@genai-fi/nanogpt 1.1.1 → 1.1.3

package/dist/ops/webgpu/transpose16_program.js

@@ -1,12 +1,12 @@
-import { i as e, o as t, s as n } from "../../webgpu_program-B0PokuRJ.js";
-import { c as r, i } from "../../webgpu_util-DfODg6XI.js";
+import { c as e, i as t } from "../../webgpu_util-DfODg6XI.js";
+import { i as n, o as r, s as i } from "../../webgpu_program-B0PokuRJ.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();
+	let t = e.length;
+	if (t > 6) throw Error(`Transpose for rank ${t} is not yet supported`);
+	let n = Array(t);
+	for (let t = 0; t < e.length; t++) n[e[t]] = `coords.${r(t)}`;
+	return n.join();
 }
 var o = class {
 	variableNames = ["A"];
@@ -22,25 +22,25 @@ var o = class {
 	];
 	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, [
+	constructor(n, r) {
+		let i = Array(n.length);
+		for (let e = 0; e < i.length; e++) i[e] = n[r[e]];
+		this.outputShape = i, this.dispatchLayout = e(this.outputShape), this.dispatch = t(this.dispatchLayout, this.outputShape, this.workgroupSize, [
 			this.workPerThread,
 			1,
 			1
-		]), this.newDim = t, this.shaderKey = `transpose16_${t}`;
+		]), this.newDim = r, this.shaderKey = `transpose16_${r}`;
 	}
 	getUserCode() {
-		let t = e(this.outputShape.length), r = a(this.newDim);
+		let e = n(this.outputShape.length), t = a(this.newDim);
 		return `
-      ${n("index")} {
+      ${i("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(
-              ${t}(${r}), uniforms.aShape)];
+              ${e}(${t}), uniforms.aShape)];
           }
         }
       }

package/dist/ops/webgpu/transpose16_shared_program.js

@@ -1,6 +1,6 @@
 import { Ps as e } from "../../dist-Da20xy8E.js";
-import { s as t } from "../../webgpu_program-B0PokuRJ.js";
-import { i as n } from "../../webgpu_util-DfODg6XI.js";
+import { i as t } from "../../webgpu_util-DfODg6XI.js";
+import { s as n } from "../../webgpu_program-B0PokuRJ.js";
 //#region lib/ops/webgpu/transpose16_shared_program.ts
 var r = class {
 	variableNames = ["A"];
@@ -13,10 +13,10 @@ var r = class {
 		8,
 		1
 	];
-	constructor(e, t) {
+	constructor(e, n) {
 		let r = e.length, i = Array(r), a = e.slice();
 		a[a.length - 1] *= 2;
-		for (let e = 0; e < i.length; e++) i[e] = a[t[e]];
+		for (let e = 0; e < i.length; e++) i[e] = a[n[e]];
 		i[i.length - 1] /= 2, this.outputShape = i, this.dispatchLayout = r === 2 ? {
 			x: [0],
 			y: [1]
@@ -24,27 +24,27 @@ var r = class {
 			x: [1],
 			y: [2],
 			z: [0]
-		}, this.dispatch = n(this.dispatchLayout, this.outputShape, this.workgroupSize, [
+		}, this.dispatch = t(this.dispatchLayout, this.outputShape, this.workgroupSize, [
 			2,
 			1,
 			1
 		]), this.shaderKey = `transposeShared16_${r}`;
 	}
 	getUserCode() {
-		let n = this.outputShape.length;
+		let t = this.outputShape.length;
 		e(this.workgroupSize[0] === this.workgroupSize[1], () => `Must be a square tile, current tile shape is ${this.workgroupSize[0]} x ${this.workgroupSize[1]}`);
 		let r = this.workgroupSize[0] * 2;
 		return `
       var<workgroup> tile : array<array<f32, ${r + 1}>, ${r}>;
-      ${t()} {
+      ${n()} {
         var x = i32(workgroupId.x) * ${r / 2} + i32(localId.x);
         var y = i32(workgroupId.y) * ${r} + i32(localId.y);
-        let batch = ${n === 3 ? "i32(workgroupId.z)" : "0"};
-        let batchOffsetA = ${n === 3 ? "batch * uniforms.aShapeStrides[0]" : "0"};
-        let batchOffsetOut = ${n === 3 ? "batch * uniforms.outShapeStrides[0]" : "0"};
+        let batch = ${t === 3 ? "i32(workgroupId.z)" : "0"};
+        let batchOffsetA = ${t === 3 ? "batch * uniforms.aShapeStrides[0]" : "0"};
+        let batchOffsetOut = ${t === 3 ? "batch * uniforms.outShapeStrides[0]" : "0"};
 
