@genai-fi/nanogpt 0.9.1 → 0.10.1

package/dist/ops/webgl/log.js

@@ -1,25 +1,10 @@
-import { f as e, aj as r } from "../../index-BzFyqcy-.js";
-import { u as s, l as N } from "../../kernel_funcs_utils-DKLK0Mg3.js";
-import { y as l } from "../../shared-DuP7ue-R.js";
-/**
- * @license
- * Copyright 2020 Google LLC. All Rights Reserved.
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- * =============================================================================
- */
-const a = N + `
+import "../../index-ZyQhjEPo.js";
+import { u as r, l as e } from "../../kernel_funcs_utils-Dg_-E44D.js";
+import { y as s } from "../../shared-BRksrJb3.js";
+import { r as N, L as l } from "../../tensor_util-DV-FP5Q3.js";
+const t = e + `
   return x < 0.0 ? NAN : log(x);
-`, t = `
+`, a = `
   vec4 result = log(x);
   bvec4 isNaN = isnan(x);
   result.r = isNaN.r ? x.r : (x.r < 0.0 ? NAN : result.r);
@@ -27,13 +12,13 @@ const a = N + `
   result.b = isNaN.b ? x.b : (x.b < 0.0 ? NAN : result.b);
   result.a = isNaN.a ? x.a : (x.a < 0.0 ? NAN : result.a);
   return result;
-`, n = s({
-  opSnippet: a,
-  packedOpSnippet: t,
-  cpuKernelImpl: l
+`, n = r({
+  opSnippet: t,
+  packedOpSnippet: a,
+  cpuKernelImpl: s
 }), o = {
-  kernelName: r,
+  kernelName: l,
   backendName: "webgl",
   kernelFunc: n
 };
-e(o);
+N(o);

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

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

package/dist/ops/webgl/matMul16.js

@@ -0,0 +1,37 @@
+import { isPackedTensor as k } from "../../utilities/packed.js";
+import { m as M, b as m } from "../../index-ZyQhjEPo.js";
+import { matMulMul as g } from "../matMulMul.js";
+import { matMulGelu as N } from "../matMulGelu.js";
+import { r as U } from "../../tensor_util-DV-FP5Q3.js";
+import { m as G } from "../../mat_mul-DeAh4uTH.js";
+import { r as w } from "../../reshape-DevtBWtf.js";
+import { t as h } from "../../transpose-DKELTqhe.js";
+function P(p) {
+  const { A: r, B: o } = p.inputs, { transposeA: l, transposeB: c, scale: u, activation: A, scaleA: f, scaleB: d, forceOutputShape: t, perm: n } = p.attrs, B = !k(r), v = !k(o);
+  if (B && v) {
+    const a = f !== void 0 ? M(r, m(f)) : r, i = d !== void 0 ? M(o, m(d)) : o;
+    let e;
+    if (u !== void 0 ? e = g(a, i, m(u), l, c) : A === "gelu" ? e = N(a, i) : e = G(a, i, l, c), n)
+      if (t) {
+        const s = w(e, t);
+        e.dispose();
+        const b = h(s, n);
+        return s.dispose(), b;
+      } else {
+        const s = h(e, n);
+        return e.dispose(), s;
+      }
+    else if (t) {
+      const s = w(e, t);
+      return e.dispose(), s;
+    } else
+      return e;
+  }
+  throw new Error("Not implemented: matMul16 with packed inputs");
+}
+const C = {
+  kernelName: "MatMul16",
+  backendName: "webgl",
+  kernelFunc: P
+};
+U(C);

