@genai-fi/nanogpt 0.17.4 → 0.18.0

package/dist/{matMulGelu-JNLZqKQp.js → matMulGelu-CsZnh18H.js}

@@ -1,11 +1,11 @@
-import { c as C, t as R, h as I, U as G, _ as L, x as U, am as F } from "./index-CUXkjxiT.js";
-import { r as M } from "./Reshape-BW__R4mZ.js";
-import { u as H } from "./gpgpu_math-BwvV12df.js";
-import { m as B } from "./mat_mul-DhG0Newp.js";
+import { a as C, t as R, e as I, N as G, _ as L, v as F, am as U } from "./index-DSGwv2Yx.js";
+import { r as M } from "./Reshape-BZC-ebeR.js";
+import { u as H } from "./gpgpu_math-3bCb5ooU.js";
+import { m as B } from "./mat_mul-DP86qZtZ.js";
 class W {
   constructor(e, s, a, n = !1, o = !1, r = !1, i = null, u = !1, l = !1) {
     this.variableNames = ["matrixA", "matrixB"], this.packedInputs = !0, this.packedOutput = !0, this.outputShape = a, this.enableShapeUniforms = H(this.outputShape.length);
-    const p = n ? e[1] : e[2], h = Math.ceil(p / 2), d = n ? "i * 2, rc.y" : "rc.y, i * 2", b = o ? "rc.z, i * 2" : "i * 2, rc.z", x = n ? ["a.xxyy", "a.zzww"] : ["a.xxzz", "a.yyww"], m = o ? ["b.xzxz", "b.ywyw"] : ["b.xyxy", "b.zwzw"];
+    const p = n ? e[1] : e[2], h = Math.ceil(p / 2), d = n ? "i * 2, rc.y" : "rc.y, i * 2", x = o ? "rc.z, i * 2" : "i * 2, rc.z", b = n ? ["a.xxyy", "a.zzww"] : ["a.xxzz", "a.yyww"], m = o ? ["b.xzxz", "b.ywyw"] : ["b.xyxy", "b.zwzw"];
     let c = "", g = "";
     i && (u ? c = `vec4 activation(vec4 a) {
           vec4 b = getPreluActivationWeightsAtOutCoords();
@@ -30,12 +30,12 @@ class W {
         int batchB = ${v};
         for (int i = 0; i < ${h}; i++) {
           vec4 a = getMatrixA(batchA, ${d});
-          vec4 b = getMatrixB(batchB, ${b});
+          vec4 b = getMatrixB(batchB, ${x});
 
