@genai-fi/nanogpt 0.17.4 → 0.18.0

package/dist/training/LRScheduler.d.ts

@@ -5,6 +5,7 @@ export default class LRScheduler {
     private step;
     private startLearningRate;
     constructor(learningRate: number, config: LRSchedulerConfig);
+    serializeConfig(): LRSchedulerConfig;
     updateConfig(newConfig: Partial<LRSchedulerConfig>, learningRate?: number): void;
     get lr(): number;
     getNextLR(): number;

package/dist/training/LRScheduler.js

@@ -1,28 +1,34 @@
-class h {
-  constructor(t, i) {
-    this.learningRate = t, this.config = i, this.startLearningRate = t;
+class o {
+  constructor(i, t) {
+    this.learningRate = i, this.config = t, this.startLearningRate = i, t.step !== void 0 && (this.step = t.step);
   }
   step = 0;
   startLearningRate;
-  updateConfig(t, i) {
-    this.config = { ...this.config, ...t }, i !== void 0 && (this.startLearningRate = i);
+  serializeConfig() {
+    return {
+      ...this.config,
+      step: this.step
+    };
+  }
+  updateConfig(i, t) {
+    this.config = { ...this.config, ...i }, t !== void 0 && (this.startLearningRate = t);
   }
   get lr() {
     return this.learningRate;
   }
   getNextLR() {
-    const t = this.step;
-    if (this.config.warmupSteps > 0 && t < this.config.warmupSteps) {
-      const r = (t + 1) / this.config.warmupSteps, e = this.startLearningRate * r;
+    const i = this.step;
+    if (this.config.warmupSteps > 0 && i < this.config.warmupSteps) {
+      const r = (i + 1) / this.config.warmupSteps, e = this.startLearningRate * r;
       return this.learningRate = e, this.step++, e;
     }
-    const i = this.config.epochSteps * this.config.decayEpochs;
-    if (t >= i || i <= this.config.warmupSteps)
+    const t = this.config.epochSteps * this.config.decayEpochs;
+    if (i >= t || t <= this.config.warmupSteps)
       return this.learningRate = this.config.minLearningRate, this.step++, this.config.minLearningRate;
-    const n = (t - this.config.warmupSteps) / (i - this.config.warmupSteps), a = 0.5 * (1 + Math.cos(Math.PI * n)), s = this.config.minLearningRate + a * (this.startLearningRate - this.config.minLearningRate);
+    const n = (i - this.config.warmupSteps) / (t - this.config.warmupSteps), a = 0.5 * (1 + Math.cos(Math.PI * n)), s = this.config.minLearningRate + a * (this.startLearningRate - this.config.minLearningRate);
     return this.learningRate = s, this.step++, s;
   }
 }
 export {
-  h as default
+  o as default
 };

package/dist/training/PreTrainer.js

@@ -1,6 +1,6 @@
 import r from "./BasicTrainer.js";
-import { DatasetBuilder as s } from "./DatasetBuilder.js";
-const n = {
+import { DatasetBuilder as n } from "./DatasetBuilder.js";
+const s = {
   decayEpochs: 100,
   epochSteps: 1e4,
   warmupSteps: 1e3,
@@ -11,7 +11,7 @@ const n = {
 };
 class m extends r {
   constructor(i, e, t, a) {
-    super(i, e, { ...n, ...t }, a), this.tokenizer = e, this.optimizerConfig.minLearningRate = this.optimizerConfig.learningRate / 10, this.updateOptimizer(), this.datasetBuilder = new s(e, i.config.blockSize);
+    super(i, e, { ...s, ...t }, a), this.tokenizer = e, this.optimizerConfig.minLearningRate = t?.minLearningRate ?? this.optimizerConfig.learningRate / 20, this.updateOptimizer(), this.datasetBuilder = new n(e, i.config.blockSize);
   }
   datasetBuilder;
 }

