@genai-fi/nanogpt 0.19.1 → 0.20.0

package/dist/ops/webgl/matMul16.js

@@ -1,45 +0,0 @@
-import { isPackedTensor as h } from "../../utilities/packed.js";
-import { c as b, m as A, a as c, n as G } from "../../index-CUXkjxiT.js";
-import { matMulMul as P } from "../matMulMul.js";
-import { matMulGelu as R } from "../matMulGelu.js";
-import { m as f } from "../../mat_mul-DhG0Newp.js";
-import { r as B } from "../../relu-J_X6MUzx.js";
-import { r as v } from "../../reshape-BDOuCSNW.js";
-import { t as g } from "../../transpose-9kRxIXWR.js";
-function q(d) {
-  const { A: i, B: m } = d.inputs, { transposeA: o, transposeB: a, scale: M, activation: u, scaleA: k, scaleB: w, forceOutputShape: n, perm: p } = d.attrs, N = !h(i), U = !h(m);
-  if (N && U) {
-    const t = k !== void 0 ? A(i, c(k)) : i, r = w !== void 0 ? A(m, c(w)) : m;
-    let e;
-    if (M !== void 0)
-      e = P(t, r, c(M), o, a);
-    else if (u === "gelu")
-      e = R(t, r);
-    else if (u === "relu2") {
-      const s = f(t, r, o, a), l = B(s);
-      s.dispose(), e = G(l), l.dispose();
-    } else u === "relu" ? e = B(f(t, r, o, a)) : e = f(t, r, o, a);
-    if (p)
-      if (n) {
-        const s = v(e, n);
-        e.dispose();
-        const l = g(s, p);
-        return s.dispose(), l;
-      } else {
-        const s = g(e, p);
-        return e.dispose(), s;
-      }
-    else if (n) {
-      const s = v(e, n);
-      return e.dispose(), s;
-    } else
-      return e;
-  }
-  throw new Error("Not implemented: matMul16 with packed inputs");
-}
-const C = {
-  kernelName: "MatMul16",
-  backendName: "webgl",
-  kernelFunc: q
-};
-b(C);

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

@@ -1,21 +0,0 @@
-import { TensorInfo } from '@tensorflow/tfjs-core';
-import { MathBackendWebGL } from '@tensorflow/tfjs-backend-webgl';
-export declare const MATMUL_SHARED_DIM_THRESHOLD = 1000;
-interface BatchMatMulConfig {
-    a: TensorInfo;
-    b: TensorInfo;
-    transposeA: boolean;
-    transposeB: boolean;
-    backend: MathBackendWebGL;
-    activationSnippet?: string;
-    multiplier?: TensorInfo;
-}
-export declare function batchMatMulGeluImpl({ a, b, transposeA, transposeB, backend, activationSnippet, multiplier, }: BatchMatMulConfig): TensorInfo;
-export declare function batchMatMulKernel(args: {
-    inputs: {
-        x: TensorInfo;
-        kernel: TensorInfo;
-    };
-    backend: MathBackendWebGL;
-}): TensorInfo;
-export {};

package/dist/ops/webgl/matMulGelu.js

@@ -1,9 +0,0 @@
-import "../../index-CUXkjxiT.js";
-import "../../Reshape-BW__R4mZ.js";
-import { a as m, b as o, c as p } from "../../matMulGelu-JNLZqKQp.js";
-import "../../mat_mul-DhG0Newp.js";
-export {
-  m as MATMUL_SHARED_DIM_THRESHOLD,
-  o as batchMatMulGeluImpl,
-  p as batchMatMulKernel
-};

