { e as default };
+import e from "./loadTransformers.js";
+import { AdamWOptimizer as t } from "../training/AdamW.js";
+//#region lib/loader/newZipLoad.ts
+async function n(n, r) {
+	let i = await n.file("config.json")?.async("string");
+	if (!i) throw Error("Config file not found in the zip archive");
+	let a = JSON.parse(i), o = await n.file("tokeniser.json")?.async("string");
+	if (!o) throw Error("Tokeniser file not found in the zip archive");
+	let s = JSON.parse(o), c = await n.file("model.safetensors")?.async("arraybuffer");
+	if (!c) throw Error("Model weights not found in the zip archive");
+	let l = await n.file("optimizer_config.json")?.async("string"), u;
+	if (l) {
+		let e = JSON.parse(l), r = await n.file("optimizer.safetensors")?.async("arraybuffer");
+		if (!r) throw Error("Optimizer weights not found in the zip archive");
+		u = new t(e), await u.loadMoments(r);
+	}
+	let d = await n.file("training_log.json")?.async("string"), f;
+	return d && (f = JSON.parse(d)), {
+		...await e(a, s, r, c),
+		optimizer: u,
+		log: f
+	};
+}
+//#endregion
+export { n as default };

package/dist/loader/oldZipLoad.d.ts

@@ -1,5 +1,5 @@
 import { default as zip } from 'jszip';
-import { ITokeniser } from '../../main';
+import { ITokeniser } from '../../tokeniser/type';
 import { default as Model, ModelForwardAttributes } from '../../models/model';
 import { TransformersMetadata } from './types';
 export default function loadOldModel(zipFile: zip, metaData: TransformersMetadata): Promise<{

package/dist/loader/oldZipLoad.js

@@ -1,2 +1,37 @@
-import { m as e } from "../main-BSaDGH7I.js";
-export { e as default };
+import e from "../tokeniser/CharTokeniser.js";
+import t from "../tokeniser/bpe.js";
+import { ri as n } from "../dist-Da20xy8E.js";
+import { importWeights as r } from "../utilities/weights.js";
+import { dummyPassAsync as i } from "../utilities/dummy.js";
+import a from "../models/factory.js";
+//#region lib/loader/oldZipLoad.ts
+async function o(o, s) {
+	let c = /* @__PURE__ */ new Map(), l = await o.file("manifest.json")?.async("string");
+	if (!l) throw Error("Manifest file not found in the zip archive");
+	let u = JSON.parse(l);
+	for (let [e, t] of Object.entries(u.weightSpec)) c.set(e, {
+		spec: t,
+		data: new Float32Array()
+	});
+	let d = await o.file("tokeniser.json")?.async("string");
+	if (!d) throw Error("Tokeniser file not found in the zip archive");
+	let f = JSON.parse(d), p = (f.type ?? "char") === "char" ? new e(f.vocab) : new t(f.vocab, f.merges), m = /* @__PURE__ */ new Map();
+	for (let e of Object.keys(o.files)) if (e.endsWith(".bin")) {
+		let t = e.replace(".bin", ""), n = await o.file(e).async("arraybuffer"), i = new Float32Array(n), a = c.get(t) || {
+			spec: [],
+			data: new Float32Array()
+		};
+		a.data = i, c.set(t, a);
+		let s = await r(a);
+		m.set(t, s);
+	}
+	n();
+	let h = a(u.config);
+	return h.metaData = s, await i(h), h.weightStore.loadWeights(m, !!s.url), {
+		model: h,
+		tokeniser: p,
+		metaData: s
+	};
+}
+//#endregion
+export { o as default };

package/dist/loader/save.js

@@ -1,2 +1,75 @@
-import { s as e } from "../main-BSaDGH7I.js";
-export { e as saveModel };
+import { i as e } from "../chunk-CWhphoD1.js";
+import t from "../tokeniser/CharTokeniser.js";
+import { t as n } from "../jszip.min-DKa1Rjyn.js";
+import { save_safetensors as r } from "../utilities/safetensors.js";
+import "./load.js";
+//#region lib/loader/save.ts
+var i = /* @__PURE__ */ e(n(), 1);
+function a(e) {
+	if (e.length > 1e3) {
+		let t = Math.ceil(e.length / 1e3);
+		return e.filter((n, r) => r % t === 0 || r === e.length - 1);
+	}
+	return e;
+}
+async function o(e, n, o, s) {
+	let c = /* @__PURE__ */ new Map();
+	e.weightStore.saveWeights(c);
+	let l = new i.default();
+	if (s?.optimizer) {
+		let e = await s.optimizer.saveMoments();
+		l.file("optimizer.safetensors", e, { binary: !0 }), l.file("optimizer_config.json", JSON.stringify(s.optimizer.serializeConfig()), { binary: !1 });
+	}
+	s?.trainingLog && l.file("training_log.json", JSON.stringify(a(s.trainingLog), void 0, 4), { binary: !1 });
+	let u = {};
+	c.forEach((e, t) => {
+		e.length === 1 && (u[t] = e[0]);
+	});
+	let d = await r(u, o?.quantize);
+	l.file("model.safetensors", d, { binary: !0 });
+	let f = e.config.modelType, p;
+	p = f === "GenAI_NanoGPT_v1" ? {
+		model_type: "GenAI_NanoGPT_v1",
+		vocab_size: n.getVocab().length,
+		hidden_size: e.config.nEmbed,
+		num_hidden_layers: e.config.nLayer,
+		num_attention_heads: e.config.nHead,
+		block_size: e.config.blockSize,
+		mlpFactor: e.config.mlpFactor,
+		useRope: e.config.useRope
+	} : {
+		model_type: "GenAI_NanoGPT_v2",
+		vocab_size: n.getVocab().length,
+		hidden_size: e.config.nEmbed,
+		num_hidden_layers: e.config.nLayer,
+		num_attention_heads: e.config.nHead,
+		block_size: e.config.blockSize,
+		mlpFactor: e.config.mlpFactor,
+		loraConfig: e.config.loraConfig ? Object.fromEntries(e.config.loraConfig) : void 0,
+		loraName: e.config.loraName,
+		windowSize: e.config.windowSize
+	}, l.file("config.json", JSON.stringify(p, void 0, 4), { binary: !1 });
+	let m = {
+		version: 2,
+		application: "@genai-fi/nanogpt",
+		meta: o?.metadata,
+		name: o?.name,
+		training: e.metaData?.training || void 0,
+		reference: e.metaData?.url || void 0,
+		mode: e.metaData?.mode || void 0,
+		pretrainingData: e.metaData?.pretrainingData || void 0,
+		pretrainingSettings: e.metaData?.pretrainingSettings || void 0,
+		generationSettings: e.metaData?.generationSettings || void 0,
+		actionLog: e.metaData?.actionLog || void 0
+	};
+	if (l.file("meta.json", JSON.stringify(m, void 0, 4), { binary: !1 }), l.file("tokeniser.json", JSON.stringify({
+		type: n instanceof t ? "char" : "bpe",
+		vocab: n.getVocab(),
+		merges: n.getMerges(),
+		datasetID: n.datasetID,
+		id: n.id
+	}), { binary: !1 }), o?.files) for (let [e, t] of Object.entries(o.files)) l.file(e, JSON.stringify(t), { binary: !1 });
+	return l.generateAsync({ type: "blob" });
+}
+//#endregion
+export { o as saveModel };