@genai-fi/nanogpt 0.2.8 → 0.2.9

package/dist/layers/TiedEmbedding.js

@@ -1,7 +1,7 @@
-import { o as h, c as i, E as o, y as V, D as X, I as Y, F as Z, N as ee, H as te, J as se, K as ne, O as re, Q as ue, g as L, x as ae, T as A, m as ie, U as oe, u as le, b as q, l as C, V as P, w as U, _ as H } from "../index-DQfEAU9u.js";
-import { s as ce, r as f } from "../sum-B-O33dgG.js";
-import { m } from "../mat_mul-CuHB58-H.js";
-import { c as pe } from "../complex-CeoYJn2o.js";
+import { o as h, c as i, E as o, D as V, F as X, I as Y, H as Z, N as ee, J as te, K as se, O as ne, Q as re, T as ue, h as L, y as ae, U as A, m as ie, V as oe, v as le, d as q, n as C, W as P, x as U, _ as H } from "../index-Dsg28SG6.js";
+import { s as ce, r as f } from "../sum-NWazHI7f.js";
+import { m } from "../mat_mul-BAYDrXvE.js";
+import { c as pe } from "../complex-Cd8sqiBC.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -169,7 +169,7 @@ function Me(t) {
   const s = { x: i(t, "x", "relu") };
   return o.runKernel(ne, s);
 }
-const we = /* @__PURE__ */ h({ relu_: Me });
+const We = /* @__PURE__ */ h({ relu_: Me });
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
@@ -186,11 +186,11 @@ const we = /* @__PURE__ */ h({ relu_: Me });
  * limitations under the License.
  * =============================================================================
  */
-function We(t) {
+function we(t) {
   const s = { x: i(t, "x", "relu6") };
   return o.runKernel(re, s);
 }
-const ze = /* @__PURE__ */ h({ relu6_: We });
+const ze = /* @__PURE__ */ h({ relu6_: we });
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -273,7 +273,7 @@ function Te(t, e, s, n) {
   if (e === "linear")
     return t;
   if (e === "relu")
-    return we(t);
+    return We(t);
   if (e === "elu")
     return me(t);
   if (e === "relu6")
@@ -310,14 +310,14 @@ function Ne({ a: t, b: e, transposeA: s = !1, transposeB: n = !1, bias: r, activ
   }
   let u = i(t, "a", "fused matMul"), a = i(e, "b", "fused matMul");
   [u, a] = q(u, a);
-  const D = s ? u.shape[u.rank - 2] : u.shape[u.rank - 1], b = n ? a.shape[a.rank - 1] : a.shape[a.rank - 2], w = s ? u.shape[u.rank - 1] : u.shape[u.rank - 2], W = n ? a.shape[a.rank - 2] : a.shape[a.rank - 1], T = u.shape.slice(0, -2), y = a.shape.slice(0, -2), B = C(T), N = C(y);
+  const D = s ? u.shape[u.rank - 2] : u.shape[u.rank - 1], b = n ? a.shape[a.rank - 1] : a.shape[a.rank - 2], W = s ? u.shape[u.rank - 1] : u.shape[u.rank - 2], w = n ? a.shape[a.rank - 2] : a.shape[a.rank - 1], T = u.shape.slice(0, -2), y = a.shape.slice(0, -2), B = C(T), N = C(y);
   L(D === b, () => `Error in fused matMul: inner shapes (${D}) and (${b}) of Tensors with shapes ${u.shape} and ${a.shape} and transposeA=${s} and transposeB=${n} must match.`);
