@genai-fi/nanogpt 0.9.1 → 0.10.1

package/dist/layers/RoPECache.js

@@ -1,16 +1,16 @@
-import { b as t, x as h, t as n, k as p } from "../index-BzFyqcy-.js";
-import { r as c } from "../reciprocal-C4rNcM-S.js";
-import { c as f, s as m } from "../sin-gpDNRxE0.js";
-import { r as a } from "../range-CWcz7xFA.js";
-class D {
-  constructor(o) {
-    this.config = o;
-    const e = this.config.nEmbed / this.config.nHead;
-    if (this.rotaryDim = e, this.rotaryDim % 2 !== 0)
+import { b as n, p as a, t as p, k as r } from "../index-ZyQhjEPo.js";
+import { r as c } from "../reciprocal-CTmshQ9J.js";
+import { c as l, s as f } from "../sin-BGfy2HZo.js";
+import { r as h } from "../range-BMS52eQi.js";
+class g {
+  constructor(s) {
+    this.config = s;
+    const i = this.config.nEmbed / this.config.nHead;
+    if (this.rotaryDim = i, this.rotaryDim % 2 !== 0)
       throw new Error("rotaryDim must be even");
     this.ropeBase = 1e4;
-    const i = a(0, this.rotaryDim, 2, "float32"), s = i.div(t(this.rotaryDim, "float32")), r = h(t(this.ropeBase, "float32"), s);
-    this.ropeInvFreq = c(r), s.dispose(), r.dispose(), i.dispose(), this.config.useRope === !1 ? (this.ropeCos = null, this.ropeSin = null, this.ropeCacheLen = 0) : n(() => {
+    const o = h(0, this.rotaryDim, 2, "float32"), e = o.div(n(this.rotaryDim, "float32")), t = a(n(this.ropeBase, "float32"), e);
+    this.ropeInvFreq = c(t), e.dispose(), t.dispose(), o.dispose(), this.config.useRope === !1 ? (this.ropeCos = null, this.ropeSin = null, this.ropeNegSin = null, this.ropeCacheLen = 0) : p(() => {
       this.ensureRopeCache(this.config.blockSize * 4);
     });
   }
@@ -21,13 +21,15 @@ class D {
   // [cacheLen, rotaryDim/2]
   ropeSin = null;
   // [cacheLen, rotaryDim/2]
+  ropeNegSin = null;
+  // [cacheLen, rotaryDim/2]
   ropeCacheLen = 0;
-  ensureRopeCache(o) {
-    n(() => {
-      if (o <= this.ropeCacheLen) return;
+  ensureRopeCache(s) {
+    p(() => {
+      if (s <= this.ropeCacheLen) return;
       this.ropeCos && this.ropeCos.dispose(), this.ropeSin && this.ropeSin.dispose();
-      const e = Math.max(o, this.ropeCacheLen + this.config.blockSize * 4), s = a(0, e, 1, "float32").expandDims(1).mul(this.ropeInvFreq.expandDims(0));
-      this.ropeCos = p(f(s).expandDims(-1)), this.ropeSin = p(m(s).expandDims(-1)), this.ropeCacheLen = e;
+      const i = Math.max(s, this.ropeCacheLen + this.config.blockSize * 4), e = h(0, i, 1, "float32").expandDims(1).mul(this.ropeInvFreq.expandDims(0));
+      this.ropeCos = r(l(e).expandDims(-1)), this.ropeSin = r(f(e).expandDims(-1)), this.ropeNegSin = r(this.ropeSin.neg()), this.ropeCacheLen = i;
     });
   }
   getCos() {
@@ -36,10 +38,13 @@ class D {
   getSin() {
     return this.ropeSin;
   }
+  getNegSin() {
+    return this.ropeNegSin;
+  }
   dispose() {
     this.ropeCos && this.ropeCos.dispose(), this.ropeSin && this.ropeSin.dispose(), this.ropeInvFreq.dispose();
   }
 }
 export {
-  D as default
+  g as default
 };