-        let inputWidth = uniforms.outShape[${n === 3 ? "1" : "0"}] / 2; // Output height
-        let inputHeight = uniforms.outShape[${n === 3 ? "2" : "1"}] * 2; // Output width
+        let inputWidth = uniforms.outShape[${t === 3 ? "1" : "0"}] / 2; // Output height
+        let inputHeight = uniforms.outShape[${t === 3 ? "2" : "1"}] * 2; // Output width
         if (x < inputWidth && y < inputHeight) {
             let unpackedA = unpack2x16float(u32(A[batchOffsetA + y * inputWidth + x]));
             tile[localId.y][localId.x * 2] = unpackedA.x;
@@ -59,8 +59,8 @@ var r = class {
         }
         workgroupBarrier();
 
-        let outputWidth = uniforms.outShape[${n === 3 ? "2" : "1"}]; // Output width
-        let outputHeight = uniforms.outShape[${n === 3 ? "1" : "0"}] * 2; // Output height
+        let outputWidth = uniforms.outShape[${t === 3 ? "2" : "1"}]; // Output width
+        let outputHeight = uniforms.outShape[${t === 3 ? "1" : "0"}] * 2; // Output height
         x = i32(workgroupId.y) * ${r / 2} + i32(localId.x);
         y = i32(workgroupId.x) * ${r} + i32(localId.y);
         if (x < outputWidth && y < outputHeight) {

package/dist/ops/webgpu/unpack16.js

@@ -1,6 +1,6 @@
 import { Ii as e } from "../../dist-Da20xy8E.js";
-import { s as t } from "../../webgpu_program-B0PokuRJ.js";
-import { c as n, i as r } from "../../webgpu_util-DfODg6XI.js";
+import { c as t, i as n } from "../../webgpu_util-DfODg6XI.js";
+import { s as r } from "../../webgpu_program-B0PokuRJ.js";
 //#region lib/ops/webgpu/unpack16.ts
 var i = class {
 	outputShape;
@@ -19,7 +19,7 @@ var i = class {
 	variableComponents = [2];
 	scaling = !1;
 	constructor(e) {
-		this.outputShape = [...e.slice(0, -1), e[e.length - 1] * 2], this.dispatchLayout = n(this.outputShape), this.dispatch = r(this.dispatchLayout, this.outputShape, this.workgroupSize, [
+		this.outputShape = [...e.slice(0, -1), e[e.length - 1] * 2], this.dispatchLayout = t(this.outputShape), this.dispatch = n(this.dispatchLayout, this.outputShape, this.workgroupSize, [
 			4,
 			1,
 			1
@@ -30,7 +30,7 @@ var i = class {
 	}
 	getUserCode() {
 		return `
-                ${t("index")} {
+                ${r("index")} {
                     let outIndex = index;
                     if (outIndex < uniforms.size) {
                         let xvec2 = x[index];

