@genai-fi/nanogpt 1.2.5 → 1.4.0

package/dist/api/responses.js

@@ -29,13 +29,10 @@ var r = class {
 	}
 	async _runForRecord(e, t, n) {
 		let r = t.previous_response_id ? this._responses.get(t.previous_response_id) : void 0;
-		if (r) {
-			let e = r.generator.getConversation();
-			t.input && Array.isArray(t.input) ? t.input = [...e, ...t.input] : t.input && typeof t.input == "string" ? t.input = [...e, {
-				role: t.nonConversational ? "text" : "user",
-				content: t.input
-			}] : t.input = e;
-		}
+		r && (t.input && Array.isArray(t.input) || t.input && typeof t.input == "string" && (t.input = [{
+			role: t.nonConversational ? "text" : "user",
+			content: t.input
+		}]));
 		let i = r ? r.generator : this._responses.get(e).generator, a = n ? () => n({
 			output: i.getConversation(),
 			id: e,
@@ -63,11 +60,17 @@ var r = class {
 				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();
+				o.done = !0, t.done = !0, t.output = n, this._hookedResponses.delete(e), this._resumeWaiters.delete(e);
+				let r = n[0]?._id || e;
+				n.forEach((e) => {
+					e._id === void 0 && (e._id = r);
+				}), 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(), {
+		let l = await c, u = l[0]?._id || e;
+		return l.forEach((e) => {
+			e._id === void 0 && (e._id = u);
+		}), 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,
 			id: e,
 			done: !0

package/dist/api/training.js

@@ -108,7 +108,7 @@ var o = class {
 		return s.trainer.log = t, s.trainer.resumeFromLog(t[t.length - 1]), this._jobs.set(s.id, s), s;
 	}
 	launchJob(e) {
-		e.trainDataset && (this.setState(e, "running"), e.trainer.trainOnDataset(e.trainDataset, e.options, e.validationDataset, (t) => {
+		e.trainDataset && (this.setState(e, "running"), e.trainer.prepareToTrain(e.options), e.trainer.trainOnDataset(e.trainDataset, e.options, e.validationDataset, (t) => {
 			e.history = e.trainer.log, e.progress = t.totalTokens / e.totalTokens, e.remaining = Math.max(0, (e.totalTokens - t.totalTokens) / t.totalTokens * t.duration), e.breakOnLog.size > 0 && e.state === "running" && (this.setState(e, "pausing"), e.trainer.stop()), this.ee.emit("progress", e);
 		}).then(() => {
 			e.state === "running" ? this.setState(e, "completed") : e.state === "pausing" ? this.setState(e, "paused") : e.state === "cancelling" && this.setState(e, "cancelled");

package/dist/inference/Generator.js

@@ -187,7 +187,9 @@ var y = class extends e {
 		this.reset();
 	}
 	initialise(e, n) {
-		if (this.cache && n?.noCache && this.reset(), this.initialPrompt = e || null, this.lastToken === -1 ? this.outputConversation = (this.initialPrompt || []).slice() : e && e.length > this.outputConversation.length && (this.outputConversation = (this.initialPrompt || []).slice(), this.resetCache()), !this.cache && !n?.noCache && (this.model.config.modelType !== "GenAI_NanoGPT_v1" || this.model.config.useRope)) {
+		if (this.cache && n?.noCache && this.reset(), this.initialPrompt = e || null, this.lastToken === -1 && this.outputConversation.length === 0 ? this.outputConversation = this.initialPrompt || [] : e && e.length > 0 && (e.forEach((e) => {
+			this.outputConversation.push(e);
+		}), this.resetCache()), !this.cache && !n?.noCache && (this.model.config.modelType !== "GenAI_NanoGPT_v1" || this.model.config.useRope)) {
 			let e = Array(this.model.config.nLayer);
 			for (let t = 0; t < this.model.config.nLayer; t++) e[t] = {
 				k: void 0,