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

@@ -1,14 +0,0 @@
-import { TensorInfo } from '@tensorflow/tfjs-core';
-import { MathBackendWebGL } from '@tensorflow/tfjs-backend-webgl';
-export declare function batchMatMulKernel(args: {
-    inputs: {
-        x: TensorInfo;
-        kernel: TensorInfo;
-        y: TensorInfo;
-    };
-    attrs: {
-        transposeA: boolean;
-        transposeB: boolean;
-    };
-    backend: MathBackendWebGL;
-}): TensorInfo;

package/dist/ops/webgl/matMulMul.js

@@ -1,28 +0,0 @@
-import { c as u } from "../../index-CUXkjxiT.js";
-import { b as c } from "../../matMulGelu-JNLZqKQp.js";
-const M = `
-    return a * b;
-`;
-function p(r) {
-  const { inputs: n, backend: a, attrs: o } = r, { x: t, kernel: e, y: l } = n, { transposeA: s, transposeB: i } = o;
-  if (t === void 0 || e === void 0)
-    throw new Error("BatchMatMul requires two input tensors.");
-  return c({
-    a: t,
-    b: e,
-    transposeA: s,
-    transposeB: i,
-    backend: a,
-    activationSnippet: M,
-    multiplier: l
-  });
-}
-const m = {
-  kernelName: "MatMulMul",
-  backendName: "webgl",
-  kernelFunc: p
-};
-u(m);
-export {
-  p as batchMatMulKernel
-};

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

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

package/dist/ops/webgl/mulDropout.js

@@ -1,41 +0,0 @@
-import { c as m } from "../../index-CUXkjxiT.js";
-class f {
-  variableNames = ["a", "b"];
-  outputShape;
-  userCode;
-  customUniforms = [
-    { name: "dropoutRate", type: "float" },
-    { name: "seed", type: "float" }
-  ];
-  constructor(t, r, o) {
-    this.outputShape = [t, r, o, o], 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 a = getA(coords.x, coords.y, coords.z, coords.w);
-            float b = getB(coords.x, coords.y, coords.z, coords.w);
-            
-            float keepProb = 1.0 - dropoutRate;
-            float rand = random(coords);
-            float mask = step(rand, keepProb);
-            setOutput(a * b * mask / keepProb);
-        }
-        `;
-  }
-}
-function b(e) {
-  const { inputs: t, attrs: r } = e, { a: o, b: s } = t, { dropoutRate: a, seed: c } = r, n = e.backend, d = o.shape[0], u = o.shape[2], p = o.shape[1], l = new f(d, p, u);
-  return n.runWebGLProgram(l, [o, s], "float32", [
-    [a ?? 0],
-    [c ?? Math.random() * 1e4]
-  ]);
-}
-const i = {
-  kernelName: "MulDropout",
-  backendName: "webgl",
-  kernelFunc: b
-};
-m(i);