package/dist/ops/webgpu/utils/binary_op.js

@@ -1,6 +1,6 @@
 import { In as e, Js as t } from "../../../dist-Da20xy8E.js";
-import { s as n } from "../../../webgpu_program-B0PokuRJ.js";
-import { c as r, i } from "../../../webgpu_util-DfODg6XI.js";
+import { c as n, i as r } from "../../../webgpu_util-DfODg6XI.js";
+import { s as i } from "../../../webgpu_program-B0PokuRJ.js";
 import { n as a, t as o } from "../../../binary_op_util-CrYk9LXL.js";
 //#region lib/ops/webgpu/utils/binary_op.ts
 var s = class {
@@ -14,17 +14,17 @@ var s = class {
 	variableNames = ["A", "B"];
 	workgroupSize;
 	variableComponents;
-	constructor(n, a, o) {
-		this.outputShape = e(a, o), this.dispatchLayout = r(this.outputShape), this.op = n;
+	constructor(i, a, o) {
+		this.outputShape = e(a, o), this.dispatchLayout = n(this.outputShape), this.op = i;
 		let s = a.length > 0 && a[a.length - 1] % 4 == 0, c = o.length > 0 && o[o.length - 1] % 4 == 0;
 		if (s && c) this.outputComponent = 4, this.variableComponents = [4, 4];
 		else if (s && (t(o) || o[o.length - 1] === 1) || c && (t(a) || a[a.length - 1] === 1)) throw Error("Cannot broadcast 16-bit float binary ops with mixed vector sizes");
 		else throw Error("16-bit float binary ops require inner dimension to be multiple of 4");
-		this.shaderKey = `binary_${n}_${this.variableComponents}`, this.workgroupSize = [
+		this.shaderKey = `binary_${i}_${this.variableComponents}`, this.workgroupSize = [
 			128,
 			1,
 			1
-		], this.dispatch = i(this.dispatchLayout, this.outputShape, this.workgroupSize, [
+		], this.dispatch = r(this.dispatchLayout, this.outputShape, this.workgroupSize, [
 			this.outputComponent,
 			1,
 			1
@@ -38,7 +38,7 @@ var s = class {
       ${a(this.op, this.outputComponent === 4)}
     };
     `}
-       ${n("index")} {
+       ${i("index")} {
          if (index < uniforms.size) {
             let a = A[index];
             let b = B[index];
@@ -87,13 +87,13 @@ var s = class {
 	workgroupSize;
 	variableComponents;
 	constructor(e, t) {
-		if (this.outputShape = t, this.dispatchLayout = r(this.outputShape), this.op = e, t.length > 0 && t[t.length - 1] % 4 == 0) this.outputComponent = 4, this.variableComponents = [4, 1];
+		if (this.outputShape = t, this.dispatchLayout = n(this.outputShape), this.op = e, t.length > 0 && t[t.length - 1] % 4 == 0) this.outputComponent = 4, this.variableComponents = [4, 1];
 		else throw Error("16-bit float binary ops require inner dimension to be multiple of 4");
 		this.shaderKey = `binary_scal_${e}_${this.variableComponents}`, this.workgroupSize = [
 			128,
 			1,
 			1
-		], this.dispatch = i(this.dispatchLayout, this.outputShape, this.workgroupSize, [
+		], this.dispatch = r(this.dispatchLayout, this.outputShape, this.workgroupSize, [
 			this.outputComponent,
 			1,
 			1
@@ -107,7 +107,7 @@ var s = class {
       ${a(this.op, this.outputComponent === 4)}
     };
     `}
-       ${n("index")} {
+       ${i("index")} {
          if (index < uniforms.size) {
             let a = A[index];
             let b = B[0];

package/dist/ops/webgpu/utils/reductions.js

@@ -1,7 +1,7 @@
 import { ii as e, kt as t, nc as n, oc as r, wn as i } from "../../../dist-Da20xy8E.js";
 import { reshape16 as a } from "../../reshape16.js";
-import { s as o } from "../../../webgpu_program-B0PokuRJ.js";
-import { c as s } from "../../../webgpu_util-DfODg6XI.js";
+import { c as o } from "../../../webgpu_util-DfODg6XI.js";
+import { s } from "../../../webgpu_program-B0PokuRJ.js";
 //#region lib/ops/webgpu/utils/reductions.ts
 function c(e, t, n, r) {
 	return e && !t ? "\n            bestValue = subgroupAdd(bestValue);\n        " : e ? `
@@ -50,7 +50,7 @@ function l(e) {
            ${n}
            ${e.utilityFunctions ?? ""}
     
