@genai-fi/nanogpt 0.4.2 → 0.4.4

package/dist/ops/cpu/gatherSub.js

@@ -1,6 +1,6 @@
-import { o as u, h as c, E as g, L as h, r as m, s as p } from "../../index-C4JCoBvj.js";
-import { r as l } from "../../range-9AzeApCc.js";
-import { s as N } from "../../stack-D1YjmgKN.js";
+import { o as u, h as c, E as g, N as h, r as m, s as p } from "../../index--6vO-cOz.js";
+import { r as N } from "../../range-C_vpUjBu.js";
+import { s as l } from "../../stack-CmqSdsfs.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -23,7 +23,7 @@ function f(e, s) {
 }
 const b = /* @__PURE__ */ u({ gatherND_: f });
 function d(e) {
-  const { values: s, labels: n, logits: r } = e.inputs, t = n.shape[0], o = l(0, t, 1, "int32"), a = N([o, n], 1), i = b(r, a);
+  const { values: s, labels: n, logits: r } = e.inputs, t = n.shape[0], o = N(0, t, 1, "int32"), a = l([o, n], 1), i = b(r, a);
   return p(s, i);
 }
 const k = {

package/dist/ops/cpu/gelu.js

@@ -1,4 +1,4 @@
-import { r as t, t as d } from "../../index-C4JCoBvj.js";
+import { r as t, t as d } from "../../index--6vO-cOz.js";
 const o = 0.7978845608028654, c = 0.044715;
 function m(u) {
   const { inputs: l } = u, { x: e } = l, n = e;

package/dist/ops/cpu/matMulGelu.d.ts

@@ -0,0 +1 @@
+export {};

package/dist/ops/cpu/matMulGelu.js

@@ -0,0 +1,40 @@
+import { r as a, t as i } from "../../index--6vO-cOz.js";
+const c = 0.7978845608028654, m = 0.044715;
+function M(o) {
+  const { inputs: s } = o, { x: t, kernel: l } = s, e = t, u = l;
+  return i(() => {
+    const n = e.matMul(u), d = n.mul(n).mul(n), r = n.add(d.mul(m)).mul(c).tanh().add(1).mul(0.5);
+    return n.mul(r);
+  });
+}
+const b = {
+  kernelName: "MatMulGelu",
+  backendName: "cpu",
+  kernelFunc: M
+};
+a(b);
+const w = {
+  kernelName: "MatMulGelu",
+  backendName: "tensorflow",
+  kernelFunc: M
+};
+a(w);
+function f(o) {
+  const { dy: s, x: t, kernel: l } = o.inputs;
+  return i(() => {
+    const e = t.matMul(l), u = e.square(), n = u.mul(e), r = e.add(n.mul(m)).mul(c).tanh(), G = r.square().neg().add(1), g = u.mul(3 * m).add(1), p = r.add(1).mul(0.5), N = e.mul(G).mul(c).mul(g).mul(0.5), K = p.add(N), k = s.mul(K), h = k.matMul(l.transpose()), F = t.transpose().matMul(k);
+    return [h, F];
+  });
+}
+const x = {
+  kernelName: "MatMulGeluGrad",
+  backendName: "cpu",
+  kernelFunc: f
+};
+a(x);
+const C = {
+  kernelName: "MatMulGeluGrad",
+  backendName: "tensorflow",
+  kernelFunc: f
+};
+a(C);

package/dist/ops/cpu/mulDropout.js

@@ -1,4 +1,4 @@
-import { r as e, b as u } from "../../index-C4JCoBvj.js";
+import { r as e, b as u } from "../../index--6vO-cOz.js";
 function n(o) {
   const { inputs: r } = o, { a: l, b: t } = r;
   return console.warn("Using fallback mulDrop implementation without dropout."), u(l, t);