           // These swizzled products need to be separately added.
           // See: https://github.com/tensorflow/tfjs/issues/1735
-          result += (${x[0]} * ${m[0]});
-          result += (${x[1]} * ${m[1]});
+          result += (${b[0]} * ${m[0]});
+          result += (${b[1]} * ${m[1]});
         }
         return result;
       }
@@ -81,7 +81,7 @@ const S = 0.7978845608028654, w = 0.044715, j = `
     vec4 dgelu = 0.5 * (1.0 + t) + 0.5 * a * sech2 * du_dx;
     return dgelu * b;
 `, se = 1e3;
-function O({
+function N({
   a: t,
   b: e,
   transposeA: s,
@@ -90,24 +90,24 @@ function O({
   activationSnippet: o,
   multiplier: r
 }) {
-  const i = t.shape.length, u = e.shape.length, l = s ? t.shape[i - 2] : t.shape[i - 1], p = a ? e.shape[u - 1] : e.shape[u - 2], h = s ? t.shape[i - 1] : t.shape[i - 2], d = a ? e.shape[u - 2] : e.shape[u - 1], b = t.shape.slice(0, -2), x = e.shape.slice(0, -2), m = G(b), c = G(x), $ = L(t.shape.slice(0, -2), e.shape.slice(0, -2)).concat([h, d]);
-  U(
+  const i = t.shape.length, u = e.shape.length, l = s ? t.shape[i - 2] : t.shape[i - 1], p = a ? e.shape[u - 1] : e.shape[u - 2], h = s ? t.shape[i - 1] : t.shape[i - 2], d = a ? e.shape[u - 2] : e.shape[u - 1], x = t.shape.slice(0, -2), b = e.shape.slice(0, -2), m = G(x), c = G(b), $ = L(t.shape.slice(0, -2), e.shape.slice(0, -2)).concat([h, d]);
+  F(
     l === p,
     () => `Error in matMul: inner shapes (${l}) and (${p}) of Tensors with shapes ${t.shape} and ${e.shape} and transposeA=${s} and transposeB=${a} must match.`
   );
-  const f = s ? [m, l, h] : [m, h, l], v = a ? [c, d, p] : [c, p, d], A = M({ inputs: { x: t }, backend: n, attrs: { shape: f } }), y = M({ inputs: { x: e }, backend: n, attrs: { shape: v } }), D = [A, y], _ = Math.max(m, c), E = o, N = F(t.dtype, e.dtype), T = new W(
+  const f = s ? [m, l, h] : [m, h, l], v = a ? [c, d, p] : [c, p, d], A = M({ inputs: { x: t }, backend: n, attrs: { shape: f } }), y = M({ inputs: { x: e }, backend: n, attrs: { shape: v } }), D = [A, y], O = Math.max(m, c), _ = o, E = U(t.dtype, e.dtype), T = new W(
     f,
     v,
-    [_, h, d],
+    [O, h, d],
     s,
     a,
     !1,
-    E,
+    _,
     !!r,
     !1
   ), k = [A, y];
   r && k.push(r);
-  const z = n.runWebGLProgram(T, k, N), K = M({ inputs: { x: z }, backend: n, attrs: { shape: $ } });
+  const z = n.runWebGLProgram(T, k, E), K = M({ inputs: { x: z }, backend: n, attrs: { shape: $ } });
   D.push(z);
   for (const P of D)
     n.disposeIntermediateTensorInfo(P);
@@ -117,7 +117,7 @@ function J(t) {
   const { inputs: e, backend: s } = t, { x: a, kernel: n } = e;
   if (a === void 0 || n === void 0)
     throw new Error("BatchMatMul requires two input tensors.");
-  return O({
+  return N({
     a,
     b: n,
     transposeA: !1,
@@ -136,7 +136,7 @@ function V(t) {
   const { dy: e, x: s, kernel: a } = t.inputs, n = t.backend;
   return R(() => {
     const o = I().makeTensorFromTensorInfo(
-      O({
+      N({
         a: s,
         b: a,
         transposeA: !1,
@@ -158,6 +158,6 @@ C(X);
 export {
   W as M,
   se as a,
-  O as b,
+  N as b,
   J as c
 };

package/dist/mat_mul-DP86qZtZ.js

@@ -0,0 +1,11 @@
+import { o as m, n as s, z as c, E as M, B as p } from "./index-DSGwv2Yx.js";
+function f(e, n, o = !1, l = !1) {
+  let a = s(e, "a", "matMul"), t = s(n, "b", "matMul");
+  [a, t] = c(a, t);
+  const r = { a, b: t }, u = { transposeA: o, transposeB: l };
+  return M.runKernel(p, r, u);
+}
+const i = /* @__PURE__ */ m({ matMul_: f });
+export {
+  i as m
+};

package/dist/mod-BXjLYwvM.js

@@ -0,0 +1,11 @@
+import { o as r, n, z as m, E as c, M as d } from "./index-DSGwv2Yx.js";
+function p(s, t) {
+  let o = n(s, "a", "mod"), a = n(t, "b", "mod");
+  [o, a] = m(o, a);
+  const e = { a: o, b: a };
+  return c.runKernel(d, e);
+}
+const b = /* @__PURE__ */ r({ mod_: p });
+export {
+  b as m
+};

