@genai-fi/nanogpt 0.19.1 → 0.20.1

package/dist/ops/webgpu/softmax16.js

@@ -1,21 +1,18 @@
-import { c as h, U as S, h as b } from "../../index-CUXkjxiT.js";
-import { reshape16 as d } from "../reshape16.js";
-import x from "./softmax16_program.js";
-import k from "./softmax16_subgroup_program.js";
-import l from "./utils/deviceInfo.js";
-import { r as z } from "../../reshape-BDOuCSNW.js";
-function F(a) {
-  const { inputs: t, backend: o, attrs: p } = a, { logits: e } = t, { dim: r } = p, m = o.subgroupMinSize, i = o.subgroupMaxSize, c = l(o).subgroupsSupported, s = z(e, [
-    S(e.shape) / e.shape[r],
-    e.shape[r]
-  ]), u = c ? new k(s.shape, m, i) : new x(s.shape), f = o.runWebGPUProgram(u, [s], "packedF16");
-  s.dispose();
-  const n = b().makeTensorFromTensorInfo(f), g = d(n, e.shape);
-  return n.dispose(), g;
+import { Ii as e, br as t, ii as n, oc as r } from "../../dist-BewPQWjc.js";
+import { reshape16 as i } from "../reshape16.js";
+import a from "./utils/deviceInfo.js";
+import o from "./softmax16_subgroup_program.js";
+import s from "./softmax16_program.js";
+//#region lib/ops/webgpu/softmax16.ts
+function c(e) {
+	let { inputs: c, backend: l, attrs: u } = e, { logits: d } = c, { dim: f } = u, p = l.subgroupMinSize, m = l.subgroupMaxSize, h = a(l).subgroupsSupported, g = t(d, [r(d.shape) / d.shape[f], d.shape[f]]), _ = h ? new o(g.shape, p, m) : new s(g.shape), v = l.runWebGPUProgram(_, [g], "packedF16");
+	g.dispose();
+	let y = n().makeTensorFromTensorInfo(v), b = i(y, d.shape);
+	return y.dispose(), b;
 }
-const I = {
-  kernelName: "Softmax16",
-  backendName: "webgpu",
-  kernelFunc: F
-};
-h(I);
+e({
+	kernelName: "Softmax16",
+	backendName: "webgpu",
+	kernelFunc: c
+});
+//#endregion

package/dist/ops/webgpu/softmax16_program.js

@@ -1,17 +1,34 @@
-import { e } from "../../webgpu_program-B4HmApL1.js";
-import { f as o } from "../../webgpu_util-DYlGSwOJ.js";
-class i {
-  variableNames = ["logits"];
-  outputShape;
-  shaderKey;
-  dispatchLayout;
-  dispatch;
-  workgroupSize;
-  constructor(r) {
-    this.outputShape = r, this.dispatchLayout = o(this.outputShape), this.dispatch = [this.outputShape[0], 1, 1], this.outputShape[1] >= 4096 ? this.workgroupSize = [256, 1, 1] : this.outputShape[1] < 64 ? this.workgroupSize = [32, 1, 1] : this.workgroupSize = [64, 1, 1], this.shaderKey = "softmax16";
-  }
-  getUserCode() {
-    return `
+import { s as e } from "../../webgpu_program-WOyIVMlZ.js";
+import { c as t } from "../../webgpu_util-B_F3SShA.js";
+//#region lib/ops/webgpu/softmax16_program.ts
+var n = class {
+	variableNames = ["logits"];
+	outputShape;
+	shaderKey;
+	dispatchLayout;
+	dispatch;
+	workgroupSize;
+	constructor(e) {
+		this.outputShape = e, this.dispatchLayout = t(this.outputShape), this.dispatch = [
+			this.outputShape[0],
+			1,
+			1
+		], this.outputShape[1] >= 4096 ? this.workgroupSize = [
+			256,
+			1,
+			1
+		] : this.outputShape[1] < 64 ? this.workgroupSize = [
+			32,
+			1,
+			1
+		] : this.workgroupSize = [
+			64,
+			1,
+			1
+		], this.shaderKey = "softmax16";
+	}
+	getUserCode() {
+		return `
         var<workgroup> buf : array<f32, ${this.workgroupSize[0]}>;
         const blockSize = ${this.workgroupSize[0]};
         ${e("index")} {
@@ -66,8 +83,7 @@ class i {
             }
         }
       `;
-  }
-}
-export {
-  i as default
+	}
 };
+//#endregion
+export { n as default };