package/dist/ops/cpu/qkv.js

@@ -1,6 +1,6 @@
-import { r as q } from "../../index-C4JCoBvj.js";
-import { r as o } from "../../reshape-Boe4DuIO.js";
-import { s as x } from "../../split-DbcNm1-i.js";
+import { r as q } from "../../index--6vO-cOz.js";
+import { r as o } from "../../reshape-z51Eu-re.js";
+import { s as x } from "../../split-B_k_jwud.js";
 function v(p) {
   const { x: c, kernel: K } = p.inputs, { heads: n } = p.attrs, [s, e, t] = c.shape, a = o(c, [s * e, t]), i = a.dot(K);
   a.dispose();

package/dist/ops/cpu/rope.js

@@ -1,8 +1,8 @@
-import { r as S } from "../../index-C4JCoBvj.js";
-import { r as F } from "../../range-9AzeApCc.js";
-import { g as I } from "../../gather-ZYRWhmXR.js";
-import { s as E } from "../../stack-D1YjmgKN.js";
-import { c as T } from "../../concat-CuRsVY-K.js";
+import { r as S } from "../../index--6vO-cOz.js";
+import { r as F } from "../../range-C_vpUjBu.js";
+import { g as I } from "../../gather-C5D8PxwA.js";
+import { s as E } from "../../stack-CmqSdsfs.js";
+import { c as T } from "../../concat-DvWM7HGZ.js";
 function U(t, c, p, o, r) {
   const n = o.shape[3], s = p;
   if (s > n) return o;

package/dist/ops/cpu/scatterSub.js

@@ -1,7 +1,7 @@
-import { o as l, k, h, E as g, a2 as w, r as $, s as d, b as m } from "../../index-C4JCoBvj.js";
-import { r as b } from "../../range-9AzeApCc.js";
-import { s as E } from "../../stack-D1YjmgKN.js";
-import { o as D } from "../../ones-Bf3YR48P.js";
+import { o as l, k, h, E as g, a5 as w, r as $, s as d, b as m } from "../../index--6vO-cOz.js";
+import { r as b } from "../../range-C_vpUjBu.js";
+import { s as E } from "../../stack-CmqSdsfs.js";
+import { o as D } from "../../ones-D6kB8bdY.js";
 function N(n, r, t) {
   const s = r.rank > 1 ? r.shape[r.rank - 1] : 1, e = r.rank > 1 ? r.rank - 1 : 1, o = `Must have updates.shape = indices.shape[:batchDim] + shape[sliceDim:], got updates.shape: ${t.shape}, indices.shape: ${r.shape}, shape: ${n}, sliceDim: ${s}, and batchDim: ${e}.`;
   if (t.rank < e)

package/dist/ops/fusedSoftmax.js

@@ -1,4 +1,4 @@
-import { e as t } from "../index-C4JCoBvj.js";
+import { e as t } from "../index--6vO-cOz.js";
 import "./cpu/fusedSoftmax.js";
 import "./webgl/fusedSoftmax.js";
 import "./grads/fusedSoftmax.js";

package/dist/ops/gatherSub.js

@@ -1,4 +1,4 @@
-import { e as n } from "../index-C4JCoBvj.js";
+import { e as n } from "../index--6vO-cOz.js";
 import "./cpu/gatherSub.js";
 import "./webgl/gatherSub.js";
 function f(r, e, t) {

package/dist/ops/gelu.js

@@ -1,7 +1,7 @@
-import "../index-C4JCoBvj.js";
+import "../index--6vO-cOz.js";
 import "./cpu/gelu.js";
 import "./webgl/gelu.js";
-import { d as e, g as i } from "../gelu-CnCt17Lk.js";
+import { d as e, g as i } from "./grads/gelu.js";
 export {
   e as dGelu,
   i as gelu

package/dist/ops/grads/attentionMask.js

@@ -1,4 +1,4 @@
-import { g as i } from "../../index-C4JCoBvj.js";
+import { g as i } from "../../index--6vO-cOz.js";
 const u = {
   kernelName: "AttentionMask",
   inputsToSave: ["q", "k"],

