@genai-fi/nanogpt 0.11.0 → 0.12.0

package/dist/training/tasks/StartSentenceTask.js

@@ -0,0 +1,42 @@
+import { Task as e } from "./Task.js";
+class a extends e {
+  rawText;
+  index = 0;
+  get length() {
+    return this.rawText.length;
+  }
+  constructor(t) {
+    super(), this.rawText = t;
+  }
+  hasMoreConversations() {
+    return this.index < this.rawText.length;
+  }
+  nextConversation() {
+    if (this.index >= this.rawText.length)
+      return null;
+    const t = this.rawText[this.index];
+    return this.index++, this.conversationFromString(t);
+  }
+  conversationFromString(t) {
+    const n = t.indexOf(".");
+    return n === -1 ? [{
+      role: "assistant",
+      content: this.rawText[this.index]
+    }] : [
+      {
+        role: "user",
+        content: t.slice(0, n + 1).trim()
+      },
+      {
+        role: "assistant",
+        content: t.slice(n + 1).trim()
+      }
+    ];
+  }
+  async estimateTokens(t) {
+    return (await t.encodeConversation(this.conversationFromString(this.rawText[0]))).length * this.length;
+  }
+}
+export {
+  a as default
+};

package/dist/training/tasks/Task.d.ts

@@ -0,0 +1,8 @@
+import { Conversation, ITokeniser } from '../../main';
+export declare abstract class Task {
+    abstract get length(): number;
+    abstract hasMoreConversations(): boolean;
+    abstract nextConversation(): Conversation[] | null;
+    abstract estimateTokens(tokeniser: ITokeniser): Promise<number>;
+}
+export declare function tokensFromTasks(tasks: Task[], tokenizer: ITokeniser): Promise<Uint16Array>;

package/dist/training/tasks/Task.js

@@ -0,0 +1,41 @@
+class g {
+}
+function h(f, a, l, e, r) {
+  for (let i = 0; i < f.length; i++) {
+    const c = f[i].nextConversation();
+    if (c) {
+      const o = l.encodeConversation(c), s = a[a.length - 1];
+      if (e.offset + o.length > s.length) {
+        const n = s.length - e.offset;
+        s.set(o.slice(0, n), e.offset);
+        const t = new Uint16Array(Math.floor(r * 0.1) + 100);
+        t.set(o.slice(n), 0), a.push(t), e.offset = o.length - n;
+      } else
+        s.set(o, e.offset), e.offset += o.length;
+    }
+  }
+}
+async function w(f, a) {
+  const l = (await Promise.all(f.map((n) => n.estimateTokens(a)))).reduce(
+    (n, t) => n + t,
+    0
+  ), e = [new Uint16Array(l)], r = {
+    offset: 0
+  };
+  let i = performance.now();
+  for (; r.offset < l && (h(f, e, a, r, l), !f.every((t) => !t.hasMoreConversations())); )
+    performance.now() - i > 40 && (await new Promise(requestAnimationFrame), i = performance.now());
+  if (e.length === 1)
+    return e[0].subarray(0, r.offset);
+  const c = e.reduce((n, t) => n + t.length, 0) - (e[e.length - 1].length - r.offset), o = new Uint16Array(c);
+  let s = 0;
+  for (let n = 0; n < e.length; n++) {
+    const t = e[n];
+    n === e.length - 1 ? (o.set(t.subarray(0, r.offset), s), s += r.offset) : (o.set(t, s), s += t.length);
+  }
+  return o;
+}
+export {
+  g as Task,
+  w as tokensFromTasks
+};

package/dist/{transpose-JawVKyZy.js → transpose-BUkQCJp9.js}

@@ -1,5 +1,5 @@
-import { q as u, u as i, E as o, ap as $, aq as g, ar as m, y as l, t as x, as as p } from "./index-D0RBWjq8.js";
-import { c as k } from "./complex-DClmWqJt.js";
+import { o as u, q as i, E as o, ap as $, aq as g, ar as x, x as l, t as m, as as p } from "./index-Duu1Lvvv.js";
+import { c as k } from "./complex-Cyg-eQeZ.js";
 function K(r) {
   const e = { input: i(r, "input", "imag") };
   return o.runKernel($, e);
@@ -12,25 +12,25 @@ function E(r) {
 const _ = /* @__PURE__ */ u({ neg_: E });
 function b(r) {
   const e = { input: i(r, "input", "real") };
-  return o.runKernel(m, e);
+  return o.runKernel(x, e);
 }
 const d = /* @__PURE__ */ u({ real_: b });
-function y(r, t, e) {
+function N(r, t, e) {
   const n = i(r, "x", "transpose");
   if (t == null && (t = n.shape.map((s, a) => a).reverse()), l(n.rank === t.length, () => `Error in transpose: rank of input ${n.rank} must match length of perm ${t}.`), t.forEach((s) => {
     l(s >= 0 && s < n.rank, () => `All entries in 'perm' must be between 0 and ${n.rank - 1} but got ${t}`);
   }), n.rank <= 1)
     return n.clone();
   const f = { x: n }, c = { perm: t };
-  return n.dtype === "complex64" ? x(() => {
+  return n.dtype === "complex64" ? m(() => {
     let s = d(n), a = h(n);
     return s = o.runKernel(p, { x: s }, c), a = o.runKernel(p, { x: a }, c), e && (a = _(a)), k(s, a);
   }) : o.runKernel(p, f, c);
 }
-const q = /* @__PURE__ */ u({ transpose_: y });
+const v = /* @__PURE__ */ u({ transpose_: N });
 export {
   h as i,
   _ as n,
   d as r,
-  q as t
+  v as t
 };

package/dist/{unsorted_segment_sum-LAbmE9G4.js → unsorted_segment_sum-BljxHhCY.js}

