@genai-fi/nanogpt 0.19.0 → 0.20.0

package/dist/ops/webgpu/attentionMask.js

@@ -1,26 +0,0 @@
-import { c as d, a6 as b } from "../../index-CUXkjxiT.js";
-import { isPackedTensor as p } from "../../utilities/packed.js";
-import { b as l } from "../../matMul16-BcVC_E62.js";
-import M from "./attentionMask32_program.js";
-function w(n) {
-  const { q: t, k: e } = n.inputs, { divisor: a, pastLen: o } = n.attrs, m = n.backend;
-  if (p(t) && p(e))
-    return l(t, e, !1, !0, { causalMask: !0, pastLen: o, scale: a });
-  const s = t.shape[0], k = t.shape[2], r = e.shape[2], i = t.shape[1], c = t.shape[3];
-  if (b(e.shape, [s, i, r, c], "Error in AttentionMask: "), a === 0)
-    throw new Error("Divisor must be non-zero in AttentionMask");
-  if (o < 0)
-    throw new Error("pastLen must be non-negative in AttentionMask");
-  const u = new M(s, i, k, r, c), f = [
-    { type: "float32", data: [a] },
-    { type: "int32", data: [o] },
-    { type: "float32", data: [Number.NEGATIVE_INFINITY] }
-  ], h = t.dtype;
-  return m.runWebGPUProgram(u, [t, e], h, f);
-}
-const A = {
-  kernelName: "AttentionMask",
-  backendName: "webgpu",
-  kernelFunc: w
-};
-d(A);

package/dist/ops/webgpu/attentionMask32_program.d.ts

@@ -1,19 +0,0 @@
-import { WebGPUProgram } from '@tensorflow/tfjs-backend-webgpu';
-export default class AttentionMaskProgram32 implements WebGPUProgram {
-    variableNames: string[];
-    outputShape: number[];
-    shaderKey: string;
-    dispatchLayout: {
-        x: number[];
-    };
-    dispatch: [number, number, number];
-    uniforms: string;
-    workgroupSize: [number, number, number];
-    size: boolean;
-    hs: number;
-    nh: number;
-    T1: number;
-    T2: number;
-    constructor(batch: number, nh: number, T1: number, T2: number, hs: number);
-    getUserCode(): string;
-}

package/dist/ops/webgpu/attentionMask32_program.js

@@ -1,54 +0,0 @@
-import { e as r } from "../../webgpu_program-B4HmApL1.js";
-import { f as a, c as u } from "../../webgpu_util-DYlGSwOJ.js";
-class p {
-  variableNames = ["q", "k"];
-  outputShape;
-  shaderKey = "AttentionMask";
-  dispatchLayout;
-  dispatch;
-  uniforms = "divisor: f32, pastLen: i32, inf: f32";
-  workgroupSize = [64, 1, 1];
-  size = !0;
-  hs;
-  nh;
-  T1;
-  T2;
-  constructor(e, i, s, o, t) {
-    if (this.shaderKey = `AttentionMask_${t}`, this.outputShape = [e, i, s, o], this.hs = t, this.nh = i, this.T1 = s, this.T2 = o, this.dispatchLayout = a(this.outputShape), this.dispatch = u(this.dispatchLayout, this.outputShape, this.workgroupSize), t % 4 !== 0)
-      throw new Error("Head size must be a multiple of 4 for AttentionMaskProgram");
-  }
-  getUserCode() {
-    return `
-            ${r("index")} {
-                
-                let coords = getCoordsFromIndex(index);
-                let b = coords[0];
-                let h = coords[1];
-                let t1 = coords[2];
-                let t2 = coords[3];
-
-                if (index < uniforms.size) {
-                    if (t2 > t1 + uniforms.pastLen) {
-                        setOutputAtIndex(index, uniforms.inf);
-                        return;
-                    }
-
-                    let q0 = getIndexFromCoords4D(vec4<i32>(b, h, t1, 0), uniforms.qShape);
-                    let k0 = getIndexFromCoords4D(vec4<i32>(b, h, t2, 0), uniforms.kShape);
-                    
-                    var sum: f32 = 0.0;
-                    for (var i: i32 = 0; i < ${this.hs}; i = i + 4) {
-                        let qv = vec4<f32>(q[q0 + i], q[q0 + i + 1], q[q0 + i + 2], q[q0 + i + 3]);
-                        let kv = vec4<f32>(k[k0 + i], k[k0 + i + 1], k[k0 + i + 2], k[k0 + i + 3]);
-                        sum = sum + dot(qv, kv);
-                    }
-                    let scaled = sum * uniforms.divisor;
-                    setOutputAtIndex(index, scaled);
-                }
-            }
-        `;
-  }
-}
-export {
-  p as default
-};

