@genai-fi/nanogpt 1.2.2 → 1.2.4

package/dist/inference/Beamer.js

@@ -79,8 +79,9 @@ var f = class {
 			text: "",
 			terminated: !1,
 			context: a.context,
-			contextLength: a.contextLength
-		}], s = t.maxBeamLength, c = t.endOnWhiteSpace === !0 ? Math.max(s, t.maxLength ?? s + this.model.config.blockSize) : s, l = Date.now();
+			contextLength: a.contextLength,
+			candidates: 0
+		}], s = t.maxBeamLength, c = t.endOnWhiteSpace === !0 ? Math.max(s, t.maxLength ?? s + this.model.config.blockSize) : s, l = Date.now(), u = 0;
 		for (let e = 0; e < c && this.active; e++) {
 			let e = o.filter((e) => !e.terminated);
 			if (e.length === 0) break;
@@ -100,33 +101,49 @@ var f = class {
 					});
 				}
 			}
-			let c = o.filter((e) => e.terminated).slice();
+			let c = o.filter((e) => e.terminated), f = c.slice();
 			a.sort((e, t) => t.score - e.score);
 			for (let e of a) {
-				if (c.length >= r) break;
+				if (f.length >= r) break;
 				let t = e.terminatedByToken || e.terminatedByWhitespace, n = !e.dropLeadingWhitespaceTerminator, i = n ? e.parent.tokens.concat(e.token) : e.parent.tokens.slice(), a = e.terminatedByToken || e.dropLeadingWhitespaceTerminator ? e.parent.text : e.parent.text + e.tokenText, o = n ? this.appendTokenToContext(e.parent.context, e.parent.contextLength, e.token) : {
 					context: e.parent.context.clone(),
 					contextLength: e.parent.contextLength
-				};
-				c.push({
+				}, s = e.parent._output?.slice() ?? [];
+				n && s.push({
+					token: e.token,
+					score: e.score,
+					text: e.tokenText,
+					terminated: t,
+					confidence: null,
+					logits: null,
+					scores: null,
+					hiddenStates: null,
+					attention: null,
+					loss: null,
+					multinomialRand: null
+				}), f.push({
 					tokens: i,
 					score: e.score,
 					text: a,
 					terminated: t,
 					context: o.context,
-					contextLength: o.contextLength
+					contextLength: o.contextLength,
+					_output: s,
+					candidates: u
 				});
 			}
-			if (o.forEach((e) => e.context.dispose()), c.length === 0) break;
-			o = c, n && Date.now() - l >= 40 && (n(o.slice(0, t.beams)), l = Date.now());
+			if (u += a.length - f.length + c.length, o.forEach((e) => e.context.dispose()), f.length === 0) break;
+			o = f, n && Date.now() - l >= 40 && (n(o.slice(0, t.beams)), l = Date.now());
 		}
-		this.active = !1;
-		let u = o.slice().sort((e, t) => t.score - e.score), f = [], p = /* @__PURE__ */ new Set();
-		for (let e of u) {
+		this.active = !1, u += o.length;
+		let f = o.slice().sort((e, t) => t.score - e.score), p = [], m = /* @__PURE__ */ new Set();
+		for (let e of f) {
 			let n = e.text.trim();
-			if (p.has(n) || (p.add(n), f.push(e)), f.length >= t.beams) break;
+			if (m.has(n) || (m.add(n), p.push(e)), p.length >= t.beams) break;
 		}
-		return o.forEach((e) => e.context.dispose()), f;
+		return o.forEach((e) => {
+			e.context.dispose(), e.candidates = u;
+		}), p;
 	}
 };
 //#endregion

package/dist/inference/tokenisePrompt.js

@@ -2,8 +2,8 @@ import { I as e } from "../dist-Da20xy8E.js";
 //#region lib/inference/tokenisePrompt.ts
 async function t(t, n, r, i) {
 	if (r) {
-		let a = r.length > 0 && r[r.length - 1].role === "text", o;
-		return o = i?.nonConversational ? a && i?.continuation ? [t.bosToken, ...t.encode(r[r.length - 1].content)] : t.encodeAsSequence(r, !0) : t.encodeConversation(r, !0), o.length > n && (o = o.slice(-n)), e([o], [1, o.length], "int32");
+		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");
 	} 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

@@ -1,7 +1,6 @@
 import { Conversation } from '../tokeniser/type';
 import { Tensor } from '@tensorflow/tfjs-core';
 export interface IGeneratorOutput {
-    outputTensor: Tensor;
     token: number;
     text: string;
     confidence: number | null;
@@ -14,6 +13,9 @@ export interface IGeneratorOutput {
     multinomialRand: number | null;
     terminated: boolean;
 }
+export interface IGeneratorOutputInternal extends IGeneratorOutput {
+    outputTensor: Tensor;
+}
 export interface GeneratorConversation extends Conversation {
     _completed?: boolean;
     _timestamp?: number;
@@ -59,4 +61,6 @@ export interface IBeam {
     tokens: number[];
     score: number;
     text: string;
+    candidates: number;
+    _output?: IGeneratorOutput[];
 }

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@genai-fi/nanogpt",
-    "version": "1.2.2",
+    "version": "1.2.4",
     "type": "module",
     "main": "dist/main.js",
     "types": "dist/main.d.ts",