@genai-fi/nanogpt 0.19.0 → 0.20.0

package/dist/training/validation.js

@@ -1,84 +0,0 @@
-import "../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 "../index-DvYrXKkX.js";
-import "../ops/cpu/adamAdjust.js";
-import "../ops/webgl/adamAdjust.js";
-import "../ops/cpu/adamMoments.js";
-import "../ops/webgl/adamMoments.js";
-import "../ops/cpu/gatherSub.js";
-import "../ops/webgl/gatherSub.js";
-import "../ops/cpu/scatterSub.js";
-import "../ops/webgl/scatterSub.js";
-import { PAGE_FACTOR as a, shuffle as k } from "./DatasetBuilder.js";
-import "../papaparse.min-C0cScC2i.js";
-import { tokensFromTasks as y } from "./tasks/Task.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";
-async function Mt(n, x, o, l, c = 0.1, d) {
-  const r = n instanceof Uint16Array ? n : await y(n, x, void 0, d), i = r instanceof Uint16Array ? r : r.tokens, f = r instanceof Uint16Array ? void 0 : r.mask, m = /* @__PURE__ */ new Set();
-  if (c > 0) {
-    const t = Math.floor(i.length / (o.blockSize * a)), s = Math.max(1, Math.floor(t * c));
-    for (; m.size < s; ) {
-      const M = Math.floor(Math.random() * t);
-      m.add(M);
-    }
-  }
-  const p = new Uint32Array(
-    i.length - m.size * o.blockSize * a
-  ), e = new Uint32Array(m.size * o.blockSize * a);
-  let h = 0, g = 0;
-  for (let t = 0; t < i.length; t++) {
-    const s = Math.floor(t / (o.blockSize * a));
-    m.has(s) ? g < e.length && (e[g++] = t) : h < p.length && (p[h++] = t);
-  }
-  const { dataset: v, state: w } = await o.createTextDataset(
-    i,
-    l,
-    k(p),
-    f || void 0
-  ), { dataset: z, state: A } = await o.createTextDataset(
-    i,
-    l,
-    k(e)
-  );
-  return { trainDataset: v, validationDataset: z, size: i.length, validationState: A, trainState: w };
-}
-export {
-  Mt as createTrainValidationSplit
-};

package/dist/transpose-9kRxIXWR.js

@@ -1,36 +0,0 @@
-import { o as u, q as i, E as o, ao as $, ap as g, aq as x, x as l, t as m, ar as p } from "./index-CUXkjxiT.js";
-import { c as k } from "./complex-DI35Q-gW.js";
-function K(r) {
-  const e = { input: i(r, "input", "imag") };
-  return o.runKernel($, e);
-}
-const h = /* @__PURE__ */ u({ imag_: K });
-function E(r) {
-  const e = { x: i(r, "x", "neg") };
-  return o.runKernel(g, e);
-}
-const _ = /* @__PURE__ */ u({ neg_: E });
-function b(r) {
-  const e = { input: i(r, "input", "real") };
-  return o.runKernel(x, e);
-}
-const d = /* @__PURE__ */ u({ real_: b });
-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" ? 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 v = /* @__PURE__ */ u({ transpose_: N });
-export {
-  h as i,
-  _ as n,
-  d as r,
-  v as t
-};

