@genai-fi/nanogpt 0.6.0 → 0.6.2

package/dist/ops/webgl/matMulGelu.js

@@ -1,129 +1,62 @@
-import { r as C, t as R, e as I, q as G, Q as L, l as U, U as F } from "../../index-bMBtI-WR.js";
-import { r as S } from "../../Reshape-CLOrdpve.js";
-import { u as H } from "../../gpgpu_math-Ctc31slO.js";
-import { m as B } from "../../mat_mul-BsrLfy81.js";
-/**
- * @license
- * Copyright 2018 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.
- * =============================================================================
- */
-class W {
-  constructor(e, s, n, a = !1, c = !1, o = !1, r = null, u = !1, l = !1) {
-    this.variableNames = ["matrixA", "matrixB"], this.packedInputs = !0, this.packedOutput = !0, this.outputShape = n, this.enableShapeUniforms = H(this.outputShape.length);
-    const h = a ? e[1] : e[2], p = Math.ceil(h / 2), d = a ? "i * 2, rc.y" : "rc.y, i * 2", $ = c ? "rc.z, i * 2" : "i * 2, rc.z", x = a ? ["a.xxyy", "a.zzww"] : ["a.xxzz", "a.yyww"], m = c ? ["b.xzxz", "b.ywyw"] : ["b.xyxy", "b.zwzw"];
-    let i = "", b = "";
-    r && (u ? i = `vec4 activation(vec4 a) {
-          vec4 b = getPreluActivationWeightsAtOutCoords();
-          ${r}
-        }` : l ? i = `vec4 activation(vec4 a) {
-          vec4 b = getLeakyreluAlphaAtOutCoords();
-          ${r}
-        }` : i = `vec4 activation(vec4 x) {
-          ${r}
-        }`, b = "result = activation(result);");
-    const M = o ? "result += getBiasAtOutCoords();" : "";
-    o && this.variableNames.push("bias"), u && this.variableNames.push("preluActivationWeights"), l && this.variableNames.push("leakyreluAlpha");
-    let f = "rc.x", v = "rc.x";
-    e[0] < s[0] ? f = `imod(rc.x, ${e[0]})` : s[0] < e[0] && (v = `imod(rc.x, ${s[0]})`), this.userCode = `
-      ${i}
-      // Don't use uniform for sharedDimensionPacked for performance.
-      const float sharedDimension = ${p}.0;
-
-      vec4 dot2x2ARowBCol(ivec3 rc) {
-        vec4 result = vec4(0);
-        int batchA = ${f};
-        int batchB = ${v};
-        for (int i = 0; i < ${p}; i++) {
-          vec4 a = getMatrixA(batchA, ${d});
-          vec4 b = getMatrixB(batchB, ${$});
-
-          // These swizzled products need to be separately added.
-          // See: https://github.com/tensorflow/tfjs/issues/1735
-          result += (${x[0]} * ${m[0]});
-          result += (${x[1]} * ${m[1]});
-        }
-        return result;
-      }
-
-      void main() {
-        ivec3 rc = getOutputCoords();
-        vec4 result = dot2x2ARowBCol(rc);
-
-        ${M}
-
-        ${b}
-
-        setOutput(result);
-      }
-    `;
-  }
-}
-const g = 0.7978845608028654, w = 0.044715, q = `
+import { r as _, t as R, e as C, g as A, h as N, i as H, u as O } from "../../index-BAzbokzv.js";
+import { r as f } from "../../Reshape-t7Kcikjk.js";
+import { M as U } from "../../mulmat_packed_gpu-D4nKF7Je.js";
+import { m as E } from "../../mat_mul-7121rsJk.js";
+const M = 0.7978845608028654, x = 0.044715, q = `
     vec4 x3 = x * x * x;
-    vec4 inner = x + ${w} * x3;
-    inner = ${g} * inner;
+    vec4 inner = x + ${x} * x3;
+    inner = ${M} * inner;
     inner = tanh(inner);
     inner = 0.5 * (1.0 + inner);
     vec4 result = x * inner;
     return result;
-`, j = `
+`, z = `
     vec4 a2 = a * a;
     vec4 a3 = a2 * a;
-    vec4 u  = ${g} * (a + ${w} * a3);
+    vec4 u  = ${M} * (a + ${x} * a3);
     vec4 t  = tanh(u);
     vec4 sech2 = 1.0 - t * t;
-    vec4 du_dx = ${g} * (1.0 + 3.0 * ${w} * a2);
+    vec4 du_dx = ${M} * (1.0 + 3.0 * ${x} * a2);
     vec4 dgelu = 0.5 * (1.0 + t) + 0.5 * a * sech2 * du_dx;
     return dgelu * b;
-`, se = 1e3;
-function O({
-  a: t,
-  b: e,
+`, te = 1e3;
+function w({
+  a: e,
+  b: t,
   transposeA: s,
