@genai-fi/nanogpt 0.7.2 → 0.8.0

package/dist/loader/oldZipLoad.js

@@ -1,30 +1,14 @@
-import d from "../NanoGPTModel.js";
-import "../jszip.min-CjP2V1VV.js";
-import h from "../tokeniser/CharTokeniser.js";
-import { a0 as k } from "../index-BoWRt-10.js";
-import b from "../tokeniser/bpe.js";
-import { dummyPassAsync as u } from "../utilities/dummy.js";
-import "../Generator.js";
-import "../index-Dwqa6Zy2.js";
-import "../dataset-CtdBYwjo.js";
-import "../index-Tf7vU29b.js";
-import "../ops/cpu/adamAdjust.js";
-import "../ops/webgl/adamAdjust.js";
-import "../ops/cpu/adamMoments.js";
-import "../ops/webgl/adamMoments.js";
-import "../papaparse.min-C8l2Kvo1.js";
-import "../ops/cpu/scatterSub.js";
-import "../ops/webgl/scatterSub.js";
-import "../ops/cpu/gatherSub.js";
-import "../ops/webgl/gatherSub.js";
+import { a1 as y } from "../index-CUQrfsw_.js";
 import "../ops/cpu/attentionMask.js";
 import "../ops/webgl/attentionMask.js";
 import "../ops/grads/attentionMask.js";
 import "../ops/cpu/qkv.js";
 import "../ops/webgl/qkv.js";
 import "../ops/grads/qkv.js";
-import "../random_width-sZORGo5k.js";
-import "../register_all_kernels-BwDSRN-f.js";
+import "../random_width-D8Pwy_na.js";
+import "../register_all_kernels-DUshvVWP.js";
+import "../index-Tf7vU29b.js";
+import "../dataset-CJmEGu6D.js";
 import "../ops/cpu/rope.js";
 import "../ops/webgl/rope.js";
 import "../ops/grads/rope.js";
@@ -36,45 +20,52 @@ import "../ops/grads/fusedSoftmax.js";
 import "../ops/cpu/matMulGelu.js";
 import "../ops/webgl/matMulGelu.js";
 import "../ops/grads/matMulGelu.js";
-import "../ops/cpu/gelu.js";
-import "../ops/webgl/gelu.js";
-import "../gelu-C-dPj6Ku.js";
 import "../ops/cpu/normRMS.js";
 import "../ops/webgl/normRMS.js";
 import "../ops/grads/normRMS.js";
+import "../ops/cpu/gatherSub.js";
+import "../ops/webgl/gatherSub.js";
+import "../ops/cpu/scatterSub.js";
+import "../ops/webgl/scatterSub.js";
+import "../jszip.min-CjP2V1VV.js";
+import h from "../tokeniser/CharTokeniser.js";
+import k from "../tokeniser/bpe.js";
+import { dummyPassAsync as g } from "../utilities/dummy.js";
+import b from "../models/factory.js";
+import "../Generator.js";
+import "../index-Dwqa6Zy2.js";
+import "../ops/cpu/adamAdjust.js";
+import "../ops/webgl/adamAdjust.js";
+import "../ops/cpu/adamMoments.js";
+import "../ops/webgl/adamMoments.js";
+import "../papaparse.min-C8l2Kvo1.js";
+import "../ops/cpu/gelu.js";
+import "../ops/webgl/gelu.js";
+import "../gelu-Bd3UBBxg.js";
 import "../ops/webgl/log.js";
-import { importWeights as O } from "../utilities/weights.js";
-async function yt(o) {
-  const n = /* @__PURE__ */ new Map(), s = await o.file("manifest.json")?.async("string");
-  if (!s)
+import { importWeights as u } from "../utilities/weights.js";
+async function dt(r) {
+  const e = /* @__PURE__ */ new Map(), a = await r.file("manifest.json")?.async("string");
+  if (!a)
     throw new Error("Manifest file not found in the zip archive");
-  const m = JSON.parse(s);
-  for (const [t, r] of Object.entries(m.weightSpec))
-    n.set(t, { spec: r, data: new Float32Array() });
-  const p = await o.file("tokeniser.json")?.async("string");
+  const m = JSON.parse(a);
+  for (const [t, o] of Object.entries(m.weightSpec))
+    e.set(t, { spec: o, data: new Float32Array() });
+  const p = await r.file("tokeniser.json")?.async("string");
   if (!p)
     throw new Error("Tokeniser file not found in the zip archive");
-  const i = JSON.parse(p), l = (i.type ?? "char") === "char" ? new h(i.vocab) : new b(i.vocab, i.merges), c = /* @__PURE__ */ new Map();
-  for (const t of Object.keys(o.files))
+  const i = JSON.parse(p), f = (i.type ?? "char") === "char" ? new h(i.vocab) : new k(i.vocab, i.merges), c = /* @__PURE__ */ new Map();
+  for (const t of Object.keys(r.files))
     if (t.endsWith(".bin")) {
-      const r = t.replace(".bin", ""), w = await o.file(t).async("arraybuffer"), g = new Float32Array(w), a = n.get(r) || { spec: [], data: new Float32Array() };
-      a.data = g, n.set(r, a);
-      const y = await O(a);
-      c.set(r, y);
-    }
-  k();
-  const e = new d(m.config);
-  await u(e), e.loadWeights(c);
-  const f = await o.file("log.json")?.async("string");
-  if (f)
-    try {
-      const t = JSON.parse(f);
-      e.log = t;
-    } catch (t) {
-      throw console.error("Error parsing training log:", t), new Error(`Failed to parse training log: ${t}`);
+      const o = t.replace(".bin", ""), l = await r.file(t).async("arraybuffer"), w = new Float32Array(l), s = e.get(o) || { spec: [], data: new Float32Array() };
+      s.data = w, e.set(o, s);
+      const d = await u(s);
+      c.set(o, d);
     }
-  return { model: e, tokeniser: l };
+  y();
+  const n = b(m.config);
+  return await g(n), n.loadWeights(c), { model: n, tokeniser: f };
 }
 export {
-  yt as default
+  dt as default
 };