package/dist/models/NanoGPTV1.js

@@ -2,11 +2,11 @@ import u from "../layers/TransformerBlock.js";
 import f from "../layers/TiedEmbedding.js";
 import g from "../layers/RoPECache.js";
 import b from "../layers/RMSNorm.js";
-import { t as r, k as d } from "../index-CUXkjxiT.js";
+import { t as r, k as d } from "../index-DSGwv2Yx.js";
 import k from "./model.js";
 import w from "../layers/PositionEmbedding.js";
 import { packingSupported as E } from "../utilities/packed.js";
-import { p as _, u as C } from "../pack16-bqltoUlR.js";
+import { p as _, u as C } from "../pack16-WlOSOuZA.js";
 const P = {
   modelType: "GenAI_NanoGPT_v1",
   vocabSize: 2e3,

package/dist/models/NanoGPTV2.js

@@ -2,10 +2,10 @@ import u from "../layers/TransformerBlock.js";
 import f from "../layers/TiedEmbedding.js";
 import g from "../layers/RoPECache.js";
 import b from "../layers/RMSNorm.js";
-import { t as r, k as d } from "../index-CUXkjxiT.js";
+import { t as r, k as d } from "../index-DSGwv2Yx.js";
 import k from "./model.js";
 import { packingSupported as w } from "../utilities/packed.js";
-import { p as C, u as _ } from "../pack16-bqltoUlR.js";
+import { p as C, u as _ } from "../pack16-WlOSOuZA.js";
 const E = {
   modelType: "GenAI_NanoGPT_v2",
   vocabSize: 2e3,

package/dist/models/model.d.ts

@@ -10,11 +10,13 @@ export interface ModelForwardAttributes extends ForwardAttributes {
     seed?: number;
     skipLogits?: boolean;
 }
-interface TrainingState {
+export interface TrainingState {
     steps: number;
     learningRate: number;
     batchSize: number;
     loss: number;
+    tokensProcessed: number;
+    duration: number;
 }
 export default abstract class Model<T extends ModelForwardAttributes, C extends GPTConfig = GPTConfig> extends BaseLayer<T, C> {
     lossScaling: number;
@@ -39,4 +41,3 @@ export default abstract class Model<T extends ModelForwardAttributes, C extends
     getNumParams(): number;
     protected validateInput(idx: Tensor): void;
 }
-export {};

package/dist/models/model.js

@@ -1,23 +1,23 @@
 import i from "../layers/BaseLayer.js";
-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";
@@ -31,22 +31,22 @@ 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 "../papaparse.min-C0cScC2i.js";
 import { estimateParameterCount as a } from "../utilities/parameters.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/{not_equal-hurPF26l.js → not_equal-CkQKkKZy.js}