-           ${o("index")} {
+           ${s("index")} {
                 let outputIndex = index / ${t};
                 let offset = outputIndex * uniforms.reduceSize;
                 var bestValue = 0.0f;
@@ -92,7 +92,7 @@ function u(e) {
             ${n}
             ${e.utilityFunctions ?? ""}
     
-           ${o("index")} {
+           ${s("index")} {
                 let outputIndex = index / ${t};
                 let offset1 = outputIndex * 2 * uniforms.reduceSize;
                 let offset2 = offset1 + uniforms.reduceSize;
@@ -139,7 +139,7 @@ function f(e) {
            ${n}
               ${e.utilityFunctions ?? ""}
     
-           ${o("index")} {
+           ${s("index")} {
                 let outputIndex = index / ${e.workgroupSizeX};
                 let offset = outputIndex * uniforms.reduceSize;
                 var bestValue = 0.0f;
@@ -205,7 +205,7 @@ var m = class {
 			Math.min(i, e.subgroupMaxSize),
 			1,
 			1
-		], this.subgroups = !0, e.variableSubgroups && (this.subgroupBuiltins = !0)) : this.workgroupSize[0] = i, this.outputShape = n.elementwise ? [t.batchSize, t.inSize] : r ? [t.outSize / 2] : [t.outSize], this.dispatchLayout = s(this.outputShape), this.dispatch = [
+		], this.subgroups = !0, e.variableSubgroups && (this.subgroupBuiltins = !0)) : this.workgroupSize[0] = i, this.outputShape = n.elementwise ? [t.batchSize, t.inSize] : r ? [t.outSize / 2] : [t.outSize], this.dispatchLayout = o(this.outputShape), this.dispatch = [
 			n.elementwise ? t.batchSize : r ? t.batchSize / 2 : t.batchSize,
 			1,
 			1

package/dist/patches/webgpu_backend.js

@@ -1,7 +1,7 @@
 import { Ms as e, Ps as t, Vs as n, oc as r } from "../dist-Da20xy8E.js";
-import { l as i } from "../webgpu_program-B0PokuRJ.js";
-import { compileProgram as a } from "./webgpu_program.js";
-import { t as o } from "../webgpu-B19Tw8Nd.js";
+import { t as i } from "../webgpu-IAn9VotJ.js";
+import { l as a } from "../webgpu_program-B0PokuRJ.js";
+import { compileProgram as o } from "./webgpu_program.js";
 //#region lib/patches/webgpu_backend.ts
 var s = (e, n) => {
 	let r = e.limits.maxComputeWorkgroupsPerDimension, i = n.dispatchLayout, a = n.dispatch;
@@ -17,16 +17,16 @@ var s = (e, n) => {
 		o,
 		1
 	];
-}, c = class extends o {
+}, c = class extends i {
 	subgroupMaxSize;
 	subgroupMinSize;
 	constructor(e, t) {
 		super(e, t), this.subgroupMaxSize = t?.subgroupMaxSize ?? 0, this.subgroupMinSize = t?.subgroupMinSize ?? 0;
 	}
-	runWebGPUProgram(t, o, c, l, u) {
+	runWebGPUProgram(t, i, c, l, u) {
 		if (u ||= this.makeTensorInfo(t.outputShape, c), r(u.shape) === 0) return this.tensorMap.get(u.dataId).values = n(u.dtype, 0), u;
 		this.uploadToGPU(u.dataId), t.dispatch = s(this.device, t);
-		let d = o.map((e, n) => {
+		let d = i.map((e, n) => {
 			if (e.dtype === "complex64") throw Error("GPGPUProgram does not support complex64 input. For complex64 dtypes, please separate the program into real and imaginary parts.");
 			return this.uploadToGPU(e.dataId), {
 				dtype: this.tensorMap.get(e.dataId).dtype,
@@ -34,9 +34,9 @@ var s = (e, n) => {
 				name: t.variableNames[n]
 			};
 		});
-		t.shaderKey = i(t, d, u);
+		t.shaderKey = a(t, d, u);
 		let f = e().getBool("WEBGPU_ENGINE_COMPILE_ONLY");
-		return t.shaderKey in this.pipelineCache || (this.pipelineCache[t.shaderKey] = a(this.device, t, d, u, f)), t.pipeline = this.pipelineCache[t.shaderKey], f || this.recordAndSubmit(t, u, o, l), u;
+		return t.shaderKey in this.pipelineCache || (this.pipelineCache[t.shaderKey] = o(this.device, t, d, u, f)), t.pipeline = this.pipelineCache[t.shaderKey], f || this.recordAndSubmit(t, u, i, l), u;
 	}
 };
 //#endregion