package/dist/loader/save.d.ts

@@ -0,0 +1,8 @@
+import { ITokeniser } from '../tokeniser/type';
+import { default as Model, ModelForwardAttributes } from '../models/model';
+export interface SaveOptions {
+    name?: string;
+    metadata?: Record<string, unknown>;
+    files?: Record<string, unknown>;
+}
+export declare function saveModel(model: Model<ModelForwardAttributes>, tokeniser: ITokeniser, options?: SaveOptions): Promise<Blob>;

package/dist/loader/save.js

@@ -0,0 +1,62 @@
+import { j as g } from "../jszip.min-CjP2V1VV.js";
+import b from "../tokeniser/CharTokeniser.js";
+import { save_safetensors as p } from "../utilities/safetensors.js";
+import { VERSION as y } from "./load.js";
+async function v(e, n, a) {
+  const o = /* @__PURE__ */ new Map();
+  e.saveWeights(o);
+  const i = new g(), r = {};
+  o.forEach((t, s) => {
+    t.length === 1 && (r[s] = t[0]);
+  });
+  const f = await p(r);
+  i.file("model.safetensors", f, { binary: !0 });
+  const c = {
+    model_type: e.config.modelType || e.getClassName(),
+    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,
+    dropout: e.config.dropout,
+    biasInLinear: e.config.biasInLinear,
+    biasInLayerNorm: e.config.biasInLayerNorm,
+    mlpFactor: e.config.mlpFactor,
+    useRope: e.config.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: a?.metadata,
+        name: a?.name,
+        training: e.trainingState || void 0
+      },
+      void 0,
+      4
+    ),
+    {
+      binary: !1
+    }
+  ), i.file(
+    "tokeniser.json",
+    JSON.stringify({
+      type: n instanceof b ? "char" : "bpe",
+      vocab: n.getVocab(),
+      merges: await n.getMerges()
+    }),
+    {
+      binary: !1
+    }
+  ), a?.files)
+    for (const [t, s] of Object.entries(a.files))
+      i.file(t, JSON.stringify(s), { binary: !1 });
+  return i.generateAsync({ type: "blob" });
+}
+export {
+  v as saveModel
+};