package/dist/ops/webgpu/softmax16_subgroup_program.js

@@ -1,25 +1,38 @@
-import { e as o } from "../../webgpu_program-B4HmApL1.js";
-import { f as u } from "../../webgpu_util-DYlGSwOJ.js";
-class i {
-  variableNames = ["logits"];
-  outputShape;
-  shaderKey;
-  dispatchLayout;
-  dispatch;
-  workgroupSize;
-  minSubgroupSize;
-  maxSubgroupSize;
-  subgroups = !0;
-  subgroupBuiltins = !1;
-  constructor(e, a, r) {
-    this.minSubgroupSize = a, this.maxSubgroupSize = r, this.outputShape = e, this.dispatchLayout = u(this.outputShape), this.dispatch = [this.outputShape[0], 1, 1], this.workgroupSize = [Math.min(64, r), 1, 1], a !== r && (this.subgroupBuiltins = !0, this.workgroupSize = [64, 1, 1]), this.shaderKey = "softmax16subgroup";
-  }
-  getUserCode() {
-    const e = this.maxSubgroupSize !== this.minSubgroupSize;
-    return `
-        ${e ? `var<workgroup> bestValues : array<f32, ${this.workgroupSize[0]}>;` : ""}
+import { s as e } from "../../webgpu_program-WOyIVMlZ.js";
+import { c as t } from "../../webgpu_util-B_F3SShA.js";
+//#region lib/ops/webgpu/softmax16_subgroup_program.ts
+var n = class {
+	variableNames = ["logits"];
+	outputShape;
+	shaderKey;
+	dispatchLayout;
+	dispatch;
+	workgroupSize;
+	minSubgroupSize;
+	maxSubgroupSize;
+	subgroups = !0;
+	subgroupBuiltins = !1;
+	constructor(e, n, r) {
+		this.minSubgroupSize = n, this.maxSubgroupSize = r, this.outputShape = e, this.dispatchLayout = t(this.outputShape), this.dispatch = [
+			this.outputShape[0],
+			1,
+			1
+		], this.workgroupSize = [
+			Math.min(64, r),
+			1,
+			1
+		], n !== r && (this.subgroupBuiltins = !0, this.workgroupSize = [
+			64,
+			1,
+			1
+		]), this.shaderKey = "softmax16subgroup";
+	}
+	getUserCode() {
+		let t = this.maxSubgroupSize !== this.minSubgroupSize;
+		return `
+        ${t ? `var<workgroup> bestValues : array<f32, ${this.workgroupSize[0]}>;` : ""}
         const blockSize = ${this.workgroupSize[0]};
-        ${o("index")} {
+        ${e("index")} {
             let row = index / blockSize;
             let tid = i32(localId.x);
             let cols = uniforms.outShape[1];
@@ -32,17 +45,7 @@ class i {
             }
 
             threadMax = subgroupMax(threadMax);
-            ${e ? `
-                let lane = localId.x % subgroupSize;
-                if (lane == 0) {
-                    bestValues[localId.x / subgroupSize] = threadMax;
-                }
-                workgroupBarrier();
-                let numSubgroups = blockSize / subgroupSize;
-                threadMax = select(-3.402823e+38f, bestValues[lane], lane < numSubgroups);
-                threadMax = subgroupMax(threadMax);
-                workgroupBarrier();    
-            ` : ""}
+            ${t ? "\n                let lane = localId.x % subgroupSize;\n                if (lane == 0) {\n                    bestValues[localId.x / subgroupSize] = threadMax;\n                }\n                workgroupBarrier();\n                let numSubgroups = blockSize / subgroupSize;\n                threadMax = select(-3.402823e+38f, bestValues[lane], lane < numSubgroups);\n                threadMax = subgroupMax(threadMax);\n                workgroupBarrier();    \n            " : ""}
 