package/dist/patches/webgpu_base.d.ts

@@ -18,4 +18,5 @@ export interface GPUOptions {
     powerPreference?: 'low-power' | 'high-performance';
     disableSubgroups?: boolean;
 }
+export declare function getBackendDevice(): GPUDevice | undefined;
 export declare function registerWebGPUBackend(options?: GPUOptions): void;

package/dist/patches/webgpu_base.js

@@ -1,22 +1,28 @@
 import { li as e } from "../dist-Da20xy8E.js";
 import t from "./webgpu_backend.js";
 //#region lib/patches/webgpu_base.ts
-function n(n) {
+var n;
+function r() {
+	return n;
+}
+function i(r) {
 	e("webgpu", async () => {
-		let e = { powerPreference: n?.powerPreference ?? "high-performance" };
+		let e = { powerPreference: r?.powerPreference ?? "high-performance" };
 		console.log("Using custom WebGPU backend with power preference:", e.powerPreference);
-		let r = await navigator.gpu.requestAdapter(e), i = {}, a = [];
-		r.features.has("timestamp-query") && a.push("timestamp-query"), r.features.has("bgra8unorm-storage") && a.push(["bgra8unorm-storage"]), !n?.disableSubgroups && r.features.has("subgroups") && a.push("subgroups"), i.requiredFeatures = a;
-		let o = r.limits;
-		return i.requiredLimits = {
-			maxComputeWorkgroupStorageSize: o.maxComputeWorkgroupStorageSize,
-			maxComputeWorkgroupsPerDimension: o.maxComputeWorkgroupsPerDimension,
-			maxStorageBufferBindingSize: o.maxStorageBufferBindingSize,
-			maxBufferSize: o.maxBufferSize,
-			maxComputeWorkgroupSizeX: o.maxComputeWorkgroupSizeX,
-			maxComputeInvocationsPerWorkgroup: o.maxComputeInvocationsPerWorkgroup
-		}, new t(await r.requestDevice(i), "info" in r ? r.info : "requestAdapterInfo" in r ? await r.requestAdapterInfo() : void 0);
+		let i = await navigator.gpu.requestAdapter(e), a = {}, o = [];
+		i.features.has("timestamp-query") && o.push("timestamp-query"), i.features.has("bgra8unorm-storage") && o.push(["bgra8unorm-storage"]), !r?.disableSubgroups && i.features.has("subgroups") && o.push("subgroups"), a.requiredFeatures = o;
+		let s = i.limits;
+		a.requiredLimits = {
+			maxComputeWorkgroupStorageSize: s.maxComputeWorkgroupStorageSize,
+			maxComputeWorkgroupsPerDimension: s.maxComputeWorkgroupsPerDimension,
+			maxStorageBufferBindingSize: s.maxStorageBufferBindingSize,
+			maxBufferSize: s.maxBufferSize,
+			maxComputeWorkgroupSizeX: s.maxComputeWorkgroupSizeX,
+			maxComputeInvocationsPerWorkgroup: s.maxComputeInvocationsPerWorkgroup
+		};
+		let c = await i.requestDevice(a), l = "info" in i ? i.info : "requestAdapterInfo" in i ? await i.requestAdapterInfo() : void 0;
+		return n = c, new t(c, l);
 	}, 3);
 }
 //#endregion
-export { n as registerWebGPUBackend };
+export { r as getBackendDevice, i as registerWebGPUBackend };