package/dist/{log_sum_exp-DbjkV734.js → log_sum_exp-C142qZqY.js}

@@ -1,8 +1,8 @@
-import { B as r, C as p, E as u, ae as E, af as h, p as S, c as $, o as d } from "./index-BoWRt-10.js";
-import { e as K } from "./axis_util-BzbKo31C.js";
-import { m as T } from "./max-Ddnnb5xe.js";
-import { r as m } from "./reshape-CdBq1WJ6.js";
-import { s as _ } from "./sum-66ew2byf.js";
+import { B as r, C as p, E as u, ag as h, ah as E, p as S, c as $, q as d } from "./index-CUQrfsw_.js";
+import { e as K } from "./axis_util-DubwyOhW.js";
+import { m as T } from "./max-B3JOcNGb.js";
+import { r as m } from "./reshape-DEfQGSin.js";
+import { s as _ } from "./sum-l_0SqM4h.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -21,7 +21,7 @@ import { s as _ } from "./sum-66ew2byf.js";
  */
 function b(s) {
   const o = { x: p(s, "x", "exp") };
-  return u.runKernel(E, o);
+  return u.runKernel(h, o);
 }
 const N = /* @__PURE__ */ r({ exp_: b });
 /**
@@ -40,11 +40,11 @@ const N = /* @__PURE__ */ r({ exp_: b });
  * limitations under the License.
  * =============================================================================
  */
-function v(s) {
+function q(s) {
   const o = { x: p(s, "x", "log", "float32") };
-  return u.runKernel(h, o);
+  return u.runKernel(E, o);
 }
-const w = /* @__PURE__ */ r({ log_: v });
+const v = /* @__PURE__ */ r({ log_: q });
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
@@ -61,22 +61,22 @@ const w = /* @__PURE__ */ r({ log_: v });
  * limitations under the License.
  * =============================================================================
  */
-function A(s, n = null, o = !1) {
+function w(s, n = null, o = !1) {
   const a = p(s, "x", "logSumExp"), t = S(n, a.shape), x = T(
     a,
     t,
     !0
     /* keepDims */
-  ), i = $(a, x), l = N(i), f = _(l, t), c = w(f), e = d(m(x, c.shape), c);
+  ), i = $(a, x), l = N(i), f = _(l, t), c = v(f), e = d(m(x, c.shape), c);
   if (o) {
     const g = K(e.shape, t);
     return m(e, g);
   }
   return e;
 }
-const L = /* @__PURE__ */ r({ logSumExp_: A });
+const I = /* @__PURE__ */ r({ logSumExp_: w });
 export {
-  w as a,
+  v as a,
   N as e,
-  L as l
+  I as l
 };

package/dist/main.d.ts

@@ -2,18 +2,19 @@ import { default as CausalSelfAttention } from './layers/CausalSelfAttention';
 import { default as MLP } from './layers/MLP';
 import { default as TransformerBlock } from './layers/TransformerBlock';
 import { default as RoPECache } from './layers/RoPECache';
-export { default as NanoGPT } from './NanoGPTModel';
+export { default as NanoGPT } from './models/NanoGPTV1';
 export { default as TeachableLLM } from './TeachableLLM';
 export { default as CharTokeniser } from './tokeniser/CharTokeniser';
 export { default as BPETokeniser } from './tokeniser/bpe';
 export { default as waitForModel } from './utilities/waitForModel';
 export { default as loadTextData } from './data/textLoader';
+export { default as Generator } from './Generator';
 export type { ITrainerOptions } from './Trainer';
 export type { IGenerateOptions } from './Generator';
-export type { TrainingLogEntry } from './NanoGPTModel';
+export { type ModelForwardAttributes, default as Model } from './models/model';
 export type { ITokeniser } from './tokeniser/type';
-export type { TrainingProgress } from './training/Trainer';
-export type { GPTConfig } from './config';
+export type { TrainingProgress, TrainingLogEntry } from './training/Trainer';
+export type { GPTConfig } from './models/config';
 export { estimateParameterCount, estimateMemoryUsage, estimateTrainingMemoryUsage, estimateResources, validateConfig, } from './utilities/parameters';
 export { selectBackend } from './backend';
 export { default as performanceTest } from './utilities/performance';