   transposeB: n,
   backend: a,
   activationSnippet: c,
   multiplier: o
 }) {
-  const r = t.shape.length, u = e.shape.length, l = s ? t.shape[r - 2] : t.shape[r - 1], h = n ? e.shape[u - 1] : e.shape[u - 2], p = s ? t.shape[r - 1] : t.shape[r - 2], d = n ? e.shape[u - 2] : e.shape[u - 1], $ = t.shape.slice(0, -2), x = e.shape.slice(0, -2), m = G($), i = G(x), M = L(t.shape.slice(0, -2), e.shape.slice(0, -2)).concat([p, d]);
-  U(
-    l === h,
-    () => `Error in matMul: inner shapes (${l}) and (${h}) of Tensors with shapes ${t.shape} and ${e.shape} and transposeA=${s} and transposeB=${n} must match.`
+  const r = e.shape.length, u = t.shape.length, i = s ? e.shape[r - 2] : e.shape[r - 1], p = n ? t.shape[u - 1] : t.shape[u - 2], h = s ? e.shape[r - 1] : e.shape[r - 2], l = n ? t.shape[u - 2] : t.shape[u - 1], K = e.shape.slice(0, -2), T = t.shape.slice(0, -2), d = A(K), m = A(T), b = N(e.shape.slice(0, -2), t.shape.slice(0, -2)).concat([h, l]);
+  H(
+    i === p,
+    () => `Error in matMul: inner shapes (${i}) and (${p}) of Tensors with shapes ${e.shape} and ${t.shape} and transposeA=${s} and transposeB=${n} must match.`
   );
-  const f = s ? [m, l, p] : [m, p, l], v = n ? [i, d, h] : [i, h, d], A = S({ inputs: { x: t }, backend: a, attrs: { shape: f } }), y = S({ inputs: { x: e }, backend: a, attrs: { shape: v } }), D = [A, y], E = Math.max(m, i), N = c, T = F(t.dtype, e.dtype), _ = new W(
-    f,
+  const v = s ? [d, i, h] : [d, h, i], S = n ? [m, l, p] : [m, p, l], g = f({ inputs: { x: e }, backend: a, attrs: { shape: v } }), D = f({ inputs: { x: t }, backend: a, attrs: { shape: S } }), G = [g, D], y = Math.max(d, m), L = c, B = O(e.dtype, t.dtype), F = new U(
     v,
-    [E, p, d],
+    S,
+    [y, h, l],
     s,
     n,
     !1,
-    N,
+    L,
     !!o,
     !1
-  ), k = [A, y];
+  ), k = [g, D];
   o && k.push(o);
-  const z = a.runWebGLProgram(_, k, T), K = S({ inputs: { x: z }, backend: a, attrs: { shape: M } });
-  D.push(z);
-  for (const P of D)
+  const $ = a.runWebGLProgram(F, k, B), I = f({ inputs: { x: $ }, backend: a, attrs: { shape: b } });
+  G.push($);
+  for (const P of G)
     a.disposeIntermediateTensorInfo(P);
-  return K;
+  return I;
 }
-function Q(t) {
-  const { inputs: e, backend: s } = t, { x: n, kernel: a } = e;
+function W(e) {
+  const { inputs: t, backend: s } = e, { x: n, kernel: a } = t;
   if (n === void 0 || a === void 0)
     throw new Error("BatchMatMul requires two input tensors.");
-  return O({
+  return w({
     a: n,
     b: a,
     transposeA: !1,
@@ -132,37 +65,37 @@ function Q(t) {
     activationSnippet: q
   });
 }
-const J = {
+const j = {
   kernelName: "MatMulGelu",
   backendName: "webgl",
-  kernelFunc: Q
+  kernelFunc: W
 };
-C(J);
-function V(t) {
-  const { dy: e, x: s, kernel: n } = t.inputs, a = t.backend;
+_(j);
+function J(e) {
+  const { dy: t, x: s, kernel: n } = e.inputs, a = e.backend;
   return R(() => {
-    const c = I().makeTensorFromTensorInfo(
-      O({
+    const c = C().makeTensorFromTensorInfo(
+      w({
         a: s,
         b: n,
         transposeA: !1,
         transposeB: !1,
         backend: a,
-        activationSnippet: j,
-        multiplier: e
+        activationSnippet: z,
+        multiplier: t
       })
-    ), o = B(c, n, !1, !0), r = B(s, c, !0, !1);
+    ), o = E(c, n, !1, !0), r = E(s, c, !0, !1);
     return [o, r];
   });
 }
-const X = {
+const Q = {
   kernelName: "MatMulGeluGrad",
   backendName: "webgl",
-  kernelFunc: V
+  kernelFunc: J
 };
-C(X);
+_(Q);
 export {
-  se as MATMUL_SHARED_DIM_THRESHOLD,
-  O as batchMatMulGeluImpl,
-  Q as batchMatMulKernel
+  te as MATMUL_SHARED_DIM_THRESHOLD,
+  w as batchMatMulGeluImpl,
+  W as batchMatMulKernel
 };