package/dist/training/BasicTrainer.js

@@ -7,7 +7,7 @@ import { createTensorStatistics as d } from "../checks/weights.js";
 import f from "./configure.js";
 //#region lib/training/BasicTrainer.ts
 var p = {
-	logInterval: 1,
+	logInterval: 200,
 	maxEpochs: 100,
 	method: { type: "pretraining" },
 	batchSize: 32
@@ -212,7 +212,7 @@ var p = {
 		this.log.push(p), r && r(p), a.logStartTime = Date.now();
 	}
 	async trainOnDataset(e, t, n, i) {
-		let { logInterval: a = 10, maxEpochs: o = Infinity } = {
+		let { logInterval: a = 40, maxEpochs: o = Infinity } = {
 			...p,
 			...t
 		};
@@ -220,33 +220,39 @@ var p = {
 		let c = o * (t?.epochSteps || 1e3);
 		t.metrics && this.setMetrics(t.metrics);
 		let l = this.createEmptyState();
-		this.lastState = l, await this.dummyPass(), t?.metrics?.includes("memoryUsage") && (this.model.getProfiler() || this.model.setProfiler(new s()));
+		if (this.lastState = l, l.step >= c) return {
+			losses: l.losses,
+			validationLosses: l.validationLosses
+		};
+		await this.dummyPass(), t?.metrics?.includes("memoryUsage") && (this.model.getProfiler() || this.model.setProfiler(new s()));
 		let d = Date.now();
 		this.running = !0, l.logStartTime = d;
-		let f = n ? new u(this.model, n, this.maskedLoss) : void 0, m = await e.iterator(), h = m.next();
+		let f = d, m = n ? new u(this.model, n, this.maskedLoss) : void 0, h = await e.iterator(), g = h.next();
 		try {
 			for (; this.running;) {
-				let e = await h;
-				if (h = m.next(), e.done) break;
-				let n = e.value, r = l.step % a === 0, o = (t?.metrics?.includes("gradientStatistics") || !1) && r, s = this.trainStep(l, n, !1, o);
+				let e = await g;
+				if (g = h.next(), e.done) break;
+				let n = e.value, r = Date.now(), o = r - f >= a;
+				o && (f = r);
+				let s = (t?.metrics?.includes("gradientStatistics") || !1) && o, u = this.trainStep(l, n, !1, s);
 				if (t.debug) {
-					let e = (await s.data())[0];
+					let e = (await u.data())[0];
 					if (isNaN(e) || !isFinite(e)) throw console.error("Invalid loss value:", e), console.error("Batch xs:", await n.xs.array()), console.error("Batch ys:", await n.ys.array()), console.error("State:", l), Error("Loss is NaN or Infinity");
 					console.log(`Step ${l.step}: Loss = ${e}`);
 				}
 				n.xs.dispose(), n.ys.dispose(), l.step++, l.totalSteps++;
-				let u = l.step >= c;
-				r || u ? await this.performLogging(s, n.xs.shape[0], f, i) : (l.gradientNorm &&= (l.gradientNorm.dispose(), void 0), l.accuracy &&= (l.accuracy.dispose(), void 0)), s.dispose(), u && this.stop();
+				let d = l.step >= c;
+				o || d ? await this.performLogging(u, n.xs.shape[0], m, i) : (l.gradientNorm &&= (l.gradientNorm.dispose(), void 0), l.accuracy &&= (l.accuracy.dispose(), void 0)), u.dispose(), d && this.stop();
 			}
 		} catch (e) {
 			throw console.error("Training error:", e), r(), e;
 		}
 		r(), this.running = !1, this.model.metaData.actionLog = this.model.metaData.actionLog || [];
-		let g = Date.now();
+		let _ = Date.now();
 		return this.model.metaData.actionLog.push({
 			action: "pretrain",
-			timestamp: g,
-			duration: g - d,
+			timestamp: _,
+			duration: _ - d,
 			tokensProcessed: this.tokensProcessed,
 			options: t
 		}), {

