@genai-fi/nanogpt 0.19.1 → 0.20.0

package/dist/ops/webgpu/normRMSGrad.js

@@ -1,284 +0,0 @@
-import { c as $, a6 as y, h as L } from "../../index-CUXkjxiT.js";
-import { createReduceInfo as M } from "./utils/reductions.js";
-import { f as w } from "../../webgpu_util-DYlGSwOJ.js";
-import { e as _ } from "../../webgpu_program-B4HmApL1.js";
-import { p as x, u as R } from "../../pack16-bqltoUlR.js";
-import { isPackedTensor as h } from "../../utilities/packed.js";
-import { reshape16 as z } from "../reshape16.js";
-import { sum16 as N } from "../sum16.js";
-import { slice16 as v } from "../slice16.js";
-class Y {
-  outputShape;
-  shaderKey = "RMSNormGrad";
-  dispatchLayout;
-  dispatch;
-  workgroupSize = [64, 1, 1];
-  variableNames = ["x", "gamma", "dy"];
-  uniforms = "reduceSize : i32, batchSize: i32";
-  inputShape;
-  size = !1;
-  rowsPerWorkgroup;
-  packed = !1;
-  outputComponent;
-  constructor(e, a = 4, t = !1) {
-    if (this.packed = t, this.shaderKey = `RMSNormGrad_${a}`, this.rowsPerWorkgroup = a, this.inputShape = [e.batchSize, e.inSize], this.outputShape = [e.batchSize + e.batchSize / this.rowsPerWorkgroup, e.inSize], this.dispatchLayout = w(this.outputShape), this.dispatch = [e.batchSize / this.rowsPerWorkgroup, 1, 1], e.batchSize % this.rowsPerWorkgroup !== 0)
-      throw new Error(
-        `RMSNormGradProgram: batch size ${e.batchSize} must be divisible by rowsPerWorkgroup ${this.rowsPerWorkgroup}`
-      );
-    if (e.inSize > 1024)
-      throw new Error(`RMSNormGradProgram: inSize ${e.inSize} exceeds max of 1024`);
-  }
-  getUserCode() {
-    const e = this.workgroupSize[0], a = this.rowsPerWorkgroup, t = `
-          var<workgroup> partials : array<vec2<f32>, ${e}>;
-          var<workgroup> accumulation: array<${this.packed ? "vec2<f32>" : "f32"}, 1024>;
-        `, u = this.packed ? `
-                let X = unpack2x16float(u32(x[offset + k]));
-                let DY = unpack2x16float(u32(dy[offset + k]));
-                let G  = unpack2x16float(u32(gamma[k]));
-                sum_x2 = fma(X.x, X.x, sum_x2);
-                sum_x2 = fma(X.y, X.y, sum_x2);
-                sum_dygx = fma(DY.x * G.x, X.x, sum_dygx);
-                sum_dygx = fma(DY.y * G.y, X.y, sum_dygx);
-        ` : `
-                let X = f32(x[offset + k]);
-                let DY = f32(dy[offset + k]);
-                let G  = f32(gamma[k]);
-                sum_x2 = fma(X, X, sum_x2);
-                sum_dygx = fma(DY * G, X, sum_dygx);
-            `, m = this.packed ? `
-                let X  = unpack2x16float(u32(x[offset + k]));
-                let DY = unpack2x16float(u32(dy[offset + k]));
-                let G  = unpack2x16float(u32(gamma[k]));
-
-                let dyGamma = DY * G;
-                let dx = vec2<f32>(
-                    fma(dyGamma.x, invRMS, -X.x * scale),
-                    fma(dyGamma.y, invRMS, -X.y * scale)
-                );
-
-                result[offset + k] = i32(pack2x16float(dx));
-
-                // dGamma
-                accumulation[k] = fma(DY, X * invRMS, accumulation[k]);
-        ` : `
-                let X  = f32(x[offset + k]);
-                let DY = f32(dy[offset + k]);
-                let G  = f32(gamma[k]);
-
-                let dyGamma = DY * G;
-                let dx = fma(dyGamma, invRMS, -X * scale);
-
-                result[offset + k] = dx;
-
-                // dGamma
-                accumulation[k] = fma(DY, X * invRMS, accumulation[k]);
-        `, n = this.packed ? `
-                result[outDgBase + k] = i32(pack2x16float(accumulation[k]));
-        ` : `
-                result[outDgBase + k] = accumulation[k];