@@ -1,30 +1,30 @@
-import { q as h, u as c, E as d, bo as T, bp as q, bq as H, y as l, br as P, N as _, bs as y, bt as I, bu as W, bv as B, bw as A, bx as G, by as L, bz as O, bA as z, bB as F, D as M, $ as j, bC as J, bD as Q, bE as U, a2 as V, c as N, m as X, bF as Y, bG as Z, bH as R, bI as nn, bJ as tn, bK as sn, bL as en, bM as rn, bN as on, bO as an, bP as un, aG as cn, bQ as ln } from "./index-D0RBWjq8.js";
-import { k as C, c as g, m as D } from "./step-dXR33iOg.js";
-import { r as b } from "./reshape-CkjKPPqB.js";
-import { m as pn, a as hn, e as w } from "./log_sum_exp-VLZgbFAH.js";
-import { s as K } from "./sum-BdplSvq_.js";
+import { o as h, q as c, E as d, bo as T, bp as q, bq as H, x as l, br as P, L as _, bs as y, bt as B, bu as I, bv as W, bw as A, bx as G, by as L, bz as O, bA as z, bB as F, B as M, _ as j, bC as J, bD as Q, bE as U, a1 as V, c as N, m as X, bF as Y, bG as Z, bH as R, bI as nn, bJ as tn, bK as sn, bL as en, bM as rn, bN as on, bO as an, bP as un, aG as cn, bQ as ln } from "./index-Duu1Lvvv.js";
+import { k as C, c as g, m as D } from "./step-BS5JXRR6.js";
+import { r as b } from "./reshape-BI0yzp1T.js";
+import { m as pn, a as hn, e as w } from "./log_sum_exp-CVqLsVLl.js";
+import { s as K } from "./sum-BPUfDB2X.js";
 function fn(s, n = null, t = !1) {
-  const u = { x: c(s, "x", "all", "bool") }, o = { axis: n, keepDims: t };
-  return d.runKernel(T, u, o);
+  const i = { x: c(s, "x", "all", "bool") }, o = { axis: n, keepDims: t };
+  return d.runKernel(T, i, o);
 }
 const nt = /* @__PURE__ */ h({ all_: fn });
 function dn(s, n = null, t = !1) {
-  const u = { x: c(s, "x", "any", "bool") }, o = { axis: n, keepDims: t };
-  return d.runKernel(q, u, o);
+  const i = { x: c(s, "x", "any", "bool") }, o = { axis: n, keepDims: t };
+  return d.runKernel(q, i, o);
 }
 const tt = /* @__PURE__ */ h({ any_: dn });
 function mn(s, n = 0) {
-  const e = { x: c(s, "x", "argMax") }, u = { axis: n };
-  return d.runKernel(H, e, u);
+  const e = { x: c(s, "x", "argMax") }, i = { axis: n };
+  return d.runKernel(H, e, i);
 }
 const st = /* @__PURE__ */ h({ argMax_: mn });
-function $n(s, n, t, e, u) {
+function $n(s, n, t, e, i) {
   const o = c(s, "x", "avgPool", "float32"), p = 1;
   l(C(t, p), () => `Error in avgPool: Either strides or dilations must be 1. Got strides ${t} and dilations '${p}'`);
   let r = o, a = !1;
-  o.rank === 3 && (a = !0, r = b(o, [1, o.shape[0], o.shape[1], o.shape[2]])), l(r.rank === 4, () => `Error in avgPool: x must be rank 4 but got rank ${r.rank}.`), g("avgPool", e, u);
-  const i = { x: r }, m = { filterSize: n, strides: t, pad: e, dimRoundingMode: u };
-  let f = d.runKernel(P, i, m);
+  o.rank === 3 && (a = !0, r = b(o, [1, o.shape[0], o.shape[1], o.shape[2]])), l(r.rank === 4, () => `Error in avgPool: x must be rank 4 but got rank ${r.rank}.`), g("avgPool", e, i);
+  const u = { x: r }, m = { filterSize: n, strides: t, pad: e, dimRoundingMode: i };
+  let f = d.runKernel(P, u, m);
   return f = _(f, o.dtype), a ? b(f, [f.shape[1], f.shape[2], f.shape[3]]) : f;
 }
 const et = /* @__PURE__ */ h({ avgPool_: $n });
@@ -34,66 +34,66 @@ function bn(s) {
 }
 const rt = /* @__PURE__ */ h({ tanh_: bn });
 function xn(s, n, t) {
-  const e = c(s, "x", "batchToSpaceND"), u = n.reduce((r, a) => r * a);
-  l(e.rank >= 1 + n.length, () => `input rank is ${e.rank} but should be > than blockShape.length ${n.length}`), l(t.length === n.length, () => `crops.length is ${t.length} but should be equal to blockShape.length  ${n.length}`), l(e.shape[0] % u === 0, () => `input tensor batch is ${e.shape[0]} but is not divisible by the product of the elements of blockShape ${n.join(" * ")} === ${u}`);
+  const e = c(s, "x", "batchToSpaceND"), i = n.reduce((r, a) => r * a);
+  l(e.rank >= 1 + n.length, () => `input rank is ${e.rank} but should be > than blockShape.length ${n.length}`), l(t.length === n.length, () => `crops.length is ${t.length} but should be equal to blockShape.length  ${n.length}`), l(e.shape[0] % i === 0, () => `input tensor batch is ${e.shape[0]} but is not divisible by the product of the elements of blockShape ${n.join(" * ")} === ${i}`);
   const o = { x: e }, p = { blockShape: n, crops: t };
-  return d.runKernel(I, o, p);
+  return d.runKernel(B, o, p);
 }
 const ot = /* @__PURE__ */ h({ batchToSpaceND_: xn });
 function kn(s) {
   let n;
