@genai-fi/nanogpt 0.19.0 → 0.20.0

package/dist/ops/webgpu/slice16.js

@@ -1,69 +0,0 @@
-import { b as u, c as m, e as l } from "../../webgpu_program-B4HmApL1.js";
-import { f, c as g } from "../../webgpu_util-DYlGSwOJ.js";
-import { c as S, U as k } from "../../index-CUXkjxiT.js";
-import { p as y, a as $ } from "../../slice_util-CC35pLmT.js";
-function x(o) {
-  switch (o) {
-    case 1:
-      return "1D";
-    case 2:
-      return "2D";
-    case 3:
-      return "3D";
-    case 4:
-      return "4D";
-  }
-  return "";
-}
-class b {
-  variableNames = ["source"];
-  uniforms;
-  outputShape;
-  shaderKey;
-  rank;
-  dispatchLayout;
-  dispatch;
-  workPerThread = 1;
-  workgroupSize = [64, 1, 1];
-  start;
-  size = !0;
-  constructor(e, t) {
-    this.outputShape = t, this.rank = t.length, this.dispatchLayout = f(this.outputShape), this.dispatch = g(this.dispatchLayout, this.outputShape, this.workgroupSize, [
-      this.workPerThread,
-      1,
-      1
-    ]), this.start = e, this.uniforms = `start : ${u(e.length)}, `, this.shaderKey = "slice";
-  }
-  getUserCode() {
-    const e = u(this.rank);
-    let t;
-    return this.start.length === 1 ? t = this.outputShape.map(() => "sourceLoc = uniforms.start + coords;") : t = this.outputShape.map((r, s) => `sourceLoc.${c[s]} = uniforms.start.${m(s)} + coords.${c[s]};`), `
-      ${l("index")} {
-        if (index < uniforms.size) {
-          var sourceLoc : ${e};
-          let coords = getCoordsFromIndex(index);
-          ${t.join(`
-`)}
-          result[index] = source[getIndexFromCoords${x(this.rank)}(sourceLoc, uniforms.sourceShape)];
-        }
-      }
-    `;
-  }
-}
-const c = ["x", "y", "z", "w", "u", "v"];
-function w(o) {
-  const { inputs: e, backend: t, attrs: n } = o, { x: r } = e, { begin: s, size: p } = n, [i, a] = y(r, s, p);
-  if ($(r, i, a), k(a) === 0)
-    return t.makeTensorInfo(a, r.dtype, []);
-  const h = new b(i, a), d = [{ type: "int32", data: i }];
-  return t.runWebGPUProgram(h, [r], r.dtype, d);
-}
-const C = {
-  kernelName: "Slice16",
-  backendName: "webgpu",
-  kernelFunc: w
-};
-S(C);
-export {
-  w as slice
-};

package/dist/ops/webgpu/softmax16.d.ts

@@ -1,17 +0,0 @@
-/**
- * @license
- * Copyright 2023 Google LLC.
- * 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/webgpu/softmax16.js

@@ -1,21 +0,0 @@
-import { c as h, U as S, h as b } from "../../index-CUXkjxiT.js";
-import { reshape16 as d } from "../reshape16.js";
-import x from "./softmax16_program.js";
-import k from "./softmax16_subgroup_program.js";
-import l from "./utils/deviceInfo.js";
-import { r as z } from "../../reshape-BDOuCSNW.js";
-function F(a) {
-  const { inputs: t, backend: o, attrs: p } = a, { logits: e } = t, { dim: r } = p, m = o.subgroupMinSize, i = o.subgroupMaxSize, c = l(o).subgroupsSupported, s = z(e, [
-    S(e.shape) / e.shape[r],
-    e.shape[r]
-  ]), u = c ? new k(s.shape, m, i) : new x(s.shape), f = o.runWebGPUProgram(u, [s], "packedF16");
-  s.dispose();
-  const n = b().makeTensorFromTensorInfo(f), g = d(n, e.shape);
-  return n.dispose(), g;
-}
-const I = {
-  kernelName: "Softmax16",
-  backendName: "webgpu",
-  kernelFunc: F
-};
-h(I);