             var threadSum = 0.0f;
             for (var col = tid; col < cols; col += blockSize) {
@@ -52,24 +55,16 @@ class i {
             }
 
             threadSum = subgroupAdd(threadSum);
-            ${e ? `
-                if (lane == 0) {
-                    bestValues[localId.x / subgroupSize] = threadSum;
-                }
-                workgroupBarrier();
-                threadSum = select(0.0f, bestValues[lane], lane < numSubgroups);
-                threadSum = subgroupAdd(threadSum);    
-            ` : ""}
+            ${t ? "\n                if (lane == 0) {\n                    bestValues[localId.x / subgroupSize] = threadSum;\n                }\n                workgroupBarrier();\n                threadSum = select(0.0f, bestValues[lane], lane < numSubgroups);\n                threadSum = subgroupAdd(threadSum);    \n            " : ""}
 
-            for (var col = tid; col < cols; col += ${e ? "i32(subgroupSize)" : "blockSize"}) {
+            for (var col = tid; col < cols; col += ${t ? "i32(subgroupSize)" : "blockSize"}) {
                 let value = unpack2x16float(u32(logits[rowIdx + col]));
                 let valuePair: vec2<f32> = exp(value - threadMax) / threadSum;
                 result[rowIdx + col] = i32(pack2x16float(valuePair));
             }
         }
       `;
-  }
-}
-export {
-  i as default
+	}
 };
+//#endregion
+export { n as default };

package/dist/ops/webgpu/softmax16grad.js

@@ -1,37 +1,31 @@
-import { c as p } from "../../index-CUXkjxiT.js";
-import { createReduceInfo as m, reduce as l, ReduceProgram as i } from "./utils/reductions.js";
-import { isPackedTensor as n } from "../../utilities/packed.js";
-import k from "./utils/deviceInfo.js";
-class x extends i {
-  constructor(e, t) {
-    super(e, t, { reductionOp: "sum", elementwise: !0 }, !0), this.shaderKey = "SoftmaxGrad16", this.variableNames = ["dy", "softmaxOutput"], this.variableComponents = [1, 1];
-  }
-  getReadSnippet() {
-    return `
-            let d: vec2<f32> = unpack2x16float(u32(dy[index]));
-            let l: vec2<f32> = unpack2x16float(u32(softmaxOutput[index]));
-            return d * l;
-        `;
-  }
-  getWriteSnippet() {
-    return `
-            let d: vec2<f32> = unpack2x16float(u32(dy[offset + k]));
-            let l: vec2<f32> = unpack2x16float(u32(softmaxOutput[offset + k]));
-            let outVal = l * (d - bestValue);
-            result[offset + k] = i32(pack2x16float(outVal));
-        `;
-  }
-}
-function b(o) {
-  const { dy: e, softmaxOutput: t } = o.inputs, r = o.backend, s = k(r), u = n(e), c = n(t);
-  if (!(u && c))
-    throw new Error("softmaxGradGPU: dy and softmaxOutput must be packed tensors");
-  const a = [e, t], f = m(a, -1), d = new x(s, f);
-  return l(d, a, r);
-}
-const v = {
-  kernelName: "Softmax16Grad",
-  backendName: "webgpu",
-  kernelFunc: b
+import { Ii as e } from "../../dist-BewPQWjc.js";
+import { isPackedTensor as t } from "../../utilities/packed.js";
+import { ReduceProgram as n, createReduceInfo as r, reduce as i } from "./utils/reductions.js";
+import a from "./utils/deviceInfo.js";
+//#region lib/ops/webgpu/softmax16grad.ts
+var o = class extends n {
+	constructor(e, t) {
+		super(e, t, {
+			reductionOp: "sum",
+			elementwise: !0
+		}, !0), this.shaderKey = "SoftmaxGrad16", this.variableNames = ["dy", "softmaxOutput"], this.variableComponents = [1, 1];
+	}
+	getReadSnippet() {
+		return "\n            let d: vec2<f32> = unpack2x16float(u32(dy[index]));\n            let l: vec2<f32> = unpack2x16float(u32(softmaxOutput[index]));\n            return d * l;\n        ";
+	}
+	getWriteSnippet() {
+		return "\n            let d: vec2<f32> = unpack2x16float(u32(dy[offset + k]));\n            let l: vec2<f32> = unpack2x16float(u32(softmaxOutput[offset + k]));\n            let outVal = l * (d - bestValue);\n            result[offset + k] = i32(pack2x16float(outVal));\n        ";
+	}
 };
-p(v);
+function s(e) {
+	let { dy: n, softmaxOutput: s } = e.inputs, c = e.backend, l = a(c), u = t(n), d = t(s);
+	if (!(u && d)) throw Error("softmaxGradGPU: dy and softmaxOutput must be packed tensors");
+	let f = [n, s];
+	return i(new o(l, r(f, -1)), f, c);
+}
+e({
+	kernelName: "Softmax16Grad",
+	backendName: "webgpu",
+	kernelFunc: s
+});
+//#endregion