   return s.rank === 0 || s.rank === 1 ? n = b(s, [1, 1, 1, s.size]) : s.rank === 2 ? n = b(s, [1, 1, s.shape[0], s.shape[1]]) : s.rank === 3 ? n = b(s, [1, s.shape[0], s.shape[1], s.shape[2]]) : n = s, n;
 }
-function vn(s, n, t, e, u, o) {
+function vn(s, n, t, e, i, o) {
   o == null && (o = 1e-3);
   const p = c(s, "x", "batchNorm"), r = c(n, "mean", "batchNorm"), a = c(t, "variance", "batchNorm");
-  let i;
-  u != null && (i = c(u, "scale", "batchNorm"));
+  let u;
+  i != null && (u = c(i, "scale", "batchNorm"));
   let m;
-  e != null && (m = c(e, "offset", "batchNorm")), l(r.rank === a.rank, () => "Batch normalization gradient requires mean and variance to have equal ranks."), l(m == null || r.rank === m.rank, () => "Batch normalization gradient requires mean and offset to have equal ranks."), l(i == null || r.rank === i.rank, () => "Batch normalization gradient requires mean and scale to have equal ranks.");
+  e != null && (m = c(e, "offset", "batchNorm")), l(r.rank === a.rank, () => "Batch normalization gradient requires mean and variance to have equal ranks."), l(m == null || r.rank === m.rank, () => "Batch normalization gradient requires mean and offset to have equal ranks."), l(u == null || r.rank === u.rank, () => "Batch normalization gradient requires mean and scale to have equal ranks.");
   const x = {
     x: kn(p),
-    scale: i,
+    scale: u,
     offset: m,
     mean: r,
     variance: a
-  }, k = { varianceEpsilon: o }, $ = d.runKernel(W, x, k);
+  }, k = { varianceEpsilon: o }, $ = d.runKernel(I, x, k);
   return b($, p.shape);
 }
 const at = /* @__PURE__ */ h({ batchNorm_: vn });
-function gn(s, n, t, e, u = "NHWC", o = [1, 1], p) {
+function gn(s, n, t, e, i = "NHWC", o = [1, 1], p) {
   const r = c(s, "x", "conv2d", "float32"), a = c(n, "filter", "conv2d", "float32");
-  let i = r, m = !1;
-  r.rank === 3 && (m = !0, i = b(r, [1, r.shape[0], r.shape[1], r.shape[2]])), l(i.rank === 4, () => `Error in conv2d: input must be rank 4, but got rank ${i.rank}.`), l(a.rank === 4, () => `Error in conv2d: filter must be rank 4, but got rank ${a.rank}.`), g("conv2d", e, p);
-  const f = u === "NHWC" ? i.shape[3] : i.shape[1];
+  let u = r, m = !1;
+  r.rank === 3 && (m = !0, u = b(r, [1, r.shape[0], r.shape[1], r.shape[2]])), l(u.rank === 4, () => `Error in conv2d: input must be rank 4, but got rank ${u.rank}.`), l(a.rank === 4, () => `Error in conv2d: filter must be rank 4, but got rank ${a.rank}.`), g("conv2d", e, p);
+  const f = i === "NHWC" ? u.shape[3] : u.shape[1];
   l(f === a.shape[2], () => `Error in conv2d: depth of input (${f}) must match input depth for filter ${a.shape[2]}.`), l(C(t, o), () => `Error in conv2D: Either strides or dilations must be 1. Got strides ${t} and dilations '${o}'`), l(D(o), () => "Error in conv2D: Dilated rates should be larger than 0."), l(D(t), () => "Error in conv2D: Strides should be larger than 0.");
-  const x = { x: i, filter: a }, k = { strides: t, pad: e, dataFormat: u, dilations: o, dimRoundingMode: p }, $ = d.runKernel(B, x, k);
+  const x = { x: u, filter: a }, k = { strides: t, pad: e, dataFormat: i, dilations: o, dimRoundingMode: p }, $ = d.runKernel(W, x, k);
   return m ? b($, [$.shape[1], $.shape[2], $.shape[3]]) : $;
 }
 const S = /* @__PURE__ */ h({ conv2d_: gn });
-function Dn(s, n, t, e, u = "NWC", o = 1, p) {
+function Dn(s, n, t, e, i = "NWC", o = 1, p) {
   const r = c(s, "x", "conv1d"), a = c(n, "filter", "conv1d");
-  let i = r, m = !1;
-  r.rank === 2 && (m = !0, i = b(r, [1, r.shape[0], r.shape[1]])), l(i.rank === 3, () => `Error in conv1d: input must be rank 3, but got rank ${i.rank}.`), l(a.rank === 3, () => `Error in conv1d: filter must be rank 3, but got rank ${a.rank}.`), g("conv1d", e, p), l(i.shape[2] === a.shape[1], () => `Error in conv1d: depth of input (${i.shape[2]}) must match input depth for filter ${a.shape[1]}.`), l(C(t, o), () => `Error in conv1D: Either stride or dilation must be 1. Got stride ${t} and dilation '${o}'`), l(D(o), () => "Error in conv1D: Dilated rates should be larger than 0."), l(D(t), () => "Error in conv1D: Stride should be larger than 0."), l(u === "NWC", () => `Error in conv1d: got dataFormat of ${u} but only NWC is currently supported.`);
-  const f = b(a, [1, a.shape[0], a.shape[1], a.shape[2]]), x = b(i, [i.shape[0], 1, i.shape[1], i.shape[2]]), v = S(x, f, [1, t], e, "NHWC", [1, o], p);
+  let u = r, m = !1;