package/dist/layers/TiedEmbedding.js

@@ -1,31 +1,36 @@
-import { d as r } from "../random_width-CXVRloNK.js";
-import "../index-BzFyqcy-.js";
-import { r as a } from "../exports_initializers-B8iZMgQ0.js";
-import s from "./BaseLayer.js";
-import { v as o } from "../variable-Bm2OFwGI.js";
-import { g as m } from "../gather-Dnpgw-YQ.js";
-class S extends s {
+import "../random_width-DY6Kk2Dl.js";
+import "../index-ZyQhjEPo.js";
+import { r as s } from "../exports_initializers-xuidcwI4.js";
+import a from "./BaseLayer.js";
+import { dot16 as o } from "../ops/dot16.js";
+import { isPackedTensor as r } from "../utilities/packed.js";
+import { p as m } from "../pack16-CFUqumar.js";
+import { transpose16 as d } from "../ops/transpose16.js";
+import { v as p } from "../variable-Bhn5bHYv.js";
+import { g as h } from "../gather-DykLGqmW.js";
+class g extends a {
   vocabSize;
   embedDim;
   initializer;
   WEIGHTS;
-  constructor(i, t, e) {
-    super(i, e), this.WEIGHTS = t, this.vocabSize = i.vocabSize, this.embedDim = i.nEmbed, this.initializer = a({
+  constructor(t, i, e) {
+    super(t, e), this.WEIGHTS = i, this.vocabSize = t.vocabSize, this.embedDim = t.nEmbed, this.initializer = s({
       mean: 0,
       stddev: 0.02
-    }), this.addVariable(this.WEIGHTS, o(this.initializer.apply([this.vocabSize, this.embedDim]), !0, t));
+    }), this.addVariable(this.WEIGHTS, p(this.initializer.apply([this.vocabSize, this.embedDim]), !0, i));
   }
-  embed(i) {
-    return m(this.getVariable(this.WEIGHTS), i, 0);
+  embed(t) {
+    return h(this.getVariable(this.WEIGHTS), t, 0);
   }
-  project(i) {
-    return r(i, this.getVariable(this.WEIGHTS).transpose());
+  project(t) {
+    const i = r(t) ? m(this.getVariable(this.WEIGHTS), void 0, 32) : this.getVariable(this.WEIGHTS), e = d(i);
+    return r(t) && i.dispose(), o(t, e);
   }
   // Dummy, should not be used.
-  forward(i, t) {
-    return this.project(t);
+  forward(t, i) {
+    return this.project(i);
   }
 }
 export {
-  S as default
+  g as default
 };