package/dist/ops/webgpu/sub16.js

@@ -1,13 +1,14 @@
-import { c as t } from "../../index-CUXkjxiT.js";
-import { BinaryOpProgram as s } from "./utils/binary_op.js";
-import { B as c } from "../../binary_op_util-pKXltfxI.js";
-function p(e) {
-  const { a: r, b: n } = e.inputs, o = e.backend, a = new s(c.SUB, r.shape, n.shape);
-  return o.runWebGPUProgram(a, [r, n], "packedF16");
+import { Ii as e } from "../../dist-BewPQWjc.js";
+import { t } from "../../binary_op_util-CrYk9LXL.js";
+import { BinaryOpProgram as n } from "./utils/binary_op.js";
+//#region lib/ops/webgpu/sub16.ts
+function r(e) {
+	let { a: r, b: i } = e.inputs, a = e.backend, o = new n(t.SUB, r.shape, i.shape);
+	return a.runWebGPUProgram(o, [r, i], "packedF16");
 }
-const m = {
-  kernelName: "Sub16",
-  backendName: "webgpu",
-  kernelFunc: p
-};
-t(m);
+e({
+	kernelName: "Sub16",
+	backendName: "webgpu",
+	kernelFunc: r
+});
+//#endregion

package/dist/ops/webgpu/sum16.js

@@ -1,38 +1,29 @@
-import { createReduceInfo as h, reduce as g, ReduceProgram as x } from "./utils/reductions.js";
-import { c as k, ah as A } from "../../index-CUXkjxiT.js";
-import { isPackedTensor as P } from "../../utilities/packed.js";
-import { transpose16 as b } from "../transpose16.js";
-import I from "./utils/deviceInfo.js";
-import { s as w } from "../../sum-BpiwSWvg.js";
-import { a as D, b as K } from "../../axis_util-GTVlo58H.js";
-class v extends x {
-  shaderKey = "sum16";
-  constructor(e, t, o) {
-    super(
-      e,
-      t,
-      {
-        reductionOp: "sum",
-        elementwise: !1
-      },
-      o
-    ), o && (this.shaderKey += "_packed");
-  }
-}
-function y(r) {
-  const { x: e } = r.inputs, { axis: t, keepDims: o } = r.attrs, a = r.backend, c = [], u = I(a), m = P(e);
-  if (!m)
-    return w(e, t, o);
-  let n = A(t ?? -1, e.shape);
-  const i = D(n, e.shape.length);
-  let s = e;
-  i != null && (s = b(e, i), n = K(n.length, s.shape.length), c.push(s));
-  const p = h([s], -1), f = new v(u, p, m), d = g(f, [s], a);
-  return c.forEach((l) => l.dispose()), d;
-}
-const N = {
-  kernelName: "Sum16",
-  backendName: "webgpu",
-  kernelFunc: y
+import { Dn as e, En as t, Ii as n, mn as r, nc as i } from "../../dist-BewPQWjc.js";
+import { isPackedTensor as a } from "../../utilities/packed.js";
+import { transpose16 as o } from "../transpose16.js";
+import { ReduceProgram as s, createReduceInfo as c, reduce as l } from "./utils/reductions.js";
+import u from "./utils/deviceInfo.js";
+//#region lib/ops/webgpu/sum16.ts
+var d = class extends s {
+	shaderKey = "sum16";
+	constructor(e, t, n) {
+		super(e, t, {
+			reductionOp: "sum",
+			elementwise: !1
+		}, n), n && (this.shaderKey += "_packed");
+	}
 };
-k(N);
+function f(n) {
+	let { x: s } = n.inputs, { axis: f, keepDims: p } = n.attrs, m = n.backend, h = [], g = u(m), _ = a(s);
