@genai-fi/nanogpt 0.19.0 → 0.20.0

package/dist/loader/oldZipLoad.js

@@ -1,80 +0,0 @@
-import { aa as y } from "../index-CUXkjxiT.js";
-import "../random_width-DN5ZtQkM.js";
-import "../zeros_like-CWjDdwr-.js";
-import "../Generator.js";
-import "../index-Cp39cXWe.js";
-import "../dataset-CGGp1z9P.js";
-import "../ops/cpu/attentionMask.js";
-import "../ops/webgl/attentionMask.js";
-import "../ops/grads/attentionMask.js";
-import "../ops/cpu/rope.js";
-import "../ops/webgl/rope.js";
-import "../rope-DcrZM_e6.js";
-import "../ops/cpu/appendCache.js";
-import "../ops/webgl/appendCache.js";
-import "../ops/grads/softmax16.js";
-import "../matMul16-BcVC_E62.js";
-import "../ops/webgl/matMul16.js";
-import "../ops/cpu/matMul16.js";
-import "../pack16-bqltoUlR.js";
-import "../ops/transpose16.js";
-import "../ops/reshape16.js";
-import "../ops/cpu/qkv.js";
-import "../ops/webgl/qkv.js";
-import "../ops/grads/qkv.js";
-import "../ops/cpu/normRMS.js";
-import "../ops/webgl/normRMS.js";
-import "../ops/grads/normRMS.js";
-import "../ops/dropout16.js";
-import "../ops/webgl/dropout16.js";
-import "../ops/grads/add16.js";
-import "../jszip.min-BZhlzntC.js";
-import g from "../tokeniser/CharTokeniser.js";
-import k from "../tokeniser/bpe.js";
-import { dummyPassAsync as u } from "../utilities/dummy.js";
-import b from "../models/factory.js";
-import "../ops/cpu/adamAdjust.js";
-import "../ops/webgl/adamAdjust.js";
-import "../ops/cpu/adamMoments.js";
-import "../ops/webgl/adamMoments.js";
-import "../index-DvYrXKkX.js";
-import "../ops/cpu/gatherSub.js";
-import "../ops/webgl/gatherSub.js";
-import "../ops/cpu/scatterSub.js";
-import "../ops/webgl/scatterSub.js";
-import "../papaparse.min-C0cScC2i.js";
-import "../ops/cpu/matMulGelu.js";
-import "../matMulGelu-JNLZqKQp.js";
-import "../ops/grads/matMulGelu.js";
-import "../ops/cpu/gelu.js";
-import "../ops/webgl/gelu.js";
-import "../gelu-B220X1Go.js";
-import "../ops/webgl/log.js";
-import "../checks/normRMS.js";
-import "../checks/normRMSGrad.js";
-import { importWeights as M } from "../utilities/weights.js";
-async function jt(r, m) {
-  const p = /* @__PURE__ */ new Map(), s = await r.file("manifest.json")?.async("string");
-  if (!s)
-    throw new Error("Manifest file not found in the zip archive");
-  const a = JSON.parse(s);
-  for (const [t, o] of Object.entries(a.weightSpec))
-    p.set(t, { spec: o, data: new Float32Array() });
-  const c = await r.file("tokeniser.json")?.async("string");
-  if (!c)
-    throw new Error("Tokeniser file not found in the zip archive");
-  const i = JSON.parse(c), l = (i.type ?? "char") === "char" ? new g(i.vocab) : new k(i.vocab, i.merges), f = /* @__PURE__ */ new Map();
-  for (const t of Object.keys(r.files))
-    if (t.endsWith(".bin")) {
-      const o = t.replace(".bin", ""), w = await r.file(t).async("arraybuffer"), d = new Float32Array(w), n = p.get(o) || { spec: [], data: new Float32Array() };
-      n.data = d, p.set(o, n);
-      const h = await M(n);
-      f.set(o, h);
-    }
-  y();
-  const e = b(a.config);
-  return e.metaData = m, await u(e), e.weightStore.loadWeights(f, !!m.url), { model: e, tokeniser: l, metaData: m };
-}
-export {
-  jt as default
-};

