@genai-fi/nanogpt 0.5.6 → 0.6.1

package/dist/{sin-CPAZXNjH.js → sin-9JBrfVaB.js}

@@ -1,4 +1,4 @@
-import { o, j as t, E as c, _ as a, $ as e } from "./index-Du-bmOP8.js";
+import { q as o, w as t, E as c, a0 as a, a1 as e } from "./index-CamYe_M8.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.

package/dist/{softmax-DhWoBa7r.js → softmax-DvMvui-_.js}

@@ -1,4 +1,4 @@
-import { o as r, j as f, E as e, S as i } from "./index-Du-bmOP8.js";
+import { q as r, w as f, E as e, S as i } from "./index-CamYe_M8.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.

package/dist/{split-BCUhuU7B.js → split-DxrHrPFK.js}

@@ -1,4 +1,4 @@
-import { o as p, j as i, E as a, x as c } from "./index-Du-bmOP8.js";
+import { q as p, w as i, E as a, H as c } from "./index-CamYe_M8.js";
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
@@ -15,9 +15,9 @@ import { o as p, j as i, E as a, x as c } from "./index-Du-bmOP8.js";
  * limitations under the License.
  * =============================================================================
  */
-function e(t, s, o = 0) {
-  const n = { x: i(t, "x", "split") }, r = { numOrSizeSplits: s, axis: o };
-  return a.runKernel(c, n, r);
+function e(t, s, n = 0) {
+  const o = { x: i(t, "x", "split") }, r = { numOrSizeSplits: s, axis: n };
+  return a.runKernel(c, o, r);
 }
 const u = /* @__PURE__ */ p({ split_: e });
 export {

package/dist/{stack-BV1v7l3S.js → stack-DgaoDmnF.js}

@@ -1,4 +1,4 @@
-import { o as e, k as c, l as n, E as k, P as i } from "./index-Du-bmOP8.js";
+import { q as e, x as c, k as n, E as k, P as i } from "./index-CamYe_M8.js";
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.

package/dist/{sum-Cvq06317.js → sum-BpcpxNEh.js}

@@ -1,4 +1,4 @@
-import { o as e, j as u, D as c, E as l, F as m } from "./index-Du-bmOP8.js";
+import { q as e, w as u, N as c, E as l, O as m } from "./index-CamYe_M8.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -15,10 +15,10 @@ import { o as e, j as u, D as c, E as l, F as m } from "./index-Du-bmOP8.js";
  * limitations under the License.
  * =============================================================================
  */
-function i(t, o = null, n = !1) {
+function i(t, n = null, o = !1) {
   let s = u(t, "x", "sum");
   s.dtype === "bool" && (s = c(s, "int32"));
-  const r = { x: s }, a = { axis: o, keepDims: n };
+  const r = { x: s }, a = { axis: n, keepDims: o };
   return l.runKernel(m, r, a);
 }
 const f = /* @__PURE__ */ e({ sum_: i });

package/dist/{tensor-DgTOPY6h.js → tensor-CDz5x1mP.js}

@@ -1,4 +1,4 @@
-import { J as t, K as a } from "./index-Du-bmOP8.js";
+import { Q as t, U as a } from "./index-CamYe_M8.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.

package/dist/{tensor2d-CRWjDyUe.js → tensor2d-jO8JY5Jd.js}

@@ -1,4 +1,4 @@
-import { I as t, J as s, K as a } from "./index-Du-bmOP8.js";
+import { V as t, Q as s, U as a } from "./index-CamYe_M8.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.