@@ -1,10 +1,10 @@
-import { o as c, q as e, _ as h, E as u, aL as b, aM as E, aN as _, aO as g, ah as S, a1 as w, j as K, aP as q, B as M, aQ as T } from "./index-CUXkjxiT.js";
-import { b as d } from "./broadcast_to-eS93CCN_.js";
-import { e as N } from "./axis_util-GTVlo58H.js";
-import { r as f } from "./reshape-BDOuCSNW.js";
-import { s as A } from "./sum-BpiwSWvg.js";
+import { o as c, n as e, _ as d, E as u, aL as b, aM as E, aN as _, aO as g, ah as S, a1 as w, h as K, aP as q, z as M, aQ as T } from "./index-DSGwv2Yx.js";
+import { b as h } from "./broadcast_to-C_EJTVTZ.js";
+import { e as N } from "./axis_util-QWWgLjut.js";
+import { r as f } from "./reshape-BIN71H3p.js";
+import { s as A } from "./sum-D39FeU5h.js";
 function B(s, o, n) {
-  const a = e(o, "a", "where"), t = e(n, "b", "where"), r = e(s, "condition", "where", "bool"), p = h(h(r.shape, a.shape), t.shape), l = d(r, p), m = d(a, p), x = d(t, p), i = {
+  const a = e(o, "a", "where"), t = e(n, "b", "where"), r = e(s, "condition", "where", "bool"), p = d(d(r.shape, a.shape), t.shape), l = h(r, p), m = h(a, p), x = h(t, p), i = {
     condition: l,
     t: m,
     e: x
@@ -22,40 +22,40 @@ function L(s) {
   return u.runKernel(_, n);
 }
 const P = /* @__PURE__ */ c({ exp_: L });
-function j(s) {
+function k(s) {
   const n = { x: e(s, "x", "log", "float32") };
   return u.runKernel(g, n);
 }
-const k = /* @__PURE__ */ c({ log_: j });
-function v(s, o = null, n = !1) {
+const v = /* @__PURE__ */ c({ log_: k });
+function y(s, o = null, n = !1) {
   const a = e(s, "x", "logSumExp"), t = S(o, a.shape), r = I(
     a,
     t,
     !0
     /* keepDims */
-  ), p = w(a, r), l = P(p), m = A(l, t), x = k(m), i = K(f(r, x.shape), x);
+  ), p = w(a, r), l = P(p), m = A(l, t), x = v(m), i = K(f(r, x.shape), x);
   if (n) {
     const $ = N(i.shape, t);
     return f(i, $);
   }
   return i;
 }
-const J = /* @__PURE__ */ c({ logSumExp_: v });
-function y(s, o = null, n = !1) {
+const J = /* @__PURE__ */ c({ logSumExp_: y });
+function z(s, o = null, n = !1) {
   const t = { x: e(s, "x", "mean") }, r = { axis: o, keepDims: n };
   return u.runKernel(q, t, r);
 }
-const R = /* @__PURE__ */ c({ mean_: y });
+const R = /* @__PURE__ */ c({ mean_: z });
 function C(s, o) {
   let n = e(s, "a", "notEqual", "string_or_numeric"), a = e(o, "b", "notEqual", "string_or_numeric");
-  [n, a] = M(n, a), h(n.shape, a.shape);
+  [n, a] = M(n, a), d(n.shape, a.shape);
   const t = { a: n, b: a };
   return u.runKernel(T, t);
 }
 const U = /* @__PURE__ */ c({ notEqual_: C });
 export {
   R as a,
-  k as b,
+  v as b,
   P as e,
   J as l,
   I as m,

package/dist/{ones-BytntneX.js → ones-DbVB5N58.js}

@@ -1,6 +1,6 @@
-import { u as n, V as t, U as m, E as i } from "./index-CUXkjxiT.js";
-import { c as f } from "./complex-DI35Q-gW.js";
-import { z as c } from "./zeros-DvZpK8s6.js";
+import { q as n, Q as t, N as m, E as i } from "./index-DSGwv2Yx.js";
+import { c as f } from "./complex-3DpPEG9B.js";
+import { z as c } from "./zeros-Bw0puq_w.js";
 function l(o, r = "float32") {
   if (n(o), r === "complex64") {
     const s = l(o, "float32"), a = c(o, "float32");

package/dist/ops/adamAdjust.js

@@ -1,8 +1,8 @@
-import { h as u } from "../index-CUXkjxiT.js";
+import { e as u } from "../index-DSGwv2Yx.js";
 import "./cpu/adamAdjust.js";
 import "./webgl/adamAdjust.js";
-function s(r, t, n, e, m, o, i = 0) {
-  return u().runKernel("AdamAdjust", { moments: r, value: t }, { beta1: n, beta2: e, epsilon: m, learningRate: o, weightDecay: i });
+function s(r, t, e, n, m, o, i = 0) {
+  return u().runKernel("AdamAdjust", { moments: r, value: t }, { beta1: e, beta2: n, epsilon: m, learningRate: o, weightDecay: i });
 }
 export {
   s as adamAdjust

package/dist/ops/adamMoments.js

@@ -1,8 +1,8 @@
-import { h as t } from "../index-CUXkjxiT.js";
+import { e as t } from "../index-DSGwv2Yx.js";
 import "./cpu/adamMoments.js";
 import "./webgl/adamMoments.js";
-function s(n, r, e, m, o) {
-  return t().runKernel("AdamMoments", { moments: n, gradient: r, scaling: o }, { beta1: e, beta2: m });
+function s(e, n, r, m, o) {
+  return t().runKernel("AdamMoments", { moments: e, gradient: n, scaling: o }, { beta1: r, beta2: m });
 }
 export {
   s as adamMoments