-  const O = P(u.shape.slice(0, -2), a.shape.slice(0, -2)).concat([w, W]), F = s ? f(u, [B, D, w]) : f(u, [B, w, D]), R = n ? f(a, [N, W, b]) : f(a, [N, b, W]);
+  const O = P(u.shape.slice(0, -2), a.shape.slice(0, -2)).concat([W, w]), F = s ? f(u, [B, D, W]) : f(u, [B, W, D]), R = n ? f(a, [N, w, b]) : f(a, [N, b, w]);
   let S;
   r != null && (S = i(r, "bias", "fused matMul"), [S] = q(S, u), P(O, S.shape));
-  let G;
-  l != null && (G = i(l, "prelu weights", "fused matMul"));
-  const I = (x, M) => {
+  let v;
+  l != null && (v = i(l, "prelu weights", "fused matMul"));
+  const G = (x, M) => {
     const [g, $, k, z] = M, d = Ae(f(x, k.shape), k, c);
     let K, _;
     if (!s && !n ? (K = m(d, $, !1, !0), _ = m(g, d, !0, !1)) : !s && n ? (K = m(d, $, !1, !1), _ = m(d, g, !0, !1)) : s && !n ? (K = m($, d, !1, !0), _ = m(g, d, !1, !1)) : (K = m($, d, !0, !0), _ = m(d, g, !0, !0)), r != null) {
@@ -325,24 +325,24 @@ function Ne({ a: t, b: e, transposeA: s = !1, transposeB: n = !1, bias: r, activ
       return [K, _, Q];
     } else
       return [K, _];
-  }, v = {
+  }, I = {
     a: F,
     b: R,
     bias: S,
-    preluActivationWeights: G
+    preluActivationWeights: v
   }, j = { transposeA: s, transposeB: n, activation: c, leakyreluAlpha: p };
   return r == null ? U((M, g, $) => {
     const k = (
       // tslint:disable-next-line: no-unnecessary-type-assertion
-      o.runKernel(H, v, j)
+      o.runKernel(H, I, j)
     );
-    return $([M, g, k]), { value: f(k, O), gradFunc: I };
+    return $([M, g, k]), { value: f(k, O), gradFunc: G };
   })(F, R) : U((M, g, $, k) => {
     const z = (
       // tslint:disable-next-line: no-unnecessary-type-assertion
-      o.runKernel(H, v, j)
+      o.runKernel(H, I, j)
     );
-    return k([M, g, z, $]), { value: f(z, O), gradFunc: I };
+    return k([M, g, z, $]), { value: f(z, O), gradFunc: G };
   })(F, R, S);
 }
 const J = /* @__PURE__ */ h({ fusedMatMul_: Ne });
@@ -369,7 +369,7 @@ class E extends Error {
  * https://opensource.org/licenses/MIT.
  * =============================================================================
  */
-function Ge(t, e, s, n) {
+function ve(t, e, s, n) {
   if (t.rank < 2 || e.rank < 2)
     throw new E(`dot requires both inputs to be rank >= 2 but got x shape = ${t.shape} and y shape = ${e.shape}`);
   if (e.rank >= 3) {
@@ -425,7 +425,7 @@ class Pe {
     return this.tf.gather(this.tiedWeights, e, 0);
   }
   project(e) {
-    return Ge(e, this.tiedWeights.transpose());
+    return ve(e, this.tiedWeights.transpose());
   }
   getWeights() {
     return [this.tiedWeights];