package/dist/training/AdamExt.js

@@ -1,4 +1,4 @@
-import { A as r, b as c, f as h, s as g, e as o } from "../index-Du-bmOP8.js";
+import { A as r, b as c, f as h, s as g, e as o } from "../index-CamYe_M8.js";
 class u extends r {
   constructor(t, e, s, a, i) {
     super(t, e, s, a), this.config = i, this.startLearningRate = t;

package/dist/training/DatasetBuilder.js

@@ -1,5 +1,5 @@
-import { t as u } from "../index-Du-bmOP8.js";
-import { d as z, i as f } from "../dataset-iqT4Otvb.js";
+import { t as u } from "../index-CamYe_M8.js";
+import { d as z, i as f } from "../dataset-BFFipD1c.js";
 import "../index-Tf7vU29b.js";
 /**
  * @license

package/dist/training/FullTrainer.js

@@ -1,7 +1,7 @@
 import { generateText as T } from "../utilities/generate.js";
 import L from "./Trainer.js";
 import x from "./Evaluator.js";
-import { a as h } from "../index-Du-bmOP8.js";
+import { a as h } from "../index-CamYe_M8.js";
 import y from "../utilities/profile.js";
 const D = {
   desiredLoss: 0.01,

package/dist/training/Trainer.js

@@ -1,8 +1,8 @@
 import { DatasetBuilder as g, flattenTokens as m, PAGE_FACTOR as u } from "./DatasetBuilder.js";
 import f from "./AdamExt.js";
-import { t as y, v as z, a as c } from "../index-Du-bmOP8.js";
-import { n as S } from "../norm-01kY9I2B.js";
-import { z as p } from "../zeros-BaHhQTWf.js";
+import { t as y, v as z, a as c } from "../index-CamYe_M8.js";
+import { n as S } from "../norm-B9dQTFYn.js";
+import { z as p } from "../zeros-DUkkVccu.js";
 class R {
   constructor(t, e, s = 1e-3) {
     this.tokenizer = e, this.model = t, this.learningRate = s, this.resetOptimizer(), this.datasetBuilder = new g(e, t.config.gpt.blockSize);

package/dist/training/sparseCrossEntropy.js

@@ -1,9 +1,9 @@
 import { gatherSub as L } from "../ops/gatherSub.js";
 import { scatterSub as y } from "../ops/scatterSub.js";
-import { e as u, c as i, z as S, t as f, s as G } from "../index-Du-bmOP8.js";
-import { s as v } from "../softmax-DhWoBa7r.js";
-import { m as z } from "../max-0Xnlpv8k.js";
-import { l as k } from "../log_sum_exp-CxfBtUaG.js";
+import { e as u, c as i, z as S, t as f, s as G } from "../index-CamYe_M8.js";
+import { s as v } from "../softmax-DvMvui-_.js";
+import { m as z } from "../max-JBBv7aUf.js";
+import { l as k } from "../log_sum_exp-CV_5-TTu.js";
 function F(a, s) {
   return f(() => {
     const e = a.shape[a.shape.length - 1], o = a.shape.slice(0, -1).reduce((d, c) => d * c, 1), p = a.shape.length > 2 ? a.reshape([o, e]) : a, n = s.shape.length > 1 ? s.reshape([o]).cast("int32") : s.cast("int32"), t = z(p, -1, !0), r = G(p, t), h = k(r, -1);

package/dist/utilities/dummy.d.ts

@@ -1,3 +1,9 @@
 import { default as NanoGPT } from '../NanoGPTModel';
 export declare function dummyPassAsync(model: NanoGPT): Promise<void>;
+export interface MemoryRequirements {
+    perBatch: number;
+    tapeSize: number;
+    gradients: number;
+}
+export declare function dummyPassTrainAsync(model: NanoGPT): Promise<MemoryRequirements>;
 export declare function dummyPass(model: NanoGPT): void;