package/dist/ops/add16.js

@@ -1,4 +1,4 @@
-import { j as t, h as o } from "../index-CUXkjxiT.js";
+import { h as t, e as o } from "../index-DSGwv2Yx.js";
 import { isPackedTensor as n } from "../utilities/packed.js";
 import "./grads/add16.js";
 function m(r, e) {

package/dist/ops/appendCache.js

@@ -1,21 +1,21 @@
-import { h as e } from "../index-CUXkjxiT.js";
+import { e as a } from "../index-DSGwv2Yx.js";
 import "./cpu/appendCache.js";
 import "./webgl/appendCache.js";
 import { isPackedTensor as c } from "../utilities/packed.js";
-import { c as t } from "../concat-CAQpCret.js";
-import { z as f } from "../zeros-DvZpK8s6.js";
+import { c as t } from "../concat-Dqk7Xk7h.js";
+import { z as f } from "../zeros-Bw0puq_w.js";
 function C(r, o, n, p) {
   if (!p) {
-    const s = r.shape[2], a = c(r);
+    const e = r.shape[2], s = c(r);
     return t(
       [
         r,
-        f([r.shape[0], r.shape[1], o - s, r.shape[3]], a ? "int32" : r.dtype)
+        f([r.shape[0], r.shape[1], o - e, r.shape[3]], s ? "int32" : r.dtype)
       ],
       2
     );
   }
-  return e().runKernel("AppendCache", { cache: p, item: r }, { maxSize: o, pastLen: n });
+  return a().runKernel("AppendCache", { cache: p, item: r }, { maxSize: o, pastLen: n });
 }
 export {
   C as appendCache

package/dist/ops/attentionMask.js

@@ -1,9 +1,9 @@
-import { h as e } from "../index-CUXkjxiT.js";
+import { e as r } from "../index-DSGwv2Yx.js";
 import "./cpu/attentionMask.js";
 import "./webgl/attentionMask.js";
 import "./grads/attentionMask.js";
-function u(t, n, o, r) {
-  return e().runKernel("AttentionMask", { q: t, k: n }, { divisor: o, pastLen: r || 0 });
+function u(t, n, e, o) {
+  return r().runKernel("AttentionMask", { q: t, k: n }, { divisor: e, pastLen: o || 0 });
 }
 export {
   u as attentionMask

package/dist/ops/concat16.js

@@ -1,8 +1,8 @@
 import { isPackedTensor as o } from "../utilities/packed.js";
-import { h as c } from "../index-CUXkjxiT.js";
-import { c as e } from "../concat-CAQpCret.js";
+import { e } from "../index-DSGwv2Yx.js";
+import { c } from "../concat-Dqk7Xk7h.js";
 function p(r, n) {
-  return o(r[0]) ? c().runKernel("Concat16", r, { axis: n ?? -1 }) : e(r, n);
+  return o(r[0]) ? e().runKernel("Concat16", r, { axis: n ?? -1 }) : c(r, n);
 }
 export {
   p as concat16

package/dist/ops/cpu/adamAdjust.js

@@ -1,18 +1,18 @@
-import { c as k, i as t, j as i, m as z, l as A } from "../../index-CUXkjxiT.js";
-function C(c) {
-  const { moments: s, value: r } = c.inputs, { beta1: l, beta2: m, epsilon: u, learningRate: d } = c.attrs, e = s.shape.length, a = new Array(e).fill(0), n = s.shape.slice();
+import { a as h, d as t, h as i, m as k, i as z } from "../../index-DSGwv2Yx.js";
+function A(a) {
+  const { moments: s, value: r } = a.inputs, { beta1: l, beta2: m, epsilon: u, learningRate: d } = a.attrs, e = s.shape.length, c = new Array(e).fill(0), n = s.shape.slice();
   n[e - 1] = 1;
-  const o = a.slice();
+  const o = c.slice();
   o[e - 1] = 1;
-  const b = n.slice(), p = s.slice(a, n).squeeze([e - 1]), M = s.slice(o, b).squeeze([e - 1]), g = t(p, l), f = t(M, m);
+  const b = n.slice(), p = s.slice(c, n).squeeze([e - 1]), M = s.slice(o, b).squeeze([e - 1]), g = t(p, l), f = t(M, m);
   return i(
-    z(t(g, i(A(f), u ?? 1e-8)), -d),
+    k(t(g, i(z(f), u ?? 1e-8)), -d),
     r
   );
 }
-const h = {
+const C = {
   kernelName: "AdamAdjust",
   backendName: "cpu",
-  kernelFunc: C
+  kernelFunc: A
 };
-k(h);
+h(C);