package/dist/training/SFTDatasetBuilder.js

@@ -1,6 +1,6 @@
-import { t as y } from "../index-CUXkjxiT.js";
-import "../dataset-CGGp1z9P.js";
-import { g as I } from "../readers-iz5u3HBo.js";
+import { t as y } from "../index-DSGwv2Yx.js";
+import "../dataset-DlqAN81i.js";
+import { g as I } from "../readers-17HLdxVM.js";
 import "../index-Cp39cXWe.js";
 function w(p, o, t, l) {
   const s = [t.bosToken], a = [!1], u = {

package/dist/training/SFTTrainer.js

@@ -12,7 +12,7 @@ const n = {
 };
 class m extends r {
   constructor(t, e, i, a) {
-    super(t, e, { ...n, ...i }, a), this.tokenizer = e, this.optimizerConfig.minLearningRate = this.optimizerConfig.learningRate / 10, this.updateOptimizer(), this.datasetBuilder = new s(e, t.config.blockSize), this.maskedLoss = !0;
+    super(t, e, { ...n, ...i }, a), this.tokenizer = e, this.optimizerConfig.minLearningRate = i?.minLearningRate ?? this.optimizerConfig.learningRate / 20, this.updateOptimizer(), this.datasetBuilder = new s(e, t.config.blockSize), this.maskedLoss = !0;
   }
   datasetBuilder;
 }

package/dist/training/orthoGrad.js

@@ -1,4 +1,4 @@
-import { t as d } from "../index-CUXkjxiT.js";
+import { t as d } from "../index-DSGwv2Yx.js";
 function l(m, t, s) {
   return d(() => {
     const r = m.reshape([-1]), o = t.reshape([-1]), e = r.mul(r).sum().add(s), c = r.mul(o).sum().div(e), n = o.sub(r.mul(c)), a = o.norm(), u = n.norm().add(s);