package/dist/utilities/dummy.js

@@ -1,14 +1,35 @@
-import "../index-Du-bmOP8.js";
-import { z as n } from "../zeros-BaHhQTWf.js";
-async function c(s) {
-  const i = n([1, s.config.gpt.blockSize], "int32"), [t, o] = s.forward({ training: !1 }, i);
-  await t.data(), t.dispose(), o && o.dispose(), i.dispose();
+import { m as f, v as S, e as w } from "../index-CamYe_M8.js";
+import { z as i } from "../zeros-DUkkVccu.js";
+async function P(s) {
+  const t = i([1, s.config.gpt.blockSize], "int32"), [e, n] = s.forward({ training: !1 }, t);
+  await e.data(), e.dispose(), n && n.dispose(), t.dispose();
 }
-function d(s) {
-  const i = n([1, s.config.gpt.blockSize], "int32"), [t, o] = s.forward({ training: !1 }, i);
-  t.dispose(), o && o.dispose(), i.dispose();
+async function v(s) {
+  const t = f(), e = t.numBytesInGPUAllocated ?? t.numBytes;
+  await P(s);
+  const n = i([1, s.config.gpt.blockSize], "int32"), r = i([1, s.config.gpt.blockSize], "int32"), o = {
+    perBatch: 0,
+    tapeSize: 0,
+    gradients: s.getNumParams() * 4
+  }, g = () => {
+    const [a, l] = s.forward({ training: !0 }, n, r), d = w().state.activeTape;
+    let u = 0;
+    if (d)
+      for (const z of d)
+        u += z.saved?.reduce((B, I) => B + I.size * 4, 0) || 0;
+    return o.tapeSize = u, a.dispose(), l;
+  }, { value: c, grads: m } = S(g), p = f(), y = p.numBytesInGPUAllocated ?? p.numBytes;
+  o.perBatch = y - e - o.gradients, console.log("Dummy training memory requirements:", o), await c.data(), c.dispose();
+  for (const a in m)
+    m[a].dispose();
+  return n.dispose(), r.dispose(), o;
+}
+function A(s) {
+  const t = i([1, s.config.gpt.blockSize], "int32"), [e, n] = s.forward({ training: !1 }, t);
+  e.dispose(), n && n.dispose(), t.dispose();
 }
 export {
-  d as dummyPass,
-  c as dummyPassAsync
+  A as dummyPass,
+  P as dummyPassAsync,
+  v as dummyPassTrainAsync
 };

package/dist/utilities/generate.js

@@ -1,6 +1,6 @@
-import { t as y } from "../index-Du-bmOP8.js";
-import { t as x } from "../tensor2d-CRWjDyUe.js";
-import { c as f } from "../concat-DdKPyAtw.js";
+import { t as y } from "../index-CamYe_M8.js";
+import { t as x } from "../tensor2d-jO8JY5Jd.js";
+import { c as f } from "../concat-XOK9ANZu.js";
 async function A(o, r, a, c, T) {
   if (c <= 0)
     throw new Error("Length must be a positive integer");

package/dist/utilities/load.d.ts

@@ -1,6 +1,31 @@
+import { default as zip } from 'jszip';
 import { default as NanoGPT } from '../NanoGPTModel';
 import { ITokeniser } from '../tokeniser/type';
+export declare const VERSION = 2;
+export interface TransformersConfig {
+    model_type: string;
+    vocab_size: number;
+    hidden_size: number;
+    num_hidden_layers: number;
+    num_attention_heads: number;
+    block_size: number;
+    dropout: number;
+    biasInLinear: boolean;
+    biasInLayerNorm: boolean;
+    mlpFactor: number;
+    useRope: boolean;
+}
+export interface Metadata {
+    version: string;
+    application: string;
+    name?: string;
+}
+export declare function loadOldModel(zipFile: zip): Promise<{
+    model: NanoGPT;
+    tokeniser: ITokeniser;
+}>;
 export declare function loadModel(data: Blob | Buffer | string): Promise<{
     model: NanoGPT;
     tokeniser: ITokeniser;
+    name?: string;
 }>;