package/dist/loader/save.d.ts

@@ -1,16 +0,0 @@
-import { ITokeniser } from '../tokeniser/type';
-import { default as Model, ModelForwardAttributes } from '../models/model';
-import { AdamWOptimizer } from '../training/AdamW';
-import { TrainingLogEntry } from '../training/types';
-import { GPTConfig } from '../models/config';
-export interface SaveOptions {
-    name?: string;
-    metadata?: Record<string, unknown>;
-    files?: Record<string, unknown>;
-    includeOptimizer?: boolean;
-}
-export interface ExtraSaveItems {
-    optimizer?: AdamWOptimizer;
-    trainingLog?: TrainingLogEntry[];
-}
-export declare function saveModel(model: Model<ModelForwardAttributes, GPTConfig>, tokeniser: ITokeniser, options?: SaveOptions, extraItems?: ExtraSaveItems): Promise<Blob>;

package/dist/loader/save.js

@@ -1,90 +0,0 @@
-import { z as y } from "../jszip.min-BZhlzntC.js";
-import _ from "../tokeniser/CharTokeniser.js";
-import { save_safetensors as b } from "../utilities/safetensors.js";
-import { VERSION as h } from "./load.js";
-function m(i) {
-  if (i.length > 1e3) {
-    const n = Math.ceil(i.length / 1e3);
-    return i.filter((o, a) => a % n === 0 || a === i.length - 1);
-  }
-  return i;
-}
-async function u(i, n, o, a) {
-  const s = /* @__PURE__ */ new Map();
-  i.weightStore.saveWeights(s);
-  const e = new y();
-  if (a?.optimizer) {
-    const t = await a.optimizer.saveMoments();
-    e.file("optimizer.safetensors", t, { binary: !0 }), e.file("optimizer_config.json", JSON.stringify(a.optimizer.serializeConfig()), {
-      binary: !1
-    });
-  }
-  a?.trainingLog && e.file("training_log.json", JSON.stringify(m(a.trainingLog), void 0, 4), {
-    binary: !1
-  });
-  const g = {};
-  s.forEach((t, f) => {
-    t.length === 1 && (g[f] = t[0]);
-  });
-  const c = await b(g);
-  e.file("model.safetensors", c, { binary: !0 });
-  const p = i.config.modelType;
-  let r;
-  p === "GenAI_NanoGPT_v1" ? r = {
-    model_type: "GenAI_NanoGPT_v1",
-    vocab_size: n.getVocab().length,
-    hidden_size: i.config.nEmbed,
-    num_hidden_layers: i.config.nLayer,
-    num_attention_heads: i.config.nHead,
-    block_size: i.config.blockSize,
-    mlpFactor: i.config.mlpFactor,
-    useRope: i.config.useRope
-  } : r = {
-    model_type: "GenAI_NanoGPT_v2",
-    vocab_size: n.getVocab().length,
-    hidden_size: i.config.nEmbed,
-    num_hidden_layers: i.config.nLayer,
-    num_attention_heads: i.config.nHead,
-    block_size: i.config.blockSize,
-    mlpFactor: i.config.mlpFactor,
-    loraConfig: i.config.loraConfig ? Object.fromEntries(i.config.loraConfig) : void 0,
-    loraName: i.config.loraName,
-    windowSize: i.config.windowSize
-  }, e.file("config.json", JSON.stringify(r, void 0, 4), {
-    binary: !1
-  });
-  const l = {
-    version: h,
-    application: "@genai-fi/nanogpt",
-    meta: o?.metadata,
-    name: o?.name,
-    training: i.metaData?.training || void 0,
-    reference: i.metaData?.url || void 0,
-    phase: i.metaData?.phase || void 0,
-    pretrainingData: i.metaData?.pretrainingData || void 0,
-    pretrainingSettings: i.metaData?.pretrainingSettings || void 0,
-    generationSettings: i.metaData?.generationSettings || void 0,
-    actionLog: i.metaData?.actionLog || void 0
-  };
-  if (e.file("meta.json", JSON.stringify(l, void 0, 4), {
-    binary: !1
-  }), e.file(
-    "tokeniser.json",
-    JSON.stringify({
-      type: n instanceof _ ? "char" : "bpe",
-      vocab: n.getVocab(),
-      merges: n.getMerges(),
-      datasetID: n.datasetID,
-      id: n.id
-    }),
-    {
-      binary: !1
-    }
-  ), o?.files)
-    for (const [t, f] of Object.entries(o.files))
-      e.file(t, JSON.stringify(f), { binary: !1 });
-  return e.generateAsync({ type: "blob" });
-}
-export {
-  u as saveModel
-};