package/dist/{webgpu-B19Tw8Nd.js → webgpu-IAn9VotJ.js}

@@ -1,7 +1,7 @@
 import { Ai as e, Ms as t, Oi as n, Ps as r, Rs as i, Vs as a, Xr as o, Ys as s, fc as c, ii as l, ki as u, lc as d, oc as f, pc as p, zs as m } from "./dist-Da20xy8E.js";
 import { s as h } from "./complex_util-CkazZsaH.js";
-import { l as g, n as _, t as v } from "./webgpu_program-B0PokuRJ.js";
-import { l as y, t as b } from "./webgpu_util-DfODg6XI.js";
+import { l as g, t as _ } from "./webgpu_util-DfODg6XI.js";
+import { l as v, n as y, t as b } from "./webgpu_program-B0PokuRJ.js";
 //#region node_modules/@tensorflow/tfjs-backend-webgpu/dist/flags_webgpu.js
 var x = t();
 x.registerFlag("WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE", () => 15), x.registerFlag("WEBGPU_CPU_FORWARD", () => !0), x.registerFlag("WEBGPU_MATMUL_PROGRAM_TYPE", () => -1), x.registerFlag("WEBGPU_USE_NAIVE_CONV2D_TRANSPOSE", () => !0), x.registerFlag("WEBGPU_USE_LOW_POWER_GPU", () => !1), x.registerFlag("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD", () => 1e3), x.registerFlag("WEBGPU_USE_PROFILE_TOOL", () => !1), x.registerFlag("WEBGPU_IMPORT_EXTERNAL_TEXTURE", () => !0), x.registerFlag("WEBGPU_USE_NAIVE_CONV2D_DEBUG", () => !1), x.registerFlag("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL", () => -1), x.registerFlag("WEBGPU_CONV_SEPARATE_IM2COL_SHADER", () => !1), x.registerFlag("WEBGPU_PRINT_SHADER", () => ""), x.registerFlag("WEBGPU_ENGINE_COMPILE_ONLY", () => !1);
@@ -141,7 +141,7 @@ var O = t().getNumber("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD"), k = (e, t) => {
 		return x.nextDataId++;
 	}
 	constructor(e, n) {
-		if (super(), this.commandQueueOwnedIds = /* @__PURE__ */ new WeakSet(), this.dispatchCountInPass = 0, this.disposed = !1, this.downloadWaitMs = 0, this.tensorDataPendingDisposal = [], this.queryResolveBuffer = null, this.querySet = null, this.querySetCount = 2, this.stagingPendingDisposal = [], this.uniformPendingDisposal = [], this.uploadWaitMs = 0, this.hasReadSyncWarned = !1, this.hasTimestampQueryWarned = !1, !y()) throw Error("WebGPU is not supported on this device");
+		if (super(), this.commandQueueOwnedIds = /* @__PURE__ */ new WeakSet(), this.dispatchCountInPass = 0, this.disposed = !1, this.downloadWaitMs = 0, this.tensorDataPendingDisposal = [], this.queryResolveBuffer = null, this.querySet = null, this.querySetCount = 2, this.stagingPendingDisposal = [], this.uniformPendingDisposal = [], this.uploadWaitMs = 0, this.hasReadSyncWarned = !1, this.hasTimestampQueryWarned = !1, !g()) throw Error("WebGPU is not supported on this device");
 		this.pipelineCache = {}, this.device = e, this.queue = e.queue, this.commandEncoder = null, this.computePassEncoder = null, this.adapterInfo = new S(n), this.supportTimestampQuery = this.device.features.has("timestamp-query"), this.thresholdToIncreaseWorkgroups = this.adapterInfo.intelGPUGeneration >= 12 ? 16 : 8, this.bufferManager = new C(this.device), this.textureManager = new T(this.device), this.tensorMap = new c(this, l()), t().getBool("WEBGPU_USE_PROFILE_TOOL") && (this.dummyCanvas = document.createElement("canvas"), this.dummyCanvas.width = 1, this.dummyCanvas.height = 1, this.dummyContext = this.dummyCanvas.getContext("webgpu"), this.dummyContext.configure({
 			device: e,
 			format: "bgra8unorm"
@@ -308,7 +308,7 @@ var O = t().getNumber("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD"), k = (e, t) => {
 			refCount: 1,
 			external: e.zeroCopy
 		});
-		let a = this.tensorMap.get(i), o = b(a.dtype) * f(a.shape);
+		let a = this.tensorMap.get(i), o = _(a.dtype) * f(a.shape);