package/dist/ops/grads/fusedSoftmax.js

@@ -1,6 +1,6 @@
-import { g as p, b as m, s as d } from "../../index-C4JCoBvj.js";
+import { g as p, b as m, s as d } from "../../index--6vO-cOz.js";
 import { mulDrop as c } from "../mulDrop.js";
-import { s as f } from "../../sum-R28pucR5.js";
+import { s as f } from "../../sum-DdkDf2MG.js";
 const g = {
   kernelName: "FusedSoftmax",
   outputsToSave: [!0],

package/dist/ops/grads/gelu.js

@@ -1,5 +1,26 @@
-import "../../index-C4JCoBvj.js";
-import { a as e } from "../../gelu-CnCt17Lk.js";
+import { g as t, e as n } from "../../index--6vO-cOz.js";
+import "../cpu/gelu.js";
+import "../webgl/gelu.js";
+const o = {
+  kernelName: "Gelu",
+  inputsToSave: ["x"],
+  outputsToSave: [],
+  gradFunc: (e, r) => {
+    const [u] = r;
+    return {
+      x: () => a(e, u)
+    };
+  }
+};
+t(o);
+function g(e) {
+  return n().runKernel("Gelu", { x: e });
+}
+function a(e, r) {
+  return n().runKernel("GeluGrad", { dy: e, x: r });
+}
 export {
-  e as geluGradConfig
+  a as d,
+  g,
+  o as geluGradConfig
 };

package/dist/ops/grads/matMulGelu.d.ts

@@ -0,0 +1 @@
+export {};

package/dist/ops/grads/matMulGelu.js

@@ -0,0 +1,17 @@
+import { g as a, e as o } from "../../index--6vO-cOz.js";
+function s(e, n, r) {
+  return o().runKernel("MatMulGeluGrad", { dy: e, x: n, kernel: r });
+}
+const d = {
+  kernelName: "MatMulGelu",
+  inputsToSave: ["x", "kernel"],
+  outputsToSave: [],
+  gradFunc: (e, n) => {
+    const [r, t] = n, [u, l] = s(e, r, t);
+    return {
+      x: () => u,
+      kernel: () => l
+    };
+  }
+};
+a(d);

package/dist/ops/grads/qkv.js

@@ -1,4 +1,4 @@
-import { g as v } from "../../index-C4JCoBvj.js";
+import { g as v } from "../../index--6vO-cOz.js";
 const g = {
   kernelName: "QKV",
   inputsToSave: ["x", "kernel"],

package/dist/ops/grads/rope.js

@@ -1,4 +1,4 @@
-import { g as a, e as i } from "../../index-C4JCoBvj.js";
+import { g as a, e as i } from "../../index--6vO-cOz.js";
 function p(n, e, s, o) {
   return i().runKernel("Rope", { x: n, sin: e, cos: s }, { pastLen: o });
 }

package/dist/ops/matMulGelu.d.ts

@@ -0,0 +1,3 @@
+import { Tensor } from '@tensorflow/tfjs-core';
+export declare function matMulGelu(x: Tensor, kernel: Tensor): Tensor;
+export declare function dMatMulGelu(dy: Tensor, x: Tensor, kernel: Tensor): Tensor;

package/dist/ops/matMulGelu.js

@@ -0,0 +1,14 @@
+import { e as u } from "../index--6vO-cOz.js";
+import "./cpu/matMulGelu.js";
+import "./webgl/matMulGelu.js";
+import "./grads/matMulGelu.js";
+function M(r, e) {
+  return u().runKernel("MatMulGelu", { x: r, kernel: e });
+}
+function a(r, e, t) {
+  return u().runKernel("MatMulGeluGrad", { dy: r, x: e, kernel: t });
+}
+export {
+  a as dMatMulGelu,
+  M as matMulGelu
+};