package/dist/{mat_mul-CuHB58-H.js → mat_mul-BAYDrXvE.js}

@@ -1,4 +1,4 @@
-import { o as c, c as s, b as m, E as M, B as p } from "./index-DQfEAU9u.js";
+import { o as c, c as s, d as m, E as M, B as p } from "./index-Dsg28SG6.js";
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
@@ -15,13 +15,13 @@ import { o as c, c as s, b as m, E as M, B as p } from "./index-DQfEAU9u.js";
  * limitations under the License.
  * =============================================================================
  */
-function b(e, o, n = !1, l = !1) {
+function f(e, o, n = !1, l = !1) {
   let a = s(e, "a", "matMul"), t = s(o, "b", "matMul");
   [a, t] = m(a, t);
   const r = { a, b: t }, u = { transposeA: n, transposeB: l };
   return M.runKernel(p, r, u);
 }
-const i = /* @__PURE__ */ c({ matMul_: b });
+const i = /* @__PURE__ */ c({ matMul_: f });
 export {
   i as m
 };

package/dist/ops/attentionMask.js

@@ -1,14 +1,14 @@
-import { engine as d } from "@tensorflow/tfjs";
-import { r as k, s as u } from "../index-DQfEAU9u.js";
-import { m as l } from "../mat_mul-CuHB58-H.js";
-class p {
+import { engine as l } from "@tensorflow/tfjs";
+import { r as u, b as k, s as d } from "../index-Dsg28SG6.js";
+import { m as p } from "../mat_mul-BAYDrXvE.js";
+class f {
   variableNames = ["q", "k", "mask"];
   outputShape;
   userCode;
   // enableShapeUniforms = true;
   customUniforms = [{ name: "divisor", type: "float" }];
-  constructor(t, e, n, a) {
-    this.outputShape = [t, e, n, n], this.userCode = `
+  constructor(s, n, e, a) {
+    this.outputShape = [s, n, e, e], this.userCode = `
         void main() {
             ivec4 coords = getOutputCoords(); // [batch, nh, t1, t2]
             int b = coords.x;
@@ -34,29 +34,49 @@ class p {
         `;
   }
 }
-function f(s) {
-  const { q: t, k: e, mask: n } = s.inputs, { divisor: a } = s.attrs, o = s.backend, c = t.shape[0], i = t.shape[2], r = t.shape[1], m = new p(c, r, i, t.shape[3]);
-  return o.runWebGLProgram(m, [t, e, n], "float32", [[a]]);
+function h(t) {
+  const { q: s, k: n, mask: e } = t.inputs, { divisor: a } = t.attrs, o = t.backend, r = s.shape[0], i = s.shape[2], c = s.shape[1], m = new f(r, c, i, s.shape[3]);
+  return o.runWebGLProgram(m, [s, n, e], "float32", [[a]]);
 }
-const h = {
+const v = {
   kernelName: "AttentionMask",
   backendName: "webgl",
-  kernelFunc: f
+  kernelFunc: h
 };
-k(h);
-function b(s) {
-  const { q: t, k: e, mask: n } = s.inputs, { divisor: a } = s.attrs, o = t.shape[2], i = l(t, e, !1, !0).mul(u(a)), r = n.slice([0, 0], [o, o]).expandDims(0).expandDims(0);
-  return i.add(r);
+u(v);
+function b(t) {
+  const { q: s, k: n, mask: e } = t.inputs, { divisor: a } = t.attrs, o = s.shape[2], i = p(s, n, !1, !0).mul(d(a)), c = e.slice([0, 0], [o, o]).expandDims(0).expandDims(0);
+  return i.add(c);
 }
-const v = {
+const M = {
   kernelName: "AttentionMask",
   backendName: "cpu",
   kernelFunc: b
 };
-k(v);
-function C(s, t, e, n) {
-  return d().runKernel("AttentionMask", { q: s, k: t, mask: e }, { divisor: n });
+u(M);
+function w(t, s, n, e) {
+  return l().runKernel("AttentionMask", { q: t, k: s, mask: n }, { divisor: e });
 }
+const g = {
+  kernelName: "AttentionMask",
+  inputsToSave: ["q", "k"],
+  outputsToSave: [],
+  gradFunc: (t, s, n) => {
+    if (Array.isArray(t))
+      throw new Error("Expected dy to be a single Tensor");
+    const [e, a] = s, { divisor: o } = n;
+    return {
+      q: () => t.matMul(a).mul(o),
+      k: () => e.transpose([0, 1, 3, 2]).matMul(t).mul(o).transpose([0, 1, 3, 2]),
+      mask: () => t,
+      divisor: () => {
+        const r = e.matMul(a, !1, !0);
+        return t.mul(r).sum();
+      }
+    };
+  }
+};
+k(g);
 export {
-  C as attentionMask
+  w as attentionMask
 };