package/dist/training/sparseCrossEntropy.js

@@ -1,24 +1,24 @@
-import { gatherSub as $ } from "../ops/gatherSub.js";
+import { gatherSub as j } from "../ops/gatherSub.js";
 import { scatterSub as T } from "../ops/scatterSub.js";
-import { t as F, a1 as G, j as A, m as x, i as U, a2 as D, a as j, a3 as q } from "../index-CUXkjxiT.js";
-import { m as J, l as K, a as E, n as N, w as O } from "../not_equal-hurPF26l.js";
-import { s as m } from "../sum-BpiwSWvg.js";
-import { s as Q } from "../softmax-D4q1LJN7.js";
-function R(e, l, a, g, v, t = 0) {
-  return F(() => {
-    const S = e.shape[e.shape.length - 1], i = v || e.shape.slice(0, -1), f = i.reduce((n, p) => n * p, 1), h = e.shape.length > 2 ? e.reshape([f, S]) : e, y = l.shape.length > 1 ? l.reshape([f]).cast("int32") : l.cast("int32"), u = J(h, -1, !0), o = G(h, u), r = K(o, -1), d = $(r, y, o);
+import { t as H, a1 as G, h as A, m as x, d as U, a2 as D, b as q, a3 as F } from "../index-DSGwv2Yx.js";
+import { m as J, l as K, a as E, n as N, w as O } from "../not_equal-CkQKkKZy.js";
+import { s as m } from "../sum-D39FeU5h.js";
+import { s as Q } from "../softmax-BLGJqdwx.js";
+function R(e, h, t, g, v, a = 0) {
+  return H(() => {
+    const S = e.shape[e.shape.length - 1], i = v || e.shape.slice(0, -1), l = i.reduce((n, p) => n * p, 1), f = e.shape.length > 2 ? e.reshape([l, S]) : e, b = h.shape.length > 1 ? h.reshape([l]).cast("int32") : h.cast("int32"), u = J(f, -1, !0), o = G(f, u), r = K(o, -1), d = j(r, b, o);
     let s;
-    if (t > 0) {
+    if (a > 0) {
       const n = G(r, E(o, -1));
-      s = A(x(d, 1 - t), x(n, t));
+      s = A(x(d, 1 - a), x(n, a));
     } else
       s = d;
-    if (a)
-      if (s = x(s, a), g) {
-        const n = m(a.reshape(i), -1);
+    if (t)
+      if (s = x(s, t), g) {
+        const n = m(t.reshape(i), -1);
         s = U(m(s.reshape(i), -1), n);
       } else {
-        const n = m(a);
+        const n = m(t);
         s = U(m(s), n);
       }
     else
@@ -26,44 +26,44 @@ function R(e, l, a, g, v, t = 0) {
     return s;
   });
 }
-function as(e, l, a = 0) {
+function ts(e, h, t = 0) {
   return D(
     // @ts-expect-error Invalid params
-    (t, S, i) => {
-      const f = t.shape[t.shape.length - 1], h = t.shape.slice(0, -1), y = h.reduce((p, c) => p * c, 1), u = t.reshape([y, f]), o = S.reshape([y]).cast("int32");
+    (a, S, i) => {
+      const l = a.shape[a.shape.length - 1], f = a.shape.slice(0, -1), b = f.reduce((p, c) => p * c, 1), u = a.reshape([b, l]), o = S.reshape([b]).cast("int32");
       let r, d = null;
       if (e) {
-        const p = j(-100, "int32"), c = N(o, p);
-        d = c.cast("float32"), r = O(c, o, q(o));
+        const p = q(-100, "int32"), c = N(o, p);
+        d = c.cast("float32"), r = O(c, o, F(o));
       } else
         r = o;
       const s = R(
         u,
         r,
         d || void 0,
-        l,
         h,
-        a
+        f,
+        t
       );
-      return i(d ? [u, r, d] : [u, r]), u.dispose(), o.dispose(), { value: s, gradFunc: (p, c) => F(() => {
-        const C = c[0], z = c[1], b = e ? c[2] : void 0, w = Q(C), H = b ? m(b) : j(C.shape[0], "float32"), I = p.div(H).broadcastTo([C.shape[0]]), M = b && e ? x(I, b) : I;
+      return i(d ? [u, r, d] : [u, r]), u.dispose(), o.dispose(), { value: s, gradFunc: (p, c) => H(() => {
+        const C = c[0], z = c[1], y = e ? c[2] : void 0, w = Q(C), P = y ? m(y) : q(C.shape[0], "float32"), I = p.div(P).broadcastTo([C.shape[0]]), M = y && e ? x(I, y) : I;
         let L;
-        if (a > 0) {
-          const V = a / f, W = G(w, V);
+        if (t > 0) {
+          const W = t / l, $ = G(w, W);
           L = T(
-            W,
+            $,
             z,
-            x(M, 1 - a)
+            x(M, 1 - t)
           );
         } else
           L = T(w, z, M);
-        const P = q(S);
-        return [L.reshape(t.shape), P];
+        const V = F(S);
+        return [L.reshape(a.shape), V];
       }) };
     }
   );
 }
 export {
-  as as createSoftmaxCrossEntropyWithGrad,
+  ts as createSoftmaxCrossEntropyWithGrad,
   R as sparseSoftmaxCrossEntropy
 };

package/dist/training/types.d.ts

@@ -19,21 +19,23 @@ export interface TrainingLogEntry {
     weightNorm?: number;
     weightStatistics?: Map<string, TensorStatistics>;
     memoryUsage?: number;
-    tokensPerSecond?: number;
+    tokensPerSecond: number;
     duration: number;
-    totalSamples: number;
-    samplesPerSecond: number;
+    totalTokens: number;
 }
 export interface LRSchedulerConfig {
     warmupSteps: number;
     decayEpochs: number;
     minLearningRate: number;
     epochSteps: number;
+    step?: number;
 }
 export interface AdamWOptimizerConfig extends LRSchedulerConfig {
     learningRate: number;
     beta1: number;
     beta2: number;
+    accBeta1?: number;
+    accBeta2?: number;
     epsilon?: number;
     weightDecay: number;
     lossScaling: number;