package/dist/ops/webgl/matMulGelu.js

@@ -1,7 +1,9 @@
-import { f as E, t as R, e as C, j as $, l as N, n as H, u as O } from "../../index-BzFyqcy-.js";
-import { r as f } from "../../Reshape-Bowtk9BP.js";
-import { M as U } from "../../mulmat_packed_gpu-BT60jmzP.js";
-import { m as A } from "../../mat_mul-DzjTFx-u.js";
+import { t as R, e as C, j as N } from "../../index-ZyQhjEPo.js";
+import { r as f } from "../../Reshape-_kILl6tK.js";
+import { M as H } from "../../mulmat_packed_gpu-BMFhLwta.js";
+import { r as E, u as O } from "../../tensor_util-DV-FP5Q3.js";
+import { m as $ } from "../../mat_mul-DeAh4uTH.js";
+import { s as A, a as U } from "../../tensor-DdQUJZlz.js";
 const M = 0.7978845608028654, g = 0.044715, j = `
     vec4 x3 = x * x * x;
     vec4 inner = x + ${g} * x3;
@@ -29,25 +31,25 @@ const M = 0.7978845608028654, g = 0.044715, j = `
     vec4 du_dx = ${M} * (1.0 + 3.0 * ${g} * a2);
     vec4 dgelu = 0.5 * (1.0 + t) + 0.5 * a * sech2 * du_dx;
     return dgelu * b;
-`, ne = 1e3;
+`, se = 1e3;
 function _({
   a: e,
   b: n,
   transposeA: s,
   transposeB: t,
   backend: a,
-  activationSnippet: c,
+  activationSnippet: i,
   multiplier: o
 }) {
-  const r = e.shape.length, i = n.shape.length, u = s ? e.shape[r - 2] : e.shape[r - 1], h = t ? n.shape[i - 1] : n.shape[i - 2], p = s ? e.shape[r - 1] : e.shape[r - 2], l = t ? n.shape[i - 2] : n.shape[i - 1], w = e.shape.slice(0, -2), K = n.shape.slice(0, -2), d = $(w), m = $(K), T = N(e.shape.slice(0, -2), n.shape.slice(0, -2)).concat([p, l]);
-  H(
-    u === h,
-    () => `Error in matMul: inner shapes (${u}) and (${h}) of Tensors with shapes ${e.shape} and ${n.shape} and transposeA=${s} and transposeB=${t} must match.`
+  const r = e.shape.length, c = n.shape.length, u = s ? e.shape[r - 2] : e.shape[r - 1], p = t ? n.shape[c - 1] : n.shape[c - 2], h = s ? e.shape[r - 1] : e.shape[r - 2], l = t ? n.shape[c - 2] : n.shape[c - 1], w = e.shape.slice(0, -2), K = n.shape.slice(0, -2), m = A(w), d = A(K), T = N(e.shape.slice(0, -2), n.shape.slice(0, -2)).concat([h, l]);
+  U(
+    u === p,
+    () => `Error in matMul: inner shapes (${u}) and (${p}) of Tensors with shapes ${e.shape} and ${n.shape} and transposeA=${s} and transposeB=${t} must match.`
   );
-  const v = s ? [d, u, p] : [d, p, u], x = t ? [m, l, h] : [m, h, l], S = f({ inputs: { x: e }, backend: a, attrs: { shape: v } }), b = f({ inputs: { x: n }, backend: a, attrs: { shape: x } }), D = [S, b], y = Math.max(d, m), L = c, B = O(e.dtype, n.dtype), F = new U(
+  const v = s ? [m, u, h] : [m, h, u], x = t ? [d, l, p] : [d, p, l], S = f({ inputs: { x: e }, backend: a, attrs: { shape: v } }), b = f({ inputs: { x: n }, backend: a, attrs: { shape: x } }), D = [S, b], y = Math.max(m, d), L = i, B = O(e.dtype, n.dtype), F = new H(
     v,
     x,
-    [y, p, l],
+    [y, h, l],
     s,
     t,
     !1,
@@ -84,7 +86,7 @@ E(W);
 function J(e) {
   const { dy: n, x: s, kernel: t } = e.inputs, a = e.backend;
   return R(() => {
-    const c = C().makeTensorFromTensorInfo(
+    const i = C().makeTensorFromTensorInfo(
       _({
         a: s,
         b: t,
@@ -94,7 +96,7 @@ function J(e) {
         activationSnippet: q,
         multiplier: n
       })
-    ), o = A(c, t, !1, !0), r = A(s, c, !0, !1);
+    ), o = $(i, t, !1, !0), r = $(s, i, !0, !1);
     return [o, r];
   });
 }
@@ -105,7 +107,7 @@ const Q = {
 };
 E(Q);
 export {
-  ne as MATMUL_SHARED_DIM_THRESHOLD,
+  se as MATMUL_SHARED_DIM_THRESHOLD,
   _ as batchMatMulGeluImpl,
   z as batchMatMulKernel
 };