package/dist/main.js

@@ -1,11 +1,13 @@
-import { default as W } from "./NanoGPTModel.js";
+import { default as W } from "./models/NanoGPTV1.js";
 import { default as Y } from "./TeachableLLM.js";
 import { default as _ } from "./tokeniser/CharTokeniser.js";
 import { default as oo } from "./tokeniser/bpe.js";
 import { default as to } from "./utilities/waitForModel.js";
 import { default as eo } from "./data/textLoader.js";
-import { estimateMemoryUsage as po, estimateParameterCount as ao, estimateResources as fo, estimateTrainingMemoryUsage as so, validateConfig as lo } from "./utilities/parameters.js";
-import "./index-BoWRt-10.js";
+import { default as io } from "./Generator.js";
+import { default as fo } from "./models/model.js";
+import { estimateMemoryUsage as lo, estimateParameterCount as xo, estimateResources as no, estimateTrainingMemoryUsage as uo, validateConfig as co } from "./utilities/parameters.js";
+import "./index-CUQrfsw_.js";
 import "./ops/cpu/scatterSub.js";
 import "./ops/webgl/scatterSub.js";
 import "./ops/cpu/gatherSub.js";
@@ -16,10 +18,10 @@ import "./ops/grads/attentionMask.js";
 import "./ops/cpu/qkv.js";
 import "./ops/webgl/qkv.js";
 import "./ops/grads/qkv.js";
-import "./random_width-sZORGo5k.js";
-import "./register_all_kernels-BwDSRN-f.js";
+import "./random_width-D8Pwy_na.js";
+import "./register_all_kernels-DUshvVWP.js";
 import "./index-Tf7vU29b.js";
-import "./dataset-CtdBYwjo.js";
+import "./dataset-CJmEGu6D.js";
 import "./ops/cpu/rope.js";
 import "./ops/webgl/rope.js";
 import "./ops/grads/rope.js";
@@ -33,7 +35,7 @@ import "./ops/webgl/matMulGelu.js";
 import "./ops/grads/matMulGelu.js";
 import "./ops/cpu/gelu.js";
 import "./ops/webgl/gelu.js";
-import "./gelu-C-dPj6Ku.js";
+import "./gelu-Bd3UBBxg.js";
 import "./ops/cpu/normRMS.js";
 import "./ops/webgl/normRMS.js";
 import "./ops/grads/normRMS.js";
@@ -42,13 +44,13 @@ import "./ops/cpu/adamMoments.js";
 import "./ops/webgl/adamMoments.js";
 import "./ops/cpu/adamAdjust.js";
 import "./ops/webgl/adamAdjust.js";
-import { selectBackend as xo } from "./backend.js";
-import { default as co } from "./utilities/performance.js";
+import { selectBackend as Mo } from "./backend.js";
+import { default as Co } from "./utilities/performance.js";
 import o from "./layers/CausalSelfAttention.js";
 import r from "./layers/MLP.js";
 import t from "./layers/TransformerBlock.js";
 import m from "./layers/RoPECache.js";
-import { default as ko } from "./training/AdamExt.js";
+import { default as go } from "./training/AdamExt.js";
 const O = {
   CausalSelfAttention: o,
   MLP: r,
@@ -56,19 +58,21 @@ const O = {
   RoPECache: m
 };
 export {
-  ko as AdamExt,
+  go as AdamExt,
   oo as BPETokeniser,
   _ as CharTokeniser,
+  io as Generator,
+  fo as Model,
   W as NanoGPT,
   Y as TeachableLLM,
-  po as estimateMemoryUsage,
-  ao as estimateParameterCount,
-  fo as estimateResources,
-  so as estimateTrainingMemoryUsage,
+  lo as estimateMemoryUsage,
+  xo as estimateParameterCount,
+  no as estimateResources,
+  uo as estimateTrainingMemoryUsage,
   O as layers,
   eo as loadTextData,
-  co as performanceTest,
-  xo as selectBackend,
-  lo as validateConfig,
+  Co as performanceTest,
+  Mo as selectBackend,
+  co as validateConfig,
   to as waitForModel
 };