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

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

package/dist/ops/webgpu/clipScale.js

@@ -1,58 +0,0 @@
-import { reduce as d, ReduceProgram as g } from "./utils/reductions.js";
-import { c as S } from "../../index-CUXkjxiT.js";
-import h from "./utils/deviceInfo.js";
-class k extends g {
-  shaderKey = "clipscale";
-  constructor(e, i, n) {
-    super(
-      e,
-      i,
-      {
-        reductionOp: "sum",
-        elementwise: !1,
-        forceWorkgroupSize: n
-      },
-      !1
-    ), this.uniforms += "scaling: f32, clipNorm: f32";
-  }
-  getReadSnippet() {
-    return `
-            return bitcast<f32>(u32(x[index]));
-        `;
-  }
-  getWriteSnippet() {
-    return `
-            if (tid == 0) {
-                let cnorm = uniforms.clipNorm;
-                let gradNorm = sqrt(max(bestValue, 0.0));
-                result[0] = (cnorm / max(cnorm, gradNorm)) * uniforms.scaling;
-                result[1] = gradNorm;
-            }
-        `;
-  }
-}
-function w(o) {
-  const { x: e } = o.inputs, { invLossScaling: i, clipNorm: n } = o.attrs, c = o.backend, a = [];
-  if (e.shape.length !== 1)
-    throw new Error(`ClipScaleProgram requires 1D input, but got shape ${e.shape}`);
-  const l = h(c);
-  let r = 128, s = 1;
-  const t = e.shape[0];
-  t <= 16 ? r = 16 : t <= 32 ? r = 32 : t <= 64 && (r = 64), t > 128 && (s = Math.ceil(t / 128));
-  const p = {
-    inSize: r * s,
-    outSize: 2,
-    batchSize: 1,
-    windowSize: r
-  }, u = new k(l, p, r), m = d(u, [e], c, [
-    { type: "float32", data: [i] },
-    { type: "float32", data: [n] }
-  ]);
-  return a.forEach((f) => f.dispose()), m;
-}
-const b = {
-  kernelName: "ClipScale",
-  backendName: "webgpu",
-  kernelFunc: w
-};
-S(b);

package/dist/ops/webgpu/concat16.d.ts

@@ -1,19 +0,0 @@
-import { KernelConfig } from '@tensorflow/tfjs-core';
-import { WebGPUProgram } from '@tensorflow/tfjs-backend-webgpu/dist/webgpu_program';
-export declare class ConcatProgram implements WebGPUProgram {
-    outputShape: number[];
-    shaderKey: string;
-    dispatchLayout: {
-        x: number[];
-    };
-    dispatch: [number, number, number];
-    variableNames: string[];
-    uniforms: string;
-    workPerThread: number;
-    workgroupSize: [number, number, number];
-    size: boolean;
-    offsetLength: number;
-    constructor(shapes: [number, number][]);
-    getUserCode(): string;
-}
-export declare const concatConfig: KernelConfig;