package/dist/loader/types.d.ts

@@ -1,67 +0,0 @@
-import { GenerateOptions } from '../inference/types';
-import { LoRAConfig } from '../models/config';
-import { default as Model, ModelForwardAttributes, TrainingState } from '../models/model';
-import { ITokeniser } from '../tokeniser/type';
-import { AdamWOptimizer } from '../training/AdamW';
-import { TrainingLogEntry, TrainingOptions } from '../training/types';
-export interface TransformersConfigBase {
-    model_type: 'GenAI_NanoGPT_v1' | 'GenAI_NanoGPT_v2';
-    vocab_size: number;
-    hidden_size: number;
-    num_hidden_layers: number;
-    num_attention_heads: number;
-    block_size: number;
-    mlpFactor: number;
-    loraConfig?: Record<string, LoRAConfig>;
-    loraName?: string;
-}
-export interface TransformersConfigV1 extends TransformersConfigBase {
-    model_type: 'GenAI_NanoGPT_v1';
-    useRope: boolean;
-}
-export interface TransformersConfigV2 extends TransformersConfigBase {
-    model_type: 'GenAI_NanoGPT_v2';
-    windowSize?: string;
-}
-export type TransformersConfig = TransformersConfigV1 | TransformersConfigV2;
-export interface TransformersTokeniser {
-    type: 'char' | 'bpe';
-    vocab: string[];
-    merges: [string, string][];
-    datasetID?: string;
-    id?: string;
-}
-export type ModelPhase = 'untrained' | 'pretrained' | 'finetuned';
-export interface DatasetMetadata {
-    id: string;
-    name: string;
-}
-export interface ActionLogEntry {
-    action: 'pretrain' | 'generate' | 'finetune';
-    timestamp: number;
-    duration: number;
-    tokensProcessed: number;
-    options: TrainingOptions | GenerateOptions;
-}
-export interface TransformersMetadata {
-    name?: string;
-    version: number;
-    application: string;
-    training?: TrainingState;
-    reference?: string;
-    id?: string;
-    url?: string;
-    phase?: ModelPhase;
-    pretrainingData?: DatasetMetadata[];
-    pretrainingSettings?: TrainingOptions;
-    generationSettings?: GenerateOptions;
-    actionLog?: ActionLogEntry[];
-    [key: string]: unknown;
-}
-export interface LoadResult {
-    model: Model<ModelForwardAttributes>;
-    tokeniser: ITokeniser;
-    metaData: TransformersMetadata;
-    optimizer?: AdamWOptimizer;
-    log?: TrainingLogEntry[];
-}

package/dist/loader/types.js

@@ -1 +0,0 @@
-

package/dist/main.d.ts

@@ -1,50 +0,0 @@
-import { default as PretrainingTask } from './training/tasks/PretrainingTask';
-import { default as StartSentenceTask } from './training/tasks/StartSentenceTask';
-import { default as ConversationTask } from './training/tasks/ConversationTask';
-import { pack16 } from './ops/pack16';
-import { unpack16 } from './ops/unpack16';
-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 './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 generateDatasetID } from './utilities/datasetID';
-export { default as loadTextData } from './data/textLoader';
-export type { DatasetMetadata } from './loader/types';
-export { default as Generator, type IGenerator } from './Generator';
-export { default as Evaluator } from './training/Evaluator';
-export { default as Trainer } from './Trainer';
-export type { IGenerateOptions } from './Generator';
-export { type ModelForwardAttributes, default as Model } from './models/model';
-export type { ITokeniser, Conversation, Roles } from './tokeniser/type';
-export type { TrainingOptions, TrainingLogEntry } from './training/types';
-export type { GPTConfig } from './models/config';
-export { estimateParameterCount, estimateMemoryUsage, estimateTrainingMemoryUsage, estimateResources, validateConfig, } from './utilities/parameters';
-export { default as topP } from './utilities/topP';
-export { Task, tokensFromTasks } from './training/tasks/Task';
-export declare const tasks: {
-    PretrainingTask: typeof PretrainingTask;
-    StartSentenceTask: typeof StartSentenceTask;
-    ConversationTask: typeof ConversationTask;
-};
-declare const ops: {
-    pack16: typeof pack16;
-    unpack16: typeof unpack16;
-};
-export { ops };
-export { selectBackend } from './backend';
-export { default as performanceTest } from './utilities/performance';
-export declare const layers: {
-    CausalSelfAttention: typeof CausalSelfAttention;
-    MLP: typeof MLP;
-    TransformerBlock: typeof TransformerBlock;
-    RoPECache: typeof RoPECache;
-};
-export { AdamWOptimizer } from './training/AdamW';
-export { default as checks } from './checks';
-export type { TensorStatistics } from './checks/weights';
-export { sentenceEmbeddings, sentenceEmbeddingsTensor } from './utilities/sentences';