-        `;
-    return `
-          fn DIV_CEIL(a : u32, b : u32) -> u32 {
-            return ((a - 1u) / b + 1u);
-          }
-
-          ${t}
-
-          ${_("index")} {
-            // One workgroup per row (batch).
-            let Length = uniforms.reduceSize;
-            let BatchSize = uniforms.batchSize;
-            for (var k = i32(localId.x); k < Length; k = k + ${e}) {
-                accumulation[k] = ${this.packed ? "vec2<f32>(0.0f)" : "0.0f"};
-            }
-
-            for (var rowOff = 0; rowOff < ${a}; rowOff = rowOff + 1) {
-                let row = i32(workgroupId.x) * ${a} + rowOff;
-                let offset = row * Length;
-                
-                var sum_x2 = 0.0f;
-                var sum_dygx = 0.0f;
-
-                for (var k = i32(localId.x); k < Length; k = k + ${e}) {
-                    ${u}
-                }
-
-                partials[localId.x] = vec2<f32>(sum_x2, sum_dygx);
-                workgroupBarrier();
-
-                var reduceSize = min(u32(Length), ${e}u);
-                for (var currentSize = reduceSize / 2u; reduceSize > 1u; currentSize = reduceSize / 2u) {
-                    let interval = DIV_CEIL(reduceSize, 2u);
-                    if (localId.x < currentSize) {
-                        partials[localId.x] = partials[localId.x] + partials[localId.x + interval];
-                    }
-                    reduceSize = interval;
-                    workgroupBarrier();
-                }
-
-                let invN = 1.0f / f32(${this.packed ? "Length * 2" : "Length"});
-                let mean_x2   = fma(partials[0].x, invN, 1e-8);
-                let mean_dygx = partials[0].y * invN;
-
-                let invRMS = inverseSqrt(mean_x2);
-                let scale = (mean_dygx / (mean_x2)) * invRMS;
-
-                // write dx and dGamma.
-                for (var k = i32(localId.x); k < Length; k = k + ${e}) {
-                    ${m}
-                }
-
-                workgroupBarrier();
-            }
-
-            // Write out the partially accumulated dGamma
-            let outDgBase = BatchSize * Length + i32(workgroupId.x) * Length;
-            for (var k = i32(localId.x); k < Length; k = k + ${e}) {
-                ${n}
-            }
-          }
-        `;
-  }
-}
-class I {
-  outputShape;
-  shaderKey = "RMSNormGrad";
-  dispatchLayout;
-  dispatch;
-  workgroupSize = [64, 1, 1];
-  variableNames = ["x", "dy"];
-  uniforms = "reduceSize : i32, batchSize: i32";
-  inputShape;
-  size = !1;
-  packed = !1;
-  outputComponent;
-  constructor(e, a = !1) {
-    this.packed = a, this.shaderKey = "RMSNormGrad_NoGamma", this.inputShape = [e.batchSize, e.inSize], this.outputShape = [e.batchSize, e.inSize], this.dispatchLayout = w(this.outputShape), this.dispatch = [e.batchSize, 1, 1];
-  }
-  getUserCode() {
-    const e = this.workgroupSize[0], a = `
-          var<workgroup> partials : array<vec2<f32>, ${e}>;
-        `, t = this.packed ? `
-                let X = unpack2x16float(u32(x[offset + k]));
-                let DY = unpack2x16float(u32(dy[offset + k]));
-                sum_x2 = fma(X.x, X.x, sum_x2);
-                sum_x2 = fma(X.y, X.y, sum_x2);
-                sum_dygx = fma(DY.x, X.x, sum_dygx);
-                sum_dygx = fma(DY.y, X.y, sum_dygx);
-        ` : `
-                let X = f32(x[offset + k]);
-                let DY = f32(dy[offset + k]);
-                sum_x2 = fma(X, X, sum_x2);
-                sum_dygx = fma(DY, X, sum_dygx);
-            `, u = this.packed ? `
-                let X  = unpack2x16float(u32(x[offset + k]));