package/dist/ops/webgpu/softmax16_program.d.ts

@@ -1,13 +0,0 @@
-import { WebGPUProgram } from '@tensorflow/tfjs-backend-webgpu';
-export default class SoftmaxProgram implements WebGPUProgram {
-    variableNames: string[];
-    outputShape: number[];
-    shaderKey: string;
-    dispatchLayout: {
-        x: number[];
-    };
-    dispatch: [number, number, number];
-    workgroupSize: [number, number, number];
-    constructor(outputShape: number[]);
-    getUserCode(): string;
-}

package/dist/ops/webgpu/softmax16_program.js

@@ -1,73 +0,0 @@
-import { e } from "../../webgpu_program-B4HmApL1.js";
-import { f as o } from "../../webgpu_util-DYlGSwOJ.js";
-class i {
-  variableNames = ["logits"];
-  outputShape;
-  shaderKey;
-  dispatchLayout;
-  dispatch;
-  workgroupSize;
-  constructor(r) {
-    this.outputShape = r, this.dispatchLayout = o(this.outputShape), this.dispatch = [this.outputShape[0], 1, 1], this.outputShape[1] >= 4096 ? this.workgroupSize = [256, 1, 1] : this.outputShape[1] < 64 ? this.workgroupSize = [32, 1, 1] : this.workgroupSize = [64, 1, 1], this.shaderKey = "softmax16";
-  }
-  getUserCode() {
-    return `
-        var<workgroup> buf : array<f32, ${this.workgroupSize[0]}>;
-        const blockSize = ${this.workgroupSize[0]};
-        ${e("index")} {
-            let row = index / blockSize;
-            let tid = i32(localId.x);
-            let cols = uniforms.outShape[1];
-            let rowIdx = row * cols;
-
-            var threadMax = -3.402823e+38f;
-            for (var col = tid; col < cols; col += blockSize) {
-                let value = unpack2x16float(u32(logits[rowIdx + col]));
-                threadMax = max(threadMax, max(value.x, value.y));
-            }
-            buf[tid] = threadMax;
-            workgroupBarrier();
-
-            for (var currSize = blockSize >> 1;  currSize > 0; currSize = currSize >> 1) {
-                if (tid < currSize) {
-                    buf[tid] = max(buf[tid], buf[tid + currSize]);
-                }
-                workgroupBarrier();
-            }
-
-            let rowMaxShared: f32 = buf[0];
-            workgroupBarrier();
-
-            var threadSum = 0.0f;
-            for (var col = tid; col < cols; col += blockSize) {
-                let value = unpack2x16float(u32(logits[rowIdx + col]));
-                let subExp = exp(value.x - rowMaxShared);
-                threadSum += subExp;
-                let subExpY = exp(value.y - rowMaxShared);
-                threadSum += subExpY;
-            }
-            buf[tid] = threadSum;
-            workgroupBarrier();
-
-            for (var currSize = blockSize >> 1;  currSize > 0; currSize = currSize >> 1) {
-                if (tid < currSize) {
-                    buf[tid] = buf[tid] + buf[tid + currSize];
-                }
-                workgroupBarrier();
-            }
-
-            let rowSumShared: f32 = buf[0];
-
-            for (var col = tid; col < cols; col += blockSize) {
-                let value = unpack2x16float(u32(logits[rowIdx + col]));
-                let value1: f32 = exp(value.x - rowMaxShared) / rowSumShared;
-                let value2: f32 = exp(value.y - rowMaxShared) / rowSumShared;
-                result[rowIdx + col] = i32(pack2x16float(vec2<f32>(value1, value2)));
-            }
-        }
-      `;
-  }
-}
-export {
-  i as default
-};

package/dist/ops/webgpu/softmax16_subgroup_program.d.ts

@@ -1,17 +0,0 @@
-import { WebGPUProgram } from '@tensorflow/tfjs-backend-webgpu';
-export default class SoftmaxSubgroupProgram implements WebGPUProgram {
-    variableNames: string[];
-    outputShape: number[];
-    shaderKey: string;
-    dispatchLayout: {
-        x: number[];
-    };
-    dispatch: [number, number, number];
-    workgroupSize: [number, number, number];
-    minSubgroupSize: number;
-    maxSubgroupSize: number;
-    subgroups: boolean;
-    subgroupBuiltins: boolean;
-    constructor(outputShape: number[], minSubgroupSize: number, maxSubgroupSize: number);
-    getUserCode(): string;
-}