package/dist/layers/TransformerBlock.js

@@ -1,32 +1,33 @@
-import l from "./CausalSelfAttention.js";
-import p from "./MLP.js";
-import o from "./RMSNorm.js";
-import m from "./BaseLayer.js";
-import { k as n, t as h } from "../index-BzFyqcy-.js";
-class k extends m {
+import p from "./CausalSelfAttention.js";
+import m from "./MLP.js";
+import d from "./RMSNorm.js";
+import h from "./BaseLayer.js";
+import { k as n, t as u } from "../index-ZyQhjEPo.js";
+import { add16 as l } from "../ops/add16.js";
+class g extends h {
   ln1;
   attn;
   ln2;
   mlp;
   index;
   skipped = !1;
-  constructor(i, s, e) {
-    super(s, e), this.index = i, this.ln1 = new o(s, `block_${this.index}_rms1`, this), this.attn = new l(this.index, s, this), this.ln2 = new o(s, `block_${this.index}_rms2`, this), this.mlp = new p(this.index, s, this);
+  constructor(e, s, i) {
+    super(s, i), this.index = e, this.ln1 = new d(s, `block_${this.index}_rms1`, this), this.attn = new p(this.index, s, this), this.ln2 = new d(s, `block_${this.index}_rms2`, this), this.mlp = new m(this.index, s, this);
   }
-  getMLPOutput(i, s) {
-    const e = this.ln2.call({ training: s.training }, i), t = this.mlp.call({ training: s.training }, e);
-    s.outputEmbeddings ? (n(e), s.embeddings.push({ name: `block_ln2_${this.index}`, tensor: e })) : e.dispose();
-    const d = i.add(t);
-    return i.dispose(), s.outputEmbeddings ? (n(t), s.embeddings.push({ name: `block_mlp_out_${this.index}`, tensor: t })) : t.dispose(), d;
+  getMLPOutput(e, s) {
+    const i = this.ln2.call({ training: s.training }, e), t = this.mlp.call(s, i);
+    s.outputEmbeddings ? (n(i), s.embeddings.push({ name: `block_ln2_${this.index}`, tensor: i })) : i.dispose();
+    const o = l(e, t);
+    return e.dispose(), s.outputEmbeddings ? (n(t), s.embeddings.push({ name: `block_mlp_out_${this.index}`, tensor: t })) : t.dispose(), o;
   }
-  forward(i, s) {
-    return h(() => {
+  forward(e, s) {
+    return u(() => {
       if (this.skipped)
         return s;
-      const e = this.ln1.call(i, s), t = this.attn.call(i, e);
-      i.outputEmbeddings ? (n(e), i.embeddings.push({ name: `block_ln1_${this.index}`, tensor: e })) : e.dispose();
-      const d = s.add(t);
-      return i.outputEmbeddings ? (n(t), i.embeddings.push({ name: `block_attn_out_${this.index}`, tensor: t })) : t.dispose(), this.getMLPOutput(d, i);
+      const i = this.ln1.call(e, s), t = this.attn.call(e, i);
+      e.outputEmbeddings ? (n(i), e.embeddings.push({ name: `block_ln1_${this.index}`, tensor: i })) : i.dispose();
+      const o = l(s, t);
+      return e.outputEmbeddings ? (n(t), e.embeddings.push({ name: `block_attn_out_${this.index}`, tensor: t })) : t.dispose(), this.getMLPOutput(o, e);
     });
   }
   dispose() {
@@ -34,5 +35,5 @@ class k extends m {
   }
 }
 export {
-  k as default
+  g as default
 };

package/dist/loader/load.js

@@ -1,4 +1,4 @@
-import { j as o } from "../jszip.min-CjP2V1VV.js";
+import { j as o } from "../jszip.min-Bz5-11Bk.js";
 import i from "./oldZipLoad.js";
 import s from "./newZipLoad.js";
 import n from "./loadHF.js";

package/dist/loader/loadTransformers.js

@@ -1,7 +1,7 @@
 import l from "../tokeniser/CharTokeniser.js";
 import c from "../tokeniser/bpe.js";
 import { load_safetensors as b } from "../utilities/safetensors.js";
-import { a2 as y } from "../index-BzFyqcy-.js";
+import { H as y } from "../index-ZyQhjEPo.js";
 import { dummyPassAsync as u } from "../utilities/dummy.js";
 import _ from "../models/factory.js";
 async function L(e, a, r, t) {