package/dist/unsorted_segment_sum-DJvk5xnh.js

@@ -1,277 +0,0 @@
-import { o as h, q as c, E as d, bm as T, bn as q, bo as H, x as l, bp as P, L as _, bq as y, br as B, bs as I, bt as W, bu as A, bv as L, bw as G, bx as O, by as z, bz as F, B as M, _ as j, bA as J, bB as U, bC as V, a2 as Q, a1 as N, m as X, bD as Y, bE as Z, bF as R, bG as nn, bH as tn, bI as sn, bJ as en, bK as rn, bL as on, bM as an, bN as un, aE as cn, bO as ln } from "./index-CUXkjxiT.js";
-import { k as C, c as g, m as D } from "./step-BTxPtq1r.js";
-import { r as b } from "./reshape-BDOuCSNW.js";
-import { m as pn, b as hn, e as w } from "./not_equal-hurPF26l.js";
-import { s as K } from "./sum-BpiwSWvg.js";
-function fn(s, n = null, t = !1) {
-  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 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") }, i = { axis: n };
-  return d.runKernel(H, e, i);
-}
-const st = /* @__PURE__ */ h({ argMax_: mn });
-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, 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 });
-function bn(s) {
-  const t = { x: c(s, "x", "tanh", "float32") };
-  return d.runKernel(y, t);
-}
-const rt = /* @__PURE__ */ h({ tanh_: bn });
-function xn(s, n, t) {
-  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(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, i, o) {
-  o == null && (o = 1e-3);
-  const p = c(s, "x", "batchNorm"), r = c(n, "mean", "batchNorm"), a = c(t, "variance", "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(u == null || r.rank === u.rank, () => "Batch normalization gradient requires mean and scale to have equal ranks.");
-  const x = {
-    x: kn(p),
-    scale: u,
-    offset: m,
-    mean: r,
-    variance: a
-  }, k = { varianceEpsilon: o }, $ = d.runKernel(I, x, k);
-  return b($, p.shape);
-}
-const at = /* @__PURE__ */ h({ batchNorm_: vn });
-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 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: 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, i = "NWC", o = 1, p) {
-  const r = c(s, "x", "conv1d"), a = c(n, "filter", "conv1d");
-  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 ${i} but only NWC is currently supported.`);
-  const f = b(a, [1, a.shape[0], a.shape[1], a.shape[2]]), x = b(u, [u.shape[0], 1, u.shape[1], u.shape[2]]), v = S(x, f, [1, t], e, "NHWC", [1, o], p);
-  return m ? b(v, [v.shape[2], v.shape[3]]) : b(v, [v.shape[0], v.shape[2], v.shape[3]]);
-}
-const it = /* @__PURE__ */ h({ conv1d_: Dn });
-function Cn(s, n, t, e, i, o = "NHWC", p) {
-  l(s.length === n.rank, () => `Length of inShape (${s.length}) and rank of dy (${n.rank}) must match`);
-  let r = s, a = n, 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", 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, i, o) {
-  const p = c(s, "x", "conv2dTranspose"), r = c(n, "filter", "conv2dTranspose");
-  return En(t, p, r, e, i, "NHWC", o);
-}
-const ut = /* @__PURE__ */ h({ conv2dTranspose_: Nn });
-function _n(s) {
-  const t = { x: c(s, "x", "cos", "float32") };
-  return d.runKernel(L, t);
-}
-const ct = /* @__PURE__ */ h({ cos_: _n });
-function wn(s) {
-  const t = { x: c(s, "x", "cosh", "float32") };
-  return d.runKernel(G, t);
-}
-const lt = /* @__PURE__ */ h({ cosh_: wn });
-function Kn(s, n = 0, t = !1, e = !1) {
-  const o = { x: c(s, "x", "cumprod") }, p = { axis: n, exclusive: t, reverse: e };
-  return d.runKernel(O, o, p);
-}
-const pt = /* @__PURE__ */ h({ cumprod_: Kn });
-function Sn(s, n = 0, t = !1, e = !1) {
-  const o = { x: c(s, "x", "cumsum") }, p = { axis: n, exclusive: t, reverse: e };
-  return d.runKernel(z, o, p);
-}
-const ht = /* @__PURE__ */ h({ cumsum_: Sn });
-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 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: 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 i = { a: t, b: e };
-  return d.runKernel(J, i);
-}
-const ft = /* @__PURE__ */ h({ equal_: Hn });
-function Pn(s) {
-  let n = c(s, "x", "erf");
-  l(n.dtype === "int32" || n.dtype === "float32", () => "Input dtype must be `int32` or `float32`."), n.dtype === "int32" && (n = _(n, "float32"));
-  const t = { x: n };
-  return d.runKernel(U, t);
-}
-const dt = /* @__PURE__ */ h({ erf_: Pn });
-function yn(s) {
-  const t = { x: c(s, "x", "softplus") };
-  return d.runKernel(V, t);
-}
-const mt = /* @__PURE__ */ h({ softplus_: yn });
-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 Q((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_: Bn });
-function In(s) {
-  const t = { x: c(s, "x", "logicalNot", "bool") };
-  return d.runKernel(Y, t);
-}
-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, 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_: 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: i, depth: n, onValue: t, offValue: e };
-  return d.runKernel(R, p, r);
-}
-const kt = /* @__PURE__ */ h({ oneHot_: An });
-function Ln(s) {
-  const t = { x: c(s, "x", "onesLike") };
-  return d.runKernel(nn, t);
-}
-const vt = /* @__PURE__ */ h({ onesLike_: Ln });
-function Gn(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 i = { paddings: n, constantValue: t }, o = { x: e };
-  return d.runKernel(tn, o, i);
-}
-const gt = /* @__PURE__ */ h({ pad_: Gn });
-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 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") }, i = { dims: n };
-  return d.runKernel(en, e, i);
-}
-const Ct = /* @__PURE__ */ h({ reverse_: zn });
-function Fn(s) {
-  const t = { x: c(s, "x", "rsqrt", "float32") };
-  return d.runKernel(rn, t);
-}
-const Et = /* @__PURE__ */ h({ rsqrt_: Fn });
-function Mn(s) {
-  const t = { x: c(s, "x", "selu") };
-  return d.runKernel(on, t);
-}
-const Nt = /* @__PURE__ */ h({ selu_: Mn });
-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(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(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 });
-function Jn(s) {
-  const t = { x: c(s, "x", "sin", "float32") };
-  return d.runKernel(an, t);
-}
-const wt = /* @__PURE__ */ h({ sin_: Jn });
-function Un(s) {
-  const t = { x: c(s, "x", "sinh") };
-  return d.runKernel(un, t);
-}
-const Kt = /* @__PURE__ */ h({ sinh_: Un });
-function Vn(s, n, t) {
-  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: i }, p = { numSegments: t };
-  return d.runKernel(ln, o, p);
-}
-const St = /* @__PURE__ */ h({ unsortedSegmentSum_: Vn });
-export {
-  Et as A,
-  Nt as B,
-  _t as C,
-  Kt as D,
-  rt as E,
-  St as F,
-  En as G,
-  mt as a,
-  Dt as b,
-  ct as c,
-  ot as d,
-  ft as e,
-  et as f,
-  nt as g,
-  tt as h,
-  st as i,
-  at as j,
-  it as k,
-  bt as l,
-  xt as m,
-  ut as n,
-  S as o,
-  lt as p,
-  pt as q,
-  Ct as r,
-  wt as s,
-  ht as t,
-  qn as u,
-  dt as v,
-  $t as w,
-  kt as x,
-  vt as y,
-  gt as z
-};