package/dist/training/validation.js

@@ -1,22 +1,22 @@
-import "../index-CUXkjxiT.js";
-import "../random_width-DN5ZtQkM.js";
-import "../zeros_like-CWjDdwr-.js";
+import "../index-DSGwv2Yx.js";
+import "../random_width-B_fVXhGx.js";
+import "../zeros_like-rOHr54NY.js";
 import "../Generator.js";
 import "../index-Cp39cXWe.js";
-import "../dataset-CGGp1z9P.js";
+import "../dataset-DlqAN81i.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-0j_f1TPm.js";
+import "../rope-CC5RjmKU.js";
 import "../ops/cpu/appendCache.js";
 import "../ops/webgl/appendCache.js";
 import "../ops/grads/softmax16.js";
-import "../matMul16-BcVC_E62.js";
+import "../matMul16-BIT70Vya.js";
 import "../ops/webgl/matMul16.js";
 import "../ops/cpu/matMul16.js";
-import "../pack16-bqltoUlR.js";
+import "../pack16-WlOSOuZA.js";
 import "../ops/transpose16.js";
 import "../ops/reshape16.js";
 import "../ops/cpu/qkv.js";
@@ -30,23 +30,23 @@ import "../ops/webgl/dropout16.js";
 import "../ops/grads/add16.js";
 import "../jszip.min-BZhlzntC.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 "../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 m, shuffle as h } from "./DatasetBuilder.js";
 import { tokensFromTasks as k } from "./tasks/Task.js";
 import "../papaparse.min-C0cScC2i.js";
 import "../ops/cpu/matMulGelu.js";
-import "../matMulGelu-JNLZqKQp.js";
+import "../matMulGelu-CsZnh18H.js";
 import "../ops/grads/matMulGelu.js";
 import "../ops/cpu/gelu.js";
 import "../ops/webgl/gelu.js";
-import "../gelu-B220X1Go.js";
+import "../gelu-DpTCC3eB.js";
 import "../ops/webgl/log.js";
 import "../checks/normRMS.js";
 import "../checks/normRMSGrad.js";

package/dist/{transpose-9kRxIXWR.js → transpose-Qxz-4os3.js}

@@ -1,5 +1,5 @@
-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";
+import { o as u, n as i, E as o, ao as $, ap as g, aq as m, v as l, t as x, ar as p } from "./index-DSGwv2Yx.js";
+import { c as k } from "./complex-3DpPEG9B.js";
 function K(r) {
   const e = { input: i(r, "input", "imag") };
   return o.runKernel($, 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(x, e);
+  return o.runKernel(m, e);
 }
 const d = /* @__PURE__ */ u({ real_: b });
-function N(r, t, e) {
+function v(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(() => {
+  return n.dtype === "complex64" ? x(() => {
     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 });
+const y = /* @__PURE__ */ u({ transpose_: v });
 export {
   h as i,
   _ as n,
   d as r,
-  v as t
+  y as t
 };

package/dist/{unsorted_segment_sum-DJvk5xnh.js → unsorted_segment_sum-BfFVV9Zm.js}