package/dist/loader/oldZipLoad.js

@@ -1,73 +1,79 @@
-import { a2 as y } from "../index-BzFyqcy-.js";
+import "../utilities/packed.js";
+import { H as y } from "../index-ZyQhjEPo.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-CXVRloNK.js";
-import "../register_all_kernels-DIGpEwcf.js";
-import "../index-Tf7vU29b.js";
-import "../dataset-DlZtKmBq.js";
+import "../random_width-DY6Kk2Dl.js";
+import "../register_all_kernels-Bwu1PTuU.js";
+import "../index-Cp39cXWe.js";
+import "../dataset-0xP8GjwI.js";
 import "../ops/cpu/rope.js";
 import "../ops/webgl/rope.js";
-import "../ops/grads/rope.js";
+import "../rope-B5UUMsPi.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/grads/softmax16.js";
+import "../matMul16--R5hOwDG.js";
+import "../ops/webgl/matMul16.js";
+import "../ops/cpu/matMul16.js";
+import "../pack16-CFUqumar.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/grads/add16.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 "../jszip.min-Bz5-11Bk.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 "../index-DvYrXKkX.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 "../papaparse.min-C0cScC2i.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-Bp_-935b.js";
+import "../gelu-CNLFZWea.js";
 import "../ops/webgl/log.js";
 import "../checks/normRMS.js";
 import "../checks/normRMSGrad.js";
 import { importWeights as u } from "../utilities/weights.js";
-async function ht(r) {
-  const e = /* @__PURE__ */ new Map(), a = await r.file("manifest.json")?.async("string");
-  if (!a)
+async function Ot(r) {
+  const e = /* @__PURE__ */ new Map(), p = await r.file("manifest.json")?.async("string");
+  if (!p)
     throw new Error("Manifest file not found in the zip archive");
-  const m = JSON.parse(a);
-  for (const [t, o] of Object.entries(m.weightSpec))
+  const s = JSON.parse(p);
+  for (const [t, o] of Object.entries(s.weightSpec))
     e.set(t, { spec: o, data: new Float32Array() });
-  const p = await r.file("tokeniser.json")?.async("string");
-  if (!p)
+  const a = await r.file("tokeniser.json")?.async("string");
+  if (!a)
     throw new Error("Tokeniser file not found in the zip archive");
-  const i = JSON.parse(p), f = (i.type ?? "char") === "char" ? new h(i.vocab) : new k(i.vocab, i.merges), c = /* @__PURE__ */ new Map();
+  const i = JSON.parse(a), 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 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);
+      const o = t.replace(".bin", ""), l = await r.file(t).async("arraybuffer"), w = new Float32Array(l), n = e.get(o) || { spec: [], data: new Float32Array() };
+      n.data = w, e.set(o, n);
+      const d = await u(n);
       c.set(o, d);
     }
   y();
-  const n = b(m.config);
-  return await g(n), n.loadWeights(c), { model: n, tokeniser: f };
+  const m = b(s.config);
+  return await g(m), m.loadWeights(c), { model: m, tokeniser: f };
 }
 export {
-  ht as default
+  Ot as default
 };

package/dist/loader/save.js

@@ -1,4 +1,4 @@
-import { j as g } from "../jszip.min-CjP2V1VV.js";
+import { j as g } from "../jszip.min-Bz5-11Bk.js";
 import b from "../tokeniser/CharTokeniser.js";
 import { save_safetensors as p } from "../utilities/safetensors.js";
 import { VERSION as y } from "./load.js";

package/dist/log_sum_exp-DWI-76TI.js

@@ -0,0 +1,41 @@
+import { A as e, B as x, E as p, c as E, n as $ } from "./index-ZyQhjEPo.js";
+import { e as d } from "./axis_util-BvHEw88j.js";
+import { y as h, z as S, L as K } from "./tensor_util-DV-FP5Q3.js";
+import { r as m } from "./reshape-DevtBWtf.js";
+import { s as _ } from "./sum-_fzj5ZTB.js";
+import { p as T } from "./tensor-DdQUJZlz.js";
+function b(s, o = null, n = !1) {
+  const t = { x: x(s, "x", "max") }, r = { reductionIndices: o, keepDims: n };
+  return p.runKernel(h, t, r);
+}
+const A = /* @__PURE__ */ e({ max_: b });
+function I(s) {
+  const n = { x: x(s, "x", "exp") };
+  return p.runKernel(S, n);
+}
+const L = /* @__PURE__ */ e({ exp_: I });
+function M(s) {
+  const n = { x: x(s, "x", "log", "float32") };
+  return p.runKernel(K, n);
+}
+const N = /* @__PURE__ */ e({ log_: M });
+function v(s, o = null, n = !1) {
+  const a = x(s, "x", "logSumExp"), t = T(o, a.shape), r = A(
+    a,
+    t,
+    !0
+    /* keepDims */
+  ), i = E(a, r), l = L(i), f = _(l, t), u = N(f), c = $(m(r, u.shape), u);
+  if (n) {
+    const g = d(c.shape, t);
+    return m(c, g);
+  }
+  return c;
+}
+const P = /* @__PURE__ */ e({ logSumExp_: v });
+export {
+  N as a,
+  L as e,
+  P as l,
+  A as m
+};