+	if (!_) return r(s, f, p);
+	let v = i(f ?? -1, s.shape), y = t(v, s.shape.length), b = s;
+	y != null && (b = o(s, y), v = e(v.length, b.shape.length), h.push(b));
+	let x = l(new d(g, c([b], -1), _), [b], m);
+	return h.forEach((e) => e.dispose()), x;
+}
+n({
+	kernelName: "Sum16",
+	backendName: "webgpu",
+	kernelFunc: f
+});
+//#endregion

package/dist/ops/webgpu/transpose16.js

@@ -1,34 +1,37 @@
-import { isPackedTensor as d } from "../../utilities/packed.js";
-import { c as k, h as l } from "../../index-CUXkjxiT.js";
-import { reshape16 as u } from "../reshape16.js";
-import P from "./transpose16_shared_program.js";
-import b from "./transpose16_program.js";
-import { t as T } from "../../transpose-9kRxIXWR.js";
-function w(p) {
-  const { inputs: h, attrs: i } = p, { x: e } = h, { perm: r } = i, a = p.backend, c = d(e);
-  if (c && r[r.length - 1] !== e.shape.length - 1) {
-    const o = e.shape.length, m = o === 4 ? r.map((s) => s - 1).filter((s) => s >= 0) : r, t = o === 4 ? u(e, [e.shape[0] * e.shape[1], e.shape[2], e.shape[3]]) : e, f = new P(t.shape, m), n = a.runWebGPUProgram(f, [t], "packedF16");
-    if (o === 4) {
-      t.dispose();
-      const s = l().makeTensorFromTensorInfo(n), g = u(s, [
-        e.shape[0],
-        e.shape[1],
-        n.shape[1],
-        n.shape[2]
-      ]);
-      return s.dispose(), g;
-    }
-    return n;
-  }
-  if (c) {
-    const o = new b(e.shape, r);
-    return a.runWebGPUProgram(o, [e], "packedF16");
-  } else
-    return T(e, r);
+import { Ii as e, ii as t, w as n } from "../../dist-BewPQWjc.js";
+import { isPackedTensor as r } from "../../utilities/packed.js";
+import { reshape16 as i } from "../reshape16.js";
+import a from "./transpose16_shared_program.js";
+import o from "./transpose16_program.js";
+//#region lib/ops/webgpu/transpose16.ts
+function s(e) {
+	let { inputs: s, attrs: c } = e, { x: l } = s, { perm: u } = c, d = e.backend, f = r(l);
+	if (f && u[u.length - 1] !== l.shape.length - 1) {
+		let e = l.shape.length, n = e === 4 ? u.map((e) => e - 1).filter((e) => e >= 0) : u, r = e === 4 ? i(l, [
+			l.shape[0] * l.shape[1],
+			l.shape[2],
+			l.shape[3]
+		]) : l, o = new a(r.shape, n), s = d.runWebGPUProgram(o, [r], "packedF16");
+		if (e === 4) {
+			r.dispose();
+			let e = t().makeTensorFromTensorInfo(s), n = i(e, [
+				l.shape[0],
+				l.shape[1],
+				s.shape[1],
+				s.shape[2]
+			]);
+			return e.dispose(), n;
+		}
+		return s;
+	}
+	if (f) {
+		let e = new o(l.shape, u);
+		return d.runWebGPUProgram(e, [l], "packedF16");
+	} else return n(l, u);
 }
-const F = {
-  kernelName: "Transpose16",
-  backendName: "webgpu",
-  kernelFunc: w
-};
-k(F);
+e({
+	kernelName: "Transpose16",
+	backendName: "webgpu",
+	kernelFunc: s
+});
+//#endregion