package/dist/ops/mulDrop.js

@@ -1,4 +1,4 @@
-import { e as t } from "../index-C4JCoBvj.js";
+import { e as t } from "../index--6vO-cOz.js";
 import "./cpu/mulDropout.js";
 import "./webgl/mulDropout.js";
 function m(r, o, e, n) {

package/dist/ops/node/sparseCrossEntropy.js

@@ -1,4 +1,4 @@
-import { r as o } from "../../index-C4JCoBvj.js";
+import { r as o } from "../../index--6vO-cOz.js";
 function r(e) {
   const { logits: t, labels: n } = e.inputs;
   return e.backend.executeMultipleOutputs("SparseSoftmaxCrossEntropyWithLogits", [], [t, n], 2);

package/dist/ops/qkv.js

@@ -1,4 +1,4 @@
-import { e as o } from "../index-C4JCoBvj.js";
+import { e as o } from "../index--6vO-cOz.js";
 import "./cpu/qkv.js";
 import "./webgl/qkv.js";
 import "./grads/qkv.js";

package/dist/ops/scatterSub.js

@@ -1,4 +1,4 @@
-import { e as i } from "../index-C4JCoBvj.js";
+import { e as i } from "../index--6vO-cOz.js";
 import "./cpu/scatterSub.js";
 import "./webgl/scatterSub.js";
 function c(t, r, e) {

package/dist/ops/webgl/appendCache.js

@@ -1,12 +1,12 @@
-import { r as h } from "../../index-C4JCoBvj.js";
+import { r as p } from "../../index--6vO-cOz.js";
 class m {
   variableNames = ["cache", "item"];
   outputShape;
   userCode;
   customUniforms = [{ name: "cacheT", type: "int" }];
-  constructor(t, a, o, s, n) {
-    const c = Math.min(o + 1, n);
-    this.outputShape = [t, a, c, s], this.userCode = `
+  constructor(t, a, n, o, c) {
+    const s = Math.min(n + 1, c);
+    this.outputShape = [t, a, s, o], this.userCode = `
         void main() {
             ivec4 coords = getOutputCoords(); // [b, h, t, d]
             int b = coords.x;
@@ -15,7 +15,7 @@ class m {
             int d = coords.w;
 
             int itemT = 1;
-            int maxSize = ${n};
+            int maxSize = ${c};
             int totalT = cacheT + itemT;
             int start = totalT >= maxSize ? 1 : 0;
 
@@ -23,21 +23,22 @@ class m {
             float val = 0.0;
             if (srcT < cacheT) {
                 val = getCache(b, h, srcT, d);
-            } else {
+            } else if (srcT == cacheT) {
                 val = getItem(b, h, 0, d);
-            }
+            } else {
+                val = 0.0;}
             setOutput(val);
         }
         `;
   }
 }
-function p(e) {
-  const { cache: t, item: a } = e.inputs, { maxSize: o } = e.attrs, s = e.backend, n = t.shape[0], c = t.shape[2], r = t.shape[1], i = new m(n, r, c, a.shape[3], o);
-  return s.runWebGLProgram(i, [t, a], "float32", [[c]]);
+function d(e) {
+  const { cache: t, item: a } = e.inputs, { maxSize: n, pastLen: o } = e.attrs, c = e.backend, s = t.shape[0], r = t.shape[2], i = t.shape[1], h = new m(s, i, r, a.shape[3], n);
+  return c.runWebGLProgram(h, [t, a], "float32", [[o]]);
 }
-const d = {
+const l = {
   kernelName: "AppendCache",
   backendName: "webgl",
-  kernelFunc: p
+  kernelFunc: d
 };
-h(d);
+p(l);

package/dist/ops/webgl/attentionMask.js

@@ -1,4 +1,4 @@
-import { r as h } from "../../index-C4JCoBvj.js";
+import { r as h } from "../../index--6vO-cOz.js";
 class l {
   variableNames = ["q", "k"];
   outputShape;