+  r.rank === 2 && (m = !0, u = b(r, [1, r.shape[0], r.shape[1]])), l(u.rank === 3, () => `Error in conv1d: input must be rank 3, but got rank ${u.rank}.`), l(a.rank === 3, () => `Error in conv1d: filter must be rank 3, but got rank ${a.rank}.`), g("conv1d", e, p), l(u.shape[2] === a.shape[1], () => `Error in conv1d: depth of input (${u.shape[2]}) must match input depth for filter ${a.shape[1]}.`), l(C(t, o), () => `Error in conv1D: Either stride or dilation must be 1. Got stride ${t} and dilation '${o}'`), l(D(o), () => "Error in conv1D: Dilated rates should be larger than 0."), l(D(t), () => "Error in conv1D: Stride should be larger than 0."), l(i === "NWC", () => `Error in conv1d: got dataFormat of ${i} but only NWC is currently supported.`);
+  const f = b(a, [1, a.shape[0], a.shape[1], a.shape[2]]), x = b(u, [u.shape[0], 1, u.shape[1], u.shape[2]]), v = S(x, f, [1, t], e, "NHWC", [1, o], p);
   return m ? b(v, [v.shape[2], v.shape[3]]) : b(v, [v.shape[0], v.shape[2], v.shape[3]]);
 }
-const ut = /* @__PURE__ */ h({ conv1d_: Dn });
-function Cn(s, n, t, e, u, o = "NHWC", p) {
+const it = /* @__PURE__ */ h({ conv1d_: Dn });
+function Cn(s, n, t, e, i, o = "NHWC", p) {
   l(s.length === n.rank, () => `Length of inShape (${s.length}) and rank of dy (${n.rank}) must match`);
-  let r = s, a = n, i = !1;
-  n.rank === 3 && (i = !0, a = b(n, [1, n.shape[0], n.shape[1], n.shape[2]]), r = [1, s[0], s[1], s[2]]), l(r.length === 4, () => `Error in conv2dDerInput: inShape must be length 4, but got length ${r.length}.`), l(a.rank === 4, () => `Error in conv2dDerInput: dy must be rank 4, but got rank ${a.rank}`), l(t.rank === 4, () => `Error in conv2dDerInput: filter must be rank 4, but got rank ${t.rank}`);
+  let r = s, a = n, u = !1;
+  n.rank === 3 && (u = !0, a = b(n, [1, n.shape[0], n.shape[1], n.shape[2]]), r = [1, s[0], s[1], s[2]]), l(r.length === 4, () => `Error in conv2dDerInput: inShape must be length 4, but got length ${r.length}.`), l(a.rank === 4, () => `Error in conv2dDerInput: dy must be rank 4, but got rank ${a.rank}`), l(t.rank === 4, () => `Error in conv2dDerInput: filter must be rank 4, but got rank ${t.rank}`);
   const m = o === "NHWC" ? r[3] : r[1], f = o === "NHWC" ? a.shape[3] : a.shape[1];
-  l(m === t.shape[2], () => `Error in conv2dDerInput: depth of input (${m}) must match input depth for filter ${t.shape[2]}.`), l(f === t.shape[3], () => `Error in conv2dDerInput: depth of output (${f}) must match output depth for filter ${t.shape[3]}.`), g("conv2dDerInput", u, p);
-  const x = { dy: a, filter: t }, k = { strides: e, pad: u, dataFormat: o, dimRoundingMode: p, inputShape: r }, $ = d.runKernel(A, x, k);
-  return i ? b($, [$.shape[1], $.shape[2], $.shape[3]]) : $;
+  l(m === t.shape[2], () => `Error in conv2dDerInput: depth of input (${m}) must match input depth for filter ${t.shape[2]}.`), l(f === t.shape[3], () => `Error in conv2dDerInput: depth of output (${f}) must match output depth for filter ${t.shape[3]}.`), g("conv2dDerInput", i, p);
+  const x = { dy: a, filter: t }, k = { strides: e, pad: i, dataFormat: o, dimRoundingMode: p, inputShape: r }, $ = d.runKernel(A, x, k);
+  return u ? b($, [$.shape[1], $.shape[2], $.shape[3]]) : $;
 }
 const En = /* @__PURE__ */ h({ conv2DBackpropInput_: Cn });
-function Nn(s, n, t, e, u, o) {
+function Nn(s, n, t, e, i, o) {
   const p = c(s, "x", "conv2dTranspose"), r = c(n, "filter", "conv2dTranspose");
-  return En(t, p, r, e, u, "NHWC", o);
+  return En(t, p, r, e, i, "NHWC", o);
 }
-const it = /* @__PURE__ */ h({ conv2dTranspose_: Nn });
+const ut = /* @__PURE__ */ h({ conv2dTranspose_: Nn });
 function _n(s) {
   const t = { x: c(s, "x", "cos", "float32") };
   return d.runKernel(G, t);
@@ -114,21 +114,21 @@ function Sn(s, n = 0, t = !1, e = !1) {
   return d.runKernel(z, o, p);
 }
 const ht = /* @__PURE__ */ h({ cumsum_: Sn });
-function Tn(s, n, t, e, u = "NHWC", o = [1, 1], p) {
+function Tn(s, n, t, e, i = "NHWC", o = [1, 1], p) {
   const r = c(s, "x", "depthwiseConv2d", "float32"), a = c(n, "filter", "depthwiseConv2d", "float32");
-  let i = r, m = !1;
-  r.rank === 3 && (m = !0, i = b(r, [1, r.shape[0], r.shape[1], r.shape[2]])), l(i.rank === 4, () => `Error in depthwiseConv2d: input must be rank 4, but got rank ${i.rank}.`), l(a.rank === 4, () => `Error in depthwiseConv2d: filter must be rank 4, but got rank ${a.rank}.`);