package/dist/ops/cpu/adamMoments.js

@@ -1,11 +1,11 @@
-import { c as p } from "../../index-CUXkjxiT.js";
-import { s as b } from "../../stack-DudVrtmG.js";
+import { a as p } from "../../index-DSGwv2Yx.js";
+import { s as b } from "../../stack-z6QE7kmP.js";
 function f(t) {
-  const { moments: n, gradient: c } = t.inputs, { beta1: o, beta2: m } = t.attrs, e = n.shape.length, a = new Array(e).fill(0), s = n.shape.slice();
+  const { moments: n, gradient: c } = t.inputs, { beta1: o, beta2: a } = t.attrs, e = n.shape.length, m = new Array(e).fill(0), s = n.shape.slice();
   s[e - 1] = 1;
-  const i = a.slice();
+  const i = m.slice();
   i[e - 1] = 1;
-  const r = s.slice(), l = n.slice(a, s).squeeze([e - 1]), u = n.slice(i, r).squeeze([e - 1]), M = l.mul(o).add(c.mul(1 - o)), d = u.mul(m).add(c.square().mul(1 - m));
+  const r = s.slice(), l = n.slice(m, s).squeeze([e - 1]), u = n.slice(i, r).squeeze([e - 1]), M = l.mul(o).add(c.mul(1 - o)), d = u.mul(a).add(c.square().mul(1 - a));
   return b([M, d], -1);
 }
 const g = {

package/dist/ops/cpu/appendCache.js

@@ -1,5 +1,5 @@
-import { c as d } from "../../index-CUXkjxiT.js";
-import { c as h } from "../../concat-CAQpCret.js";
+import { a as d } from "../../index-DSGwv2Yx.js";
+import { c as h } from "../../concat-Dqk7Xk7h.js";
 function u(p) {
   const { cache: n, item: s } = p.inputs, { maxSize: i, pastLen: c } = p.attrs, t = n.shape[0], o = n.shape[1], a = n.shape[3], e = s.shape[2];
   if (c + e <= i) {

package/dist/ops/cpu/attentionMask.js

@@ -1,9 +1,9 @@
-import { c as o, f as d, a as u } from "../../index-CUXkjxiT.js";
-import { l as N } from "../../ops-CsXeTq1P.js";
-import { o as b } from "../../ones-BytntneX.js";
-import { z as A } from "../../zeros-DvZpK8s6.js";
-import { w as I } from "../../not_equal-hurPF26l.js";
-import { m as g } from "../../mat_mul-DhG0Newp.js";
+import { a as o, f as d, b as u } from "../../index-DSGwv2Yx.js";
+import { l as N } from "../../ops-CURIZSVt.js";
+import { o as b } from "../../ones-DbVB5N58.js";
+import { z as A } from "../../zeros-Bw0puq_w.js";
+import { w as I } from "../../not_equal-CkQKkKZy.js";
+import { m as g } from "../../mat_mul-DP86qZtZ.js";
 function a(n) {
   const { q: s, k: e } = n.inputs, { divisor: r } = n.attrs, c = s.shape[2], t = e.shape[2], m = N.bandPart(b([t, t]), -1, 0).cast("bool"), i = A([t, t]), l = d([t, t], Number.NEGATIVE_INFINITY), f = I(m, i, l), k = g(s, e, !1, !0).mul(u(r)), p = f.slice([0, 0], [c, t]).expandDims(0).expandDims(0);
   return k.add(p);