package/dist/ops/webgl/matMulMul.js

@@ -1,4 +1,4 @@
-import { r as u } from "../../index-bMBtI-WR.js";
+import { r as u } from "../../index-BAzbokzv.js";
 import { batchMatMulGeluImpl as c } from "./matMulGelu.js";
 const M = `
     return a * b;

package/dist/ops/webgl/mulDropout.js

@@ -1,4 +1,4 @@
-import { r as m } from "../../index-bMBtI-WR.js";
+import { r as m } from "../../index-BAzbokzv.js";
 class f {
   variableNames = ["a", "b"];
   outputShape;

package/dist/ops/webgl/normRMS.js

@@ -1,5 +1,5 @@
-import { r as p, e as G } from "../../index-bMBtI-WR.js";
-import { s as x } from "../../sum-CutF5lj2.js";
+import { r as p, e as G } from "../../index-BAzbokzv.js";
+import { s as x } from "../../sum-BE_Irnim.js";
 class y {
   variableNames = ["x", "meanSquare", "gamma"];
   outputShape;

package/dist/ops/webgl/qkv.js

@@ -1,4 +1,4 @@
-import { r as i } from "../../index-bMBtI-WR.js";
+import { r as i } from "../../index-BAzbokzv.js";
 class l {
   variableNames = ["x", "kernel"];
   outputShape;

package/dist/ops/webgl/rope.js

@@ -1,4 +1,4 @@
-import { r as u } from "../../index-bMBtI-WR.js";
+import { r as u } from "../../index-BAzbokzv.js";
 class l {
   variableNames = ["x", "sin", "cos"];
   outputShape;

package/dist/ops/webgl/scatterSub.js

@@ -1,4 +1,4 @@
-import { r as i } from "../../index-bMBtI-WR.js";
+import { r as i } from "../../index-BAzbokzv.js";
 class u {
   variableNames = ["labels", "softmaxProbs", "dy"];
   outputShape;