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

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

package/dist/ops/webgl/normRMS.js

@@ -1,93 +0,0 @@
-import { c, h as G } from "../../index-CUXkjxiT.js";
-import { s as x } from "../../sum-BpiwSWvg.js";
-class y {
-  variableNames = ["x", "meanSquare"];
-  outputShape;
-  userCode;
-  constructor(a, e, o, t = !0) {
-    t && this.variableNames.push("gamma"), this.outputShape = [a, e, o], this.userCode = `
-        void main() {
-            ivec3 coords = getOutputCoords();
-            float x = getXAtOutCoords();
-            float meanSquare = getMeanSquare(coords.x, coords.y, 0);
-            ${t ? "float gamma = getGammaAtOutCoords();" : ""}
-            float invRms = inversesqrt(meanSquare + 1e-8);
-            float normalized = x * invRms;
-            float outVal = normalized ${t ? " * gamma" : ""};
-            setOutput(outVal);
-        }
-        `;
-  }
-}
-function S(r) {
-  const { x: a, gamma: e } = r.inputs, o = r.backend, t = a.shape[0], m = a.shape[1], u = a.shape[2], s = a.square().mean(-1, !0), n = new y(t, m, u, e !== void 0);
-  return o.runWebGLProgram(n, e ? [a, s, e] : [a, s], "float32");
-}
-const N = {
-  kernelName: "RMSNorm",
-  backendName: "webgl",
-  kernelFunc: S
-}, b = {
-  kernelName: "RMSNormNoGamma",
-  backendName: "webgl",
-  kernelFunc: S
-};
-c(N);
-c(b);
-class C {
-  variableNames = ["x", "meanSquare", "dyGamma", "dyXMean"];
-  outputShape;
-  userCode;
-  constructor(a, e, o) {
-    this.outputShape = [a, e, o], this.userCode = `
-        void main() {
-            ivec3 coords = getOutputCoords();
-            float x = getXAtOutCoords();
-            float meanSquare = getMeanSquare(coords.x, coords.y, 0) + 1e-8;
-            float dyGamma = getDyGammaAtOutCoords();
-            float dyXMean = getDyXMean(coords.x, coords.y, 0) / ${o}.0;
-            float invRms = inversesqrt(meanSquare);
-            float dx = dyGamma * invRms - x * dyXMean * invRms / meanSquare;
-            setOutput(dx);
-        }
-        `;
-  }
-}
-class k {
-  variableNames = ["x", "meanSquare", "dy"];
-  outputShape;
-  userCode;
-  constructor(a, e, o) {
-    this.outputShape = [a, e, o], this.userCode = `
-        void main() {
-            ivec3 coords = getOutputCoords();
-            float x = getXAtOutCoords();
-            float meanSquare = getMeanSquare(coords.x, coords.y, 0) + 1e-8;
-            float dy = getDyAtOutCoords();
-            float invRms = inversesqrt(meanSquare);
-            float dGamma = dy * (x * invRms);
-            setOutput(dGamma);
-        }
-        `;
-  }
-}
-function M(r) {
-  const { dy: a, x: e, gamma: o } = r.inputs, t = r.backend, m = e.shape[0], u = e.shape[1], s = e.shape[2], n = o ? a.mul(o) : a, i = n.mul(e), l = i.sum(-1, !0);
-  i.dispose();
-  const g = e.square(), d = g.mean(-1, !0);
-  g.dispose();
-  const h = new C(m, u, s), p = t.runWebGLProgram(h, [e, d, n, l], "float32");
-  if (o && n.dispose(), l.dispose(), o) {
-    const q = new k(m, u, s), f = t.runWebGLProgram(q, [e, d, a], "float32");
-    d.dispose();
-    const v = x(G().makeTensorFromTensorInfo(f), [0, 1]);
-    return t.disposeIntermediateTensorInfo(f), [p, v];
-  } else
-    return d.dispose(), [p];
-}
-const R = {
-  kernelName: "RMSNormGrad",
-  backendName: "webgl",
-  kernelFunc: M
-};
-c(R);

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

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

package/dist/ops/webgl/qkv.js

@@ -1,46 +0,0 @@
-import { c as i } from "../../index-CUXkjxiT.js";
-class l {
-  variableNames = ["x", "kernel"];
-  outputShape;
-  userCode;
-  // enableShapeUniforms = true;
-  customUniforms = [{ name: "mode", type: "int" }];
-  constructor(e, t, s, o) {
-    const n = o / t;
-    this.outputShape = [e, t, s, n], 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;
-
-            // Compute output channel index in fused kernel
-            int out_offset = mode * ${t} * ${n} + h * ${n} + d;
-
-            float sum = 0.0;
-            for (int c = 0; c < ${o}; ++c) {
-                float xval = getX(b, t, c); // fetch from x
-                float kval = getKernel(c, out_offset); // fetch from kernel
-                sum += xval * kval;
-            }
-
-            setOutput(sum);
-        }
-        `;
-  }
-}
-function m(r) {
-  const { x: e, kernel: t } = r.inputs, { heads: s } = r.attrs, o = r.backend, n = e.shape[0], c = e.shape[1], u = e.shape[2], a = new l(n, s, c, u);
-  return [
-    o.runWebGLProgram(a, [e, t], "float32", [[0]]),
-    o.runWebGLProgram(a, [e, t], "float32", [[1]]),
-    o.runWebGLProgram(a, [e, t], "float32", [[2]])
-  ];
-}
-const d = {
-  kernelName: "QKV",
-  backendName: "webgl",
-  kernelFunc: m
-};
-i(d);