package/dist/{mat_mul-8m8pfdcx.js → mat_mul-DMkduNJu.js}

@@ -1,4 +1,4 @@
-import { B as m, C as s, I as c, E as M, J as p } from "./index-BoWRt-10.js";
+import { B as m, C as s, N as c, E as M, Q as p } from "./index-CUQrfsw_.js";
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.

package/dist/{max-Ddnnb5xe.js → max-B3JOcNGb.js}

@@ -1,4 +1,4 @@
-import { B as r, C as e, E as x, U as c } from "./index-BoWRt-10.js";
+import { B as r, C as e, E as x, X as c } from "./index-CUQrfsw_.js";
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.

package/dist/mod-uUuj4gSb.js

@@ -0,0 +1,27 @@
+import { B as r, C as s, N as m, E as c, Y as d } from "./index-CUQrfsw_.js";
+/**
+ * @license
+ * Copyright 2020 Google LLC. All Rights Reserved.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ * =============================================================================
+ */
+function p(t, e) {
+  let o = s(t, "a", "mod"), a = s(e, "b", "mod");
+  [o, a] = m(o, a);
+  const n = { a: o, b: a };
+  return c.runKernel(d, n);
+}
+const b = /* @__PURE__ */ r({ mod_: p });
+export {
+  b as m
+};

package/dist/models/NanoGPTV1.d.ts

@@ -0,0 +1,15 @@
+import { GPTConfig } from './config';
+import { Tensor } from '@tensorflow/tfjs-core';
+import { default as Model, ModelForwardAttributes } from './model';
+export default class NanoGPT extends Model<ModelForwardAttributes> {
+    private wte;
+    private wpe?;
+    private blocks;
+    private lnF;
+    private ropeCache?;
+    constructor(config?: Partial<GPTConfig>);
+    getClassName(): string;
+    private inputPhase;
+    forward(attrs: ModelForwardAttributes, idx: Tensor, targets?: Tensor): Tensor[];
+    dispose(): void;
+}

package/dist/models/NanoGPTV1.js

@@ -0,0 +1,71 @@
+import { defaultConfig as m } from "./config.js";
+import f from "../layers/TransformerBlock.js";
+import u from "../layers/TiedEmbedding.js";
+import g from "../layers/RoPECache.js";
+import b from "../layers/RMSNorm.js";
+import { t as l, k as p } from "../index-CUQrfsw_.js";
+import w from "./model.js";
+import k from "../layers/PositionEmbedding.js";
+class R extends w {
+  wte;
+  // Token embeddings
+  wpe;
+  // Position embeddings
+  // private drop: layers.Layer; // Dropout
+  blocks;
+  lnF;
+  // Final layer norm
+  ropeCache;
+  constructor(e = {}) {
+    super({ ...m, ...e }), this.wte = new u(this.config, "token_embedding", this), this.config.useRope === !1 ? this.wpe = new k(this.config, "positional_embedding", this) : this.ropeCache = new g(this.config), this.blocks = [];
+    for (let i = 0; i < this.config.nLayer; i++)
+      this.blocks.push(new f(i, this.config, this));
+    this.lnF = new b(this.config, "final_rms_norm", this);
+  }
+  getClassName() {
+    return "GenAI_NanoGPT_v1";
+  }
+  inputPhase(e, i) {
+    return l(() => {
+      const s = this.wte.embed(e);
+      if (this.config.useRope === !1) {
+        const o = this.wpe.call(i, s);
+        if (Array.isArray(o))
+          throw new Error("PositionEmbedding output should not be an array");
+        return o;
+      }
+      return s;
+    });
+  }
+  forward(e, i, s) {
+    return this.validateInput(i), e.ropeCache = this.ropeCache, e.outputEmbeddings && (e.embeddings = []), l(() => {
+      this.startMemory();
+      let o = this.inputPhase(i, e);
+      if (e.cache && e.cache.length !== this.blocks.length)
+        throw console.error("Cache", e.cache), new Error(
+          `Cache length ${e.cache.length} does not match number of blocks ${this.blocks.length}`
+        );
+      for (let t = 0; t < this.blocks.length; t++) {
+        const c = this.blocks[t], d = Math.random() * 1e9, r = {
+          ropeCache: e.ropeCache,
+          training: e.training,
+          seed: d,
+          attentionScores: e.attentionScores,
+          pastKV: e.cache ? e.cache[t] : void 0
+        }, a = e.checkpointing && e.training ? c.callCheckpoint(r, o) : c.call(r, o);
+        e.outputEmbeddings ? (p(o), e.embeddings.push(o)) : o.dispose(), o = a;
+      }
+      o = this.lnF.call(e, o);
+      const n = this.wte.project(o);
+      e.outputEmbeddings ? (p(o), e.embeddings.push(o)) : o.dispose();
+      let h;
+      return s && (h = this.calculateLoss(n, s)), this.endMemory("Forward"), h ? [n, h] : [n];
+    });
+  }
+  dispose() {
+    this.wte.dispose(), this.wpe && this.wpe.dispose(), this.blocks.forEach((e) => e.dispose()), this.lnF.dispose();
+  }
+}
+export {
+  R as default
+};