package/dist/main.js

@@ -1,109 +0,0 @@
-import "./index-CUXkjxiT.js";
-import "./random_width-DN5ZtQkM.js";
-import "./zeros_like-CWjDdwr-.js";
-import { default as io } from "./Generator.js";
-import "./index-Cp39cXWe.js";
-import "./dataset-CGGp1z9P.js";
-import { default as fo } from "./models/NanoGPTV1.js";
-import { default as lo } from "./TeachableLLM.js";
-import { default as uo } from "./tokeniser/CharTokeniser.js";
-import { default as ko } from "./tokeniser/bpe.js";
-import { default as go } from "./utilities/waitForModel.js";
-import { default as Co } from "./utilities/datasetID.js";
-import { default as Eo } from "./data/textLoader.js";
-import { default as Bo } from "./training/Evaluator.js";
-import { default as vo } from "./Trainer.js";
-import { default as Do } from "./models/model.js";
-import { estimateMemoryUsage as So, estimateParameterCount as Ao, estimateResources as Fo, estimateTrainingMemoryUsage as Go, validateConfig as Ro } from "./utilities/parameters.js";
-import { default as wo } from "./utilities/topP.js";
-import { Task as Io, tokensFromTasks as No } from "./training/tasks/Task.js";
-import o from "./training/tasks/PretrainingTask.js";
-import r from "./training/tasks/StartSentenceTask.js";
-import t from "./training/tasks/ConversationTask.js";
-import "./ops/cpu/scatterSub.js";
-import "./ops/webgl/scatterSub.js";
-import "./ops/cpu/gatherSub.js";
-import "./ops/webgl/gatherSub.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 "./ops/cpu/rope.js";
-import "./ops/webgl/rope.js";
-import "./rope-DcrZM_e6.js";
-import "./ops/cpu/appendCache.js";
-import "./ops/webgl/appendCache.js";
-import "./ops/cpu/matMulGelu.js";
-import "./matMulGelu-JNLZqKQp.js";
-import "./ops/grads/matMulGelu.js";
-import "./ops/cpu/gelu.js";
-import "./ops/webgl/gelu.js";
-import "./gelu-B220X1Go.js";
-import "./ops/cpu/normRMS.js";
-import "./ops/webgl/normRMS.js";
-import "./ops/grads/normRMS.js";
-import "./ops/webgl/log.js";
-import "./ops/cpu/adamMoments.js";
-import "./ops/webgl/adamMoments.js";
-import "./ops/cpu/adamAdjust.js";
-import "./ops/webgl/adamAdjust.js";
-import { u as e, p as m } from "./pack16-bqltoUlR.js";
-import "./ops/grads/softmax16.js";
-import "./matMul16-BcVC_E62.js";
-import "./ops/webgl/matMul16.js";
-import "./ops/cpu/matMul16.js";
-import "./ops/transpose16.js";
-import { selectBackend as Wo } from "./backend.js";
-import { default as qo } from "./utilities/performance.js";
-import a from "./layers/CausalSelfAttention.js";
-import p from "./layers/MLP.js";
-import i from "./layers/TransformerBlock.js";
-import s from "./layers/RoPECache.js";
-import { AdamWOptimizer as Jo } from "./training/AdamW.js";
-import { default as Qo } from "./checks/index.js";
-import { sentenceEmbeddings as Xo, sentenceEmbeddingsTensor as Yo } from "./utilities/sentences.js";
-const to = {
-  PretrainingTask: o,
-  StartSentenceTask: r,
-  ConversationTask: t
-}, eo = {
-  pack16: m,
-  unpack16: e
-}, mo = {
-  CausalSelfAttention: a,
-  MLP: p,
-  TransformerBlock: i,
-  RoPECache: s
-};
-export {
-  Jo as AdamWOptimizer,
-  ko as BPETokeniser,
-  uo as CharTokeniser,
-  Bo as Evaluator,
-  io as Generator,
-  Do as Model,
-  fo as NanoGPT,
-  Io as Task,
-  lo as TeachableLLM,
-  vo as Trainer,
-  Qo as checks,
-  So as estimateMemoryUsage,
-  Ao as estimateParameterCount,
-  Fo as estimateResources,
-  Go as estimateTrainingMemoryUsage,
-  Co as generateDatasetID,
-  mo as layers,
-  Eo as loadTextData,
-  eo as ops,
-  qo as performanceTest,
-  Wo as selectBackend,
-  Xo as sentenceEmbeddings,
-  Yo as sentenceEmbeddingsTensor,
-  to as tasks,
-  No as tokensFromTasks,
-  wo as topP,
-  Ro as validateConfig,
-  go as waitForModel
-};