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

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

package/dist/ops/webgl/rope.js

@@ -1,56 +0,0 @@
-import { c as h } from "../../index-CUXkjxiT.js";
-class g {
-  variableNames = ["x", "sin", "cos"];
-  outputShape;
-  userCode;
-  // enableShapeUniforms = true;
-  customUniforms = [{ name: "pastLen", type: "int" }];
-  constructor(t, s, e, n) {
-    this.outputShape = [t, s, e, n], 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 rotaryDim = ${n};
-
-            float outVal = 0.0;
-
-            if (d < rotaryDim) {
-                int pairIdx = d / 2;
-                float cos = getCos(t + pastLen, pairIdx, 0);
-                float sin = getSin(t + pastLen, pairIdx, 0);
-
-                if (d % 2 == 0) {
-                    // even index
-                    float even = getX(b, h, t, d);
-                    float odd = getX(b, h, t, d + 1);
-                    outVal = even * cos - odd * sin;
-                } else {
-                    // odd index
-                    float even = getX(b, h, t, d - 1);
-                    float odd = getX(b, h, t, d);
-                    outVal = even * sin + odd * cos;
-                }
-            } else {
-                // pass through for non-rotary dims
-                outVal = getX(b, h, t, d);
-            }
-
-            setOutput(outVal);
-        }
-        `;
-  }
-}
-function f(o) {
-  const { x: t } = o.inputs, { pastLen: s, ropeCache: e, negSin: n } = o.attrs, a = n ? e.getNegSin() : e.getSin(), r = e.getCos(), d = o.backend, i = t.shape[0], c = t.shape[1], p = t.shape[2], u = t.shape[3], l = new g(i, c, p, u);
-  return d.runWebGLProgram(l, [t, a, r], "float32", [[s]]);
-}
-const m = {
-  kernelName: "Rope",
-  backendName: "webgl",
-  kernelFunc: f
-};
-h(m);

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

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

package/dist/ops/webgl/scatterSub.js

@@ -1,27 +0,0 @@
-import { c as i } from "../../index-CUXkjxiT.js";
-class u {
-  variableNames = ["labels", "softmaxProbs", "dy"];
-  outputShape;
-  userCode;
-  constructor(e, t) {
-    this.outputShape = [e, t], this.userCode = `
-            void main() {
-                ivec2 coords = getOutputCoords();
-                int index = int(getLabels(coords.x));
-                float prob = getSoftmaxProbsAtOutCoords();
-                float dy = getDy(coords.x);
-                setOutput((index == coords.y ? prob - 1.0 : prob) * dy);
-            }
-            `;
-  }
-}
-function d(o) {
-  const { logits: e, labels: t, dy: r } = o.inputs, s = o.backend, n = t.shape[0], a = e.shape[1], c = new u(n, a);
-  return s.runWebGLProgram(c, [t, e, r], "float32");
-}
-const b = {
-  kernelName: "EfficientScatterSub",
-  backendName: "webgl",
-  kernelFunc: d
-};
-i(b);

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

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

package/dist/ops/webgpu/adamAdjust.js

@@ -1,57 +0,0 @@
-import { e as d } from "../../webgpu_program-B4HmApL1.js";
-import { f as l, c as f } from "../../webgpu_util-DYlGSwOJ.js";
-import { c as h, a6 as c } from "../../index-CUXkjxiT.js";
-class v {
-  variableNames = ["moments", "value"];
-  outputShape;
-  shaderKey = "AdamAdjust";
-  dispatchLayout;
-  dispatch;
-  workgroupSize = [64, 1, 1];
-  size = !0;
-  uniforms = "invbeta1: f32, invbeta2: f32, learningRate: f32, epsilon: f32";
-  outputComponent = 1;
-  variableComponents = [2, 1];
-  useWeightDecay;
-  constructor(a, t) {
-    this.outputShape = a, this.dispatchLayout = l(this.outputShape), this.dispatch = f(this.dispatchLayout, this.outputShape, this.workgroupSize), this.useWeightDecay = t, t && (this.uniforms += ", weightDecay: f32");
-  }
-  getUserCode() {
-    return `
-        ${d("index")} {
-            if (index < uniforms.size) {
-                let moments: vec2<f32> = moments[index];
-                let value: f32 = value[index];
-
-                let m1Hat = moments.x * uniforms.invbeta1;
-                let m2Hat = moments.y * uniforms.invbeta2;
-
-                let invSqrt = inverseSqrt(max(m2Hat, 1e-30));
-                let invDenom = invSqrt / fma(uniforms.epsilon, invSqrt, 1.0);
-                var adjustedValue = fma(-uniforms.learningRate * m1Hat, invDenom, value);
-
-                ${this.useWeightDecay ? "adjustedValue = adjustedValue - uniforms.learningRate * uniforms.weightDecay * value;" : ""}
-
-                setOutputAtIndex(index, adjustedValue);
-            }
-        }
-    `;
-  }
-}
-function g(e) {
-  const { moments: a, value: t } = e.inputs, { beta1: i, beta2: o, learningRate: u, epsilon: r, weightDecay: s } = e.attrs, m = e.backend;
-  c(a.shape, [...t.shape, 2], "Error in AdamAdjust: ");
-  const p = new v(t.shape, s > 0), n = [
-    { type: "float32", data: [1 / i] },
-    { type: "float32", data: [1 / o] },
-    { type: "float32", data: [u] },
-    { type: "float32", data: [r] }
-  ];
-  return s > 0 && n.push({ type: "float32", data: [s] }), m.runWebGPUProgram(p, [a, t], "float32", n);
-}
-const y = {
-  kernelName: "AdamAdjust",
-  backendName: "webgpu",
-  kernelFunc: g
-};
-h(y);