package/dist/utilities/arrayClose.d.ts

@@ -1 +0,0 @@
-export declare function arraysClose(a: unknown, b: unknown): number;

package/dist/utilities/arrayClose.js

@@ -1,20 +0,0 @@
-function i(r, e) {
-  let t = 0;
-  if ((Array.isArray(r) || r instanceof Float32Array) && (Array.isArray(e) || e instanceof Float32Array)) {
-    if (r.length !== e.length) return Number.POSITIVE_INFINITY;
-    for (let n = 0; n < r.length; ++n)
-      t = Math.max(t, i(r[n], e[n]));
-    return t;
-  } else if (typeof r == "number" && typeof e == "number") {
-    if (isNaN(r) && isNaN(e))
-      return 0;
-    if (!isFinite(r) || !isFinite(e))
-      return r === e ? 0 : Number.POSITIVE_INFINITY;
-    const n = Math.abs(r - e);
-    return t = Math.max(t, n), t;
-  } else
-    return Number.POSITIVE_INFINITY;
-}
-export {
-  i as arraysClose
-};

package/dist/utilities/datasetID.d.ts

@@ -1,2 +0,0 @@
-import { DatasetMetadata } from '../loader/types';
-export default function generateDatasetID(datasets: DatasetMetadata[]): string;

package/dist/utilities/datasetID.js

@@ -1,21 +0,0 @@
-function h(s) {
-  const c = s.map((n) => String(n.id)).sort();
-  let e = 2166136261, o = 2654435769;
-  const t = (n) => {
-    e ^= n & 255, e = Math.imul(e, 16777619), o ^= n & 255, o = Math.imul(o, 2246822507);
-  }, g = (n) => {
-    const r = n.length >>> 0;
-    t(r & 255), t(r >>> 8 & 255), t(r >>> 16 & 255), t(r >>> 24 & 255);
-    for (let a = 0; a < n.length; a++) {
-      const l = n.charCodeAt(a);
-      t(l & 255), t(l >>> 8 & 255);
-    }
-  }, i = c.length >>> 0;
-  t(i & 255), t(i >>> 8 & 255), t(i >>> 16 & 255), t(i >>> 24 & 255);
-  for (const n of c)
-    g(n);
-  return "dataset__" + (e >>> 0).toString(36) + "_" + (o >>> 0).toString(36);
-}
-export {
-  h as default
-};