package/dist/ops/webgpu/concat16.js

@@ -1,126 +0,0 @@
-import { c as x, ah as I, U as c } from "../../index-CUXkjxiT.js";
-import { e as D } from "../../webgpu_program-B4HmApL1.js";
-import { f as $, c as F } from "../../webgpu_util-DYlGSwOJ.js";
-import { r as g } from "../../Reshape-CPBkTIH2.js";
-import { a as L, c as d } from "../../concat_util-D18dJ4fD.js";
-class T {
-  outputShape;
-  shaderKey;
-  dispatchLayout;
-  dispatch;
-  variableNames;
-  uniforms = "";
-  workPerThread = 1;
-  workgroupSize = [64, 1, 1];
-  size = !0;
-  offsetLength;
-  constructor(t) {
-    this.outputShape = d(
-      t,
-      1
-      /* axis */
-    ), this.variableNames = t.map((e, a) => `T${a}`), this.dispatchLayout = $(this.outputShape), this.dispatch = F(this.dispatchLayout, this.outputShape, this.workgroupSize, [
-      this.workPerThread,
-      1,
-      1
-    ]), this.offsetLength = t.length - 1;
-    for (let e = 0; e < this.offsetLength; e++)
-      this.uniforms += `offset${e} : i32,`;
-    this.shaderKey = "concat16";
-  }
-  getUserCode() {
-    const t = [];
-    if (this.offsetLength > 0) {
-      t.push(
-        "if (yC < uniforms.offset0){ result[getIndexFromCoords2D(coords, uniforms.outShape)] = T0[getIndexFromCoords2D(vec2<i32>(yR, yC), uniforms.t0Shape)]; }"
-      );
-      for (let s = 1; s < this.offsetLength; s++)
-        t.push(
-          `else if (yC < uniforms.offset${[s]}){ result[getIndexFromCoords2D(coords, uniforms.outShape)] = T${s}[getIndexFromCoords2D(vec2<i32>(yR, yC - uniforms.offset${s - 1}), uniforms.t${s}Shape)]; }`
-        );
-      const a = this.offsetLength, i = this.offsetLength - 1;
-      t.push(
-        `else { result[getIndexFromCoords2D(coords, uniforms.outShape)] = T${a}[getIndexFromCoords2D(vec2<i32>(yR, yC - uniforms.offset${i}), uniforms.t${a}Shape)]; }`
-      );
-    } else
-      t.push(
-        "result[getIndexFromCoords2D(coords, uniforms.outShape)] = T0[getIndexFromCoords2D(vec2<i32>(yR, yC), uniforms.t0Shape)];"
-      );
-    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);
-            let yR = coords.x;
-            let yC = coords.y;
-
-            ${t.join(`
-        `)}
-          }
-        }
-      }
-    `;
-  }
-}
-function m(n, t, e) {
-  const a = e.device.limits.maxStorageBuffersPerShaderStage - 1;
-  if (n.length > a) {
-    const o = [];
-    for (let p = 0; p < n.length; p += a) {
-      const C = n.slice(p, p + a);
-      o.push(m(C, t, e));
-    }
-    const S = m(o, t, e);
-    for (const p of o)
-      e.disposeData(p.dataId);
-    return S;
-  }
-  const { tensors2D: i, outShape: s } = P(n, t, e), h = i.map((o) => o.shape), u = new T(h), f = [], r = new Array(h.length - 1);
-  if (r.length > 0) {
-    r[0] = h[0][1], f.push({ type: "int32", data: [r[0]] });
-    for (let o = 1; o < r.length; o++)
-      r[o] = r[o - 1] + h[o][1], f.push({ type: "int32", data: [r[o]] });
-  }
-  const l = e.runWebGPUProgram(u, i, i[0].dtype, f);
-  i.forEach((o) => e.disposeData(o.dataId));
-  const y = g({ inputs: { x: l }, backend: e, attrs: { shape: s } });
-  return e.disposeData(l.dataId), y;
-}
-function P(n, t, e) {
-  const a = d(
-    n.map((s) => s.shape),
-    t
-  );
-  return { tensors2D: n.map(
-    (s) => g({
-      inputs: { x: s },
-      backend: e,
-      attrs: {
-        shape: [c(s.shape.slice(0, t)), c(s.shape.slice(t))]
-      }
-    })
-  ), outShape: a };
-}
-function w(n) {
-  const { inputs: t, backend: e, attrs: a } = n, { axis: i } = a, s = I(i, t[0].shape)[0], h = t.map((r) => r.shape);
-  L(h, s);
-  const u = d(
-    t.map((r) => r.shape),
-    s
-  );
-  if (c(u) === 0)
-    return e.makeTensorInfo(u, t[0].dtype, []);
-  const f = t.filter((r) => c(r.shape) > 0);
-  return m(f, s, e);
-}
-const v = {
-  kernelName: "Concat16",
-  backendName: "webgpu",
-  kernelFunc: w
-};
-x(v);
-export {
-  T as ConcatProgram,
-  v as concatConfig
-};