package/dist/ops/webgl/matMulMul.js

@@ -1,28 +1,29 @@
-import { f as u } from "../../index-BzFyqcy-.js";
-import { batchMatMulGeluImpl as c } from "./matMulGelu.js";
-const M = `
+import "../../index-ZyQhjEPo.js";
+import { batchMatMulGeluImpl as u } from "./matMulGelu.js";
+import { r as c } from "../../tensor_util-DV-FP5Q3.js";
+const p = `
     return a * b;
 `;
-function p(r) {
-  const { inputs: n, backend: o, attrs: a } = r, { x: t, kernel: e, y: l } = n, { transposeA: i, transposeB: s } = a;
+function M(r) {
+  const { inputs: n, backend: o, attrs: a } = r, { x: t, kernel: e, y: i } = n, { transposeA: l, transposeB: s } = a;
   if (t === void 0 || e === void 0)
     throw new Error("BatchMatMul requires two input tensors.");
-  return c({
+  return u({
     a: t,
     b: e,
-    transposeA: i,
+    transposeA: l,
     transposeB: s,
     backend: o,
-    activationSnippet: M,
-    multiplier: l
+    activationSnippet: p,
+    multiplier: i
   });
 }
 const m = {
   kernelName: "MatMulMul",
   backendName: "webgl",
-  kernelFunc: p
+  kernelFunc: M
 };
-u(m);
+c(m);
 export {
-  p as batchMatMulKernel
+  M as batchMatMulKernel
 };

package/dist/ops/webgl/mulDropout.js

@@ -1,4 +1,5 @@
-import { f as m } from "../../index-BzFyqcy-.js";
+import "../../index-ZyQhjEPo.js";
+import { r as m } from "../../tensor_util-DV-FP5Q3.js";
 class f {
   variableNames = ["a", "b"];
   outputShape;
@@ -7,8 +8,8 @@ class f {
     { name: "dropoutRate", type: "float" },
     { name: "seed", type: "float" }
   ];
-  constructor(t, r, o) {
-    this.outputShape = [t, r, o, o], this.userCode = `
+  constructor(r, t, o) {
+    this.outputShape = [r, t, 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);
@@ -26,16 +27,16 @@ class f {
         `;
   }
 }
-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);
+function i(e) {
+  const { inputs: r, attrs: t } = e, { a: o, b: s } = r, { dropoutRate: a, seed: c } = t, 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 = {
+const b = {
   kernelName: "MulDropout",
   backendName: "webgl",
-  kernelFunc: b
+  kernelFunc: i
 };
-m(i);
+m(b);

package/dist/ops/webgl/normRMS.js

@@ -1,5 +1,6 @@
-import { f as p, e as G } from "../../index-BzFyqcy-.js";
-import { s as x } from "../../sum-CJ0ULhmt.js";
+import { e as G } from "../../index-ZyQhjEPo.js";
+import { r as g } from "../../tensor_util-DV-FP5Q3.js";
+import { s as x } from "../../sum-_fzj5ZTB.js";
 class y {
   variableNames = ["x", "meanSquare", "gamma"];
   outputShape;
@@ -28,7 +29,7 @@ const C = {
   backendName: "webgl",
   kernelFunc: v
 };
-p(C);
+g(C);
 class b {
   variableNames = ["x", "meanSquare", "dyGamma", "dyXMean"];
   outputShape;
@@ -73,14 +74,14 @@ function M(t) {
   l.dispose();
   const f = new b(n, m, u), S = r.runWebGLProgram(f, [e, d, s, i], "float32");
   s.dispose(), i.dispose();
-  const h = new N(n, m, u), g = r.runWebGLProgram(h, [e, d, a], "float32");
+  const h = new N(n, m, u), p = r.runWebGLProgram(h, [e, d, a], "float32");
   d.dispose();
-  const q = x(G().makeTensorFromTensorInfo(g), [0, 1]);
-  return r.disposeIntermediateTensorInfo(g), [S, q];
+  const q = x(G().makeTensorFromTensorInfo(p), [0, 1]);
+  return r.disposeIntermediateTensorInfo(p), [S, q];
 }
 const k = {
   kernelName: "RMSNormGrad",
   backendName: "webgl",
   kernelFunc: M
 };
-p(k);
+g(k);

package/dist/ops/webgl/qkv.js

@@ -1,5 +1,6 @@
-import { f as i } from "../../index-BzFyqcy-.js";
-class l {
+import "../../index-ZyQhjEPo.js";
+import { r as i } from "../../tensor_util-DV-FP5Q3.js";
+class m {
   variableNames = ["x", "kernel"];
   outputShape;
   userCode;
@@ -30,8 +31,8 @@ class l {
         `;
   }
 }
-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);
+function l(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 m(n, s, c, u);
   return [
     o.runWebGLProgram(a, [e, t], "float32", [[0]]),
     o.runWebGLProgram(a, [e, t], "float32", [[1]]),
@@ -41,6 +42,6 @@ function m(r) {
 const d = {
   kernelName: "QKV",
   backendName: "webgl",
-  kernelFunc: m
+  kernelFunc: l
 };
 i(d);