package/dist/utilities/load.js

@@ -1,47 +1,99 @@
-import { j as k } from "../jszip.min-CjP2V1VV.js";
-import { importWeights as b } from "./weights.js";
-import u from "../tokeniser/CharTokeniser.js";
-import F from "../NanoGPTModel.js";
-import { dummyPassAsync as j } from "./dummy.js";
-import { d as T } from "../index-Du-bmOP8.js";
-import E from "../tokeniser/bpe.js";
-async function A(t) {
-  const o = await fetch(t);
-  if (!o.ok)
-    throw new Error(`Failed to fetch ${t}: ${o.statusText}`);
-  return o.arrayBuffer();
+import { j as v } from "../jszip.min-CjP2V1VV.js";
+import { importWeights as F } from "./weights.js";
+import h from "../tokeniser/CharTokeniser.js";
+import b from "../NanoGPTModel.js";
+import { dummyPassAsync as u } from "./dummy.js";
+import { d as k } from "../index-CamYe_M8.js";
+import j from "../tokeniser/bpe.js";
+import { load_safetensors as N } from "./safetensors.js";
+const I = 2;
+async function O(t) {
+  const s = await fetch(t);
+  if (!s.ok)
+    throw new Error(`Failed to fetch ${t}: ${s.statusText}`);
+  return s.arrayBuffer();
 }
-async function P(t) {
-  const o = typeof t == "string" ? await A(t) : t, n = await k.loadAsync(o), i = /* @__PURE__ */ new Map(), f = await n.file("manifest.json")?.async("string");
-  if (!f)
+async function S(t) {
+  const s = /* @__PURE__ */ new Map(), r = await t.file("manifest.json")?.async("string");
+  if (!r)
     throw new Error("Manifest file not found in the zip archive");
-  const l = JSON.parse(f);
-  for (const [e, r] of Object.entries(l.weightSpec))
-    i.set(e, { spec: r, data: new Float32Array() });
-  const p = await n.file("tokeniser.json")?.async("string");
-  if (!p)
+  const p = JSON.parse(r);
+  for (const [o, a] of Object.entries(p.weightSpec))
+    s.set(o, { spec: a, data: new Float32Array() });
+  const e = await t.file("tokeniser.json")?.async("string");
+  if (!e)
     throw new Error("Tokeniser file not found in the zip archive");
-  const s = JSON.parse(p), y = (s.type ?? "char") === "char" ? new u(s.vocab) : new E(s.vocab, s.merges), w = /* @__PURE__ */ new Map();
-  for (const e of Object.keys(n.files))
-    if (e.endsWith(".bin")) {
-      const r = e.replace(".bin", ""), g = await n.file(e).async("arraybuffer"), h = new Float32Array(g), c = i.get(r) || { spec: [], data: new Float32Array() };
-      c.data = h, i.set(r, c);
-      const d = await b(c);
-      w.set(r, d);
+  const i = JSON.parse(e), c = (i.type ?? "char") === "char" ? new h(i.vocab) : new j(i.vocab, i.merges), d = /* @__PURE__ */ new Map();
+  for (const o of Object.keys(t.files))
+    if (o.endsWith(".bin")) {
+      const a = o.replace(".bin", ""), w = await t.file(o).async("arraybuffer"), g = new Float32Array(w), l = s.get(a) || { spec: [], data: new Float32Array() };
+      l.data = g, s.set(a, l);
+      const n = await F(l);
+      d.set(a, n);
     }
-  T();
-  const a = new F(l.config);
-  await j(a), a.loadWeights(w);
-  const m = await n.file("log.json")?.async("string");
+  k();
+  const f = new b(p.config);
+  await u(f), f.loadWeights(d);
+  const m = await t.file("log.json")?.async("string");
   if (m)
     try {
-      const e = JSON.parse(m);
-      a.log = e;
-    } catch (e) {
-      throw console.error("Error parsing training log:", e), new Error(`Failed to parse training log: ${e}`);
+      const o = JSON.parse(m);
+      f.log = o;
+    } catch (o) {
+      throw console.error("Error parsing training log:", o), new Error(`Failed to parse training log: ${o}`);
     }
-  return { model: a, tokeniser: y };
+  return { model: f, tokeniser: c };
+}
+async function R(t) {
+  const s = typeof t == "string" ? await O(t) : t, r = await v.loadAsync(s);
+  if (r.file("manifest.json"))
+    return S(r);
+  {
+    const p = await r.file("config.json")?.async("string");
+    if (!p)
+      throw new Error("Config file not found in the zip archive");
+    const e = JSON.parse(p), i = {
+      vocabSize: e.vocab_size,
+      blockSize: e.block_size,
+      nLayer: e.num_hidden_layers,
+      nHead: e.num_attention_heads,
+      nEmbed: e.hidden_size,
+      dropout: e.dropout,
+      biasInLinear: e.biasInLinear,
+      biasInLayerNorm: e.biasInLayerNorm,
+      mlpFactor: e.mlpFactor,
+      useRope: e.useRope
+    }, y = await r.file("tokeniser.json")?.async("string");
+    if (!y)
+      throw new Error("Tokeniser file not found in the zip archive");
+    const c = JSON.parse(y), f = (c.type ?? "char") === "char" ? new h(c.vocab) : new j(c.vocab, c.merges), m = await N(await r.file("model.safetensors").async("arraybuffer")), o = /* @__PURE__ */ new Map();
+    for (const [n, E] of Object.entries(m))
+      o.set(n, [E]);
+    k();
+    const a = new b(i);
+    await u(a), a.loadWeights(o);
+    const w = await r.file("meta.json")?.async("string");
+    let g;
+    if (w)
+      try {
+        const n = JSON.parse(w);
+        n.name && (g = n.name);
+      } catch (n) {
+        console.error("Error parsing meta file:", n);
+      }
+    const l = await r.file("log.json")?.async("string");
+    if (l)
+      try {
+        const n = JSON.parse(l);
+        a.log = n;
+      } catch (n) {
+        throw console.error("Error parsing training log:", n), new Error(`Failed to parse training log: ${n}`);
+      }
+    return { model: a, tokeniser: f, name: g };
+  }
 }
 export {
-  P as loadModel
+  I as VERSION,
+  R as loadModel,
+  S as loadOldModel
 };