-  const f = u === "NHWC" ? i.shape[3] : i.shape[1];
+  let u = r, m = !1;
+  r.rank === 3 && (m = !0, u = b(r, [1, r.shape[0], r.shape[1], r.shape[2]])), l(u.rank === 4, () => `Error in depthwiseConv2d: input must be rank 4, but got rank ${u.rank}.`), l(a.rank === 4, () => `Error in depthwiseConv2d: filter must be rank 4, but got rank ${a.rank}.`);
+  const f = i === "NHWC" ? u.shape[3] : u.shape[1];
   l(f === a.shape[2], () => `Error in depthwiseConv2d: number of input channels (${f}) must match the inChannels dimension in filter ${a.shape[2]}.`), g("depthwiseConv2d", e, p);
-  const x = { x: i, filter: a }, k = { strides: t, pad: e, dataFormat: u, dilations: o, dimRoundingMode: p }, $ = d.runKernel(F, x, k);
+  const x = { x: u, filter: a }, k = { strides: t, pad: e, dataFormat: i, dilations: o, dimRoundingMode: p }, $ = d.runKernel(F, x, k);
   return m ? b($, [$.shape[1], $.shape[2], $.shape[3]]) : $;
 }
 const qn = /* @__PURE__ */ h({ depthwiseConv2d_: Tn });
 function Hn(s, n) {
   let t = c(s, "a", "equal", "string_or_numeric"), e = c(n, "b", "equal", "string_or_numeric");
   [t, e] = M(t, e), j(t.shape, e.shape);
-  const u = { a: t, b: e };
-  return d.runKernel(J, u);
+  const i = { a: t, b: e };
+  return d.runKernel(J, i);
 }
 const ft = /* @__PURE__ */ h({ equal_: Hn });
 function Pn(s) {
@@ -143,36 +143,36 @@ function yn(s) {
   return d.runKernel(U, t);
 }
 const mt = /* @__PURE__ */ h({ softplus_: yn });
-function In(s, n = -1) {
+function Bn(s, n = -1) {
   const t = c(s, "logits", "logSoftmax");
   if (n === -1 && (n = t.rank - 1), n !== t.rank - 1)
     throw Error(`Log Softmax along a non-last dimension is not yet supported. Logits was rank ${t.rank} and axis was ${n}`);
-  return V((u, o) => {
-    const r = pn(u, n, !0), a = N(u, r), i = N(_(a, "float32"), hn(K(w(a), n, !0)));
-    return o([i]), { value: i, gradFunc: (f, x) => {
+  return V((i, o) => {
+    const r = pn(i, n, !0), a = N(i, r), u = N(_(a, "float32"), hn(K(w(a), n, !0)));
+    return o([u]), { value: u, gradFunc: (f, x) => {
       const [k] = x, $ = !0, E = w(k);
       return N(f, X(K(f, n, $), E));
     } };
   })(t);
 }
-const $t = /* @__PURE__ */ h({ logSoftmax_: In });
-function Wn(s) {
+const $t = /* @__PURE__ */ h({ logSoftmax_: Bn });
+function In(s) {
   const t = { x: c(s, "x", "logicalNot", "bool") };
   return d.runKernel(Y, t);
 }
-const bt = /* @__PURE__ */ h({ logicalNot_: Wn });
-function Bn(s, n, t, e, u) {
+const bt = /* @__PURE__ */ h({ logicalNot_: In });
+function Wn(s, n, t, e, i) {
   const o = c(s, "x", "maxPool"), p = 1;
   let r = o, a = !1;
-  o.rank === 3 && (a = !0, r = b(o, [1, o.shape[0], o.shape[1], o.shape[2]])), l(r.rank === 4, () => `Error in maxPool: input must be rank 4 but got rank ${r.rank}.`), l(C(t, p), () => `Error in maxPool: Either strides or dilations must be 1. Got strides ${t} and dilations '${p}'`), g("maxPool", e, u);
-  const i = { x: r }, m = { filterSize: n, strides: t, pad: e, dimRoundingMode: u }, f = d.runKernel(Z, i, m);
+  o.rank === 3 && (a = !0, r = b(o, [1, o.shape[0], o.shape[1], o.shape[2]])), l(r.rank === 4, () => `Error in maxPool: input must be rank 4 but got rank ${r.rank}.`), l(C(t, p), () => `Error in maxPool: Either strides or dilations must be 1. Got strides ${t} and dilations '${p}'`), g("maxPool", e, i);
+  const u = { x: r }, m = { filterSize: n, strides: t, pad: e, dimRoundingMode: i }, f = d.runKernel(Z, u, m);
   return a ? b(f, [f.shape[1], f.shape[2], f.shape[3]]) : f;
 }
-const xt = /* @__PURE__ */ h({ maxPool_: Bn });
-function An(s, n, t = 1, e = 0, u = "int32") {
+const xt = /* @__PURE__ */ h({ maxPool_: Wn });
+function An(s, n, t = 1, e = 0, i = "int32") {
   if (n < 2)
     throw new Error(`Error in oneHot: depth must be >=2, but it is ${n}`);
-  const p = { indices: c(s, "indices", "oneHot", "int32") }, r = { dtype: u, depth: n, onValue: t, offValue: e };
+  const p = { indices: c(s, "indices", "oneHot", "int32") }, r = { dtype: i, depth: n, onValue: t, offValue: e };
   return d.runKernel(R, p, r);
 }
 const kt = /* @__PURE__ */ h({ oneHot_: An });
@@ -185,20 +185,20 @@ function Ln(s, n, t = 0) {
   const e = c(s, "x", "pad");
   if (e.rank === 0)
     throw new Error("pad(scalar) is not defined. Pass non-scalar to pad");
-  const u = { paddings: n, constantValue: t }, o = { x: e };
-  return d.runKernel(tn, o, u);
