@genai-fi/nanogpt 0.19.1 → 0.20.0

package/dist/ops/reshape16.js

@@ -1,41 +0,0 @@
-import { e as p, c as s, h as c } from "../index-CUXkjxiT.js";
-import { r as u } from "../reshape-BDOuCSNW.js";
-const i = {
-  kernelName: "Reshape16",
-  inputsToSave: ["x"],
-  gradFunc: (e, r) => {
-    const [n] = r;
-    if (Array.isArray(e))
-      throw new Error("Reshape16 gradient does not support multiple outputs.");
-    return { x: () => m(e, n.shape) };
-  }
-};
-p(i);
-function a(e) {
-  const { inputs: r, attrs: n } = e, { x: t } = r, { shape: o } = n;
-  return u(t, o);
-}
-const l = {
-  kernelName: "Reshape16",
-  backendName: "webgpu",
-  kernelFunc: a
-};
-s(l);
-const h = {
-  kernelName: "Reshape16",
-  backendName: "webgl",
-  kernelFunc: a
-};
-s(h);
-const g = {
-  kernelName: "Reshape16",
-  backendName: "cpu",
-  kernelFunc: a
-};
-s(g);
-function m(e, r) {
-  return c().runKernel("Reshape16", { x: e }, { shape: r });
-}
-export {
-  m as reshape16
-};

package/dist/ops/rope.d.ts

@@ -1,3 +0,0 @@
-import { default as RoPECache } from '../layers/RoPECache';
-import { Tensor } from '@tensorflow/tfjs-core';
-export declare function rope(x: Tensor, cache: RoPECache, pastLength: number, negSin?: boolean): Tensor;

package/dist/ops/rope.js

@@ -1,7 +0,0 @@
-import "../index-CUXkjxiT.js";
-import "./cpu/rope.js";
-import "./webgl/rope.js";
-import { r as i } from "../rope-DcrZM_e6.js";
-export {
-  i as rope
-};

package/dist/ops/scatterSub.d.ts

@@ -1,2 +0,0 @@
-import { Tensor } from '@tensorflow/tfjs-core';
-export declare function scatterSub(probabilities: Tensor, labels: Tensor, scale: Tensor): Tensor;

package/dist/ops/scatterSub.js

@@ -1,9 +0,0 @@
-import { h as i } from "../index-CUXkjxiT.js";
-import "./cpu/scatterSub.js";
-import "./webgl/scatterSub.js";
-function c(t, r, e) {
-  return i().runKernel("EfficientScatterSub", { logits: t, labels: r, dy: e }, {});
-}
-export {
-  c as scatterSub
-};

package/dist/ops/slice16.d.ts

@@ -1,2 +0,0 @@
-import { Rank, Tensor } from '@tensorflow/tfjs-core';
-export declare function slice16<R extends Rank = Rank>(x: Tensor<R>, begin: number | number[], size: number | number[]): Tensor<R>;

package/dist/ops/slice16.js

@@ -1,9 +0,0 @@
-import { isPackedTensor as n } from "../utilities/packed.js";
-import { h as c } from "../index-CUXkjxiT.js";
-import { s as i } from "../slice-BzS11Qh0.js";
-function a(r, e, o) {
-  return n(r) ? c().runKernel("Slice16", { x: r }, { begin: e, size: o }) : i(r, e, o);
-}
-export {
-  a as slice16
-};

package/dist/ops/softmax16.d.ts

@@ -1,2 +0,0 @@
-import { Tensor } from '@tensorflow/tfjs-core';
-export declare function softmax16(logits: Tensor): Tensor;

package/dist/ops/softmax16.js

@@ -1,9 +0,0 @@
-import { h as n } from "../index-CUXkjxiT.js";
-import "./grads/softmax16.js";
-import { isPackedTensor as e } from "../utilities/packed.js";
-function t(r) {
-  return e(r) ? n().runKernel("Softmax16", { logits: r }, { dim: r.rank - 1 }) : n().runKernel("Softmax", { logits: r }, { dim: r.rank - 1 });
-}
-export {
-  t as softmax16
-};