package/dist/utilities/dummy.d.ts

@@ -1,9 +0,0 @@
-import { default as Model, ModelForwardAttributes } from '../models/model';
-export declare function dummyPassAsync(model: Model<ModelForwardAttributes>): Promise<void>;
-export interface MemoryRequirements {
-    perBatch: number;
-    tapeSize: number;
-    gradients: number;
-}
-export declare function dummyPassTrainAsync(model: Model<ModelForwardAttributes>): Promise<MemoryRequirements>;
-export declare function dummyPass(model: Model<ModelForwardAttributes>): void;

package/dist/utilities/dummy.js

@@ -1,43 +0,0 @@
-import { b as y, h as I, v as P } from "../index-CUXkjxiT.js";
-import { z as c } from "../zeros-DvZpK8s6.js";
-async function w(s) {
-  const t = c([1, s.config.blockSize], "int32"), n = s.forward({ training: !1 }, t);
-  await n.data(), n.dispose(), t.dispose();
-}
-async function b(s) {
-  const t = y(), n = t.numBytesInGPUAllocated ?? t.numBytesAllocatedInGPU ?? t.numBytes;
-  await w(s);
-  const a = c([1, s.config.blockSize], "int32"), e = {
-    perBatch: 0,
-    tapeSize: 0,
-    gradients: s.getNumParams() * 4
-  };
-  try {
-    const i = () => {
-      const o = s.forward({ training: !0 }, a), u = I().state.activeTape;
-      let p = 0;
-      if (u)
-        for (const g of u)
-          p += g.saved?.reduce((B, z) => B + z.size * 4, 0) || 0;
-      e.tapeSize = p;
-      const l = o.mean();
-      return o.dispose(), l;
-    }, { value: d, grads: m } = P(i), r = y(), f = r.numBytesInGPUAllocated ?? r.numBytesAllocatedInGPU ?? r.numBytes;
-    e.perBatch = f - n - e.gradients, await d.data(), d.dispose();
-    for (const o in m)
-      m[o].dispose();
-    a.dispose();
-  } catch (i) {
-    console.error("Error during dummy training pass:", i), a.dispose();
-  }
-  return e;
-}
-function v(s) {
-  const t = c([1, s.config.blockSize], "int32");
-  s.forward({ training: !1 }, t).dispose(), t.dispose();
-}
-export {
-  v as dummyPass,
-  w as dummyPassAsync,
-  b as dummyPassTrainAsync
-};

package/dist/utilities/multinomialCPU.d.ts

@@ -1,2 +0,0 @@
-import { Tensor2D } from '@tensorflow/tfjs-core';
-export default function multinomialCPU(probs: number[], randomNumber?: number): Tensor2D;

package/dist/utilities/multinomialCPU.js

@@ -1,13 +0,0 @@
-import "../index-CUXkjxiT.js";
-import { t as e } from "../tensor2d-BnXMKScO.js";
-function f(n, i) {
-  let r = 0;
-  const o = i ?? Math.random();
-  for (let t = 0; t < n.length; t++)
-    if (r += n[t], o < r)
-      return e([[t]], [1, 1], "int32");
-  return e([[n.length - 1]], [1, 1], "int32");
-}
-export {
-  f as default
-};

package/dist/utilities/naming.d.ts

@@ -1,4 +0,0 @@
-export interface VariableNaming {
-    blockIndex?: number;
-    component?: 'mlp' | 'attn' | 'rms' | 'embeddings';
-}

package/dist/utilities/naming.js

@@ -1 +0,0 @@
-

package/dist/utilities/packed.d.ts

@@ -1,4 +0,0 @@
-import { Tensor } from '@tensorflow/tfjs-core';
-export declare function packingSupported(): boolean;
-export declare function isPackableTensor(tensor: Tensor): boolean;
-export declare function isPackedTensor(tensor: Tensor): boolean;

package/dist/utilities/packed.js

@@ -1,15 +0,0 @@
-import { h as n } from "../index-CUXkjxiT.js";
-function o() {
-  return n().backendName === "webgpu";
-}
-function r(e) {
-  return e.dtype === "packedF16";
-}
-function a(e) {
-  return r(e);
-}
-export {
-  r as isPackableTensor,
-  a as isPackedTensor,
-  o as packingSupported
-};