package/dist/ops/webgl/rope.js

@@ -1,12 +1,13 @@
-import { f as u } from "../../index-BzFyqcy-.js";
-class l {
+import "../../index-ZyQhjEPo.js";
+import { r as h } from "../../tensor_util-DV-FP5Q3.js";
+class g {
   variableNames = ["x", "sin", "cos"];
   outputShape;
   userCode;
   // enableShapeUniforms = true;
   customUniforms = [{ name: "pastLen", type: "int" }];
-  constructor(t, n, s, o) {
-    this.outputShape = [t, n, s, o], this.userCode = `
+  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;
@@ -14,7 +15,7 @@ class l {
             int t = coords.z;
             int d = coords.w;
 
-            int rotaryDim = ${o};
+            int rotaryDim = ${n};
 
             float outVal = 0.0;
 
@@ -44,13 +45,13 @@ class l {
         `;
   }
 }
-function h(e) {
-  const { x: t, sin: n, cos: s } = e.inputs, { pastLen: o } = e.attrs, a = e.backend, r = t.shape[0], d = t.shape[1], i = t.shape[2], c = t.shape[3], p = new l(r, d, i, c);
-  return a.runWebGLProgram(p, [t, n, s], "float32", [[o]]);
+function m(o) {
+  const { x: t } = o.inputs, { pastLen: s, ropeCache: e, negSin: n } = o.attrs, a = n ? e.getNegSin() : e.getSin(), r = e.getCos(), i = o.backend, d = t.shape[0], c = t.shape[1], p = t.shape[2], u = t.shape[3], l = new g(d, c, p, u);
+  return i.runWebGLProgram(l, [t, a, r], "float32", [[s]]);
 }
 const f = {
   kernelName: "Rope",
   backendName: "webgl",
-  kernelFunc: h
+  kernelFunc: m
 };
-u(f);
+h(f);

package/dist/ops/webgl/scatterSub.js

@@ -1,10 +1,11 @@
-import { f as i } from "../../index-BzFyqcy-.js";
+import "../../index-ZyQhjEPo.js";
+import { r as i } from "../../tensor_util-DV-FP5Q3.js";
 class u {
   variableNames = ["labels", "softmaxProbs", "dy"];
   outputShape;
   userCode;
-  constructor(e, t) {
-    this.outputShape = [e, t], this.userCode = `
+  constructor(t, e) {
+    this.outputShape = [t, e], this.userCode = `
             void main() {
                 ivec2 coords = getOutputCoords();
                 int index = int(getLabels(coords.x));
@@ -16,8 +17,8 @@ class u {
   }
 }
 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 { logits: t, labels: e, dy: r } = o.inputs, s = o.backend, n = e.shape[0], a = t.shape[1], c = new u(n, a);
+  return s.runWebGLProgram(c, [e, t, r], "float32");
 }
 const b = {
   kernelName: "EfficientScatterSub",

package/dist/ops/webgpu/adamAdjust.js

@@ -1,6 +1,8 @@
-import { g as p } from "../../webgpu_program-DkQJOJSd.js";
-import { f as d, c as l } from "../../webgpu_util-pLEV9tks.js";
-import { f, a4 as c } from "../../index-BzFyqcy-.js";
+import { e as p } from "../../webgpu_program-Cigz-7RF.js";
+import { f as d, c as l } from "../../webgpu_util-BBCnKm2X.js";
+import "../../index-ZyQhjEPo.js";
+import { j as f } from "../../tensor-DdQUJZlz.js";
+import { r as c } from "../../tensor_util-DV-FP5Q3.js";
 class h {
   variableNames = ["moments", "value"];
   outputShape;
@@ -36,19 +38,19 @@ class h {
   }
 }
 function v(t) {
-  const { moments: e, value: a } = t.inputs, { beta1: n, beta2: s, learningRate: o, epsilon: i } = t.attrs, r = t.backend;
-  c(e.shape, [...a.shape, 2], "Error in AdamAdjust: ");
-  const u = new h(a.shape), m = [
+  const { moments: e, value: a } = t.inputs, { beta1: n, beta2: o, learningRate: s, epsilon: i } = t.attrs, r = t.backend;
+  f(e.shape, [...a.shape, 2], "Error in AdamAdjust: ");
+  const m = new h(a.shape), u = [
     { type: "float32", data: [1 / n] },
-    { type: "float32", data: [1 / s] },
-    { type: "float32", data: [o] },
+    { type: "float32", data: [1 / o] },
+    { type: "float32", data: [s] },
     { type: "float32", data: [i] }
   ];
-  return r.runWebGPUProgram(u, [e, a], "float32", m);
+  return r.runWebGPUProgram(m, [e, a], "float32", u);
 }
 const b = {
   kernelName: "AdamAdjust",
   backendName: "webgpu",
   kernelFunc: v
 };
-f(b);
+c(b);