package/dist/ops/sub16.d.ts

@@ -1,2 +0,0 @@
-import { Tensor } from '@tensorflow/tfjs-core';
-export declare function sub16(a: Tensor, b: Tensor): Tensor;

package/dist/ops/sub16.js

@@ -1,8 +0,0 @@
-import { a1 as s, h as t } from "../index-CUXkjxiT.js";
-import { isPackedTensor as n } from "../utilities/packed.js";
-function f(r, e) {
-  return !n(r) && !n(e) ? s(r, e) : t().runKernel("Sub16", { a: r, b: e });
-}
-export {
-  f as sub16
-};

package/dist/ops/sum16.d.ts

@@ -1,2 +0,0 @@
-import { Tensor } from '@tensorflow/tfjs-core';
-export declare function sum16(x: Tensor, axis?: number | number[], keepDims?: boolean): Tensor;

package/dist/ops/sum16.js

@@ -1,13 +0,0 @@
-import { h as e } from "../index-CUXkjxiT.js";
-import { isPackedTensor as s } from "../utilities/packed.js";
-import { s as n } from "../sum-BpiwSWvg.js";
-function p(r, t, o = !1) {
-  if (!s(r))
-    return n(r, t, o);
-  if (o)
-    throw new Error("sum16 with keepDims=true not supported for packed tensors");
-  return e().runKernel("Sum16", { x: r }, { axis: t ?? -1, keepDims: o });
-}
-export {
-  p as sum16
-};

package/dist/ops/transpose16.d.ts

@@ -1,3 +0,0 @@
-import { GradConfig, Tensor } from '@tensorflow/tfjs-core';
-export declare const transpose16GradConfig: GradConfig;
-export declare function transpose16(x: Tensor, perm?: number[]): Tensor;

package/dist/ops/transpose16.js

@@ -1,40 +0,0 @@
-import { e as i, c as p, h as u } from "../index-CUXkjxiT.js";
-import { forcePacked as l, forceFloat as m } from "./grads/utils.js";
-import { g } from "../axis_util-GTVlo58H.js";
-import { isPackedTensor as f } from "../utilities/packed.js";
-import { t as a } from "../transpose-9kRxIXWR.js";
-const d = {
-  kernelName: "Transpose16",
-  gradFunc: (r, s, t) => {
-    if (Array.isArray(r))
-      throw new Error("Transpose16 gradient does not support multiple outputs.");
-    const n = t, { perm: e } = n, o = g(e);
-    return { x: () => T(r, o) };
-  }
-};
-i(d);
-function c(r) {
-  const { inputs: s, attrs: t } = r, { x: n } = s, { perm: e } = t, o = f(n);
-  if (o && e[e.length - 1] !== n.shape.length - 1)
-    throw new Error("Transpose16 currently only supports the last axis being unchanged.");
-  return o ? l(a(m(n), e)) : a(n, e);
-}
-const k = {
-  kernelName: "Transpose16",
-  backendName: "webgl",
-  kernelFunc: c
-};
-p(k);
-const h = {
-  kernelName: "Transpose16",
-  backendName: "cpu",
-  kernelFunc: c
-};
-p(h);
-function T(r, s) {
-  return s == null && (s = r.shape.map((n, e) => e).reverse()), u().runKernel("Transpose16", { x: r }, { perm: s });
-}
-export {
-  T as transpose16,
-  d as transpose16GradConfig
-};

package/dist/ops/unpack16.d.ts

@@ -1,2 +0,0 @@
-import { Tensor } from '@tensorflow/tfjs-core';
-export declare function unpack16(x: Tensor, scaling?: number, disposeArg?: boolean): Tensor;

package/dist/ops/unpack16.js

@@ -1,6 +0,0 @@
-import "../index-CUXkjxiT.js";
-import { u as t } from "../pack16-bqltoUlR.js";
-import "../utilities/packed.js";
-export {
-  t as unpack16
-};