package/dist/ops/webgpu/softmax16_subgroup_program.js

@@ -1,75 +0,0 @@
-import { e as o } from "../../webgpu_program-B4HmApL1.js";
-import { f as u } from "../../webgpu_util-DYlGSwOJ.js";
-class i {
-  variableNames = ["logits"];
-  outputShape;
-  shaderKey;
-  dispatchLayout;
-  dispatch;
-  workgroupSize;
-  minSubgroupSize;
-  maxSubgroupSize;
-  subgroups = !0;
-  subgroupBuiltins = !1;
-  constructor(e, a, r) {
-    this.minSubgroupSize = a, this.maxSubgroupSize = r, this.outputShape = e, this.dispatchLayout = u(this.outputShape), this.dispatch = [this.outputShape[0], 1, 1], this.workgroupSize = [Math.min(64, r), 1, 1], a !== r && (this.subgroupBuiltins = !0, this.workgroupSize = [64, 1, 1]), this.shaderKey = "softmax16subgroup";
-  }
-  getUserCode() {
-    const e = this.maxSubgroupSize !== this.minSubgroupSize;
-    return `
-        ${e ? `var<workgroup> bestValues : array<f32, ${this.workgroupSize[0]}>;` : ""}
-        const blockSize = ${this.workgroupSize[0]};
-        ${o("index")} {
-            let row = index / blockSize;
-            let tid = i32(localId.x);
-            let cols = uniforms.outShape[1];
-            let rowIdx = row * cols;
-
-            var threadMax = -3.402823e+38f;
-            for (var col = tid; col < cols; col += blockSize) {
-                let value = unpack2x16float(u32(logits[rowIdx + col]));
-                threadMax = max(threadMax, max(value.x, value.y));
-            }
-
-            threadMax = subgroupMax(threadMax);
-            ${e ? `
-                let lane = localId.x % subgroupSize;
-                if (lane == 0) {
-                    bestValues[localId.x / subgroupSize] = threadMax;
-                }
-                workgroupBarrier();
-                let numSubgroups = blockSize / subgroupSize;
-                threadMax = select(-3.402823e+38f, bestValues[lane], lane < numSubgroups);
-                threadMax = subgroupMax(threadMax);
-                workgroupBarrier();    
-            ` : ""}
-
-            var threadSum = 0.0f;
-            for (var col = tid; col < cols; col += blockSize) {
-                let value = unpack2x16float(u32(logits[rowIdx + col]));
-                let subExp = exp(value - threadMax);
-                threadSum += subExp.x + subExp.y;
-            }
-
-            threadSum = subgroupAdd(threadSum);
-            ${e ? `
-                if (lane == 0) {
-                    bestValues[localId.x / subgroupSize] = threadSum;
-                }
-                workgroupBarrier();
-                threadSum = select(0.0f, bestValues[lane], lane < numSubgroups);
-                threadSum = subgroupAdd(threadSum);    
-            ` : ""}
-
-            for (var col = tid; col < cols; col += ${e ? "i32(subgroupSize)" : "blockSize"}) {
-                let value = unpack2x16float(u32(logits[rowIdx + col]));
-                let valuePair: vec2<f32> = exp(value - threadMax) / threadSum;
-                result[rowIdx + col] = i32(pack2x16float(valuePair));
-            }
-        }
-      `;
-  }
-}
-export {
-  i as default
-};