package/dist/matMul16-BcVC_E62.js

@@ -1,80 +0,0 @@
-import { e as y, h } from "./index-CUXkjxiT.js";
-import "./ops/webgl/matMul16.js";
-import "./ops/cpu/matMul16.js";
-import { isPackedTensor as v } from "./utilities/packed.js";
-import { p as g } from "./pack16-bqltoUlR.js";
-import { d as k } from "./gelu-B220X1Go.js";
-import { transpose16 as S } from "./ops/transpose16.js";
-import { reshape16 as w } from "./ops/reshape16.js";
-import { mul16 as D } from "./ops/mul16.js";
-const G = {
-  kernelName: "MatMul16",
-  inputsToSave: ["A", "B"],
-  outputsToSave: [],
-  gradFunc: (r, o, l) => {
-    const [s, t] = o;
-    if (Array.isArray(r))
-      throw new Error("Expected dy to be a single Tensor");
-    let e = r;
-    const { transposeA: u, transposeB: i, scale: a, activation: p, originalShape: d, perm: m } = l;
-    if (m && d) {
-      const f = new Array(m.length);
-      for (let M = 0; M < m.length; ++M)
-        f[m[M]] = M;
-      const c = e;
-      e = S(e, f), c.dispose();
-    }
-    if (d) {
-      const f = e;
-      e = w(e, d), f.dispose();
-    }
-    if (p === "gelu") {
-      const f = e, c = n(s, t, u, i);
-      e = k(f, c), f.dispose(), c.dispose();
-    } else if (p === "relu2") {
-      const f = e, c = n(s, t, u, i, { activation: "relu", scale: 2 });
-      e = D(f, c), f.dispose(), c.dispose();
-    }
-    if (!u && !i)
-      return {
-        A: () => a !== void 0 ? B(e, t, a, !1, !0) : n(e, t, !1, !0),
-        B: () => a !== void 0 ? A(s, e, a, !0, !1) : n(s, e, !0, !1)
-      };
-    if (!u && i)
-      return {
-        A: () => a !== void 0 ? B(e, t, a, !1, !1) : n(e, t, !1, !1),
-        B: () => a !== void 0 ? A(s, e, a, !0, !1) : n(s, e, !0, !1)
-      };
-    if (u && !i)
-      return {
-        A: () => a !== void 0 ? A(t, e, a, !1, !0) : n(t, e, !1, !0),
-        B: () => a !== void 0 ? A(s, e, a, !1, !1) : n(s, e, !1, !1)
-      };
-    throw new Error("Gradient for transposeA=true and transposeB=true is not supported yet.");
-  }
-};
-y(G);
-function n(r, o, l = !1, s = !1, t = {}) {
-  const e = v(r), u = v(o), i = e || u, a = !i || e ? r : g(r), p = !i || u ? o : g(o), d = h().runKernel("MatMul16", { A: a, B: p }, { transposeA: l, transposeB: s, ...t });
-  return i && !e && a.dispose(), i && !u && p.dispose(), d;
-}
-function j(r, o, l, s = !1, t = !1) {
-  return n(r, o, s, t, { scale: l });
-}
-function B(r, o, l, s = !1, t = !1) {
-  return n(r, o, s, t, { scaleA: l });
-}
-function A(r, o, l, s = !1, t = !1) {
-  return n(r, o, s, t, { scaleB: l });
-}
-function q(r, o, l = !1, s = !1) {
-  return n(r, o, l, s, { activation: "gelu" });
-}
-export {
-  G as a,
-  n as b,
-  q as c,
-  B as d,
-  A as e,
-  j as m
-};