package/dist/ops/webgl/adamAdjust.d.ts

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

package/dist/ops/webgl/adamAdjust.js

@@ -1,49 +0,0 @@
-import { r } from "../../Reshape-BW__R4mZ.js";
-import { c as v } from "../../index-CUXkjxiT.js";
-class g {
-  variableNames = ["moments", "value"];
-  outputShape;
-  userCode;
-  customUniforms = [
-    { name: "invBeta1", type: "float" },
-    { name: "invBeta2", type: "float" },
-    { name: "learningRate", type: "float" },
-    { name: "epsilon", type: "float" }
-  ];
-  constructor(s, t) {
-    this.outputShape = s, t && this.customUniforms.push({ name: "weightDecay", type: "float" }), this.userCode = `
-        void main() {
-            float v = getValueAtOutCoords();
-            int coords = getOutputCoords();
-            coords *= 2;
-            float m1 = getMoments(coords);
-            float m2 = getMoments(coords + 1);
-
-            float m1Hat = m1 * invBeta1;
-            float m2Hat = m2 * invBeta2;
-
-            float invSqrt = inversesqrt(max(m2Hat, 1e-30));
-            float invDenom = invSqrt / (1.0 + epsilon * invSqrt);
-            float adjustedValue = -learningRate * m1Hat * invDenom + v;
-
-            ${t ? "adjustedValue = adjustedValue - learningRate * weightDecay * v;" : ""}
-
-            setOutput(adjustedValue);
-        }
-        `;
-  }
-}
-function b(a) {
-  const { moments: s, value: t } = a.inputs, { beta1: l, beta2: d, learningRate: p, epsilon: c, weightDecay: n } = a.attrs, e = a.backend, i = r({ inputs: { x: s }, backend: e, attrs: { shape: [-1] } }), o = r({ inputs: { x: t }, backend: e, attrs: { shape: [-1] } }), f = new g(o.shape, n > 0), m = [[1 / l], [1 / d], [p], [c]];
-  n > 0 && m.push([n]);
-  const u = e.runWebGLProgram(f, [i, o], "float32", m);
-  e.disposeIntermediateTensorInfo(i), e.disposeIntermediateTensorInfo(o);
-  const h = r({ inputs: { x: u }, backend: e, attrs: { shape: t.shape } });
-  return e.disposeIntermediateTensorInfo(u), h;
-}
-const j = {
-  kernelName: "AdamAdjust",
-  backendName: "webgl",
-  kernelFunc: b
-};
-v(j);

package/dist/ops/webgl/adamMoments.d.ts

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

package/dist/ops/webgl/adamMoments.js

@@ -1,40 +0,0 @@
-import { c as m } from "../../index-CUXkjxiT.js";
-class i {
-  variableNames = ["moments", "gradient"];
-  outputShape;
-  userCode;
-  customUniforms = [
-    { name: "beta1", type: "float" },
-    { name: "beta2", type: "float" }
-  ];
-  constructor(e) {
-    this.outputShape = e;
-    const t = e.length, n = t === 1 ? "int" : `ivec${Math.min(t, 4)}`, a = t === 1 ? "coords" : `coords[${t - 1}]`, r = t === 1 ? "coords" : e.slice(0, -1).map((s, c) => `coords[${c}]`).join(", ");
-    this.userCode = `
-        void main() {
-            float m = getMomentsAtOutCoords();
-            ${n} coords = getOutputCoords();
-
-            // Add gradient clipping here
-            float g = clamp(getGradient(${r}), -1.0, 1.0);
-            int which = ${a};
-
-            float beta = which == 0 ? beta1 : beta2;
-            float gg = which == 0 ? g : g * g;
-    
-            float newM = m * beta + gg * (1.0 - beta);
-            setOutput(newM);
-        }
-        `;
-  }
-}
-function d(o) {
-  const { moments: e, gradient: t } = o.inputs, { beta1: n, beta2: a } = o.attrs, r = o.backend, s = new i(e.shape);
-  return r.runWebGLProgram(s, [e, t], "float32", [[n], [a]]);
-}
-const g = {
-  kernelName: "AdamMoments",
-  backendName: "webgl",
-  kernelFunc: d
-};
-m(g);