package/dist/ops/webgpu/softmax16grad.d.ts

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

package/dist/ops/webgpu/softmax16grad.js

@@ -1,37 +0,0 @@
-import { c as p } from "../../index-CUXkjxiT.js";
-import { createReduceInfo as m, reduce as l, ReduceProgram as i } from "./utils/reductions.js";
-import { isPackedTensor as n } from "../../utilities/packed.js";
-import k from "./utils/deviceInfo.js";
-class x extends i {
-  constructor(e, t) {
-    super(e, t, { reductionOp: "sum", elementwise: !0 }, !0), this.shaderKey = "SoftmaxGrad16", this.variableNames = ["dy", "softmaxOutput"], this.variableComponents = [1, 1];
-  }
-  getReadSnippet() {
-    return `
-            let d: vec2<f32> = unpack2x16float(u32(dy[index]));
-            let l: vec2<f32> = unpack2x16float(u32(softmaxOutput[index]));
-            return d * l;
-        `;
-  }
-  getWriteSnippet() {
-    return `
-            let d: vec2<f32> = unpack2x16float(u32(dy[offset + k]));
-            let l: vec2<f32> = unpack2x16float(u32(softmaxOutput[offset + k]));
-            let outVal = l * (d - bestValue);
-            result[offset + k] = i32(pack2x16float(outVal));
-        `;
-  }
-}
-function b(o) {
-  const { dy: e, softmaxOutput: t } = o.inputs, r = o.backend, s = k(r), u = n(e), c = n(t);
-  if (!(u && c))
-    throw new Error("softmaxGradGPU: dy and softmaxOutput must be packed tensors");
-  const a = [e, t], f = m(a, -1), d = new x(s, f);
-  return l(d, a, r);
-}
-const v = {
-  kernelName: "Softmax16Grad",
-  backendName: "webgpu",
-  kernelFunc: b
-};
-p(v);

package/dist/ops/webgpu/sub16.d.ts

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

package/dist/ops/webgpu/sub16.js

@@ -1,13 +0,0 @@
-import { c as t } from "../../index-CUXkjxiT.js";
-import { BinaryOpProgram as s } from "./utils/binary_op.js";
-import { B as c } from "../../binary_op_util-pKXltfxI.js";
-function p(e) {
-  const { a: r, b: n } = e.inputs, o = e.backend, a = new s(c.SUB, r.shape, n.shape);
-  return o.runWebGPUProgram(a, [r, n], "packedF16");
-}
-const m = {
-  kernelName: "Sub16",
-  backendName: "webgpu",
-  kernelFunc: p
-};
-t(m);

package/dist/ops/webgpu/sum16.d.ts

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

package/dist/ops/webgpu/sum16.js

@@ -1,38 +0,0 @@
-import { createReduceInfo as h, reduce as g, ReduceProgram as x } from "./utils/reductions.js";
-import { c as k, ah as A } from "../../index-CUXkjxiT.js";
-import { isPackedTensor as P } from "../../utilities/packed.js";
-import { transpose16 as b } from "../transpose16.js";
-import I from "./utils/deviceInfo.js";
-import { s as w } from "../../sum-BpiwSWvg.js";
-import { a as D, b as K } from "../../axis_util-GTVlo58H.js";
-class v extends x {
-  shaderKey = "sum16";
-  constructor(e, t, o) {
-    super(
-      e,
-      t,
-      {
-        reductionOp: "sum",
-        elementwise: !1
-      },
-      o
-    ), o && (this.shaderKey += "_packed");
-  }
-}
-function y(r) {
-  const { x: e } = r.inputs, { axis: t, keepDims: o } = r.attrs, a = r.backend, c = [], u = I(a), m = P(e);
-  if (!m)
-    return w(e, t, o);
-  let n = A(t ?? -1, e.shape);
-  const i = D(n, e.shape.length);
-  let s = e;
-  i != null && (s = b(e, i), n = K(n.length, s.shape.length), c.push(s));
-  const p = h([s], -1), f = new v(u, p, m), d = g(f, [s], a);
-  return c.forEach((l) => l.dispose()), d;
-}
-const N = {
-  kernelName: "Sum16",
-  backendName: "webgpu",
-  kernelFunc: y
-};
-k(N);