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

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

package/dist/ops/webgpu/dropout16.js

@@ -1,51 +0,0 @@
-import { e as u } from "../../webgpu_program-B4HmApL1.js";
-import { f as d, c as i } from "../../webgpu_util-DYlGSwOJ.js";
-import { c } from "../../index-CUXkjxiT.js";
-class p {
-  variableNames = ["x"];
-  outputShape;
-  shaderKey = "Dropout16";
-  dispatchLayout;
-  dispatch;
-  workgroupSize = [64, 1, 1];
-  size = !0;
-  uniforms = "dropout: f32, seed: f32";
-  constructor(e) {
-    this.shaderKey = "Dropout16", this.outputShape = e, this.dispatchLayout = d(this.outputShape), this.dispatch = i(this.dispatchLayout, this.outputShape, this.workgroupSize);
-  }
-  getUserCode() {
-    return `
-
-        fn random(coords: vec3<i32>) -> vec2<f32> {
-            let x1 = f32(coords.x * 4096 + coords.y * 256 + coords.z * 2 * 16);
-            let x2 = f32(coords.x * 4096 + coords.y * 256 + (coords.z * 2 + 1) * 16);
-            return vec2<f32>(fract(sin(uniforms.seed + x1) * 43758.5453123), fract(sin(uniforms.seed + x2) * 43758.5453123));
-        }
-
-        ${u("index")} {
-            if (index < uniforms.size) {
-                let coords = getCoordsFromIndex(index);
-                let values = unpack2x16float(u32(x[index]));
-                let keepProb = 1.0 - uniforms.dropout;
-                let rand = random(coords);
-                let mask = step(rand, vec2<f32>(keepProb));
-                let outVal = values * mask / keepProb;
-                result[index] = i32(pack2x16float(outVal));
-            }
-        }
-        `;
-  }
-}
-function f(t) {
-  const { x: e } = t.inputs, { dropout: o, seed: r } = t.attrs, s = t.backend, a = new p(e.shape), n = [
-    { type: "float32", data: [o] },
-    { type: "float32", data: [r] }
-  ];
-  return s.runWebGPUProgram(a, [e], "packedF16", n);
-}
-const l = {
-  kernelName: "Dropout16",
-  backendName: "webgpu",
-  kernelFunc: f
-};
-c(l);