package/dist/ops/gatherSub.js

@@ -1,6 +1,6 @@
 import { engine as l } from "@tensorflow/tfjs";
-import { o as g, c as i, E as b, G as d, r as c, a as h } from "../index-DQfEAU9u.js";
-import { r as p, s as f } from "../stack-C9cTkqpq.js";
+import { o as g, c as i, E as b, G as d, r as c, a as h } from "../index-Dsg28SG6.js";
+import { r as p, s as f } from "../stack-1o648CP_.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.

package/dist/ops/node/sparseCrossEntropy.js

@@ -1,4 +1,4 @@
-import { r as o } from "../../index-DQfEAU9u.js";
+import { r as o } from "../../index-Dsg28SG6.js";
 function r(e) {
   const { logits: t, labels: n } = e.inputs;
   return e.backend.executeMultipleOutputs("SparseSoftmaxCrossEntropyWithLogits", [], [t, n], 2);

package/dist/ops/scatterSub.js

@@ -1,7 +1,7 @@
 import { engine as $ } from "@tensorflow/tfjs";
-import { j as u, k as S, l as p, E as f, n as E, o as N, c as l, p as y, r as h, a as D, m as x } from "../index-DQfEAU9u.js";
-import { c as m } from "../complex-CeoYJn2o.js";
-import { r as v, s as T } from "../stack-C9cTkqpq.js";
+import { k as u, l as S, n as p, E as f, p as E, o as N, c as l, q as y, r as h, a as D, m as x } from "../index-Dsg28SG6.js";
+import { c as m } from "../complex-Cd8sqiBC.js";
+import { r as v, s as T } from "../stack-1o648CP_.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -142,9 +142,9 @@ const F = {
   kernelFunc: A
 };
 h(F);
-function M(e, t, r) {
+function R(e, t, r) {
   return $().runKernel("EfficientScatterSub", { logits: e, labels: t, dy: r }, {});
 }
 export {
-  M as scatterSub
+  R as scatterSub
 };

package/dist/{stack-C9cTkqpq.js → stack-1o648CP_.js}

@@ -1,4 +1,4 @@
-import { E as e, R as c, o as f, f as u, g as a, P as i } from "./index-DQfEAU9u.js";
+import { E as e, R as c, o as f, g as u, h as a, P as i } from "./index-Dsg28SG6.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -15,7 +15,7 @@ import { E as e, R as c, o as f, f as u, g as a, P as i } from "./index-DQfEAU9u
  * limitations under the License.
  * =============================================================================
  */
-function l(n, s, t = 1, r = "float32") {
+function h(n, s, t = 1, r = "float32") {
   if (t === 0)
     throw new Error("Cannot have a step of zero");
   const o = { start: n, stop: s, step: t, dtype: r };
@@ -43,8 +43,8 @@ function k(n, s = 0) {
   const r = t, o = { axis: s };
   return e.runKernel(i, r, o);
 }
-const h = /* @__PURE__ */ f({ stack_: k });
+const l = /* @__PURE__ */ f({ stack_: k });
 export {
-  l as r,
-  h as s
+  h as r,
+  l as s
 };

package/dist/{sum-B-O33dgG.js → sum-NWazHI7f.js}

@@ -1,4 +1,4 @@
-import { o, c as a, E as u, h as i, i as p, S as x } from "./index-DQfEAU9u.js";
+import { o, c as a, E as u, i, j as p, S as x } from "./index-Dsg28SG6.js";
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.