package/dist/ops/webgl/appendCache.d.ts

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

package/dist/ops/webgl/appendCache.js

@@ -1,44 +0,0 @@
-import { c as p } from "../../index-CUXkjxiT.js";
-class m {
-  variableNames = ["cache", "item"];
-  outputShape;
-  userCode;
-  customUniforms = [{ name: "cacheT", type: "int" }];
-  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;
-            int h = coords.y;
-            int t = coords.z;
-            int d = coords.w;
-
-            int itemT = 1;
-            int maxSize = ${c};
-            int totalT = cacheT + itemT;
-            int start = totalT >= maxSize ? 1 : 0;
-
-            int srcT = t + start;
-            float val = 0.0;
-            if (srcT < cacheT) {
-                val = getCache(b, h, srcT, d);
-            } else if (srcT == cacheT) {
-                val = getItem(b, h, 0, d);
-            } else {
-                val = 0.0;}
-            setOutput(val);
-        }
-        `;
-  }
-}
-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 l = {
-  kernelName: "AppendCache",
-  backendName: "webgl",
-  kernelFunc: d
-};
-p(l);

package/dist/ops/webgl/attentionMask.d.ts

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

package/dist/ops/webgl/attentionMask.js

@@ -1,45 +0,0 @@
-import { c as m } from "../../index-CUXkjxiT.js";
-class h {
-  variableNames = ["q", "k"];
-  outputShape;
-  userCode;
-  customUniforms = [
-    { name: "divisor", type: "float" },
-    { name: "pastLen", type: "int" },
-    { name: "inf", type: "float" }
-  ];
-  constructor(t, e, s, n, a) {
-    this.outputShape = [t, e, s, n], this.userCode = `
-        void main() {
-            ivec4 coords = getOutputCoords(); // [batch, nh, t1, t2]
-            int b = coords.x;
-            int h = coords.y;
-            int t1 = coords.z;
-            int t2 = coords.w;
-
-            float sum = 0.0;
-            for (int i = 0; i < ${a}; ++i) {
-                float qv = getQ(b, h, t1, i);
-                float kv = getK(b, h, t2, i);
-                sum += qv * kv;
-            }
-
-            // Scale by divisor
-            float scaled = sum * divisor;
-
-            // Mask out future positions
-            setOutput((t2 > t1 + pastLen) ? inf : scaled);
-        }
-        `;
-  }
-}
-function l(o) {
-  const { q: t, k: e } = o.inputs, { divisor: s, pastLen: n } = o.attrs, a = o.backend, i = t.shape[0], r = t.shape[2], c = e.shape[2], u = t.shape[1], p = t.shape[3], d = new h(i, u, r, c, p);
-  return a.runWebGLProgram(d, [t, e], "float32", [[s], [n], [Number.NEGATIVE_INFINITY]]);
-}
-const f = {
-  kernelName: "AttentionMask",
-  backendName: "webgl",
-  kernelFunc: l
-};
-m(f);

package/dist/ops/webgl/dropout16.d.ts

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

package/dist/ops/webgl/dropout16.js

@@ -1,11 +0,0 @@
-import { c as n } from "../../index-CUXkjxiT.js";
-function o(e) {
-  const { x: r } = e.inputs;
-  return r;
-}
-const t = {
-  kernelName: "Dropout16",
-  backendName: "webgl",
-  kernelFunc: o
-};
-n(t);

package/dist/ops/webgl/fusedSoftmax.d.ts

@@ -1,11 +0,0 @@
-import { SoftmaxAttrs, SoftmaxInputs, TensorInfo } from '@tensorflow/tfjs-core';
-interface FusedSoftmaxAttrs extends SoftmaxAttrs {
-    dropoutRate?: number;
-    seed?: number;
-}
-export declare function softmax(args: {
-    inputs: SoftmaxInputs;
-    backend: unknown;
-    attrs: FusedSoftmaxAttrs;
-}): TensorInfo;
-export {};