package/dist/ops/webgpu/adamMoments.js

@@ -1,7 +1,9 @@
-import { g as m } from "../../webgpu_program-DkQJOJSd.js";
-import { f as u, c as p } from "../../webgpu_util-pLEV9tks.js";
-import { f as d, a4 as c } from "../../index-BzFyqcy-.js";
-class h {
+import { e as p } from "../../webgpu_program-Cigz-7RF.js";
+import { f as u, c as d } from "../../webgpu_util-BBCnKm2X.js";
+import "../../index-ZyQhjEPo.js";
+import { j as f } from "../../tensor-DdQUJZlz.js";
+import { r as c } from "../../tensor_util-DV-FP5Q3.js";
+class l {
   variableNames = ["moments", "gradient"];
   outputShape;
   shaderKey = "AdamMoments";
@@ -9,11 +11,11 @@ class h {
   dispatch;
   workgroupSize = [64, 1, 1];
   size = !0;
-  uniforms = "beta1: f32, beta2: f32";
+  uniforms = "beta1: f32, beta2: f32, lossScaling: f32";
   outputComponent = 2;
   variableComponents = [2, 1];
-  constructor(e) {
-    this.outputShape = e, this.dispatchLayout = u(this.outputShape.slice(0, -1)), this.dispatch = p(
+  constructor(t) {
+    this.outputShape = t, this.dispatchLayout = u(this.outputShape.slice(0, -1)), this.dispatch = d(
       this.dispatchLayout,
       this.outputShape.slice(0, -1),
       this.workgroupSize,
@@ -22,12 +24,12 @@ class h {
   }
   getUserCode() {
     return `
-        ${m("index")} {
+        ${p("index")} {
             if (index < uniforms.size) {
                 let m: vec2<f32> = moments[index];
 
                 // Add gradient clipping here
-                let g: f32 = clamp(gradient[index], -1.0, 1.0);
+                let g: f32 = clamp(gradient[index] * uniforms.lossScaling, -1.0, 1.0);
 
                 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));
@@ -38,21 +40,24 @@ class h {
     `;
   }
 }
-function f(t) {
-  const { moments: e, gradient: o } = t.inputs, { beta1: a, beta2: n } = t.attrs, r = t.backend;
-  if (c(e.shape, [...o.shape, 2], "Error in AdamMoments: "), a < 0 || a >= 1)
-    throw new Error(`Invalid beta1 value: ${a}. Must be in the range [0, 1).`);
-  if (n < 0 || n >= 1)
-    throw new Error(`Invalid beta2 value: ${n}. Must be in the range [0, 1).`);
-  const s = new h(e.shape), i = [
-    { type: "float32", data: [a] },
-    { type: "float32", data: [n] }
+function h(e) {
+  const { moments: t, gradient: a } = e.inputs, { beta1: n, beta2: o, lossScaling: r } = e.attrs, s = 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 l(t.shape), m = [
+    { type: "float32", data: [n] },
+    { type: "float32", data: [o] },
+    { type: "float32", data: [1 / r] }
   ];
-  return r.runWebGPUProgram(s, [e, o], "float32", i);
+  return s.runWebGPUProgram(i, [t, a], "float32", m);
 }
-const l = {
+const g = {
   kernelName: "AdamMoments",
   backendName: "webgpu",
-  kernelFunc: f
+  kernelFunc: h
 };
-d(l);
+c(g);