-                let DY = unpack2x16float(u32(dy[offset + k]));
-
-                let dx = vec2<f32>(
-                    fma(DY.x, invRMS, -X.x * scale),
-                    fma(DY.y, invRMS, -X.y * scale)
-                );
-
-                result[offset + k] = i32(pack2x16float(dx));
-        ` : `
-                let X  = f32(x[offset + k]);
-                let DY = f32(dy[offset + k]);
-
-                let dx = fma(DY, invRMS, -X * scale);
-
-                result[offset + k] = dx;
-        `;
-    return `
-          fn DIV_CEIL(a : u32, b : u32) -> u32 {
-            return ((a - 1u) / b + 1u);
-          }
-
-          ${a}
-
-          ${_("index")} {
-            // One workgroup per row (batch).
-            let Length = uniforms.reduceSize;
-            let BatchSize = uniforms.batchSize;
-
-                let row = i32(workgroupId.x);
-                let offset = row * Length;
-                
-                var sum_x2 = 0.0f;
-                var sum_dygx = 0.0f;
-
-                for (var k = i32(localId.x); k < Length; k = k + ${e}) {
-                    ${t}
-                }
-
-                partials[localId.x] = vec2<f32>(sum_x2, sum_dygx);
-                workgroupBarrier();
-
-                var reduceSize = min(u32(Length), ${e}u);
-                for (var currentSize = reduceSize / 2u; reduceSize > 1u; currentSize = reduceSize / 2u) {
-                    let interval = DIV_CEIL(reduceSize, 2u);
-                    if (localId.x < currentSize) {
-                        partials[localId.x] = partials[localId.x] + partials[localId.x + interval];
-                    }
-                    reduceSize = interval;
-                    workgroupBarrier();
-                }
-
-                let invN = 1.0f / f32(${this.packed ? "Length * 2" : "Length"});
-                let mean_x2   = fma(partials[0].x, invN, 1e-8);
-                let mean_dygx = partials[0].y * invN;
-
-                let invRMS = inverseSqrt(mean_x2);
-                let scale = (mean_dygx / (mean_x2)) * invRMS;
-
-                // write dx and dGamma.
-                for (var k = i32(localId.x); k < Length; k = k + ${e}) {
-                    ${u}
-                }
-          }
-        `;
-  }
-}
-function P(d) {
-  const { dy: e, x: a, gamma: t } = d.inputs, u = 4;
-  y(a.shape, e.shape, "Error in RMSNormGrad dy: ");
-  const m = h(a), n = t ? h(t) : !1, f = h(e), r = m || n || f, i = !r || m ? a : x(a), s = !r || n || !t ? t : x(t), c = !r || f ? e : x(e);
-  s && y(s.shape, [i.shape[i.shape.length - 1]], "Error in RMSNormGrad gamma: ");
-  const G = d.backend, o = M(s ? [i, s, c] : [i, c], -1), k = t ? new Y(o, u, r) : new I(o, r), X = [
-    { type: "int32", data: [k.inputShape[1]] },
-    // Reduce size
-    { type: "int32", data: [k.inputShape[0]] }
-    // Batch size
-  ];
-  if (o.inSize > 1024)
-    throw new Error(`rmsNormGradGPU: inSize ${o.inSize} exceeds max of 1024`);
-  const D = G.runWebGPUProgram(
-    k,
-    s ? [i, s, c] : [i, c],
-    r ? "packedF16" : "float32",
-    X
-  );
-  r && !m && i.dispose(), r && !n && s && s.dispose(), r && !f && c.dispose();
-  const l = L().makeTensorFromTensorInfo(D);
-  if (t) {
-    const p = v(l, [0, 0], [o.batchSize, o.inSize]), S = v(
-      l,
-      [o.batchSize, 0],
-      [o.batchSize / u, o.inSize]
-    );
-    l.dispose();
-    const b = z(p, a.shape);
-    p.dispose();
-    const g = N(S, [0]);
-    return S.dispose(), [b, !r || n ? g : R(g)];
-  } else {
-    const p = z(l, a.shape);
-    return l.dispose(), [p];
-  }
-}
-const C = {
-  kernelName: "RMSNormGrad",
-  backendName: "webgpu",
-  kernelFunc: P
-};
-$(C);