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

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

package/dist/ops/webgpu/adamMoments.js

@@ -1,60 +0,0 @@
-import { e as u } from "../../webgpu_program-B4HmApL1.js";
-import { f as p, c } from "../../webgpu_util-DYlGSwOJ.js";
-import { c as d, a6 as f } from "../../index-CUXkjxiT.js";
-class h {
-  variableNames = ["moments", "gradient", "scaling"];
-  outputShape;
-  shaderKey = "AdamMoments";
-  dispatchLayout;
-  dispatch;
-  workgroupSize = [64, 1, 1];
-  size = !0;
-  uniforms = "beta1: f32, beta2: f32";
-  outputComponent = 2;
-  variableComponents = [2, 1, 1];
-  constructor(t) {
-    this.outputShape = t, this.dispatchLayout = p(this.outputShape.slice(0, -1)), this.dispatch = c(
-      this.dispatchLayout,
-      this.outputShape.slice(0, -1),
-      this.workgroupSize,
-      [1, 1, 1]
-    );
-  }
-  getUserCode() {
-    return `
-        ${u("index")} {
-            if (index < uniforms.size) {
-                let m: vec2<f32> = moments[index];
-
-                // Loss and clip scaling.
-                let g: f32 = gradient[index] * scaling[0];
-
-                let newM1 = fma(m.x, uniforms.beta1, g * (1.0 - uniforms.beta1));
-                let newM2 = fma(m.y, uniforms.beta2, g * g * (1.0 - uniforms.beta2));
-
-                setOutputAtIndex(index, vec2<f32>(newM1, newM2));
-            }
-        }
-    `;
-  }
-}
-function l(e) {
-  const { moments: t, gradient: a, scaling: s } = e.inputs, { beta1: n, beta2: o } = e.attrs, r = e.backend;
-  if (a.dtype !== "float32")
-    throw new Error(`Gradient must be float32, but got ${a.dtype}`);
-  if (f(t.shape, [...a.shape, 2], "Error in AdamMoments: "), n < 0 || n >= 1)
-    throw new Error(`Invalid beta1 value: ${n}. Must be in the range [0, 1).`);
-  if (o < 0 || o >= 1)
-    throw new Error(`Invalid beta2 value: ${o}. Must be in the range [0, 1).`);
-  const i = new h(t.shape), m = [
-    { type: "float32", data: [n] },
-    { type: "float32", data: [o] }
-  ];
-  return r.runWebGPUProgram(i, [t, a, s], "float32", m);
-}
-const g = {
-  kernelName: "AdamMoments",
-  backendName: "webgpu",
-  kernelFunc: l
-};
-d(g);

package/dist/ops/webgpu/add16.d.ts

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

package/dist/ops/webgpu/add16.js

@@ -1,13 +0,0 @@
-import { c as t } from "../../index-CUXkjxiT.js";
-import { BinaryOpProgram as c } from "./utils/binary_op.js";
-import { B as p } from "../../binary_op_util-pKXltfxI.js";
-function s(e) {
-  const { a: r, b: n } = e.inputs, a = e.backend, o = new c(p.ADD, r.shape, n.shape);
-  return a.runWebGPUProgram(o, [r, n], "packedF16");
-}
-const m = {
-  kernelName: "Add16",
-  backendName: "webgpu",
-  kernelFunc: s
-};
-t(m);