package/dist/main.d.ts

@@ -1,3 +1,5 @@
+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';
@@ -17,6 +19,11 @@ export type { TrainingProgress, TrainingLogEntry } from './training/Trainer';
 export type { GPTConfig } from './models/config';
 export { estimateParameterCount, estimateMemoryUsage, estimateTrainingMemoryUsage, estimateResources, validateConfig, } from './utilities/parameters';
 export { default as topP } from './utilities/topP';
+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: {
@@ -28,3 +35,4 @@ export declare const layers: {
 export { default as AdamExt } from './training/AdamExt';
 export { default as checks } from './checks';
 export type { TensorStatistics } from './checks/weights';
+export { sentenceEmbeddings, sentenceEmbeddingsTensor } from './utilities/sentences';

package/dist/main.js

@@ -1,14 +1,15 @@
-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 { default as io } from "./Generator.js";
-import { default as fo } from "./models/model.js";
-import { estimateMemoryUsage as lo, estimateParameterCount as xo, estimateResources as uo, estimateTrainingMemoryUsage as no, validateConfig as co } from "./utilities/parameters.js";
-import { default as ko } from "./utilities/topP.js";
-import "./index-BzFyqcy-.js";
+import "./utilities/packed.js";
+import { default as ro } from "./models/NanoGPTV1.js";
+import { default as eo } from "./TeachableLLM.js";
+import { default as po } from "./tokeniser/CharTokeniser.js";
+import { default as ao } from "./tokeniser/bpe.js";
+import { default as fo } from "./utilities/waitForModel.js";
+import { default as no } from "./data/textLoader.js";
+import { default as uo } from "./Generator.js";
+import { default as To } from "./models/model.js";
+import { estimateMemoryUsage as go, estimateParameterCount as Mo, estimateResources as Po, estimateTrainingMemoryUsage as Co, validateConfig as Eo } from "./utilities/parameters.js";
+import { default as Bo } from "./utilities/topP.js";
+import "./index-ZyQhjEPo.js";
 import "./ops/cpu/scatterSub.js";
 import "./ops/webgl/scatterSub.js";
 import "./ops/cpu/gatherSub.js";
@@ -19,24 +20,21 @@ import "./ops/grads/attentionMask.js";
 import "./ops/cpu/qkv.js";
 import "./ops/webgl/qkv.js";
 import "./ops/grads/qkv.js";
-import "./random_width-CXVRloNK.js";
-import "./register_all_kernels-DIGpEwcf.js";
-import "./index-Tf7vU29b.js";
-import "./dataset-DlZtKmBq.js";
+import "./random_width-DY6Kk2Dl.js";
+import "./register_all_kernels-Bwu1PTuU.js";
+import "./index-Cp39cXWe.js";
+import "./dataset-0xP8GjwI.js";
 import "./ops/cpu/rope.js";
 import "./ops/webgl/rope.js";