package/dist/ops/webgl/fusedSoftmax.js

@@ -1,80 +0,0 @@
-import { m as b, s as I, r as k } from "../../RealDiv-CGwv0liw.js";
-import { r as v } from "../../Reshape-BW__R4mZ.js";
-import { c as w, ah as P } from "../../index-CUXkjxiT.js";
-import { e as S } from "../../axis_util-GTVlo58H.js";
-class T {
-  variableNames = ["logits", "maxLogits"];
-  outputShape;
-  userCode;
-  constructor(t) {
-    this.outputShape = t, this.userCode = `
-        void main() {
-        ivec4 coords = getOutputCoords(); // [batch, nh, t1, t2]
-            int b = coords.x;
-            int h = coords.y;
-            int t1 = coords.z;
-            int t2 = coords.w;
-            float x = getLogitsAtOutCoords();
-            float maxLogit = getMaxLogits(b, h, t1);
-            setOutput(exp(x - maxLogit));
-        }
-        `;
-  }
-}
-class C {
-  variableNames = ["exp", "sum"];
-  outputShape;
-  userCode;
-  customUniforms = [
-    { name: "dropoutRate", type: "float" },
-    { name: "seed", type: "float" }
-  ];
-  constructor(t) {
-    this.outputShape = t, this.userCode = `
-        float random(ivec4 coords) {
-            float x = float(coords.x * 4096 + coords.y * 256 + coords.z * 16 + coords.w);
-            return fract(sin(seed + x) * 43758.5453123);
-        }
-        void main() {
-            ivec4 coords = getOutputCoords();
-            float numerator = getExp(coords.x, coords.y, coords.z, coords.w);
-            float denominator = getSum(coords.x, coords.y, coords.z, coords.w);
-            float val = numerator / denominator;
-            float keepProb = 1.0 - dropoutRate;
-            float rand = random(coords);
-            float mask = step(rand, keepProb);
-            setOutput(val * mask / keepProb);
-        }
-        `;
-  }
-}
-function L(r) {
-  const { inputs: t, attrs: m } = r, { logits: e } = t, { dim: c, dropoutRate: n, seed: u } = m, o = r.backend;
-  if (!e)
-    throw new Error("Error in softmax: input logits is null");
-  const i = P([c], e.shape), d = b({
-    inputs: { x: e },
-    backend: o,
-    attrs: { reductionIndices: i, keepDims: !1 }
-  }), f = S(d.shape, i), l = new T(e.shape), s = o.runWebGLProgram(l, [e, d], "float32");
-  o.disposeIntermediateTensorInfo(d);
-  const p = I({ inputs: { x: s }, backend: o, attrs: { axis: i, keepDims: !1 } }), a = v({ inputs: { x: p }, backend: o, attrs: { shape: f } });
-  if (n !== void 0 && n > 0) {
-    const h = new C(e.shape), g = o.runWebGLProgram(h, [s, a], "float32", [
-      [n],
-      [u ?? Math.random() * 1e4]
-    ]);
-    return o.disposeIntermediateTensorInfo(s), o.disposeIntermediateTensorInfo(p), o.disposeIntermediateTensorInfo(a), g;
-  }
-  const x = k({ inputs: { a: s, b: a }, backend: o });
-  return o.disposeIntermediateTensorInfo(s), o.disposeIntermediateTensorInfo(p), o.disposeIntermediateTensorInfo(a), x;
-}
-const E = {
-  kernelName: "FusedSoftmax",
-  backendName: "webgl",
-  kernelFunc: L
-};
-w(E);
-export {
-  L as softmax
-};

package/dist/ops/webgl/gatherSub.d.ts

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

package/dist/ops/webgl/gatherSub.js

@@ -1,27 +0,0 @@
-import { c as l } from "../../index-CUXkjxiT.js";
-class u {
-  variableNames = ["labels", "logits", "values"];
-  outputShape;
-  userCode;
-  constructor(e) {
-    this.outputShape = [e], this.userCode = `
-          void main() {
-              int coords = getOutputCoords();
-              int index = int(getLabelsAtOutCoords());
-              float val = getValuesAtOutCoords();
-              float logit = getLogits(coords, index);
-              setOutput(val - logit);
-          }
-    `;
-  }
-}
-function i(t) {
-  const { logits: e, labels: o, values: s } = t.inputs, a = t.backend, r = o.shape[0], n = new u(r);
-  return a.runWebGLProgram(n, [o, e, s], "float32");
-}
-const c = {
-  kernelName: "EfficientGatherSub",
-  backendName: "webgl",
-  kernelFunc: i
-};
-l(c);