+  const i = { paddings: n, constantValue: t }, o = { x: e };
+  return d.runKernel(tn, o, i);
 }
 const gt = /* @__PURE__ */ h({ pad_: Ln });
 function On(s, n, t) {
   const e = c(s, "x", "spaceToBatchND");
   l(e.rank >= 1 + n.length, () => `input rank ${e.rank} should be > than [blockShape] ${n.length}`), l(t.length === n.length, () => `paddings.shape[0] ${t.length} must be equal to [blockShape] ${n.length}`), l(e.shape.reduce((p, r, a) => a > 0 && a <= n.length ? p && (r + t[a - 1][0] + t[a - 1][1]) % n[a - 1] === 0 : p, !0), () => `input spatial dimensions ${e.shape.slice(1)} with paddings ${t.toString()} must be divisible by blockShapes ${n.toString()}`);
-  const u = { x: e }, o = { blockShape: n, paddings: t };
-  return d.runKernel(sn, u, o);
+  const i = { x: e }, o = { blockShape: n, paddings: t };
+  return d.runKernel(sn, i, o);
 }
 const Dt = /* @__PURE__ */ h({ spaceToBatchND_: On });
 function zn(s, n) {
-  const e = { x: c(s, "x", "reverse") }, u = { dims: n };
-  return d.runKernel(en, e, u);
+  const e = { x: c(s, "x", "reverse") }, i = { dims: n };
+  return d.runKernel(en, e, i);
 }
 const Ct = /* @__PURE__ */ h({ reverse_: zn });
 function Fn(s) {
@@ -211,15 +211,15 @@ function Mn(s) {
   return d.runKernel(on, t);
 }
 const Nt = /* @__PURE__ */ h({ selu_: Mn });
-function jn(s, n, t, e, u, o = [1, 1], p = "NHWC") {
-  const r = c(s, "x", "separableConv2d"), a = c(n, "depthwiseFilter", "separableConv2d"), i = c(t, "pointwiseFilter", "separableConv2d");
+function jn(s, n, t, e, i, o = [1, 1], p = "NHWC") {
+  const r = c(s, "x", "separableConv2d"), a = c(n, "depthwiseFilter", "separableConv2d"), u = c(t, "pointwiseFilter", "separableConv2d");
   let m = r, f = !1;
   if (r.rank === 3 && (f = !0, m = b(r, [1, r.shape[0], r.shape[1], r.shape[2]])), p === "NCHW")
     throw new Error("separableConv2d currently does not support dataFormat NCHW; only NHWC is supported");
-  l(m.rank === 4, () => `Error in separableConv2d: input must be rank 4, but got rank ${m.rank}.`), l(a.rank === 4, () => `Error in separableConv2d: depthwise filter must be rank 4, but got rank ${a.rank}.`), l(i.rank === 4, () => `Error in separableConv2d: pointwise filter must be rank 4, but got rank ${a.rank}.`), l(i.shape[0] === 1, () => `Error in separableConv2d: the first dimension of pointwise filter  must be 1, but got ${i.shape[0]}.`), l(i.shape[1] === 1, () => `Error in separableConv2d: the second dimension of pointwise filter must be 1, but got ${i.shape[1]}.`);
+  l(m.rank === 4, () => `Error in separableConv2d: input must be rank 4, but got rank ${m.rank}.`), l(a.rank === 4, () => `Error in separableConv2d: depthwise filter must be rank 4, but got rank ${a.rank}.`), l(u.rank === 4, () => `Error in separableConv2d: pointwise filter must be rank 4, but got rank ${a.rank}.`), l(u.shape[0] === 1, () => `Error in separableConv2d: the first dimension of pointwise filter  must be 1, but got ${u.shape[0]}.`), l(u.shape[1] === 1, () => `Error in separableConv2d: the second dimension of pointwise filter must be 1, but got ${u.shape[1]}.`);
   const x = a.shape[2], k = a.shape[3];
-  l(i.shape[2] === x * k, () => `Error in separableConv2d: the third dimension of pointwise filter must be ${x * k}, but got ${i.shape[2]}.`);
-  const $ = qn(m, a, e, u, p, o), v = S($, i, 1, "valid", p);
+  l(u.shape[2] === x * k, () => `Error in separableConv2d: the third dimension of pointwise filter must be ${x * k}, but got ${u.shape[2]}.`);
+  const $ = qn(m, a, e, i, p, o), v = S($, u, 1, "valid", p);
   return f ? b(v, [v.shape[1], v.shape[2], v.shape[3]]) : v;
 }
 const _t = /* @__PURE__ */ h({ separableConv2d_: jn });
@@ -234,9 +234,9 @@ function Qn(s) {
 }
 const Kt = /* @__PURE__ */ h({ sinh_: Qn });
 function Un(s, n, t) {
-  const e = c(s, "x", "unsortedSegmentSum"), u = c(n, "segmentIds", "unsortedSegmentSum", "int32");
+  const e = c(s, "x", "unsortedSegmentSum"), i = c(n, "segmentIds", "unsortedSegmentSum", "int32");
   l(cn(t), () => "numSegments must be of dtype int");
-  const o = { x: e, segmentIds: u }, p = { numSegments: t };
+  const o = { x: e, segmentIds: i }, p = { numSegments: t };
   return d.runKernel(ln, o, p);
 }
 const St = /* @__PURE__ */ h({ unsortedSegmentSum_: Un });
@@ -258,10 +258,10 @@ export {
   tt as h,
   st as i,
   at as j,
-  ut as k,
+  it as k,
   bt as l,
   xt as m,
-  it as n,
+  ut as n,
   S as o,
   lt as p,
   pt as q,