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

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

package/dist/ops/webgpu/gatherSub.js

@@ -1,39 +0,0 @@
-import { e as u } from "../../webgpu_program-B4HmApL1.js";
-import { f as h, c as p } from "../../webgpu_util-DYlGSwOJ.js";
-import { c, a6 as r } from "../../index-CUXkjxiT.js";
-class l {
-  variableNames = ["labels", "logits", "values"];
-  outputShape;
-  shaderKey = "GatherSub";
-  dispatchLayout;
-  dispatch;
-  workgroupSize = [64, 1, 1];
-  size = !0;
-  constructor(t) {
-    this.outputShape = [t], this.dispatchLayout = h(this.outputShape), this.dispatch = p(this.dispatchLayout, this.outputShape, this.workgroupSize);
-  }
-  getUserCode() {
-    return `
-        ${u("index")} {
-            if (index < uniforms.size) {
-                let idx = i32(labels[index]);
-                let val = values[index];
-                let logit = logits[index * uniforms.logitsShape[1] + idx];
-                setOutputAtIndex(index, val - logit);
-            }
-        }
-    `;
-  }
-}
-function d(e) {
-  const { logits: t, labels: a, values: s } = e.inputs, o = e.backend, i = a.shape[0];
-  r(s.shape, [i], "Error in EfficientGatherSub: "), r(a.shape, [i], "Error in EfficientGatherSub: ");
-  const n = new l(i);
-  return o.runWebGPUProgram(n, [a, t, s], "float32");
-}
-const f = {
-  kernelName: "EfficientGatherSub",
-  backendName: "webgpu",
-  kernelFunc: d
-};
-c(f);

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

@@ -1,14 +0,0 @@
-import { WebGPUProgram } from '@tensorflow/tfjs-backend-webgpu';
-export declare class GeluProgram implements WebGPUProgram {
-    outputShape: number[];
-    shaderKey: string;
-    dispatchLayout: {
-        x: number[];
-    };
-    dispatch: [number, number, number];
-    variableNames: string[];
-    workgroupSize: [number, number, number];
-    size: boolean;
-    constructor(outputShape: number[]);
-    getUserCode(): string;
-}