-import "./ops/grads/rope.js";
+import "./rope-B5UUMsPi.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/gelu.js";
 import "./ops/webgl/gelu.js";
-import "./gelu-Bp_-935b.js";
+import "./gelu-CNLFZWea.js";
 import "./ops/cpu/normRMS.js";
 import "./ops/webgl/normRMS.js";
 import "./ops/grads/normRMS.js";
@@ -45,38 +43,51 @@ import "./ops/cpu/adamMoments.js";
 import "./ops/webgl/adamMoments.js";
 import "./ops/cpu/adamAdjust.js";
 import "./ops/webgl/adamAdjust.js";
-import { selectBackend as Po } from "./backend.js";
-import { default as go } 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 Bo } from "./training/AdamExt.js";
-import { default as Eo } from "./checks/index.js";
-const O = {
-  CausalSelfAttention: o,
-  MLP: r,
-  TransformerBlock: t,
-  RoPECache: m
+import { u as o, p as r } from "./pack16-CFUqumar.js";
+import "./ops/grads/softmax16.js";
+import "./matMul16--R5hOwDG.js";
+import "./ops/webgl/matMul16.js";
+import "./ops/cpu/matMul16.js";
+import "./ops/transpose16.js";
+import { selectBackend as yo } from "./backend.js";
+import { default as Ao } from "./utilities/performance.js";
+import t from "./layers/CausalSelfAttention.js";
+import e from "./layers/MLP.js";
+import m from "./layers/TransformerBlock.js";
+import p from "./layers/RoPECache.js";
+import { default as Ro } from "./training/AdamExt.js";
+import { default as vo } from "./checks/index.js";
+import { sentenceEmbeddings as Do, sentenceEmbeddingsTensor as Fo } from "./utilities/sentences.js";
+const Z = {
+  pack16: r,
+  unpack16: o
+}, _ = {
+  CausalSelfAttention: t,
+  MLP: e,
+  TransformerBlock: m,
+  RoPECache: p
 };
 export {
-  Bo as AdamExt,
-  oo as BPETokeniser,
-  _ as CharTokeniser,
-  io as Generator,
-  fo as Model,
-  W as NanoGPT,
-  Y as TeachableLLM,
-  Eo as checks,
-  lo as estimateMemoryUsage,
-  xo as estimateParameterCount,
-  uo as estimateResources,
-  no as estimateTrainingMemoryUsage,
-  O as layers,
-  eo as loadTextData,
-  go as performanceTest,
-  Po as selectBackend,
-  ko as topP,
-  co as validateConfig,
-  to as waitForModel
+  Ro as AdamExt,
+  ao as BPETokeniser,
+  po as CharTokeniser,
+  uo as Generator,
+  To as Model,
+  ro as NanoGPT,
+  eo as TeachableLLM,
+  vo as checks,
+  go as estimateMemoryUsage,
+  Mo as estimateParameterCount,
+  Po as estimateResources,
+  Co as estimateTrainingMemoryUsage,
+  _ as layers,
+  no as loadTextData,
+  Z as ops,
+  Ao as performanceTest,
+  yo as selectBackend,
+  Do as sentenceEmbeddings,
+  Fo as sentenceEmbeddingsTensor,
+  Bo as topP,
+  Eo as validateConfig,
+  fo as waitForModel
 };