package/dist/training/AdamExt.js

@@ -1,4 +1,4 @@
-import { A as r, m as c, s as h, a as g, e as o } from "../index-DQfEAU9u.js";
+import { A as r, m as c, s as h, a as g, e as o } from "../index-Dsg28SG6.js";
 class u extends r {
   constructor(t, e, s, a, i) {
     super(t, e, s, a), this.config = i, this.startLearningRate = t;

package/dist/training/sparseCrossEntropy.js

@@ -1,7 +1,7 @@
 import { gatherSub as w } from "../ops/gatherSub.js";
 import { scatterSub as K } from "../ops/scatterSub.js";
-import { o as l, c as d, E as f, M as _, q as z, L as I, t as N, a as E, u as M, v as T, e as m, w as g, x as $, z as S } from "../index-DQfEAU9u.js";
-import { s as F, r as b } from "../sum-B-O33dgG.js";
+import { o as l, c as d, E as f, M as _, t as z, L as I, u as N, a as E, v as M, w as T, e as m, x as g, y as $, z as S } from "../index-Dsg28SG6.js";
+import { s as F, r as b } from "../sum-NWazHI7f.js";
 /**
  * @license
  * Copyright 2017 Google LLC. All Rights Reserved.
@@ -25,7 +25,7 @@ function P(n, s, t) {
     t.indexOf(o) === -1 ? e.push(n[r++]) : e.push(s[c++]);
   return e;
 }
-function q(n, s) {
+function A(n, s) {
   const t = s.map((a) => 1);
   return P(n, t, s);
 }
@@ -45,11 +45,11 @@ function q(n, s) {
  * limitations under the License.
  * =============================================================================
  */
-function A(n, s = null, t = !1) {
+function D(n, s = null, t = !1) {
   const e = { x: d(n, "x", "max") }, r = { reductionIndices: s, keepDims: t };
   return f.runKernel(_, e, r);
 }
-const L = /* @__PURE__ */ l({ max_: A });
+const L = /* @__PURE__ */ l({ max_: D });
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -66,11 +66,11 @@ const L = /* @__PURE__ */ l({ max_: A });
  * limitations under the License.
  * =============================================================================
  */
-function D(n) {
+function O(n) {
   const t = { x: d(n, "x", "exp") };
   return f.runKernel(z, t);
 }
-const O = /* @__PURE__ */ l({ exp_: D });
+const W = /* @__PURE__ */ l({ exp_: O });
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -87,11 +87,11 @@ const O = /* @__PURE__ */ l({ exp_: D });
  * limitations under the License.
  * =============================================================================
  */
-function W(n) {
+function j(n) {
   const t = { x: d(n, "x", "log", "float32") };
   return f.runKernel(I, t);
 }
-const j = /* @__PURE__ */ l({ log_: W });
+const q = /* @__PURE__ */ l({ log_: j });
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
@@ -114,9 +114,9 @@ function B(n, s = null, t = !1) {
     e,
     !0
     /* keepDims */
-  ), c = E(a, r), o = O(c), p = F(o, e), u = j(p), i = M(b(r, u.shape), u);
+  ), c = E(a, r), o = W(c), p = F(o, e), u = q(p), i = M(b(r, u.shape), u);
   if (t) {
-    const h = q(i.shape, e);
+    const h = A(i.shape, e);
     return b(i, h);
   }
   return i;
@@ -165,7 +165,7 @@ function ss() {
     (s, t, a) => {
       const e = s.shape[s.shape.length - 1], c = s.shape.slice(0, -1).reduce((h, x) => h * x, 1), o = s.reshape([c, e]), p = t.reshape([c]).cast("int32"), u = R(o, p);
       return a([o, p]), o.dispose(), p.dispose(), { value: u, gradFunc: (h, x) => $(() => {
-        const k = x[0], y = x[1], C = Q(k), G = K(C, y, h), v = S(t);
+        const y = x[0], k = x[1], C = Q(y), G = K(C, k, h), v = S(t);
         return [G.reshape(s.shape), v];
       }) };
     }

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@genai-fi/nanogpt",
-    "version": "0.2.8",
+    "version": "0.2.9",
     "type": "module",
     "main": "dist/main.js",
     "types": "dist/main.d.ts",