package/dist/{config.d.ts → models/config.d.ts}

@@ -1,4 +1,5 @@
 export interface GPTConfig {
+    modelType?: string;
     vocabSize: number;
     blockSize: number;
     nLayer: number;

package/dist/{config.js → models/config.js}

@@ -1,4 +1,5 @@
 const e = {
+  modelType: "GenAI_NanoGPT_v1",
   vocabSize: 2e3,
   blockSize: 128,
   // Maximum sequence length

package/dist/models/factory.d.ts

@@ -0,0 +1,3 @@
+import { GPTConfig } from './config';
+import { default as Model, ModelForwardAttributes } from './model';
+export default function createModelInstance(config: GPTConfig): Model<ModelForwardAttributes>;

package/dist/models/factory.js

@@ -0,0 +1,14 @@
+import o from "./NanoGPTV1.js";
+function a(e) {
+  const n = e.modelType || "GenAI_NanoGPT_v1";
+  switch (n) {
+    case "GenAI_NanoGPT_1":
+    case "GenAI_NanoGPT_v1":
+      return new o(e);
+    default:
+      throw new Error(`Unsupported model type: ${n}`);
+  }
+}
+export {
+  a as default
+};

package/dist/models/model.d.ts

@@ -0,0 +1,26 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+import { ForwardAttributes, default as BaseLayer } from '../layers/BaseLayer';
+import { AttentionScores, KVCache } from '../layers/CausalSelfAttention';
+export interface ModelForwardAttributes extends ForwardAttributes {
+    cache?: KVCache[];
+    attentionScores?: AttentionScores;
+    seed?: number;
+    outputEmbeddings?: boolean;
+    embeddings?: Tensor[];
+}
+interface TrainingState {
+    steps: number;
+    learningRate: number;
+    batchSize: number;
+    loss: number;
+}
+export default abstract class Model<T extends ModelForwardAttributes> extends BaseLayer<T> {
+    trainingState: TrainingState | null;
+    abstract getClassName(): string;
+    abstract forward(attrs: T, idx: Tensor, targets?: Tensor): Tensor[];
+    abstract dispose(): void;
+    getNumParams(): number;
+    protected validateInput(idx: Tensor): void;
+    protected calculateLoss(logits: Tensor, targets: Tensor): Tensor;
+}
+export {};