@@ -1,8 +1,8 @@
-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";
+import { o as h, n as c, E as d, bm as T, bn as q, bo as H, v as l, bp as P, J as _, bq as y, br as I, bs as W, bt as B, bu as A, bv as G, bw as L, bx as O, by as z, bz as F, z 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-DSGwv2Yx.js";
+import { k as C, c as g, m as D } from "./step-DQY6_ABw.js";
+import { r as b } from "./reshape-BIN71H3p.js";
+import { m as pn, b as hn, e as w } from "./not_equal-CkQKkKZy.js";
+import { s as K } from "./sum-D39FeU5h.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);
@@ -37,7 +37,7 @@ 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);
+  return d.runKernel(I, o, p);
 }
 const ot = /* @__PURE__ */ h({ batchToSpaceND_: xn });
 function kn(s) {
@@ -57,7 +57,7 @@ function vn(s, n, t, e, i, o) {
     offset: m,
     mean: r,
     variance: a
-  }, k = { varianceEpsilon: o }, $ = d.runKernel(I, x, k);
+  }, k = { varianceEpsilon: o }, $ = d.runKernel(W, x, k);
   return b($, p.shape);
 }
 const at = /* @__PURE__ */ h({ batchNorm_: vn });
@@ -67,7 +67,7 @@ function gn(s, n, t, e, i = "NHWC", o = [1, 1], p) {
   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);
+  const x = { x: u, filter: a }, k = { strides: t, pad: e, dataFormat: i, dilations: o, dimRoundingMode: p }, $ = d.runKernel(B, x, k);
   return m ? b($, [$.shape[1], $.shape[2], $.shape[3]]) : $;
 }
 const S = /* @__PURE__ */ h({ conv2d_: gn });
@@ -96,12 +96,12 @@ function Nn(s, n, t, e, i, o) {
 const ut = /* @__PURE__ */ h({ conv2dTranspose_: Nn });
 function _n(s) {
   const t = { x: c(s, "x", "cos", "float32") };
-  return d.runKernel(L, t);
+  return d.runKernel(G, t);
 }
 const ct = /* @__PURE__ */ h({ cos_: _n });
 function wn(s) {
   const t = { x: c(s, "x", "cosh", "float32") };
-  return d.runKernel(G, t);
+  return d.runKernel(L, t);
 }
 const lt = /* @__PURE__ */ h({ cosh_: wn });
 function Kn(s, n = 0, t = !1, e = !1) {
@@ -143,7 +143,7 @@ function yn(s) {
   return d.runKernel(V, t);
 }
 const mt = /* @__PURE__ */ h({ softplus_: yn });
-function Bn(s, n = -1) {
+function In(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}`);
@@ -155,20 +155,20 @@ function Bn(s, n = -1) {
     } };
   })(t);
 }
-const $t = /* @__PURE__ */ h({ logSoftmax_: Bn });
-function In(s) {
+const $t = /* @__PURE__ */ h({ logSoftmax_: In });
+function Wn(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 bt = /* @__PURE__ */ h({ logicalNot_: Wn });
+function Bn(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 });
+const xt = /* @__PURE__ */ h({ maxPool_: Bn });
 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}`);
@@ -176,19 +176,19 @@ function An(s, n, t = 1, e = 0, i = "int32") {
   return d.runKernel(R, p, r);
 }
 const kt = /* @__PURE__ */ h({ oneHot_: An });
-function Ln(s) {
+function Gn(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 vt = /* @__PURE__ */ h({ onesLike_: Gn });
+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 i = { paddings: n, constantValue: t }, o = { x: e };
   return d.runKernel(tn, o, i);
 }
-const gt = /* @__PURE__ */ h({ pad_: Gn });
+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()}`);

package/dist/utilities/datasetID.d.ts

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

package/dist/utilities/datasetID.js

@@ -0,0 +1,21 @@
+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.js

@@ -1,10 +1,10 @@
-import { b as y, h as I, v as P } from "../index-CUXkjxiT.js";
-import { z as c } from "../zeros-DvZpK8s6.js";
+import { Y as y, e as I, Z as P } from "../index-DSGwv2Yx.js";
+import { z as c } from "../zeros-Bw0puq_w.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) {
+async function G(s) {
   const t = y(), n = t.numBytesInGPUAllocated ?? t.numBytesAllocatedInGPU ?? t.numBytes;
   await w(s);
   const a = c([1, s.config.blockSize], "int32"), e = {
@@ -32,12 +32,12 @@ async function b(s) {
   }
   return e;
 }
-function v(s) {
+function b(s) {
   const t = c([1, s.config.blockSize], "int32");
   s.forward({ training: !1 }, t).dispose(), t.dispose();
 }
 export {
-  v as dummyPass,
+  b as dummyPass,
   w as dummyPassAsync,
-  b as dummyPassTrainAsync
+  G as dummyPassTrainAsync
 };

package/dist/utilities/multinomialCPU.js

@@ -1,5 +1,5 @@
-import "../index-CUXkjxiT.js";
-import { t as e } from "../tensor2d-BnXMKScO.js";
+import "../index-DSGwv2Yx.js";
+import { t as e } from "../tensor2d-DTtQ1QcT.js";
 function f(n, i) {
   let r = 0;
   const o = i ?? Math.random();