package/dist/ops/webgpu/transpose16.d.ts

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

package/dist/ops/webgpu/transpose16.js

@@ -1,34 +0,0 @@
-import { isPackedTensor as d } from "../../utilities/packed.js";
-import { c as k, h as l } from "../../index-CUXkjxiT.js";
-import { reshape16 as u } from "../reshape16.js";
-import P from "./transpose16_shared_program.js";
-import b from "./transpose16_program.js";
-import { t as T } from "../../transpose-9kRxIXWR.js";
-function w(p) {
-  const { inputs: h, attrs: i } = p, { x: e } = h, { perm: r } = i, a = p.backend, c = d(e);
-  if (c && r[r.length - 1] !== e.shape.length - 1) {
-    const o = e.shape.length, m = o === 4 ? r.map((s) => s - 1).filter((s) => s >= 0) : r, t = o === 4 ? u(e, [e.shape[0] * e.shape[1], e.shape[2], e.shape[3]]) : e, f = new P(t.shape, m), n = a.runWebGPUProgram(f, [t], "packedF16");
-    if (o === 4) {
-      t.dispose();
-      const s = l().makeTensorFromTensorInfo(n), g = u(s, [
-        e.shape[0],
-        e.shape[1],
-        n.shape[1],
-        n.shape[2]
-      ]);
-      return s.dispose(), g;
-    }
-    return n;
-  }
-  if (c) {
-    const o = new b(e.shape, r);
-    return a.runWebGPUProgram(o, [e], "packedF16");
-  } else
-    return T(e, r);
-}
-const F = {
-  kernelName: "Transpose16",
-  backendName: "webgpu",
-  kernelFunc: w
-};
-k(F);

package/dist/ops/webgpu/transpose16_program.d.ts

@@ -1,16 +0,0 @@
-import { WebGPUProgram } from '@tensorflow/tfjs-backend-webgpu';
-export default class TransposeProgram16 implements WebGPUProgram {
-    variableNames: string[];
-    shaderKey: string;
-    outputShape: number[];
-    dispatchLayout: {
-        x: number[];
-    };
-    dispatch: [number, number, number];
-    workPerThread: number;
-    workgroupSize: [number, number, number];
-    newDim: number[];
-    size: boolean;
-    constructor(aShape: number[], newDim: number[]);
-    getUserCode(): string;
-}

package/dist/ops/webgpu/transpose16_program.js

@@ -1,50 +0,0 @@
-import { f as a, c as i } from "../../webgpu_util-DYlGSwOJ.js";
-import { b as h, e as d, c as n } from "../../webgpu_program-B4HmApL1.js";
-function p(r) {
-  const e = r.length;
-  if (e > 6)
-    throw Error(`Transpose for rank ${e} is not yet supported`);
-  const o = new Array(e);
-  for (let t = 0; t < r.length; t++)
-    o[r[t]] = `coords.${n(t)}`;
-  return o.join();
-}
-class l {
-  variableNames = ["A"];
-  shaderKey;
-  outputShape;
-  dispatchLayout;
-  dispatch;
-  workPerThread = 1;
-  workgroupSize = [64, 1, 1];
-  newDim;
-  size = !0;
-  constructor(e, o) {
-    const t = new Array(e.length);
-    for (let s = 0; s < t.length; s++)
-      t[s] = e[o[s]];
-    this.outputShape = t, this.dispatchLayout = a(this.outputShape), this.dispatch = i(this.dispatchLayout, this.outputShape, this.workgroupSize, [
-      this.workPerThread,
-      1,
-      1
-    ]), this.newDim = o, this.shaderKey = `transpose16_${o}`;
-  }
-  getUserCode() {
-    const e = h(this.outputShape.length), o = p(this.newDim);
-    return `
-      ${d("index")} {
-        for(var i = 0; i < ${this.workPerThread}; i = i + 1) {
-          let flatIndex = index * ${this.workPerThread} + i;
-          if(flatIndex < uniforms.size) {
-            let coords = getCoordsFromIndex(flatIndex);
-            result[flatIndex] = A[getIndexFromCoords${this.outputShape.length}D(
-              ${e}(${o}), uniforms.aShape)];
-          }
-        }
-      }
-    `;
-  }
-}
-export {
-  l as default
-};

