@genai-fi/nanogpt 1.1.2 → 1.1.4

package/dist/api/responses.js

@@ -56,13 +56,16 @@ var r = class {
 			role: t.nonConversational ? "text" : "user",
 			content: t.input
 		}], s) : i.generate(s);
-		if (t.background) return c.then(() => {
-			o.done = !0, this._hookedResponses.delete(e), this._resumeWaiters.delete(e), a && i.off("tokens", a), this.ee.emit("done", e), this._processNextJob();
-		}), {
-			output: null,
-			id: e,
-			done: !1
-		};
+		if (t.background) {
+			let t = {
+				output: null,
+				id: e,
+				done: !1
+			};
+			return c.then((n) => {
+				o.done = !0, t.done = !0, t.output = n, this._hookedResponses.delete(e), this._resumeWaiters.delete(e), a && i.off("tokens", a), this.ee.emit("done", e), this._processNextJob();
+			}), t;
+		}
 		let l = await c;
 		return a && i.off("tokens", a), o.done = !0, this._hookedResponses.delete(e), this._resumeWaiters.delete(e), this.ee.emit("done", e), this._processNextJob(), {
 			output: l,

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
 		};
@@ -227,30 +227,32 @@ var p = {
 		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/training/prepareData.js

@@ -5,17 +5,16 @@ import { createTrainValidationDatasets as r, storeFromArray as i } from "./valid
 //#region lib/training/prepareData.ts
 async function a(a, o, s, c, l, u, d) {
 	let f = a.loraName || a.loraConfig;
-	if (d && f) throw Error("Cannot specify datasets when using LoRA fine-tuning");
 	if (!d && !f) throw Error("Must specify datasets for non-LoRA training");
 	if (d) {
 		let e = !1;
 		for (let t of d) t.conversational && (e = !0);
-		o.metaData.pretrainingData = d.map((e) => ({
+		f || (o.metaData.pretrainingData = d.map((e) => ({
 			id: e.id,
 			name: e.name,
 			conversational: e.conversational,
 			url: e.url
-		})), e ? o.metaData.mode = "conversational" : o.metaData.mode !== "conversational" && (o.metaData.mode = "completion");
+		}))), e ? o.metaData.mode = "conversational" : o.metaData.mode !== "conversational" && (o.metaData.mode = "completion");
 	} else o.metaData.mode !== "conversational" && (o.metaData.mode = "completion");
 	let p = a.maskedLoss ?? a.method.type === "supervised", m, h = u;
 	if (Array.isArray(c)) if (c[0] instanceof Uint16Array) m = c;
@@ -27,6 +26,7 @@ async function a(a, o, s, c, l, u, d) {
 		m = e.trainingTokens, u || (h = e.validationTokens);
 	}
 	else m = c;
+	console.log("Training tokens", m);
 	let g = m instanceof e ? m.getTokenCount() : m.reduce((e, t) => e + t.length, 0);
 	a.epochSteps = Math.max(1, Math.floor(g / ((a?.batchSize || 32) * o.config.blockSize)));
 	let _ = new n(s, o.config.blockSize);

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@genai-fi/nanogpt",
-    "version": "1.1.2",
+    "version": "1.1.4",
     "type": "module",
     "main": "dist/main.js",
     "types": "dist/main.d.ts",