package/dist/utilities/dummy.js

@@ -1,6 +1,6 @@
-import { a as y, e as S, v as w } from "../index-D0RBWjq8.js";
-import { z as m } from "../zeros-DeiE2zTa.js";
-import { o as P } from "../ones-BX_wEgzB.js";
+import { a as y, e as S, v as w } from "../index-Duu1Lvvv.js";
+import { z as m } from "../zeros-5YROwwUH.js";
+import { o as P } from "../ones-CBI1AQjb.js";
 async function b(s) {
   const t = m([1, s.config.blockSize], "int32"), [n, o] = s.forward({ training: !1 }, t);
   await n.data(), n.dispose(), o && o.dispose(), t.dispose();

package/dist/utilities/multinomialCPU.js

@@ -1,5 +1,5 @@
-import "../index-D0RBWjq8.js";
-import { t as e } from "../tensor2d-BN1sSfQO.js";
+import "../index-Duu1Lvvv.js";
+import { t as e } from "../tensor2d-BN97fF71.js";
 function l(n) {
   let r = 0;
   const i = Math.random();

package/dist/utilities/packed.js

@@ -1,4 +1,4 @@
-import { e as n } from "../index-D0RBWjq8.js";
+import { e as n } from "../index-Duu1Lvvv.js";
 function o() {
   return n().backendName === "webgpu";
 }

package/dist/utilities/performance.js

@@ -1,4 +1,4 @@
-import { t as s } from "../index-D0RBWjq8.js";
+import { t as s } from "../index-Duu1Lvvv.js";
 async function f(e, o = 10, r = !1) {
   for (let t = 0; t < 100; t++) {
     const a = r ? await e() : s(e);

package/dist/utilities/profile.js

@@ -1,4 +1,4 @@
-import { a } from "../index-D0RBWjq8.js";
+import { a } from "../index-Duu1Lvvv.js";
 const s = 1024 * 1024;
 class l {
   log = /* @__PURE__ */ new Map();

package/dist/utilities/safetensors.js

@@ -1,5 +1,5 @@
-import "../index-D0RBWjq8.js";
-import { t as y } from "../tensor-BQqrDvpx.js";
+import "../index-Duu1Lvvv.js";
+import { t as y } from "../tensor-CEt9Nm2s.js";
 function l(t) {
   if (t === "float32") return "F32";
   if (t === "int32") return "I32";

package/dist/utilities/sentences.d.ts

@@ -1,5 +1,5 @@
 import { default as TeachableLLM } from '../TeachableLLM';
 import { Tensor2D, Tensor3D } from '@tensorflow/tfjs-core';
 export declare function meanPooling(embeddings: Tensor3D, attentionMask?: Tensor2D): Tensor2D;
-export declare function sentenceEmbeddingsTensor(model: TeachableLLM, sentences: string[], batchSize?: number): Promise<Tensor2D>;
+export declare function sentenceEmbeddingsTensor(model: TeachableLLM, sentences: string[], batchSize?: number): Tensor2D;
 export declare function sentenceEmbeddings(model: TeachableLLM, sentences: string[], batchSize?: number): Promise<number[][]>;