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

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

package/dist/ops/webgpu/appendCache.js

@@ -1,105 +0,0 @@
-import { isPackedTensor as T } from "../../utilities/packed.js";
-import { e as p } from "../../webgpu_program-B4HmApL1.js";
-import { f as d, c as u } from "../../webgpu_util-DYlGSwOJ.js";
-import { c as S, a6 as g } from "../../index-CUXkjxiT.js";
-class x {
-  variableNames = ["cache", "item"];
-  outputShape;
-  shaderKey = "AppendCache";
-  dispatchLayout;
-  dispatch;
-  workgroupSize = [64, 1, 1];
-  size = !0;
-  uniforms = "cacheT: i32";
-  constructor(e, t, a, s, i) {
-    const o = Math.min(a + 1, i);
-    this.shaderKey = `AppendCache_${o}`, this.outputShape = [e, t, o, s], this.dispatchLayout = d(this.outputShape), this.dispatch = u(this.dispatchLayout, this.outputShape, this.workgroupSize);
-  }
-  getUserCode() {
-    const e = this.outputShape[2];
-    return `
-        ${p("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 itemT = 1;
-                let maxSize = ${e};
-                let totalT = uniforms.cacheT + itemT;
-                let start = select(0, 1, totalT >= maxSize);
-
-                let srcT = t + start;
-                var val = 0.0;
-                if (srcT < uniforms.cacheT) {
-                    val = getCache(b, h, srcT, d);
-                }
-                if (srcT == uniforms.cacheT) {
-                    val = getItem(b, h, 0, d);
-                }
-
-                setOutputAtIndex(index, val);
-            }
-        }
-        `;
-  }
-}
-class b {
-  variableNames = ["cache", "item"];
-  outputShape;
-  shaderKey = "AppendCache";
-  dispatchLayout;
-  dispatch;
-  workgroupSize = [64, 1, 1];
-  size = !0;
-  uniforms = "cacheT: i32";
-  constructor(e, t, a, s, i) {
-    const o = Math.min(a + 1, i);
-    this.shaderKey = `AppendCache_${o}`, this.outputShape = [e, t, o, s], this.dispatchLayout = d(this.outputShape), this.dispatch = u(this.dispatchLayout, this.outputShape, this.workgroupSize);
-  }
-  getUserCode() {
-    const e = this.outputShape[2];
-    return `
-        ${p("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 itemT = 1;
-                let maxSize = ${e};
-                let totalT = uniforms.cacheT + itemT;
-                let start = select(0, 1, totalT >= maxSize);
-
-                let srcT = t + start;
-                var val: i32 = 0i;
-                if (srcT < uniforms.cacheT) {
-                    val = cache[getIndexFromCoords4D(vec4<i32>(b, h, srcT, d), uniforms.cacheShape)];
-                }
-                if (srcT == uniforms.cacheT) {
-                    val = item[getIndexFromCoords4D(vec4<i32>(b, h, 0, d), uniforms.itemShape)];
-                }
-
-                result[index] = val;
-            }
-        }
-        `;
-  }
-}
-function C(r) {
-  const { cache: e, item: t } = r.inputs, { maxSize: a, pastLen: s } = r.attrs, i = r.backend, o = T(e), c = e.shape[0], n = e.shape[2], h = e.shape[1];
-  if (g(t.shape, [c, h, 1, t.shape[3]], "Error in AppendCache: "), s < 0 || s > a)
-    throw new Error(`Invalid pastLen value: ${s}. Must be in the range [0, ${a}].`);
-  const m = o ? new b(c, h, n, t.shape[3], a) : new x(c, h, n, t.shape[3], a), l = [{ type: "int32", data: [s] }], f = o ? "packedF16" : e.dtype;
-  return i.runWebGPUProgram(m, [e, t], f, l);
-}
-const v = {
-  kernelName: "AppendCache",
-  backendName: "webgpu",
-  kernelFunc: C
-};
-S(v);

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

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