package/dist/ops/webgpu/pack16.d.ts

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

package/dist/ops/webgpu/pack16.js

@@ -1,18 +0,0 @@
-import { c as i } from "../../index-CUXkjxiT.js";
-import p from "./pack16_program.js";
-function m(e) {
-  const { x: n } = e.inputs, { scaling: a, padding: r } = e.attrs, c = e.backend;
-  if (n.shape[n.shape.length - 1] % 2 !== 0)
-    throw new Error("Last dimension of input tensor must be even to use Pack16.");
-  e.attrs && (e.attrs.originalShape = n.shape);
-  const t = new p(n.shape, r), o = a !== 1;
-  o && t.useScaling();
-  const s = [{ type: "float32", data: [a] }];
-  return c.runWebGPUProgram(t, [n], "packedF16", o ? s : void 0);
-}
-const k = {
-  kernelName: "Pack16",
-  backendName: "webgpu",
-  kernelFunc: m
-};
-i(k);

package/dist/ops/webgpu/pack16_program.d.ts

@@ -1,19 +0,0 @@
-import { WebGPUProgram } from '@tensorflow/tfjs-backend-webgpu';
-export default class PackProgram implements WebGPUProgram {
-    outputShape: number[];
-    shaderKey: string;
-    dispatchLayout: {
-        x: number[];
-    };
-    dispatch: [number, number, number];
-    workgroupSize: [number, number, number];
-    variableNames: string[];
-    uniforms?: string;
-    size: boolean;
-    outputComponent: number;
-    scaling: boolean;
-    padding: number;
-    constructor(outShape: number[], padding?: number);
-    useScaling(): void;
-    getUserCode(): string;
-}