package/dist/ops/webgpu/transpose16_shared_program.d.ts

@@ -1,15 +0,0 @@
-import { WebGPUProgram } from '@tensorflow/tfjs-backend-webgpu/dist/webgpu_program';
-export default class TransposeSharedProgram16 implements WebGPUProgram {
-    variableNames: string[];
-    outputShape: number[];
-    shaderKey: string;
-    dispatchLayout: {
-        x: number[];
-        y: number[];
-        z?: number[];
-    };
-    dispatch: [number, number, number];
-    workgroupSize: [number, number, number];
-    constructor(aShape: number[], newDim: number[]);
-    getUserCode(): string;
-}

package/dist/ops/webgpu/transpose16_shared_program.js

@@ -1,70 +0,0 @@
-import { c as a } from "../../webgpu_util-DYlGSwOJ.js";
-import { e as l } from "../../webgpu_program-B4HmApL1.js";
-import { x as p } from "../../index-CUXkjxiT.js";
-class n {
-  variableNames = ["A"];
-  outputShape;
-  shaderKey;
-  dispatchLayout;
-  dispatch;
-  // Note that the maximum number of workgroup invocations by webgpu is 256.
-  // Nick: Reduce to 8x8
-  workgroupSize = [8, 8, 1];
-  constructor(t, o) {
-    const i = t.length, e = new Array(i), r = t.slice();
-    r[r.length - 1] *= 2;
-    for (let u = 0; u < e.length; u++)
-      e[u] = r[o[u]];
-    e[e.length - 1] /= 2, this.outputShape = e, this.dispatchLayout = i === 2 ? { x: [0], y: [1] } : { x: [1], y: [2], z: [0] }, this.dispatch = a(this.dispatchLayout, this.outputShape, this.workgroupSize, [2, 1, 1]), this.shaderKey = `transposeShared16_${i}`;
-  }
-  getUserCode() {
-    const t = this.outputShape.length;
-    p(
-      this.workgroupSize[0] === this.workgroupSize[1],
-      () => `Must be a square tile, current tile shape is ${this.workgroupSize[0]} x ${this.workgroupSize[1]}`
-    );
-    const o = this.workgroupSize[0] * 2;
-    return `
-      var<workgroup> tile : array<array<f32, ${o + 1}>, ${o}>;
-      ${l()} {
-        var x = i32(workgroupId.x) * ${o / 2} + i32(localId.x);
-        var y = i32(workgroupId.y) * ${o} + i32(localId.y);
-        let batch = ${t === 3 ? "i32(workgroupId.z)" : "0"};
-        let batchOffsetA = ${t === 3 ? "batch * uniforms.aShapeStrides[0]" : "0"};
-        let batchOffsetOut = ${t === 3 ? "batch * uniforms.outShapeStrides[0]" : "0"};
-
-        let inputWidth = uniforms.outShape[${t === 3 ? "1" : "0"}] / 2; // Output height
-        let inputHeight = uniforms.outShape[${t === 3 ? "2" : "1"}] * 2; // Output width
-        if (x < inputWidth && y < inputHeight) {
-            let unpackedA = unpack2x16float(u32(A[batchOffsetA + y * inputWidth + x]));
-            tile[localId.y][localId.x * 2] = unpackedA.x;
-            tile[localId.y][localId.x * 2 + 1] = unpackedA.y;
-        }
-        // Second load to cover the tile
-        y = y + ${this.workgroupSize[0]};
-        if (x < inputWidth && y < inputHeight) {
-            let unpackedA = unpack2x16float(u32(A[batchOffsetA + y * inputWidth + x]));
-            tile[localId.y + ${this.workgroupSize[0]}][localId.x * 2] = unpackedA.x;
-            tile[localId.y + ${this.workgroupSize[0]}][localId.x * 2 + 1] = unpackedA.y;
-        }
-        workgroupBarrier();
-
-        let outputWidth = uniforms.outShape[${t === 3 ? "2" : "1"}]; // Output width
-        let outputHeight = uniforms.outShape[${t === 3 ? "1" : "0"}] * 2; // Output height
-        x = i32(workgroupId.y) * ${o / 2} + i32(localId.x);
-        y = i32(workgroupId.x) * ${o} + i32(localId.y);
-        if (x < outputWidth && y < outputHeight) {
-          result[batchOffsetOut + y * outputWidth + x] = i32(pack2x16float(vec2<f32>(tile[localId.x * 2][localId.y], tile[localId.x * 2 + 1][localId.y])));
-        }
-        // Second store to cover the tile
-        y = y + ${this.workgroupSize[0]};
-        if (x < outputWidth && y < outputHeight) {
-          result[batchOffsetOut + y * outputWidth + x] = i32(pack2x16float(vec2<f32>(tile[localId.x * 2][localId.y + ${this.workgroupSize[0]}], tile[localId.x * 2 + 1][localId.y + ${this.workgroupSize[0]}])));
-        }
-      }
-    `;
-  }
-}
-export {
-  n as default
-};