package/dist/ops/webgl/gelu.d.ts

@@ -1,2 +0,0 @@
-import { KernelFunc } from '@tensorflow/tfjs-core';
-export declare const gelu: KernelFunc;

package/dist/ops/webgl/gelu.js

@@ -1,50 +0,0 @@
-import { c as a } from "../../index-CUXkjxiT.js";
-import { u as s, C as i } from "../../kernel_funcs_utils-pq0CK9co.js";
-const t = 0.7978845608028654, r = 0.044715, c = i + `
-    float x3 = x * x * x;
-    float inner = x + ${r} * x3;
-    inner = ${t} * inner;
-    inner = abs(inner) > 15.0 ? sign(inner) : tanh(inner);
-    inner = 0.5 * (1.0 + inner);
-    inner = x * inner;
-    return inner;
-`, d = s({ opSnippet: c }), x = {
-  kernelName: "Gelu",
-  backendName: "webgl",
-  kernelFunc: d
-};
-a(x);
-class f {
-  // Inputs: dy, x
-  variableNames = ["dy", "x"];
-  outputShape;
-  userCode;
-  constructor(n) {
-    this.outputShape = n, this.userCode = `
-            void main() {
-                float dy = getDyAtOutCoords();
-                float x  = getXAtOutCoords();
-                float x2 = x * x;
-                float x3 = x2 * x;
-                float u  = ${t} * (x + ${r} * x3);
-                float t = abs(u) > 15.0 ? sign(u) : tanh(u);
-                float sech2 = 1.0 - t * t;
-                float du_dx = ${t} * (1.0 + 3.0 * ${r} * x2);
-                float dgelu = 0.5 * (1.0 + t) + 0.5 * x * sech2 * du_dx;
-                setOutput(dy * dgelu);
-            }`;
-  }
-}
-function g(e) {
-  const { dy: n, x: o } = e.inputs, u = e.backend, l = new f(o.shape);
-  return u.runWebGLProgram(l, [n, o], "float32");
-}
-const p = {
-  kernelName: "GeluGrad",
-  backendName: "webgl",
-  kernelFunc: g
-};
-a(p);
-export {
-  d as gelu
-};

package/dist/ops/webgl/log.d.ts

@@ -1,17 +0,0 @@
-/**
- * @license
- * Copyright 2020 Google LLC. All Rights Reserved.
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- * =============================================================================
- */
-export {};

package/dist/ops/webgl/log.js

@@ -1,23 +0,0 @@
-import { c as e, aO as r } from "../../index-CUXkjxiT.js";
-import { u as s, l as N } from "../../kernel_funcs_utils-pq0CK9co.js";
-import { y as l } from "../../shared-3agzAqQ_.js";
-const a = N + `
-  return x < 0.0 ? NAN : log(x);
-`, t = `
-  vec4 result = log(x);
-  bvec4 isNaN = isnan(x);
-  result.r = isNaN.r ? x.r : (x.r < 0.0 ? NAN : result.r);
-  result.g = isNaN.g ? x.g : (x.g < 0.0 ? NAN : result.g);
-  result.b = isNaN.b ? x.b : (x.b < 0.0 ? NAN : result.b);
-  result.a = isNaN.a ? x.a : (x.a < 0.0 ? NAN : result.a);
-  return result;
-`, n = s({
-  opSnippet: a,
-  packedOpSnippet: t,
-  cpuKernelImpl: l
-}), o = {
-  kernelName: r,
-  backendName: "webgl",
-  kernelFunc: n
-};
-e(o);

package/dist/ops/webgl/matMul16.d.ts

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