package/dist/ops/webgpu/pack16_program.js

@@ -1,92 +0,0 @@
-import { f as o, c as a } from "../../webgpu_util-DYlGSwOJ.js";
-import { e as s } from "../../webgpu_program-B4HmApL1.js";
-class h {
-  outputShape;
-  shaderKey = "Pack16";
-  dispatchLayout;
-  dispatch;
-  workgroupSize = [64, 1, 1];
-  variableNames = ["x"];
-  uniforms;
-  size = !0;
-  outputComponent = 4;
-  scaling = !1;
-  padding = 0;
-  constructor(t, e = 0) {
-    if (t[t.length - 1] % 2 !== 0 && e === 0)
-      throw new Error("Last dimension of output shape must be even to use Pack16.");
-    if (e % 4 !== 0)
-      throw new Error("Padding must be a multiple of 4 to use Pack16.");
-    if (this.outputShape = [...t.slice(0, -1), t[t.length - 1]], e > 0) {
-      this.shaderKey += `_Padded${e}`, this.padding = e;
-      for (let i = this.outputShape.length - 2; i < this.outputShape.length; i++)
-        this.outputShape[i] % this.padding !== 0 && (this.outputShape[i] += this.padding - this.outputShape[i] % this.padding);
-      this.outputComponent = 1;
-    }
-    this.outputShape[this.outputShape.length - 1] /= 2, this.outputShape[this.outputShape.length - 1] % this.outputComponent !== 0 && (this.outputComponent = 1), this.dispatchLayout = o(this.outputShape), this.dispatch = a(this.dispatchLayout, this.outputShape, this.workgroupSize, [
-      this.outputComponent,
-      1,
-      1
-    ]);
-  }
-  useScaling() {
-    this.shaderKey += "_Scaled", this.uniforms = "scaling : f32,", this.scaling = !0;
-  }
-  getUserCode() {
-    if (this.padding > 0 && this.outputComponent === 1) {
-      const t = this.outputShape.length;
-      return `
-                ${s("index")} {
-                    if (index < uniforms.size) {
-                        var coords = getCoordsFromIndex(index);
-                        coords[${t} - 1] = coords[${t} - 1] * 2;
-                        let row = coords[${t} - 2];
-                        let col = coords[${t} - 1];
-                        let width = uniforms.xShape[${t} - 1];
-                        let height = uniforms.xShape[${t} - 2];
-                    
-                        var value1 = 0.0f;
-                        if (col < width && row < height) {
-                            let baseInputIndex = getIndexFromCoords${t}D(coords, uniforms.xShape);
-                            value1 = x[baseInputIndex] ${this.scaling ? "* uniforms.scaling" : ""};
-                        }
-                        var value2 = 0.0f;
-                        if (col + 1 < width && row < height) {
-                            coords[${t} - 1] = coords[${t} - 1] + 1;
-                            let baseInputIndex = getIndexFromCoords${t}D(coords, uniforms.xShape);
-                            value2 = x[baseInputIndex] ${this.scaling ? "* uniforms.scaling" : ""};
-                        }
-                        let packed = i32(pack2x16float(vec2<f32>(value1, value2)));
-                        result[index] = packed;
-                    }
-                }`;
-    }
-    return this.outputComponent === 1 ? `
-                ${s("index")} {
-                    if (index < uniforms.size) {
-                        let baseInputIndex =  index * 2;
-                        let x1 = x[baseInputIndex] ${this.scaling ? "* uniforms.scaling" : ""};
-                        let x2 = x[baseInputIndex + 1] ${this.scaling ? "* uniforms.scaling" : ""};
-                        let packed = i32(pack2x16float(vec2<f32>(x1, x2)));
-                        result[index] = packed;
-                    }
-                }` : `
-                ${s("index")} {
-                    if (index < uniforms.size) {
-                        let baseInputIndex =  index * 2;
-                        let x1 = x[baseInputIndex] ${this.scaling ? "* uniforms.scaling" : ""};
-                        let x2 = x[baseInputIndex + 1] ${this.scaling ? "* uniforms.scaling" : ""};
-                        let packed = vec4<i32>(
-                            i32(pack2x16float(vec2<f32>(x1.x, x1.y))),
-                            i32(pack2x16float(vec2<f32>(x1.z, x1.w))),
-                            i32(pack2x16float(vec2<f32>(x2.x, x2.y))),
-                            i32(pack2x16float(vec2<f32>(x2.z, x2.w)))
-                        );
-                        result[index] = packed;
-                    }
-                }`;
-  }
-}
-export {
-  h as default
-};

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

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

package/dist/ops/webgpu/qkv.js

@@ -1,24 +0,0 @@
-import { c as h, a6 as l } from "../../index-CUXkjxiT.js";
-import { b as f } from "../../matMul16-BcVC_E62.js";
-import { slice16 as a } from "../slice16.js";
-import { isPackedTensor as u } from "../../utilities/packed.js";
-function k(i) {
-  const { x: n, kernel: c } = i.inputs, { heads: e } = i.attrs, r = n.shape[0], t = n.shape[1], s = n.shape[2], p = u(n);
-  if (l(c.shape, [p ? s * 2 : s, 3 * s], "Error in QKV: "), s % e !== 0)
-    throw new Error(`Channel dimension ${s} must be divisible by number of heads ${e} in QKV.`);
-  const o = f(n, c, !1, !1, {
-    forceOutputShape: [r, t, 3 * e, s / e],
-    perm: [0, 2, 1, 3]
-  }), m = [
-    a(o, [0, 0, 0, 0], [r, e, t, s / e]),
-    a(o, [0, e, 0, 0], [r, e, t, s / e]),
-    a(o, [0, 2 * e, 0, 0], [r, e, t, s / e])
-  ];
-  return o.dispose(), m;
-}
-const b = {
-  kernelName: "QKV",
-  backendName: "webgpu",
-  kernelFunc: k
-};
-h(b);