package/dist/matMulGelu-JNLZqKQp.js

@@ -1,163 +0,0 @@
-import { c as C, t as R, h as I, U as G, _ as L, x as U, am as F } from "./index-CUXkjxiT.js";
-import { r as M } from "./Reshape-BW__R4mZ.js";
-import { u as H } from "./gpgpu_math-BwvV12df.js";
-import { m as B } from "./mat_mul-DhG0Newp.js";
-class W {
-  constructor(e, s, a, n = !1, o = !1, r = !1, i = null, u = !1, l = !1) {
-    this.variableNames = ["matrixA", "matrixB"], this.packedInputs = !0, this.packedOutput = !0, this.outputShape = a, this.enableShapeUniforms = H(this.outputShape.length);
-    const p = n ? e[1] : e[2], h = Math.ceil(p / 2), d = n ? "i * 2, rc.y" : "rc.y, i * 2", b = o ? "rc.z, i * 2" : "i * 2, rc.z", x = n ? ["a.xxyy", "a.zzww"] : ["a.xxzz", "a.yyww"], m = o ? ["b.xzxz", "b.ywyw"] : ["b.xyxy", "b.zwzw"];
-    let c = "", g = "";
-    i && (u ? c = `vec4 activation(vec4 a) {
-          vec4 b = getPreluActivationWeightsAtOutCoords();
-          ${i}
-        }` : l ? c = `vec4 activation(vec4 a) {
-          vec4 b = getLeakyreluAlphaAtOutCoords();
-          ${i}
-        }` : c = `vec4 activation(vec4 x) {
-          ${i}
-        }`, g = "result = activation(result);");
-    const $ = r ? "result += getBiasAtOutCoords();" : "";
-    r && this.variableNames.push("bias"), u && this.variableNames.push("preluActivationWeights"), l && this.variableNames.push("leakyreluAlpha");
-    let f = "rc.x", v = "rc.x";
-    e[0] < s[0] ? f = `imod(rc.x, ${e[0]})` : s[0] < e[0] && (v = `imod(rc.x, ${s[0]})`), this.userCode = `
-      ${c}
-      // Don't use uniform for sharedDimensionPacked for performance.
-      const float sharedDimension = ${h}.0;
-
-      vec4 dot2x2ARowBCol(ivec3 rc) {
-        vec4 result = vec4(0);
-        int batchA = ${f};
-        int batchB = ${v};
-        for (int i = 0; i < ${h}; i++) {
-          vec4 a = getMatrixA(batchA, ${d});
-          vec4 b = getMatrixB(batchB, ${b});
-
-          // These swizzled products need to be separately added.
-          // See: https://github.com/tensorflow/tfjs/issues/1735
-          result += (${x[0]} * ${m[0]});
-          result += (${x[1]} * ${m[1]});
-        }
-        return result;
-      }
-
-      void main() {
-        ivec3 rc = getOutputCoords();
-        vec4 result = dot2x2ARowBCol(rc);
-
-        ${$}
-
-        ${g}
-
-        setOutput(result);
-      }
-    `;
-  }
-}
-const S = 0.7978845608028654, w = 0.044715, j = `
-    vec4 x3 = x * x * x;
-    vec4 inner = x + ${w} * x3;
-    inner = ${S} * inner;
-    inner = vec4(
-        abs(inner[0]) > 15.0 ? sign(inner[0]) : tanh(inner[0]),
-        abs(inner[1]) > 15.0 ? sign(inner[1]) : tanh(inner[1]),
-        abs(inner[2]) > 15.0 ? sign(inner[2]) : tanh(inner[2]),
-        abs(inner[3]) > 15.0 ? sign(inner[3]) : tanh(inner[3])
-    );
-    inner = 0.5 * (1.0 + inner);
-    vec4 result = x * inner;
-    return result;
-`, q = `
-    vec4 a2 = a * a;
-    vec4 a3 = a2 * a;
-    vec4 u  = ${S} * (a + ${w} * a3);
-    vec4 t = vec4(