package/dist/utilities/parameters.d.ts

@@ -1,11 +0,0 @@
-import { GPTConfig } from '../models/config';
-export declare function estimateLayerParameters(config: GPTConfig): number;
-export declare function estimateParameterCount(config: GPTConfig): number;
-export declare function estimateMemoryUsage(config: GPTConfig): number;
-export declare function estimateTrainingMemoryUsage(config: GPTConfig, batchSize: number): number;
-export declare function estimateResources(config: GPTConfig, batchSize: number): {
-    numParams: number;
-    modelMemoryMB: number;
-    trainingMemoryMB: number;
-};
-export declare function validateConfig(config: GPTConfig): void;

package/dist/utilities/parameters.js

@@ -1,57 +0,0 @@
-function b(e) {
-  const r = 4 * e.nEmbed * e.nEmbed, t = e.mlpFactor * e.nEmbed * e.nEmbed + // fc
-  e.nEmbed * e.mlpFactor * e.nEmbed;
-  return r + t;
-}
-function a(e) {
-  const r = e.vocabSize * e.nEmbed, t = e.nLayer * (4 * e.nEmbed * e.nEmbed), m = e.nLayer * (e.mlpFactor * e.nEmbed * e.nEmbed + // fc
-  e.nEmbed * e.mlpFactor * e.nEmbed);
-  return r + t + m;
-}
-function o(e) {
-  return a(e) * 4;
-}
-function E(e, r) {
-  const t = o(e), m = t * 2, n = t * 2, s = r * e.blockSize * e.nEmbed * 4;
-  return t + m + n + s;
-}
-function i(e, r) {
-  const t = a(e), m = o(e) / (1024 * 1024), n = E(e, r) / (1024 * 1024);
-  return {
-    numParams: t,
-    modelMemoryMB: m,
-    trainingMemoryMB: n
-  };
-}
-function d(e) {
-  if (e.nEmbed % e.nHead !== 0)
-    throw new Error("nEmbed_divisible_nHead");
-  if (e.blockSize <= 0)
-    throw new Error("blockSize_positive");
-  if (e.vocabSize <= 0)
-    throw new Error("vocabSize_positive");
-  if (e.nLayer <= 0)
-    throw new Error("nLayer_positive");
-  if (e.mlpFactor <= 0)
-    throw new Error("mlpFactor_positive");
-  if (e.nEmbed / e.nHead % 2 !== 0)
-    throw new Error("headDim_even");
-  if (!Number.isInteger(e.nEmbed))
-    throw new Error("nEmbed_integer");
-  if (!Number.isInteger(e.nHead))
-    throw new Error("nHead_integer");
-  if (!Number.isInteger(e.nLayer))
-    throw new Error("nLayer_integer");
-  if (!Number.isInteger(e.blockSize))
-    throw new Error("blockSize_integer");
-  if (!Number.isInteger(e.vocabSize))
-    throw new Error("vocabSize_integer");
-}
-export {
-  b as estimateLayerParameters,
-  o as estimateMemoryUsage,
-  a as estimateParameterCount,
-  i as estimateResources,
-  E as estimateTrainingMemoryUsage,
-  d as validateConfig
-};

package/dist/utilities/performance.d.ts

@@ -1,2 +0,0 @@
-import { Tensor } from '@tensorflow/tfjs-core';
-export default function performanceTest(fn: () => Tensor, iterations?: number, allowPromise?: boolean): Promise<number>;

package/dist/utilities/performance.js

@@ -1,16 +0,0 @@
-import { t as s } from "../index-CUXkjxiT.js";
-async function f(e, o = 10, r = !1) {
-  for (let t = 0; t < 100; t++) {
-    const a = r ? await e() : s(e);
-    t === 99 && await a.data(), a.dispose();
-  }
-  const n = performance.now();
-  for (let t = 0; t < o; t++) {
-    const a = r ? await e() : s(e);
-    t === o - 1 && await a.data(), a.dispose();
-  }
-  return (performance.now() - n) / o;
-}
-export {
-  f as default
-};

package/dist/utilities/profile.d.ts

@@ -1,17 +0,0 @@
-import { MemoryInfo } from '@tensorflow/tfjs-core';
-export interface ExtendedMemoryInfo extends MemoryInfo {
-    numBytesInGPUAllocated?: number;
-    numBytesAllocatedInGPU?: number;
-}
-export default class MemoryProfiler {
-    private log;
-    private maxMemory;
-    private maxLabel?;
-    private lastMemInfo;
-    private peakMemory;
-    startMemory(): void;
-    getPeakMemory(): number;
-    getMaxMemory(): number;
-    endMemory(label: string): void;
-    printSummary(): void;
-}