package/dist/utilities/packed.js

@@ -1,4 +1,4 @@
-import { h as n } from "../index-CUXkjxiT.js";
+import { e as n } from "../index-DSGwv2Yx.js";
 function o() {
   return n().backendName === "webgpu";
 }

package/dist/utilities/performance.js

@@ -1,4 +1,4 @@
-import { t as s } from "../index-CUXkjxiT.js";
+import { t as s } from "../index-DSGwv2Yx.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 { b as a } from "../index-CUXkjxiT.js";
+import { Y as a } from "../index-DSGwv2Yx.js";
 const s = 1024 * 1024;
 class l {
   log = /* @__PURE__ */ new Map();

package/dist/utilities/safetensors.js

@@ -1,5 +1,5 @@
-import "../index-CUXkjxiT.js";
-import { t as y } from "../tensor-BWFldCso.js";
+import "../index-DSGwv2Yx.js";
+import { t as y } from "../tensor-D8e0Gd7c.js";
 function l(t) {
   if (t === "float32") return "F32";
   if (t === "int32") return "I32";

package/dist/utilities/sentences.js

@@ -1,8 +1,8 @@
-import { m as w } from "../index-CUXkjxiT.js";
-import { t as f } from "../tensor2d-BnXMKScO.js";
-import { e as y } from "../expand_dims-Bkd1YD5x.js";
-import { s as g } from "../sum-BpiwSWvg.js";
-import { c as T } from "../concat-CAQpCret.js";
+import { m as w } from "../index-DSGwv2Yx.js";
+import { t as f } from "../tensor2d-DTtQ1QcT.js";
+import { e as y } from "../expand_dims-D0rBtgT1.js";
+import { s as g } from "../sum-D39FeU5h.js";
+import { c as T } from "../concat-Dqk7Xk7h.js";
 const p = 16;
 function A(o, t) {
   if (!t)

package/dist/utilities/weights.js

@@ -1,5 +1,5 @@
-import "../index-CUXkjxiT.js";
-import { t as p } from "../tensor-BWFldCso.js";
+import "../index-DSGwv2Yx.js";
+import { t as p } from "../tensor-D8e0Gd7c.js";
 function h(n) {
   const e = n.reduce((s, o) => s + o.length, 0), a = new Float32Array(e);
   let t = 0;

package/dist/{variable-Ck482e3n.js → variable-SSATClyt.js}

@@ -1,4 +1,4 @@
-import { E as i } from "./index-CUXkjxiT.js";
+import { E as i } from "./index-DSGwv2Yx.js";
 function m(r, a = !0, e, t) {
   return i.makeVariable(r, a, e, t);
 }

package/dist/{webgpu_program-B4HmApL1.js → webgpu_program-CbjdYLYk.js}

@@ -1,4 +1,4 @@
-import { ad as z, ac as F, ab as E, a9 as j, x as A } from "./index-CUXkjxiT.js";
+import { ad as z, ac as F, ab as E, a9 as j, v as A } from "./index-DSGwv2Yx.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-DYlGSwOJ.js → webgpu_util-DuofJBMo.js}

@@ -1,4 +1,4 @@
-import { x as u } from "./index-CUXkjxiT.js";
+import { v as u } from "./index-DSGwv2Yx.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 x(r, t, n = !1) {
+function M(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 M(r) {
+function h(r) {
   return { x: r.map((t, n) => n) };
 }
-function h(r) {
+function x(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 {
-  h as G,
+  x as G,
   s as M,
   d as a,
   b,
   m as c,
   p as d,
-  x as e,
-  M as f,
+  M as e,
+  h as f,
   g as i
 };

package/dist/{zeros-DvZpK8s6.js → zeros-Bw0puq_w.js}

@@ -1,5 +1,5 @@
-import { u as n, Q as m, U as i, E as c } from "./index-CUXkjxiT.js";
-import { c as f } from "./complex-DI35Q-gW.js";
+import { q as n, L as m, N as i, E as c } from "./index-DSGwv2Yx.js";
+import { c as f } from "./complex-3DpPEG9B.js";
 function e(o, r = "float32") {
   if (n(o), r === "complex64") {
     const s = e(o, "float32"), t = e(o, "float32");