package/dist/utilities/profile.d.ts

@@ -1,3 +1,7 @@
+import { MemoryInfo } from '@tensorflow/tfjs-core';
+export interface ExtendedMemoryInfo extends MemoryInfo {
+    numBytesInGPUAllocated?: number;
+}
 export default class MemoryProfiler {
     private log;
     private maxMemory;
@@ -6,6 +10,7 @@ export default class MemoryProfiler {
     private peakMemory;
     startMemory(): void;
     getPeakMemory(): number;
+    getMaxMemory(): number;
     endMemory(label: string): void;
     printSummary(): void;
 }

package/dist/utilities/profile.js

@@ -1,32 +1,35 @@
-import { m as s } from "../index-Du-bmOP8.js";
+import { m as a } from "../index-CamYe_M8.js";
 const m = 1024 * 1024;
-class M {
+class l {
   log = /* @__PURE__ */ new Map();
   maxMemory = 0;
   maxLabel;
   lastMemInfo = [];
   peakMemory = 0;
   startMemory() {
-    this.lastMemInfo.push(s());
+    this.lastMemInfo.push(a());
   }
   getPeakMemory() {
     return this.peakMemory;
   }
-  endMemory(e) {
+  getMaxMemory() {
+    return this.maxMemory;
+  }
+  endMemory(o) {
     if (this.lastMemInfo.length === 0) {
       console.warn("MemoryProfiler: endMemory called without matching startMemory");
       return;
     }
-    const o = s(), t = o.numBytes - (this.lastMemInfo.pop()?.numBytes || 0);
-    this.log.set(e, Math.max(this.log.get(e) || 0, t)), t > this.maxMemory && (this.maxMemory = t, this.maxLabel = e), this.peakMemory = Math.max(this.peakMemory, o.numBytes);
+    const e = a(), s = this.lastMemInfo.pop(), t = (e.numBytesInGPUAllocated ?? e.numBytes) - (s?.numBytes ?? s?.numBytesInGPUAllocated ?? 0);
+    this.log.set(o, Math.max(this.log.get(o) || 0, t)), t > this.maxMemory && (this.maxMemory = t, this.maxLabel = o), this.peakMemory = Math.max(this.peakMemory, e.numBytesInGPUAllocated ?? e.numBytes);
   }
   printSummary() {
     console.log("Memory Usage Summary:");
-    for (const [e, o] of this.log.entries())
-      console.log(`- ${e}: ${(o / m).toFixed(2)} MB`);
+    for (const [o, e] of this.log.entries())
+      console.log(`- ${o}: ${(e / m).toFixed(2)} MB`);
     this.maxLabel && console.log(`Peak Memory Usage: ${(this.maxMemory / m).toFixed(2)} MB at "${this.maxLabel}"`), console.log(`Overall Peak Memory Usage: ${(this.peakMemory / m).toFixed(2)} MB`);
   }
 }
 export {
-  M as default
+  l as default
 };