package/dist/inference/tokenisePrompt.js

@@ -3,7 +3,7 @@ import { I as e } from "../dist-Da20xy8E.js";
 async function t(t, n, r, i) {
 	if (r) {
 		let a = r.length > 0 && (r[r.length - 1].role === "text" || r[r.length - 1].role === "assistant"), o;
-		return i?.nonConversational ? o = a && i?.continuation ? [t.bosToken, ...t.encode(r[r.length - 1].content)] : t.encodeAsSequence(r, !0) : a && i?.continuation ? (o = t.encodeConversation(r, !0), o[o.length - 1] === t.getSpecialTokenIndex("<|assistant_start|>") && o[o.length - 2] === t.getSpecialTokenIndex("<|assistant_end|>") && (o.pop(), o.pop())) : o = t.encodeConversation(r, !0), o.length > n && (o = o.slice(-n)), e([o], [1, o.length], "int32");
+		return i?.nonConversational ? o = a && i?.continuation ? [t.bosToken, ...t.encode(r[r.length - 1].content)] : t.encodeAsSequence(r, !0) : a && i?.continuation ? (o = t.encodeConversation(r, !0), o[o.length - 1] === t.getSpecialTokenIndex("<|assistant_start|>") && o[o.length - 2] === t.getSpecialTokenIndex("<|assistant_end|>") && (o.pop(), o.pop())) : o = t.encodeConversation(r, !0), o.length > n && (o = o.slice(-n)), i?.debug && console.log("Tokenised prompt:", t.decode(o)), e([o], [1, o.length], "int32");
 	} else {
 		let n = i?.nonConversational ? void 0 : t.getSpecialTokenIndex("<|assistant_start|>"), r = n === void 0 ? [t.bosToken] : [t.bosToken, n];
 		return e([r], [1, r.length], "int32");

package/dist/inference/types.d.ts

@@ -17,6 +17,7 @@ export interface IGeneratorOutputInternal extends IGeneratorOutput {
     outputTensor: Tensor;
 }
 export interface GeneratorConversation extends Conversation {
+    _id?: string;
     _completed?: boolean;
     _timestamp?: number;
     _output?: IGeneratorOutput[];
@@ -45,6 +46,7 @@ export interface IGenerateOptions {
     continuation?: boolean; /** Continue the previous response without adding a new turn. */
     chunkSize?: number; /** Number of tokens to generate before returning a partial response. */
     background?: boolean; /** Run the generation in the background and return immediately. */
+    debug?: boolean;
     _onChunk?: (output: IGeneratorOutput) => void | Promise<void>;
 }
 export interface IGeneratorResponse {

package/dist/training/BasicTrainer.d.ts

@@ -45,6 +45,13 @@ export default class BasicTrainer {
     dispose(): void;
     private createEmptyState;
     private performLogging;
+    /** A sync function to ensure the trainer is prepared and the state is initialized before training.
+     *  It should be called before invoking trainOnDataset.
+     */
+    prepareToTrain(options: Partial<TrainingOptions>): {
+        losses: number[];
+        validationLosses: number[];
+    } | undefined;
     trainOnDataset(dataset: Dataset<{
         xs: Tensor;
         ys: Tensor;

package/dist/training/BasicTrainer.js

@@ -211,22 +211,31 @@ var p = {
 		}
 		this.log.push(p), r && r(p), a.logStartTime = Date.now();
 	}
-	async trainOnDataset(e, t, n, i) {
-		let { logInterval: a = 40, maxEpochs: o = Infinity } = {
+	prepareToTrain(e) {
+		let { maxEpochs: t = Infinity } = {
 			...p,
-			...t
+			...e
 		};
 		this.log.length > 0 && this.resumeFromLog(this.log[this.log.length - 1]);
-		let c = o * (t?.epochSteps || 1e3);
-		t.metrics && this.setMetrics(t.metrics);
-		let l = this.createEmptyState();
-		if (this.lastState = l, l.step >= c) return {
-			losses: l.losses,
-			validationLosses: l.validationLosses
+		let n = t * (e?.epochSteps || 1e3);
+		e.metrics && this.setMetrics(e.metrics);
+		let r = this.createEmptyState();
+		if (this.lastState = r, r.step >= n) return {
+			losses: r.losses,
+			validationLosses: r.validationLosses
 		};
+		this.running = !0;
+	}
+	async trainOnDataset(e, t, n, i) {
+		let { logInterval: a = 40, maxEpochs: o = Infinity } = {
+			...p,
+			...t
+		}, c = o * (t?.epochSteps || 1e3);
 		await this.dummyPass(), t?.metrics?.includes("memoryUsage") && (this.model.getProfiler() || this.model.setProfiler(new s()));
+		let l = this.lastState;
+		if (!l) throw Error("Trainer state has not been initialized. Call prepareToTrain() first.");
 		let d = Date.now();
-		this.running = !0, l.logStartTime = d;
+		l.logStartTime = d;
 		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;) {

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@genai-fi/nanogpt",
-    "version": "1.2.5",
+    "version": "1.4.0",
     "type": "module",
     "main": "dist/main.js",
     "types": "dist/main.d.ts",