package/dist/ops/webgpu/gelu.js

@@ -1,141 +0,0 @@
-import { c as d } from "../../index-CUXkjxiT.js";
-import { e as s } from "../../webgpu_program-B4HmApL1.js";
-import { f as n, c as o } from "../../webgpu_util-DYlGSwOJ.js";
-import { isPackedTensor as c } from "../../utilities/packed.js";
-const u = 0.7978845608028654, r = 0.044715;
-class l {
-  outputShape;
-  shaderKey;
-  dispatchLayout;
-  dispatch;
-  variableNames = ["A"];
-  workgroupSize;
-  size = !0;
-  constructor(e) {
-    this.workgroupSize = [128, 1, 1], this.outputShape = e, this.dispatchLayout = n(this.outputShape), this.dispatch = o(this.dispatchLayout, this.outputShape, this.workgroupSize), this.shaderKey = "unary_gelu";
-  }
-  getUserCode() {
-    return `
-            // TODO: revisit after https://github.com/gpuweb/gpuweb/issues/4458 is resolved
-            fn tanhComplete(x: f32) -> f32 {
-                return select(tanh(x), sign(x), abs(x) > 15.0);
-            }
-            fn unaryOperation(x : f32) -> f32 {
-                let x3 = x * x * x;
-                var inner = fma(${r}, x3, x);
-                inner = ${u} * inner;
-                inner = tanhComplete(inner);
-                inner = 0.5 * (1.0 + inner);
-                return x * inner;
-            }
-            ${s("index")} {
-                if (index < uniforms.size) {
-                let a = getAByOutputIndex(index);
-                setOutputAtIndex(index, unaryOperation(a));
-                }
-            }
-      `;
-  }
-}
-function x(t) {
-  const { x: e } = t.inputs, a = t.backend, i = new l(e.shape);
-  return a.runWebGPUProgram(i, [e], "float32");
-}
-const g = {
-  kernelName: "Gelu",
-  backendName: "webgpu",
-  kernelFunc: x
-};
-d(g);
-class f {
-  // Inputs: dy, x
-  variableNames = ["dy", "x"];
-  outputShape;
-  shaderKey = "GeluGrad";
-  dispatchLayout;
-  dispatch;
-  workgroupSize = [128, 1, 1];
-  size = !0;
-  constructor(e) {
-    this.outputShape = e, this.dispatchLayout = n(this.outputShape), this.dispatch = o(this.dispatchLayout, this.outputShape, this.workgroupSize);
-  }
-  getUserCode() {
-    return `
-            // TODO: revisit after https://github.com/gpuweb/gpuweb/issues/4458 is resolved
-            fn tanhComplete(x: f32) -> f32 {
-                return select(tanh(x), sign(x), abs(x) > 15.0);
-            }
-            fn activationGrad(dy: f32, X: f32) -> f32 {
-                let x2 = X * X;
-                let x3 = x2 * X;
-                let u  = ${u} * (X + ${r} * x3);
-                let t  = tanhComplete(u);
-                let sech2 = 1.0 - t * t;
-                let du_dx = ${u} * (1.0 + 3.0 * ${r} * x2);
-                let dgelu = 0.5 * (1.0 + t) + 0.5 * X * sech2 * du_dx;
-                return dy *dgelu;
-            }
-            ${s("index")} {
-                if (index < uniforms.size) {
-                    let X  = unpack2x16float(u32(x[index]));
-                    let DY = unpack2x16float(u32(dy[index]));
-                    let dgelu = vec2<f32>(
-                        activationGrad(DY.x, X.x),
-                        activationGrad(DY.y, X.y)
-                    );
-                    result[index] = i32(pack2x16float(dgelu));
-                }
-            }`;
-  }
-}
-class m {
-  // Inputs: dy, x
-  variableNames = ["dy", "x"];
-  outputShape;
-  shaderKey = "GeluGrad";
-  dispatchLayout;
-  dispatch;
-  workgroupSize = [128, 1, 1];
-  size = !0;
-  constructor(e) {
-    this.outputShape = e, this.dispatchLayout = n(this.outputShape), this.dispatch = o(this.dispatchLayout, this.outputShape, this.workgroupSize);
-  }
-  getUserCode() {
-    return `
-            // TODO: revisit after https://github.com/gpuweb/gpuweb/issues/4458 is resolved
-            fn tanhComplete(x: f32) -> f32 {
-                return select(tanh(x), sign(x), abs(x) > 15.0);
-            }
-            fn activationGrad(dy: f32, X: f32) -> f32 {
-                let x2 = X * X;
-                let x3 = x2 * X;
-                let u  = ${u} * (X + ${r} * x3);
-                let t  = tanhComplete(u);
-                let sech2 = 1.0 - t * t;
-                let du_dx = ${u} * (1.0 + 3.0 * ${r} * x2);
-                let dgelu = 0.5 * (1.0 + t) + 0.5 * X * sech2 * du_dx;
-                return dy *dgelu;
-            }
-            ${s("index")} {
-                if (index < uniforms.size) {
-                    let X  = getXByOutputIndex(index);
-                    let DY = getDyByOutputIndex(index);
-                    let dgelu = activationGrad(DY, X);
-                    setOutputAtIndex(index, dgelu);
-                }
-            }`;
-  }
-}
-function y(t) {
-  const { dy: e, x: a } = t.inputs, i = t.backend, p = c(e), h = p ? new f(a.shape) : new m(a.shape);
-  return i.runWebGPUProgram(h, [e, a], p ? "packedF16" : "float32");
-}
-const b = {
-  kernelName: "GeluGrad",
-  backendName: "webgpu",
-  kernelFunc: y
-};
-d(b);
-export {
-  l as GeluProgram
-};

package/dist/ops/webgpu/index.js

@@ -1,26 +0,0 @@
-import "./attentionMask.js";
-import "./normRMS.js";
-import "./normRMSGrad.js";
-import "./rope.js";
-import "./appendCache.js";
-import "./scatterSub.js";
-import "./gatherSub.js";
-import "./qkv.js";
-import "./gelu.js";
-import "./adamMoments.js";
-import "./adamAdjust.js";
-import "./pack16.js";
-import "./unpack16.js";
-import "./softmax16.js";
-import "./matMul16.js";
-import "./transpose16.js";
-import "./sum16.js";
-import "./slice16.js";
-import "./add16.js";
-import "./concat16.js";
-import "./mul16.js";
-import "./sub16.js";
-import "./softmax16grad.js";
-import "./norm2.js";
-import "./clipScale.js";
-import "./dropout16.js";