package/dist/ops/cpu/fusedSoftmax.js

@@ -1,5 +1,5 @@
-import { c as e } from "../../index-CUXkjxiT.js";
-import { s as m } from "../../softmax-D4q1LJN7.js";
+import { a as e } from "../../index-DSGwv2Yx.js";
+import { s as m } from "../../softmax-BLGJqdwx.js";
 function n(t) {
   const { inputs: s, attrs: a } = t, { logits: o } = s, { dim: i, dropoutRate: r } = a;
   if (!o)
@@ -18,12 +18,12 @@ const u = {
   kernelFunc: n
 };
 e(u);
-const c = {
+const l = {
   kernelName: "FusedSoftmax",
   backendName: "webgpu",
   kernelFunc: n
 };
-e(c);
+e(l);
 export {
   n as softmaxCPU
 };

package/dist/ops/cpu/gatherSub.js

@@ -1,13 +1,13 @@
-import { o as u, q as c, E as g, al as h, c as l, a1 as m } from "../../index-CUXkjxiT.js";
-import { r as p } from "../../range-C-CjF-LI.js";
-import { s as N } from "../../stack-DudVrtmG.js";
+import { o as u, n as a, E as g, al as h, a as l, a1 as m } from "../../index-DSGwv2Yx.js";
+import { r as p } from "../../range-BDxO73mk.js";
+import { s as N } from "../../stack-z6QE7kmP.js";
 function f(e, t) {
-  const n = c(t, "indices", "gatherND", "int32"), s = { params: c(e, "x", "gatherND", "string_or_numeric"), indices: n };
+  const n = a(t, "indices", "gatherND", "int32"), s = { params: a(e, "x", "gatherND", "string_or_numeric"), indices: n };
   return g.runKernel(h, s);
 }
 const b = /* @__PURE__ */ u({ gatherND_: f });
 function d(e) {
-  const { values: t, labels: n, logits: r } = e.inputs, s = n.shape[0], a = p(0, s, 1, "int32"), o = N([a, n], 1), i = b(r, o);
+  const { values: t, labels: n, logits: r } = e.inputs, s = n.shape[0], c = p(0, s, 1, "int32"), o = N([c, n], 1), i = b(r, o);
   return m(t, i);
 }
 const k = {

package/dist/ops/cpu/gelu.js

@@ -1,10 +1,10 @@
-import { c as t, t as d } from "../../index-CUXkjxiT.js";
+import { a as t, t as d } from "../../index-DSGwv2Yx.js";
 const o = 0.7978845608028654, c = 0.044715;
 function m(r) {
   const { inputs: u } = r, { x: n } = u, e = n;
   return d(() => {
-    const l = e.pow(3), s = e.add(l.mul(c)).mul(o).tanh().add(1).mul(0.5);
-    return e.mul(s);
+    const l = e.pow(3), a = e.add(l.mul(c)).mul(o).tanh().add(1).mul(0.5);
+    return e.mul(a);
   });
 }
 const p = {
@@ -22,7 +22,7 @@ t(K);
 function i(r) {
   const { dy: u, x: n } = r.inputs;
   return d(() => {
-    const e = n.square(), l = e.mul(n), a = n.add(l.mul(c)).mul(o).tanh(), f = a.square().neg().add(1), k = e.mul(3 * c).add(1), g = a.add(1).mul(0.5), G = n.mul(f).mul(o).mul(k).mul(0.5), N = g.add(G);
+    const e = n.square(), l = e.mul(n), s = n.add(l.mul(c)).mul(o).tanh(), f = s.square().neg().add(1), k = e.mul(3 * c).add(1), g = s.add(1).mul(0.5), G = n.mul(f).mul(o).mul(k).mul(0.5), N = g.add(G);
     return u.mul(N);
   });
 }

package/dist/ops/cpu/matMul16.js

@@ -1,6 +1,6 @@
 import { isPackedTensor as t } from "../../utilities/packed.js";
-import { c as p } from "../../index-CUXkjxiT.js";
-import { m as l } from "../../mat_mul-DhG0Newp.js";
+import { a as p } from "../../index-DSGwv2Yx.js";
+import { m as l } from "../../mat_mul-DP86qZtZ.js";
 function m(e) {
   const { A: n, B: r } = e.inputs, { transposeA: o, transposeB: s } = e.attrs, a = !t(n), c = !t(r);
   if (a && c)