package/dist/utilities/safetensors.d.ts

@@ -0,0 +1,3 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function save_safetensors(tensors: Record<string, Tensor>): Promise<ArrayBuffer>;
+export declare function load_safetensors(buffer: ArrayBuffer): Promise<Record<string, Tensor>>;

package/dist/utilities/safetensors.js

@@ -0,0 +1,83 @@
+import "../index-CamYe_M8.js";
+import { t as y } from "../tensor-CDz5x1mP.js";
+function l(t) {
+  if (t === "float32") return "F32";
+  if (t === "int32") return "I32";
+  throw new Error(`Unsupported dtype: ${t}`);
+}
+function h(t) {
+  if (t === "F32") return "float32";
+  if (t === "I32") return "int32";
+  throw new Error(`Unsupported dtype: ${t}`);
+}
+async function _(t) {
+  const c = {};
+  let a = 0;
+  for (const [n, s] of Object.entries(t))
+    c[n] = {
+      dtype: l(s.dtype),
+      shape: s.shape,
+      data_offsets: [a, a + s.size * 4]
+    }, a += s.size * 4;
+  const p = JSON.stringify(c);
+  let r = new TextEncoder().encode(p);
+  if (r.length % 4 !== 0) {
+    const n = 4 - r.length % 4, s = new Uint8Array(r.length + n);
+    s.set(r);
+    for (let w = r.length; w < s.length; w++)
+      s[w] = 32;
+    r = s;
+  }
+  const o = r.length, f = 8 + o + a, e = new ArrayBuffer(f);
+  new DataView(e).setUint32(0, o, !0), new Uint8Array(e, 8, o).set(r);
+  let d = 8 + o;
+  for (const n of Object.values(t)) {
+    if (n.size === 0) continue;
+    const s = await n.data();
+    if (n.dtype === "float32")
+      new Float32Array(e, d, n.size).set(s), d += n.size * 4;
+    else if (n.dtype === "int32")
+      new Int32Array(e, d, n.size).set(s), d += n.size * 4;
+    else
+      throw new Error(`Unsupported dtype: ${n.dtype}`);
+  }
+  return e;
+}
+async function U(t) {
+  const a = new DataView(t).getUint32(0, !0), p = new Uint8Array(t, 8, a), r = JSON.parse(new TextDecoder().decode(p)), o = {};
+  for (const [f, e] of Object.entries(r)) {
+    if (e.data_offsets[0] === e.data_offsets[1]) {
+      o[f] = y([], e.shape, h(e.dtype));
+      continue;
+    }
+    if (e.dtype === "F32") {
+      const i = y(
+        new Float32Array(
+          t,
+          e.data_offsets[0] + 8 + a,
+          (e.data_offsets[1] - e.data_offsets[0]) / 4
+        ),
+        e.shape,
+        h(e.dtype)
+      );
+      o[f] = i;
+    } else if (e.dtype === "I32") {
+      const i = y(
+        new Int32Array(
+          t,
+          e.data_offsets[0] + 8 + a,
+          (e.data_offsets[1] - e.data_offsets[0]) / 4
+        ),
+        e.shape,
+        h(e.dtype)
+      );
+      o[f] = i;
+    } else
+      throw new Error(`Unsupported dtype: ${e.dtype}`);
+  }
+  return o;
+}
+export {
+  U as load_safetensors,
+  _ as save_safetensors
+};