package/dist/matMul16--R5hOwDG.js

@@ -0,0 +1,77 @@
+import { e as y } from "./index-ZyQhjEPo.js";
+import "./ops/webgl/matMul16.js";
+import "./ops/cpu/matMul16.js";
+import { isPackedTensor as g, packTensor as k } from "./utilities/packed.js";
+import { p as v } from "./pack16-CFUqumar.js";
+import { d as h } from "./gelu-CNLFZWea.js";
+import { transpose16 as S } from "./ops/transpose16.js";
+import { reshape16 as w } from "./ops/reshape16.js";
+import { a as G } from "./tensor_util-DV-FP5Q3.js";
+const T = {
+  kernelName: "MatMul16",
+  inputsToSave: ["A", "B"],
+  outputsToSave: [],
+  gradFunc: (r, o, n) => {
+    const [s, t] = o;
+    if (Array.isArray(r))
+      throw new Error("Expected dy to be a single Tensor");
+    let e = r;
+    const { transposeA: f, transposeB: i, scale: a, activation: p, originalShape: c, perm: d } = n;
+    if (d && c) {
+      const u = new Array(d.length);
+      for (let A = 0; A < d.length; ++A)
+        u[d[A]] = A;
+      const m = e;
+      e = S(e, u), m.dispose();
+    }
+    if (c) {
+      const u = e;
+      e = w(e, c), u.dispose();
+    }
+    if (p === "gelu") {
+      const u = e, m = l(s, t, f, i);
+      e = h(u, m), u.dispose(), m.dispose();
+    }
+    if (!f && !i)
+      return {
+        A: () => a !== void 0 ? B(e, t, a, !1, !0) : l(e, t, !1, !0),
+        B: () => a !== void 0 ? M(s, e, a, !0, !1) : l(s, e, !0, !1)
+      };
+    if (!f && i)
+      return {
+        A: () => a !== void 0 ? B(e, t, a, !1, !1) : l(e, t, !1, !1),
+        B: () => a !== void 0 ? M(s, e, a, !0, !1) : l(s, e, !0, !1)
+      };
+    if (f && !i)
+      return {
+        A: () => a !== void 0 ? M(t, e, a, !1, !0) : l(t, e, !1, !0),
+        B: () => a !== void 0 ? M(s, e, a, !1, !1) : l(s, e, !1, !1)
+      };
+    throw new Error("Gradient for transposeA=true and transposeB=true is not supported yet.");
+  }
+};
+G(T);
+function l(r, o, n = !1, s = !1, t = {}) {
+  const e = g(r), f = g(o), i = e || f, a = !i || e ? r : v(r), p = !i || f ? o : v(o), c = y().runKernel("MatMul16", { A: a, B: p }, { transposeA: n, transposeB: s, ...t });
+  return i && !e && a.dispose(), i && !f && p.dispose(), i ? k(c) : c;
+}
+function j(r, o, n, s = !1, t = !1) {
+  return l(r, o, s, t, { scale: n });
+}
+function B(r, o, n, s = !1, t = !1) {
+  return l(r, o, s, t, { scaleA: n });
+}
+function M(r, o, n, s = !1, t = !1) {
+  return l(r, o, s, t, { scaleB: n });
+}
+function q(r, o, n = !1, s = !1) {
+  return l(r, o, n, s, { activation: "gelu" });
+}
+export {
+  T as a,
+  l as b,
+  q as c,
+  B as d,
+  M as e,
+  j as m
+};

package/dist/mat_mul-DeAh4uTH.js

@@ -0,0 +1,12 @@
+import { A as u, B as s, E as c } from "./index-ZyQhjEPo.js";
+import { m as M, B as p } from "./tensor_util-DV-FP5Q3.js";
+function f(o, e, r = !1, m = !1) {
+  let t = s(o, "a", "matMul"), a = s(e, "b", "matMul");
+  [t, a] = M(t, a);
+  const n = { a: t, b: a }, l = { transposeA: r, transposeB: m };
+  return c.runKernel(p, n, l);
+}
+const B = /* @__PURE__ */ u({ matMul_: f });
+export {
+  B as m
+};

package/dist/mod-Gt1rMB4n.js

@@ -0,0 +1,12 @@
+import { A as e, B as a, E as n } from "./index-ZyQhjEPo.js";
+import { m as p, M as c } from "./tensor_util-DV-FP5Q3.js";
+function d(m, r) {
+  let o = a(m, "a", "mod"), t = a(r, "b", "mod");
+  [o, t] = p(o, t);
+  const s = { a: o, b: t };
+  return n.runKernel(c, s);
+}
+const b = /* @__PURE__ */ e({ mod_: d });
+export {
+  b as m
+};