package/dist/ops/webgpu/unpack16.d.ts

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

package/dist/ops/webgpu/unpack16.js

@@ -1,48 +0,0 @@
-import { f as c, c as r } from "../../webgpu_util-DYlGSwOJ.js";
-import { e as u } from "../../webgpu_program-B4HmApL1.js";
-import { c as p } from "../../index-CUXkjxiT.js";
-class l {
-  outputShape;
-  shaderKey = "Unpack16";
-  dispatchLayout;
-  dispatch;
-  workgroupSize = [64, 1, 1];
-  variableNames = ["x"];
-  size = !0;
-  uniforms;
-  outputComponent = 4;
-  variableComponents = [2];
-  scaling = !1;
-  constructor(t) {
-    this.outputShape = [...t.slice(0, -1), t[t.length - 1] * 2], this.dispatchLayout = c(this.outputShape), this.dispatch = r(this.dispatchLayout, this.outputShape, this.workgroupSize, [4, 1, 1]);
-  }
-  useScaling() {
-    this.shaderKey += "_Scaled", this.uniforms = "scaling : f32,", this.scaling = !0;
-  }
-  getUserCode() {
-    return `
-                ${u("index")} {
-                    let outIndex = index;
-                    if (outIndex < uniforms.size) {
-                        let xvec2 = x[index];
-                        let v1 = vec4<f32>(
-                            unpack2x16float(u32(xvec2.x)),
-                            unpack2x16float(u32(xvec2.y))
-                        ) ${this.scaling ? "* uniforms.scaling" : ""};
-                        result[outIndex] = v1;
-                    }
-                }`;
-  }
-}
-function h(e) {
-  const { x: t } = e.inputs, { scaling: a } = e.attrs, i = e.backend, n = new l(t.shape), s = a !== 1;
-  s && n.useScaling();
-  const o = [{ type: "float32", data: [1 / a] }];
-  return i.runWebGPUProgram(n, [t], "float32", s ? o : void 0);
-}
-const d = {
-  kernelName: "Unpack16",
-  backendName: "webgpu",
-  kernelFunc: h
-};
-p(d);

package/dist/ops/webgpu/utils/binary_op.d.ts

@@ -1,35 +0,0 @@
-import { WebGPUProgram } from '@tensorflow/tfjs-backend-webgpu';
-import { BinaryOpType } from '@tensorflow/tfjs-backend-webgpu/dist/binary_op_util';
-export { BinaryOpType };
-export declare class BinaryOpProgram implements WebGPUProgram {
-    dispatch: [number, number, number];
-    dispatchLayout: {
-        x: number[];
-    };
-    outputComponent: number;
-    op: BinaryOpType;
-    outputShape: number[];
-    shaderKey: string;
-    size: boolean;
-    variableNames: string[];
-    workgroupSize: [number, number, number];
-    variableComponents: number[];
-    constructor(op: BinaryOpType, aShape: number[], bShape: number[]);
-    getUserCode(): string;
-}
-export declare class BinaryOpScalarProgram implements WebGPUProgram {
-    dispatch: [number, number, number];
-    dispatchLayout: {
-        x: number[];
-    };
-    outputComponent: number;
-    op: BinaryOpType;
-    outputShape: number[];
-    shaderKey: string;
-    size: boolean;
-    variableNames: string[];
-    workgroupSize: [number, number, number];
-    variableComponents: number[];
-    constructor(op: BinaryOpType, aShape: number[]);
-    getUserCode(): string;
-}