 		if (e.buffer.size < o) throw Error(`GPUBuffer size(${e.buffer.size}) is smaller than tensor size(${o})!`);
 		if ((e.buffer.usage & (GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC)) !== (GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC)) throw Error("GPUBuffer.usage should include GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC!");
 		return e.zeroCopy !== !0 && (r = this.copyBuffer(r)), a.resource = r, l().makeTensorFromDataId(i, t, n, this);
@@ -367,7 +367,7 @@ var O = t().getNumber("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD"), k = (e, t) => {
 	uploadToGPU(e) {
 		let t = this.tensorMap.get(e);
 		if (t.resource != null) return;
-		let n = b(t.dtype) * f(t.shape), r, i = GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC | GPUBufferUsage.COPY_DST;
+		let n = _(t.dtype) * f(t.shape), r, i = GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC | GPUBufferUsage.COPY_DST;
 		if (t.values) {
 			if (r = this.bufferManager.acquireBuffer(n, i, !0), r.mapState === "unmapped") {
 				let e = this.bufferManager.acquireBuffer(n, GPUBufferUsage.MAP_WRITE | GPUBufferUsage.COPY_SRC, !0, !1), i = e.getMappedRange();
@@ -431,9 +431,9 @@ var O = t().getNumber("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD"), k = (e, t) => {
 				name: e.variableNames[n]
 			};
 		});
-		e.shaderKey = g(e, s, o);
+		e.shaderKey = v(e, s, o);
 		let c = t().getBool("WEBGPU_ENGINE_COMPILE_ONLY");
-		return e.shaderKey in this.pipelineCache || (this.pipelineCache[e.shaderKey] = _(this.device, e, s, o, c)), e.pipeline = this.pipelineCache[e.shaderKey], c || this.recordAndSubmit(e, o, n, i), o;
+		return e.shaderKey in this.pipelineCache || (this.pipelineCache[e.shaderKey] = y(this.device, e, s, o, c)), e.pipeline = this.pipelineCache[e.shaderKey], c || this.recordAndSubmit(e, o, n, i), o;
 	}
 	recordAndSubmit(e, n, r, a) {
 		if (e.pipeline instanceof Promise) throw Error("Please call checkCompileCompletionAsync to ensure parallel compilation is done!");
@@ -497,7 +497,7 @@ var O = t().getNumber("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD"), k = (e, t) => {
 			querySet: this.querySet,
 			beginningOfPassWriteIndex: 0,
 			endOfPassWriteIndex: 1
-		}, this.computePassEncoder = this.commandEncoder.beginComputePass(p)) : this.computePassEncoder ||= this.commandEncoder.beginComputePass(p), this.computePassEncoder.setPipeline(e.pipeline), this.computePassEncoder.setBindGroup(0, u), this.computePassEncoder.dispatchWorkgroups(e.dispatch[0], e.dispatch[1], e.dispatch[2]), this.dispatchCountInPass++, (d || t().get("WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE") <= this.dispatchCountInPass || e.pixelsOpType === v.DRAW) && (this.endComputePassEncoder(), d ? this.activeTimers.push({
+		}, this.computePassEncoder = this.commandEncoder.beginComputePass(p)) : this.computePassEncoder ||= this.commandEncoder.beginComputePass(p), this.computePassEncoder.setPipeline(e.pipeline), this.computePassEncoder.setBindGroup(0, u), this.computePassEncoder.dispatchWorkgroups(e.dispatch[0], e.dispatch[1], e.dispatch[2]), this.dispatchCountInPass++, (d || t().get("WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE") <= this.dispatchCountInPass || e.pixelsOpType === b.DRAW) && (this.endComputePassEncoder(), d ? this.activeTimers.push({
 			name: e.constructor.name,
 			query: this.getQueryTime()
 		}) : this.submitQueue());

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@genai-fi/nanogpt",
-    "version": "1.1.1",
+    "version": "1.1.3",
     "type": "module",
     "main": "dist/main.js",
     "types": "dist/main.d.ts",