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

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

package/dist/ops/webgpu/matMul16.js

@@ -1,65 +0,0 @@
-import { c as V, m as U, a as g, n as X, U as y, _ as Y } from "../../index-CUXkjxiT.js";
-import { isPackedTensor as x } from "../../utilities/packed.js";
-import { reshape16 as C } from "../reshape16.js";
-import { matMulMul as Z } from "../matMulMul.js";
-import { matMulGelu as $ } from "../matMulGelu.js";
-import ee from "./matMul16_program.js";
-import { m as v } from "../../mat_mul-DhG0Newp.js";
-import { r as E } from "../../relu-J_X6MUzx.js";
-import { r as F } from "../../reshape-BDOuCSNW.js";
-import { t as W } from "../../transpose-9kRxIXWR.js";
-function se(m) {
-  const { A: s, B: t } = m.inputs, { transposeA: p, transposeB: i, scale: h, activation: c, scaleA: d, scaleB: f, forceOutputShape: r, perm: k, causalMask: w, pastLen: z } = m.attrs, I = m.backend, B = !x(s), b = !x(t);
-  if (B && b) {
-    const u = d !== void 0 ? U(s, g(d)) : s, l = f !== void 0 ? U(t, g(f)) : t;
-    if (w)
-      throw new Error("Causal mask is not supported for unpacked MatMul16.");
-    let e;
-    if (h !== void 0)
-      e = Z(u, l, g(h), p, i);
-    else if (c === "gelu")
-      e = $(u, l);
-    else if (c === "relu2") {
-      const o = v(u, l, p, i), A = E(o);
-      o.dispose(), e = X(A), A.dispose();
-    } else c === "relu" ? e = E(v(u, l, p, i)) : e = v(u, l, p, i);
-    if (k)
-      if (r) {
-        const o = F(e, r);
-        e.dispose();
-        const A = W(o, k);
-        return o.dispose(), A;
-      } else {
-        const o = W(e, k);
-        return e.dispose(), o;
-      }
-    else if (r) {
-      const o = F(e, r);
-      return e.dispose(), o;
-    } else
-      return e;
-  }
-  if (B && !b)
-    throw new Error("When using mixed precision, A must be packed if B is packed.");
-  if (!B && b)
-    throw new Error("When using mixed precision, B must be packed if A is packed.");
-  const M = s.shape.length, S = t.shape.length, N = s.shape.slice(0, -2), q = t.shape.slice(0, -2), R = y(N), D = y(q), K = Y(s.shape.slice(0, -2), t.shape.slice(0, -2)), L = Math.max(R, D), T = s.shape[M - 2], _ = t.shape[S - 2], j = s.shape[M - 1] * 2, H = t.shape[S - 1] * 2, G = C(s, [R, s.shape[M - 2], s.shape[M - 1]]), O = C(t, [D, t.shape[S - 2], t.shape[S - 1]]), a = new ee(L, T, _, j, H, p, i), n = [];
-  h !== void 0 && (a.useScale(), n.push({ type: "float32", data: [h] })), d !== void 0 && (a.useScaleA(), n.push({ type: "float32", data: [d] })), f !== void 0 && (a.useScaleB(), n.push({ type: "float32", data: [f] })), c !== void 0 && a.useActivation(c), w && (a.useCausalMask(), n.push({ type: "int32", data: [z || 0] }));
-  const P = a.outputShape.length;
-  r && (m.attrs.originalShape = a.outputShape);
-  const J = r ?? K.concat([a.outputShape[P - 2], a.outputShape[P - 1]]);
-  a.setOutputShape(J, k);
-  const Q = I.runWebGPUProgram(
-    a,
-    [G, O],
-    "packedF16",
-    n.length > 0 ? n : void 0
-  );
-  return G.dispose(), O.dispose(), Q;
-}
-const te = {
-  kernelName: "MatMul16",
-  backendName: "webgpu",
-  kernelFunc: se
-};
-V(te);