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

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

package/dist/ops/webgpu/rope.js

@@ -1,135 +0,0 @@
-import { isPackedTensor as y } from "../../utilities/packed.js";
-import { e as c } from "../../webgpu_program-B4HmApL1.js";
-import { f as x, c as l } from "../../webgpu_util-DYlGSwOJ.js";
-import { c as w, a6 as b } from "../../index-CUXkjxiT.js";
-class v {
-  variableNames = ["x", "sin", "cos"];
-  outputShape;
-  shaderKey = "Rope";
-  dispatchLayout;
-  dispatch;
-  workgroupSize = [64, 1, 1];
-  size = !0;
-  uniforms = "pastLen: i32";
-  constructor(e, o, a, t) {
-    this.shaderKey = `Rope_${t}`, this.outputShape = [e, o, a, t], this.dispatchLayout = x(this.outputShape), this.dispatch = l(this.dispatchLayout, this.outputShape, this.workgroupSize);
-  }
-  getUserCode() {
-    const e = this.outputShape[3];
-    return `
-        ${c("index")} {
-            if (index < uniforms.size) {
-                let coords = getCoordsFromIndex(index); // [b, h, t, d]
-                let b = coords[0];
-                let h = coords[1];
-                let t = coords[2];
-                let d = coords[3];
-
-                let rotaryDim = ${e};
-
-                var outVal = 0.0;
-
-                let xIdx = b * uniforms.outShapeStrides[0] +
-                    h * uniforms.outShapeStrides[1] +
-                    t * uniforms.outShapeStrides[2] +
-                    d;
-
-                if (d < rotaryDim) {
-                    let idx = (t + uniforms.pastLen) * uniforms.cosShape[1] + d / 2;
-                    let cos = cos[idx];
-                    let sin = sin[idx];
-
-                    let ownX = x[xIdx] * cos;
-                    var evenOdd = 0.0;
-
-                    if (d % 2 == 0) {
-                        // even index
-                        evenOdd = -x[xIdx + 1];
-                    } else {
-                        // odd index
-                        evenOdd = x[xIdx - 1];
-                    }
-
-                    outVal = fma(evenOdd, sin, ownX);
-                } else {
-                    // pass through for non-rotary dims
-                    outVal = x[xIdx];
-                }
-
-                setOutputAtIndex(index, outVal);
-            }
-        }
-        `;
-  }
-}
-class k {
-  variableNames = ["x", "sin", "cos"];
-  outputShape;
-  shaderKey = "Rope";
-  dispatchLayout;
-  dispatch;
-  workgroupSize = [64, 1, 1];
-  size = !0;
-  uniforms = "pastLen: i32";
-  constructor(e, o, a, t) {
-    this.shaderKey = `Rope_${t}`, this.outputShape = [e, o, a, t / 2], this.dispatchLayout = x(this.outputShape), this.dispatch = l(this.dispatchLayout, this.outputShape, this.workgroupSize);
-  }
-  getUserCode() {
-    return `
-        ${c("index")} {
-            if (index < uniforms.size) {
-                let coords = getCoordsFromIndex(index); // [b, h, t, d]
-                let b = coords[0];
-                let h = coords[1];
-                let t = coords[2];
-                let d = coords[3];
-
-                var outVal = vec2<f32>(0.0, 0.0);
-
-                let xIdx = b * uniforms.outShapeStrides[0] +
-                    h * uniforms.outShapeStrides[1] +
-                    t * uniforms.outShapeStrides[2] +
-                    d;
-
-                let idx = (t + uniforms.pastLen) * uniforms.cosShape[1] + d;
-                let cos = cos[idx];
-                let sin = sin[idx];
-
-                let xPair = unpack2x16float(u32(x[xIdx]));
-                let ownX = vec2<f32>(xPair.x * cos, xPair.y * cos);
-
-                let evenOdd = vec2<f32>(
-                    -xPair.y,
-                    xPair.x
-                );
-
-                outVal = vec2<f32>(
-                    fma(evenOdd.x, sin, ownX.x),
-                    fma(evenOdd.y, sin, ownX.y)
-                );
-
-                result[index] = i32(pack2x16float(outVal));
-            }
-        }
-        `;
-  }
-}
-function L(i) {
-  const { x: e } = i.inputs, { pastLen: o, negSin: a, ropeCache: t } = i.attrs, m = i.backend, d = y(e), p = e.shape[0], h = e.shape[1], r = e.shape[2], n = d ? e.shape[3] * 2 : e.shape[3], s = a ? t.getNegSin() : t.getSin(), u = t.getCos();
-  if (b(s.shape, u.shape, "Error in Rope: "), s.shape[0] < r + o)
-    throw new Error(
-      `Sin tensor shape ${s.shape} is not compatible with seqLength ${r} and pastLen ${o}.`
-    );
-  if (s.shape[1] * 2 < n)
-    throw new Error(`Sin tensor shape ${s.shape} is not compatible with feature dimension ${n}.`);
-  if (s.shape.length !== 3)
-    throw new Error(`Sin tensor must be 3-dimensional, but got shape ${s.shape}.`);
-  const f = d ? new k(p, h, r, n) : new v(p, h, r, n), S = [{ type: "int32", data: [o] }], g = d ? "packedF16" : e.dtype;
-  return m.runWebGPUProgram(f, [e, s, u], g, S);
-}
-const P = {
-  kernelName: "Rope",
-  backendName: "webgpu",
-  kernelFunc: L
-};
-w(P);