package/dist/utilities/save.js

@@ -1,43 +1,61 @@
-import { j as g } from "../jszip.min-CjP2V1VV.js";
-import { exportWeights as l } from "./weights.js";
-import p from "../tokeniser/CharTokeniser.js";
-const b = "1.0.0";
-async function h(t, a, i) {
-  const c = i?.includeLog ?? !0, f = /* @__PURE__ */ new Map();
-  t.saveWeights(f);
-  const e = new g(), r = {};
-  for (const [n, s] of f) {
-    const o = await l(s);
-    r[n] = o.spec, e.file(`${n}.bin`, o.data.buffer, { binary: !0 });
-  }
-  if (e.file(
-    "manifest.json",
-    JSON.stringify({
-      weightSpec: r,
-      config: t.config.gpt,
-      version: b,
-      application: "@genai-fi/nanogpt",
-      meta: i?.metadata,
-      name: i?.name
-    }),
+import { j as p } from "../jszip.min-CjP2V1VV.js";
+import b from "../tokeniser/CharTokeniser.js";
+import { save_safetensors as l } from "./safetensors.js";
+import { VERSION as y } from "./load.js";
+async function N(e, a, n) {
+  const f = n?.includeLog ?? !0, s = /* @__PURE__ */ new Map();
+  e.saveWeights(s);
+  const i = new p(), r = {};
+  s.forEach((t, o) => {
+    t.length === 1 && (r[o] = t[0]);
+  });
+  const g = await l(r);
+  i.file("model.safetensors", g, { binary: !0 });
+  const c = {
+    model_type: "GenAI_NanoGPT_1",
+    vocab_size: a.getVocab().length,
+    hidden_size: e.config.gpt.nEmbed,
+    num_hidden_layers: e.config.gpt.nLayer,
+    num_attention_heads: e.config.gpt.nHead,
+    block_size: e.config.gpt.blockSize,
+    dropout: e.config.gpt.dropout,
+    biasInLinear: e.config.gpt.biasInLinear,
+    biasInLayerNorm: e.config.gpt.biasInLayerNorm,
+    mlpFactor: e.config.gpt.mlpFactor,
+    useRope: e.config.gpt.useRope
+  };
+  if (i.file("config.json", JSON.stringify(c, void 0, 4), {
+    binary: !1
+  }), i.file(
+    "meta.json",
+    JSON.stringify(
+      {
+        version: y,
+        application: "@genai-fi/nanogpt",
+        meta: n?.metadata,
+        name: n?.name
+      },
+      void 0,
+      4
+    ),
     {
       binary: !1
     }
-  ), e.file(
+  ), i.file(
     "tokeniser.json",
     JSON.stringify({
-      type: a instanceof p ? "char" : "bpe",
+      type: a instanceof b ? "char" : "bpe",
       vocab: a.getVocab(),
       merges: await a.getMerges()
     }),
     {
       binary: !1
     }
-  ), c && e.file("log.json", JSON.stringify(t.log), { binary: !1 }), i?.files)
-    for (const [n, s] of Object.entries(i.files))
-      e.file(n, JSON.stringify(s), { binary: !1 });
-  return e.generateAsync({ type: "blob" });
+  ), f && i.file("log.json", JSON.stringify(e.log), { binary: !1 }), n?.files)
+    for (const [t, o] of Object.entries(n.files))
+      i.file(t, JSON.stringify(o), { binary: !1 });
+  return i.generateAsync({ type: "blob" });
 }
 export {
-  h as saveModel
+  N as saveModel
 };

package/dist/utilities/weights.js

@@ -1,5 +1,5 @@
-import "../index-Du-bmOP8.js";
-import { t as p } from "../tensor-DgTOPY6h.js";
+import "../index-CamYe_M8.js";
+import { t as p } from "../tensor-CDz5x1mP.js";
 function h(n) {
   const e = n.reduce((s, o) => s + o.length, 0), a = new Float32Array(e);
   let t = 0;

package/dist/{variable-DZ3fF0R2.js → variable-CLVXjN7F.js}

@@ -1,4 +1,4 @@
-import { E as i } from "./index-Du-bmOP8.js";
+import { E as i } from "./index-CamYe_M8.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.