package/dist/ops/cpu/matMulGelu.js

@@ -1,10 +1,10 @@
-import { c as e, t as m } from "../../index-CUXkjxiT.js";
-import { g as M, d as i } from "../../gelu-B220X1Go.js";
-import { m as k } from "../../mat_mul-DhG0Newp.js";
+import { a as e, t as m } from "../../index-DSGwv2Yx.js";
+import { g as M, d as i } from "../../gelu-DpTCC3eB.js";
+import { m as k } from "../../mat_mul-DP86qZtZ.js";
 function c(t) {
-  const { inputs: u } = t, { x: n, kernel: r } = u, a = n, l = r;
+  const { inputs: a } = t, { x: n, kernel: r } = a, u = n, l = r;
   return m(() => {
-    const o = k(a, l);
+    const o = k(u, l);
     return M(o);
   });
 }
@@ -27,9 +27,9 @@ const p = {
 };
 e(p);
 function s(t) {
-  const { dy: u, x: n, kernel: r } = t.inputs;
+  const { dy: a, x: n, kernel: r } = t.inputs;
   return m(() => {
-    const a = k(n, r), l = i(u, a), o = l.matMul(r.transpose()), d = n.transpose().matMul(l);
+    const u = k(n, r), l = i(a, u), o = l.matMul(r.transpose()), d = n.transpose().matMul(l);
     return [o, d];
   });
 }

package/dist/ops/cpu/matMulMul.js

@@ -1,6 +1,6 @@
-import { c as e, t as i } from "../../index-CUXkjxiT.js";
+import { a as e, t as i } from "../../index-DSGwv2Yx.js";
 function n(t) {
-  const { inputs: r, attrs: l } = t, { transposeA: o, transposeB: s } = l, { x: u, kernel: c, y: a } = r, m = u, k = c, M = a;
+  const { inputs: r, attrs: l } = t, { transposeA: o, transposeB: s } = l, { x: u, kernel: a, y: c } = r, m = u, k = a, M = c;
   return i(() => m.matMul(k, o, s).mul(M));
 }
 const p = {

package/dist/ops/cpu/mulDropout.js

@@ -1,20 +1,20 @@
-import { c as e, m as t } from "../../index-CUXkjxiT.js";
+import { a as e, m as t } from "../../index-DSGwv2Yx.js";
 function n(o) {
   const { inputs: r } = o, { a: l, b: u } = r;
   return console.warn("Using fallback mulDrop implementation without dropout."), t(l, u);
 }
-const c = {
+const a = {
   kernelName: "MulDropout",
   backendName: "cpu",
   kernelFunc: n
 };
-e(c);
-const a = {
+e(a);
+const c = {
   kernelName: "MulDropout",
   backendName: "tensorflow",
   kernelFunc: n
 };
-e(a);
+e(c);
 const m = {
   kernelName: "MulDropout",
   backendName: "webgpu",

package/dist/ops/cpu/normRMS.js

@@ -1,4 +1,4 @@
-import { c as o, t as d } from "../../index-CUXkjxiT.js";
+import { a as o, t as d } from "../../index-DSGwv2Yx.js";
 function i(t) {
   const { inputs: e } = t, { x: n, gamma: s } = e, r = n, a = s;
   return d(() => {

package/dist/ops/cpu/qkv.js

@@ -1,6 +1,6 @@
-import { c as q } from "../../index-CUXkjxiT.js";
-import { r as o } from "../../reshape-BDOuCSNW.js";
-import { s as x } from "../../split-C2Sj255c.js";
+import { a as q } from "../../index-DSGwv2Yx.js";
+import { r as o } from "../../reshape-BIN71H3p.js";
+import { s as x } from "../../split-BN9LkEgS.js";
 function v(p) {
   const { x: c, kernel: K } = p.inputs, { heads: n, packed: C } = p.attrs;
   if (C)