-        abs(u[0]) > 15.0 ? sign(u[0]) : tanh(u[0]),
-        abs(u[1]) > 15.0 ? sign(u[1]) : tanh(u[1]),
-        abs(u[2]) > 15.0 ? sign(u[2]) : tanh(u[2]),
-        abs(u[3]) > 15.0 ? sign(u[3]) : tanh(u[3])
-    );
-    vec4 sech2 = 1.0 - t * t;
-    vec4 du_dx = ${S} * (1.0 + 3.0 * ${w} * a2);
-    vec4 dgelu = 0.5 * (1.0 + t) + 0.5 * a * sech2 * du_dx;
-    return dgelu * b;
-`, se = 1e3;
-function O({
-  a: t,
-  b: e,
-  transposeA: s,
-  transposeB: a,
-  backend: n,
-  activationSnippet: o,
-  multiplier: r
-}) {
-  const i = t.shape.length, u = e.shape.length, l = s ? t.shape[i - 2] : t.shape[i - 1], p = a ? e.shape[u - 1] : e.shape[u - 2], h = s ? t.shape[i - 1] : t.shape[i - 2], d = a ? e.shape[u - 2] : e.shape[u - 1], b = t.shape.slice(0, -2), x = e.shape.slice(0, -2), m = G(b), c = G(x), $ = L(t.shape.slice(0, -2), e.shape.slice(0, -2)).concat([h, d]);
-  U(
-    l === p,
-    () => `Error in matMul: inner shapes (${l}) and (${p}) of Tensors with shapes ${t.shape} and ${e.shape} and transposeA=${s} and transposeB=${a} must match.`
-  );
-  const f = s ? [m, l, h] : [m, h, l], v = a ? [c, d, p] : [c, p, d], A = M({ inputs: { x: t }, backend: n, attrs: { shape: f } }), y = M({ inputs: { x: e }, backend: n, attrs: { shape: v } }), D = [A, y], _ = Math.max(m, c), E = o, N = F(t.dtype, e.dtype), T = new W(
-    f,
-    v,
-    [_, h, d],
-    s,
-    a,
-    !1,
-    E,
-    !!r,
-    !1
-  ), k = [A, y];
-  r && k.push(r);
-  const z = n.runWebGLProgram(T, k, N), K = M({ inputs: { x: z }, backend: n, attrs: { shape: $ } });
-  D.push(z);
-  for (const P of D)
-    n.disposeIntermediateTensorInfo(P);
-  return K;
-}
-function J(t) {
-  const { inputs: e, backend: s } = t, { x: a, kernel: n } = e;
-  if (a === void 0 || n === void 0)
-    throw new Error("BatchMatMul requires two input tensors.");
-  return O({
-    a,
-    b: n,
-    transposeA: !1,
-    transposeB: !1,
-    backend: s,
-    activationSnippet: j
-  });
-}
-const Q = {
-  kernelName: "MatMulGelu",
-  backendName: "webgl",
-  kernelFunc: J
-};
-C(Q);
-function V(t) {
-  const { dy: e, x: s, kernel: a } = t.inputs, n = t.backend;
-  return R(() => {
-    const o = I().makeTensorFromTensorInfo(
-      O({
-        a: s,
-        b: a,
-        transposeA: !1,
-        transposeB: !1,
-        backend: n,
-        activationSnippet: q,
-        multiplier: e
-      })
-    ), r = B(o, a, !1, !0), i = B(s, o, !0, !1);
-    return [r, i];
-  });
-}
-const X = {
-  kernelName: "MatMulGeluGrad",
-  backendName: "webgl",
-  kernelFunc: V
-};
-C(X);
-export {
-  W as M,
-  se as a,
-  O as b,
-  J as c
-};

package/dist/mat_mul-DhG0Newp.js

@@ -1,11 +0,0 @@
-import { o as m, q as s, B as c, E as M, D as p } from "./index-CUXkjxiT.js";
-function f(e, o, n = !1, l = !1) {
-  let a = s(e, "a", "matMul"), t = s(o, "b", "matMul");
-  [a, t] = c(a, t);
-  const r = { a, b: t }, u = { transposeA: n, transposeB: l };
-  return M.runKernel(p, r, u);
-}
-const i = /* @__PURE__ */ m({ matMul_: f });
-export {
-  i as m
-};