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

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

package/dist/ops/webgpu/scatterSub.js

@@ -1,40 +0,0 @@
-import { e as p } from "../../webgpu_program-B4HmApL1.js";
-import { f as u, c as d } from "../../webgpu_util-DYlGSwOJ.js";
-import { c as h, a6 as o } from "../../index-CUXkjxiT.js";
-class b {
-  variableNames = ["labels", "softmaxProbs", "dy"];
-  outputShape;
-  shaderKey = "ScatterSub";
-  dispatchLayout;
-  dispatch;
-  workgroupSize = [64, 1, 1];
-  size = !0;
-  constructor(t, e) {
-    this.outputShape = [t, e], this.dispatchLayout = u(this.outputShape), this.dispatch = d(this.dispatchLayout, this.outputShape, this.workgroupSize);
-  }
-  getUserCode() {
-    return `
-            ${p("index")} {
-                if (index < uniforms.size) {
-                    let coords = getCoordsFromIndex(index); // [batch, depth]
-                    let idx = i32(labels[coords[0]]);
-                    let prob = softmaxProbs[index];
-                    let dy = dy[coords[0]];
-                    setOutputAtIndex(index, select(prob, prob - 1.0, idx == coords[1]) * dy);
-                }
-            }
-            `;
-  }
-}
-function f(a) {
-  const { logits: t, labels: e, dy: s } = a.inputs, c = a.backend, r = e.shape[0], i = t.shape[1];
-  o(s.shape, [r], "Error in EfficientScatterSub dy: "), o(t.shape, [r, i], "Error in EfficientScatterSub logits: "), o(e.shape, [r], "Error in EfficientScatterSub labels: ");
-  const n = new b(r, i);
-  return c.runWebGPUProgram(n, [e, t, s], "float32");
-}
-const l = {
-  kernelName: "EfficientScatterSub",
-  backendName: "webgpu",
-  kernelFunc: f
-};
-h(l);

package/dist/ops/webgpu/slice16.d.ts

@@ -1,7 +0,0 @@
-import { WebGPUBackend } from '@tensorflow/tfjs-backend-webgpu';
-import { SliceAttrs, SliceInputs, TensorInfo } from '@tensorflow/tfjs-core';
-export declare function slice(args: {
-    inputs: SliceInputs;
-    backend: WebGPUBackend;
-    attrs: SliceAttrs;
-}): TensorInfo;