package/dist/ops/webgpu/transpose16_program.js

@@ -1,50 +1,51 @@
-import { f as a, c as i } from "../../webgpu_util-DYlGSwOJ.js";
-import { b as h, e as d, c as n } from "../../webgpu_program-B4HmApL1.js";
-function p(r) {
-  const e = r.length;
-  if (e > 6)
-    throw Error(`Transpose for rank ${e} is not yet supported`);
-  const o = new Array(e);
-  for (let t = 0; t < r.length; t++)
-    o[r[t]] = `coords.${n(t)}`;
-  return o.join();
+import { i as e, o as t, s as n } from "../../webgpu_program-WOyIVMlZ.js";
+import { c as r, i } from "../../webgpu_util-B_F3SShA.js";
+//#region lib/ops/webgpu/transpose16_program.ts
+function a(e) {
+	let n = e.length;
+	if (n > 6) throw Error(`Transpose for rank ${n} is not yet supported`);
+	let r = Array(n);
+	for (let n = 0; n < e.length; n++) r[e[n]] = `coords.${t(n)}`;
+	return r.join();
 }
-class l {
-  variableNames = ["A"];
-  shaderKey;
-  outputShape;
-  dispatchLayout;
-  dispatch;
-  workPerThread = 1;
-  workgroupSize = [64, 1, 1];
-  newDim;
-  size = !0;
-  constructor(e, o) {
-    const t = new Array(e.length);
-    for (let s = 0; s < t.length; s++)
-      t[s] = e[o[s]];
-    this.outputShape = t, this.dispatchLayout = a(this.outputShape), this.dispatch = i(this.dispatchLayout, this.outputShape, this.workgroupSize, [
-      this.workPerThread,
-      1,
-      1
-    ]), this.newDim = o, this.shaderKey = `transpose16_${o}`;
-  }
-  getUserCode() {
-    const e = h(this.outputShape.length), o = p(this.newDim);
-    return `
-      ${d("index")} {
+var o = class {
+	variableNames = ["A"];
+	shaderKey;
+	outputShape;
+	dispatchLayout;
+	dispatch;
+	workPerThread = 1;
+	workgroupSize = [
+		64,
+		1,
+		1
+	];
+	newDim;
+	size = !0;
+	constructor(e, t) {
+		let n = Array(e.length);
+		for (let r = 0; r < n.length; r++) n[r] = e[t[r]];
+		this.outputShape = n, this.dispatchLayout = r(this.outputShape), this.dispatch = i(this.dispatchLayout, this.outputShape, this.workgroupSize, [
+			this.workPerThread,
+			1,
+			1
+		]), this.newDim = t, this.shaderKey = `transpose16_${t}`;
+	}
+	getUserCode() {
+		let t = e(this.outputShape.length), r = a(this.newDim);
+		return `
+      ${n("index")} {
         for(var i = 0; i < ${this.workPerThread}; i = i + 1) {
           let flatIndex = index * ${this.workPerThread} + i;
           if(flatIndex < uniforms.size) {
             let coords = getCoordsFromIndex(flatIndex);
             result[flatIndex] = A[getIndexFromCoords${this.outputShape.length}D(
-              ${e}(${o}), uniforms.aShape)];
+              ${t}(${r}), uniforms.aShape)];
           }
         }
       }
     `;
-  }
-}
-export {
-  l as default
+	}
 };
+//#endregion
+export { o as default };