package/dist/models/model.js

@@ -0,0 +1,68 @@
+import i from "../layers/BaseLayer.js";
+import "../index-CUQrfsw_.js";
+import "../ops/cpu/attentionMask.js";
+import "../ops/webgl/attentionMask.js";
+import "../ops/grads/attentionMask.js";
+import "../ops/cpu/qkv.js";
+import "../ops/webgl/qkv.js";
+import "../ops/grads/qkv.js";
+import "../random_width-D8Pwy_na.js";
+import "../register_all_kernels-DUshvVWP.js";
+import "../index-Tf7vU29b.js";
+import "../dataset-CJmEGu6D.js";
+import "../ops/cpu/rope.js";
+import "../ops/webgl/rope.js";
+import "../ops/grads/rope.js";
+import "../ops/cpu/appendCache.js";
+import "../ops/webgl/appendCache.js";
+import "../ops/cpu/fusedSoftmax.js";
+import "../ops/webgl/fusedSoftmax.js";
+import "../ops/grads/fusedSoftmax.js";
+import "../ops/cpu/matMulGelu.js";
+import "../ops/webgl/matMulGelu.js";
+import "../ops/grads/matMulGelu.js";
+import "../ops/cpu/normRMS.js";
+import "../ops/webgl/normRMS.js";
+import "../ops/grads/normRMS.js";
+import "../jszip.min-CjP2V1VV.js";
+import "../index-Dwqa6Zy2.js";
+import "../Generator.js";
+import "../ops/cpu/adamAdjust.js";
+import "../ops/webgl/adamAdjust.js";
+import "../ops/cpu/adamMoments.js";
+import "../ops/webgl/adamMoments.js";
+import "../papaparse.min-C8l2Kvo1.js";
+import { estimateParameterCount as p } from "../utilities/parameters.js";
+import "../ops/cpu/scatterSub.js";
+import "../ops/webgl/scatterSub.js";
+import "../ops/cpu/gatherSub.js";
+import "../ops/webgl/gatherSub.js";
+import "../ops/cpu/gelu.js";
+import "../ops/webgl/gelu.js";
+import "../gelu-Bd3UBBxg.js";
+import "../ops/webgl/log.js";
+import { createSoftmaxCrossEntropyWithGrad as m } from "../training/sparseCrossEntropy.js";
+class Y extends i {
+  trainingState = null;
+  getNumParams() {
+    return p(this.config);
+  }
+  validateInput(t) {
+    if (t.shape.length !== 2)
+      throw new Error(`Invalid input shape: expected [batch_size, sequence_length], got ${t.shape}`);
+    if (t.shape[1] > this.config.blockSize)
+      throw new Error(`Input sequence length ${t.shape[1]} isn't block size ${this.config.blockSize}`);
+    if (t.dtype !== "int32")
+      throw new Error(`Input tensor must be of type int32, got ${t.dtype}`);
+  }
+  calculateLoss(t, o) {
+    try {
+      return m()(t, o).mean();
+    } catch (r) {
+      throw console.error("Error computing loss:", r), new Error(`Loss computation failed: ${r}`);
+    }
+  }
+}
+export {
+  Y as default
+};

package/dist/{mulmat_packed_gpu-VSekgsNv.js → mulmat_packed_gpu-Cm2gw-c8.js}

@@ -1,4 +1,4 @@
-import { u as z } from "./gpgpu_math-DGNLNL4I.js";
+import { u as z } from "./gpgpu_math-TFLxaLkw.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.

package/dist/{ones-Dj0SDhHf.js → ones-ZdgQGBCP.js}

@@ -1,5 +1,5 @@
-import { F as n, H as t, j as m, E as i } from "./index-BoWRt-10.js";
-import { z as c, c as f } from "./zeros--BdLQ3oG.js";
+import { L as n, M as t, j as m, E as i } from "./index-CUQrfsw_.js";
+import { z as c, c as f } from "./zeros-VZ72lWXM.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.

package/dist/ops/adamAdjust.js

@@ -1,4 +1,4 @@
-import { e as i } from "../index-BoWRt-10.js";
+import { e as i } from "../index-CUQrfsw_.js";
 import "./cpu/adamAdjust.js";
 import "./webgl/adamAdjust.js";
 function p(r, t, e, n, m, o) {