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

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

package/dist/ops/webgpu/add16.js

@@ -0,0 +1,14 @@
+import "../../index-ZyQhjEPo.js";
+import { BinaryOpProgram as p } from "./utils/binary_op.js";
+import { B as m } from "../../binary_op_util-pKXltfxI.js";
+import { r as c } from "../../tensor_util-DV-FP5Q3.js";
+function i(r) {
+  const { a: e, b: n } = r.inputs, a = r.backend, t = new p(m.ADD, e.shape, n.shape), o = a.runWebGPUProgram(t, [e, n], "int32");
+  return o.packed = !0, o;
+}
+const s = {
+  kernelName: "Add16",
+  backendName: "webgpu",
+  kernelFunc: i
+};
+c(s);

package/dist/ops/webgpu/appendCache.js

@@ -1,7 +1,10 @@
-import { g as u } from "../../webgpu_program-DkQJOJSd.js";
-import { f as d, c as m } from "../../webgpu_util-pLEV9tks.js";
-import { f as l, a4 as f } from "../../index-BzFyqcy-.js";
-class g {
+import { isPackedTensor as S } from "../../utilities/packed.js";
+import { e as d } from "../../webgpu_program-Cigz-7RF.js";
+import { f as u, c as m } from "../../webgpu_util-BBCnKm2X.js";
+import "../../index-ZyQhjEPo.js";
+import { j as g } from "../../tensor-DdQUJZlz.js";
+import { r as x } from "../../tensor_util-DV-FP5Q3.js";
+class b {
   variableNames = ["cache", "item"];
   outputShape;
   shaderKey = "AppendCache";
@@ -10,14 +13,14 @@ class g {
   workgroupSize = [64, 1, 1];
   size = !0;
   uniforms = "cacheT: i32";
-  constructor(e, t, s, a, c) {
-    const o = Math.min(s + 1, c);
-    this.shaderKey = `AppendCache_${o}`, this.outputShape = [e, t, o, a], this.dispatchLayout = d(this.outputShape), this.dispatch = m(this.dispatchLayout, this.outputShape, this.workgroupSize);
+  constructor(e, t, s, o, i) {
+    const a = Math.min(s + 1, i);
+    this.shaderKey = `AppendCache_${a}`, this.outputShape = [e, t, a, o], this.dispatchLayout = u(this.outputShape), this.dispatch = m(this.dispatchLayout, this.outputShape, this.workgroupSize);
   }
   getUserCode() {
     const e = this.outputShape[2];
     return `
-        ${u("index")} {
+        ${d("index")} {
             if (index < uniforms.size) {
                 let coords = getCoordsFromIndex(index); // [b, h, t, d]
                 let b = coords[0];
@@ -45,16 +48,60 @@ class g {
         `;
   }
 }
-function S(r) {
-  const { cache: e, item: t } = r.inputs, { maxSize: s, pastLen: a } = r.attrs, c = r.backend, o = e.shape[0], n = e.shape[2], i = e.shape[1];
-  if (f(t.shape, [o, i, 1, t.shape[3]], "Error in AppendCache: "), a < 0 || a > s)
-    throw new Error(`Invalid pastLen value: ${a}. Must be in the range [0, ${s}].`);
-  const h = new g(o, i, n, t.shape[3], s), p = [{ type: "int32", data: [a] }];
-  return c.runWebGPUProgram(h, [e, t], "float32", p);
+class C {
+  variableNames = ["cache", "item"];
+  outputShape;
+  shaderKey = "AppendCache";
+  dispatchLayout;
+  dispatch;
+  workgroupSize = [64, 1, 1];
+  size = !0;
+  uniforms = "cacheT: i32";
+  constructor(e, t, s, o, i) {
+    const a = Math.min(s + 1, i);
+    this.shaderKey = `AppendCache_${a}`, this.outputShape = [e, t, a, o], this.dispatchLayout = u(this.outputShape), this.dispatch = m(this.dispatchLayout, this.outputShape, this.workgroupSize);
+  }
+  getUserCode() {
+    const e = this.outputShape[2];
+    return `
+        ${d("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 v(r) {
+  const { cache: e, item: t } = r.inputs, { maxSize: s, pastLen: o } = r.attrs, i = r.backend, a = S(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: "), o < 0 || o > s)
+    throw new Error(`Invalid pastLen value: ${o}. Must be in the range [0, ${s}].`);
+  const l = a ? new C(c, h, n, t.shape[3], s) : new b(c, h, n, t.shape[3], s), f = [{ type: "int32", data: [o] }], T = a ? "int32" : e.dtype, p = i.runWebGPUProgram(l, [e, t], T, f);
+  return p.packed = a, p;
 }
-const T = {
+const z = {
   kernelName: "AppendCache",
   backendName: "webgpu",
-  kernelFunc: S
+  kernelFunc: v
 };
-l(T);
+x(z);