package/dist/utilities/sentences.js

@@ -1,26 +1,26 @@
-import { m as w } from "../index-D0RBWjq8.js";
-import { t as g } from "../tensor2d-BN1sSfQO.js";
-import { e as y } from "../expand_dims-BzfJK2uc.js";
-import { s as h } from "../sum-BdplSvq_.js";
-import { c as T } from "../concat-C6X3AAlQ.js";
+import { m as w } from "../index-Duu1Lvvv.js";
+import { t as f } from "../tensor2d-BN97fF71.js";
+import { e as y } from "../expand_dims-ChkuOp6I.js";
+import { s as g } from "../sum-BPUfDB2X.js";
+import { c as T } from "../concat-CSm2rMwe.js";
 const p = 16;
 function A(o, t) {
   if (!t)
     return o.mean(1);
-  const r = y(t, 2), i = w(o, r), e = h(i, 1), s = h(t, 1, !0), c = e.div(s.maximum(1e-9));
+  const r = y(t, 2), i = w(o, r), e = g(i, 1), s = g(t, 1, !0), c = e.div(s.maximum(1e-9));
   return r.dispose(), i.dispose(), e.dispose(), s.dispose(), c;
 }
-async function E(o, t, r = p) {
+function E(o, t, r = p) {
   const i = o.tokeniser, e = o.config.blockSize;
   let s = null, c = 0;
   for (; c < t.length; ) {
-    const m = t.slice(c, c + p), k = await i.tokenise(m, !0), l = [], d = [];
-    for (const n of k)
+    const b = t.slice(c, c + p).map((n) => i.encode(n)), l = [], d = [];
+    for (const n of b)
       n.length > e ? (l.push(n.slice(n.length - e, n.length)), d.push(new Array(e).fill(1))) : n.length < e ? (l.push(n.concat(new Array(e - n.length).fill(0))), d.push(
         new Array(n.length).fill(1).concat(new Array(e - n.length).fill(0))
       )) : (l.push(n), d.push(new Array(e).fill(1)));
-    const b = g(l, [l.length, e], "int32"), u = g(d, [d.length, e], "float32"), f = o.model.forward({ skipLogits: !0, training: !1 }, b)[0], a = A(f, u);
-    if (u.dispose(), f.dispose(), s === null)
+    const k = f(l, [l.length, e], "int32"), m = f(d, [d.length, e], "float32"), u = o.model.forward({ skipLogits: !0, training: !1 }, k)[0], a = A(u, m);
+    if (m.dispose(), u.dispose(), s === null)
       s = a;
     else {
       const n = s;

package/dist/utilities/weights.js

@@ -1,5 +1,5 @@
-import "../index-D0RBWjq8.js";
-import { t as p } from "../tensor-BQqrDvpx.js";
+import "../index-Duu1Lvvv.js";
+import { t as p } from "../tensor-CEt9Nm2s.js";
 function h(n) {
   const e = n.reduce((s, o) => s + o.length, 0), a = new Float32Array(e);
   let t = 0;

package/dist/{variable-DQ9yYgEU.js → variable-DPt_Iuog.js}

@@ -1,4 +1,4 @@
-import { E as i } from "./index-D0RBWjq8.js";
+import { E as i } from "./index-Duu1Lvvv.js";
 function m(r, a = !0, e, t) {
   return i.makeVariable(r, a, e, t);
 }

package/dist/{webgpu_program-CAE4RICo.js → webgpu_program-BpWRlghH.js}

@@ -1,4 +1,4 @@
-import { ad as z, ac as F, ab as E, a9 as j, y as A } from "./index-D0RBWjq8.js";
+import { ac as z, ab as F, aa as E, a8 as j, x as A } from "./index-Duu1Lvvv.js";
 function L(t, s) {
   if (Math.max(...t) > 5)
     throw new Error("Cannot symbolically compute strides for rank > 6 tensor.");

package/dist/{webgpu_util-BdovYhXr.js → webgpu_util-DMiKzzQM.js}

@@ -1,4 +1,4 @@
-import { y as u } from "./index-D0RBWjq8.js";
+import { x as u } from "./index-Duu1Lvvv.js";
 const c = (r) => {
   let t = 1;
   for (let n = 0; n < r.length; n++)
@@ -23,16 +23,16 @@ function p(r, t, n = !1) {
   const a = c(r.x.map((i) => t[i])), o = c(r.y.map((i) => t[i]));
   return a <= 4 ? [4, 16, 1] : o <= 4 ? [16, 4, 1] : [16, 16, 1];
 }
-function M(r, t, n = !1) {
+function x(r, t, n = !1) {
   if (n)
     return [4, 4, 1];
   const a = c(r.x.map((i) => t[i])), o = c(r.y.map((i) => t[i]));
   return a <= 4 ? [1, 2, 1] : o <= 4 ? [2, 1, 1] : [2, 2, 1];
 }
-function h(r) {
+function M(r) {
   return { x: r.map((t, n) => n) };
 }
-function x(r) {
+function h(r) {
   if (r === "float32" || r === "int32" || r === "bool" || r === "string" || r === "packedF16")
     return 4;
   if (r === "complex64")
@@ -52,13 +52,13 @@ var s;
   r[r.MatMulReduceProgram = 0] = "MatMulReduceProgram", r[r.MatMulSplitKProgram = 1] = "MatMulSplitKProgram", r[r.MatMulSmallOutputSizeProgram = 2] = "MatMulSmallOutputSizeProgram", r[r.MatMulPackedProgram = 3] = "MatMulPackedProgram", r[r.MatMulMax = 4] = "MatMulMax";
 })(s || (s = {}));
 export {
-  x as G,
+  h as G,
   s as M,
   d as a,
   b,
   m as c,
   p as d,
-  M as e,
-  h as f,
+  x as e,
+  M as f,
   g as i
 };

package/dist/{zeros-DeiE2zTa.js → zeros-5YROwwUH.js}

@@ -1,5 +1,5 @@
-import { w as n, U as m, V as i, E as c } from "./index-D0RBWjq8.js";
-import { c as f } from "./complex-DClmWqJt.js";
+import { u as n, Q as m, U as i, E as c } from "./index-Duu1Lvvv.js";
+import { c as f } from "./complex-Cyg-eQeZ.js";
 function e(o, r = "float32") {
   if (n(o), r === "complex64") {
     const s = e(o, "float32"), t = e(o, "float32");