@flaxia/node 0.3.7 → 0.5.0

package/dist/assets/ai-inference.js

@@ -1,6 +1,107 @@
 import { TextStreamer as e, env as t, pipeline as n } from "./transformers.web.js";
+//#region src/workloads/ai-model-registry.ts
+var r = 512 * 1024 * 1024, i = 4 * 1024 ** 3, a = new Map(Object.entries({
+	"Xenova/bert-base-multilingual-uncased-sentiment": {
+		repo: "Xenova/bert-base-multilingual-uncased-sentiment",
+		revision: "main",
+		maxBytes: r
+	},
+	"onnx-community/Qwen3-Embedding-0.6B-ONNX": {
+		repo: "onnx-community/Qwen3-Embedding-0.6B-ONNX",
+		revision: "main",
+		maxBytes: r
+	}
+}).map(([e, t]) => [e, { ...t }])), o = [
+	"wasm",
+	"cpu",
+	"webgpu"
+], s = [
+	"q4",
+	"q4f16",
+	"q8",
+	"int8",
+	"uint8"
+], c = {
+	max_new_tokens: {
+		min: 1,
+		max: 1024
+	},
+	temperature: {
+		min: 0,
+		max: 4
+	},
+	top_p: {
+		min: 0,
+		max: 1
+	},
+	top_k: {
+		min: 0,
+		max: 1e3
+	},
+	repetition_penalty: {
+		min: 0,
+		max: 4
+	}
+}, l = 1e5;
+function u() {
+	return [...a.keys()];
+}
+function d(e) {
+	if (typeof e != "string" || !e) throw Error(`ai-inference: model must be a non-empty string, got ${JSON.stringify(e)}`);
+	let t = a.get(e);
+	if (!t) throw Error(`ai-inference: model not allowed: ${e} (allowed: ${u().join(", ") || "none"})`);
+	return t;
+}
+function f(e) {
+	if (e != null) {
+		if (typeof e != "string" || !o.includes(e)) throw Error(`ai-inference: device not allowed: ${JSON.stringify(e)} (allowed: ${o.join(", ")})`);
+		return e;
+	}
+}
+function p(e) {
+	if (e != null) {
+		if (typeof e != "string" || !s.includes(e)) throw Error(`ai-inference: dtype not allowed: ${JSON.stringify(e)} (allowed: ${s.join(", ")})`);
+		return e;
+	}
+}
+function m(e) {
+	if (typeof e == "string") {
+		if (e.length > 1e5) throw Error(`ai-inference: input exceeds ${l} characters`);
+		return e;
+	}
+	if (Array.isArray(e)) {
+		if (e.length === 0 || e.length > 64) throw Error("ai-inference: input array must hold 1..64 items");
+		for (let t of e) {
+			if (typeof t != "string") throw Error("ai-inference: input array items must be strings");
+			if (t.length > 1e5) throw Error(`ai-inference: input exceeds ${l} characters`);
+		}
+		return e;
+	}
+	throw Error("ai-inference: input must be a string or an array of strings");
+}
+function h(e, t) {
+	for (let [n, r] of Object.entries(c)) {
+		let i = t[n];
+		if (i != null) {
+			if (typeof i != "number" || !Number.isFinite(i) || i < r.min || i > r.max) throw Error(`ai-inference: option ${n} must be a number within ${r.min}..${r.max}, got ${JSON.stringify(i)}`);
+			e[n] = i;
+		}
+	}
+	if (t.do_sample !== void 0) {
+		if (typeof t.do_sample != "boolean") throw Error("ai-inference: option do_sample must be a boolean");
+		e.do_sample = t.do_sample;
+	}
+	for (let n of ["src_lang", "tgt_lang"]) {
+		let r = t[n];
+		if (r != null) {
+			if (typeof r != "string" || !/^[a-z]{2,3}(_[A-Za-z0-9]{2,8})?$/.test(r)) throw Error(`ai-inference: option ${n} must be a language code like 'en_XX'`);
+			e[n] = r;
+		}
+	}
+}
+//#endregion
 //#region src/workloads/ai-inference.ts
-var r = [
+var g = [
 	"text-classification",
 	"token-classification",
 	"question-answering",
@@ -26,51 +127,90 @@ var r = [
 	"depth-estimation",
 	"feature-extraction",
 	"image-feature-extraction"
-], i = /* @__PURE__ */ new Map(), a = 2, o = () => {
-	i.clear();
-}, s = (e, t) => {
-	if (!t.tokenBuffer) return e;
-	let n = "", r = null, i = t.tokenBufferIntervalMs ?? 50, a = () => {
-		n &&= (e(n), ""), r = null;
+], _ = /* @__PURE__ */ new Map(), v = 2, y = () => {
+	_.clear();
+}, b = (e, t) => {
+	let n = "", r = null, i = Promise.resolve(), a, o = !1, s = !1, c = (t) => {
+		i = i.then(() => {
+			if (!o) return e(t);
+		}).then(() => void 0, (e) => {
+			o = !0, a = e;
+		});
+	}, l = () => {
+		r !== null && clearTimeout(r), r = null;
+		let e = n;
+		n = "", e && !o && c(e);
+	}, u = (e) => {
+		if (!s) {
+			if (o) throw a;
+			if (!t.tokenBuffer) {
+				c(e);
+				return;
+			}
+			n += e, r === null && (r = setTimeout(l, t.tokenBufferIntervalMs ?? 50));
+		}
+	}, d = async (e) => {
+		if (s = !0, r !== null && clearTimeout(r), r = null, e && !o ? l() : n = "", await i, o) throw a;
 	};
-	return (e) => {
-		n += e, r ||= setTimeout(a, i);
+	return {
+		callback: u,
+		finish: () => d(!0),
+		cancel: () => d(!1)
 	};
 };
-t.backends?.onnx?.wasm && (t.backends.onnx.wasm.numThreads = 1, t.backends.onnx.wasm.proxy = !1);
-var c = async (t, o) => {
-	let { task: c, model: l, input: u, options: d = {} } = t;
-	if (!r.includes(c)) throw Error(`Invalid or unsupported task: ${c}. Supported tasks are: ${r.join(", ")}`);
-	let f = `${c}:${l}`, p = i.get(f);
-	if (!p) {
-		if (i.size >= a) {
-			let e = i.keys().next().value;
-			i.delete(e);
+function x() {
+	let e = globalThis.fetch.bind(globalThis);
+	t.fetch = async (t, n) => {
+		let r = new URL(String(t));
+		if (!/(^|\.)huggingface\.co$/.test(r.hostname) && !/(^|\.)hf\.co$/.test(r.hostname)) throw Error(`ai-inference: model download blocked from ${r.hostname}`);
+		let a = await e(r, n), o = Number(a.headers.get("content-length"));
+		if (Number.isFinite(o) && o > i) throw Error(`ai-inference: model file too large (${o} bytes > ${i})`);
+		return a;
+	};
+}
+x(), t.backends?.onnx?.wasm && (t.backends.onnx.wasm.numThreads = 1, t.backends.onnx.wasm.proxy = !1);
+var S = async (t, r) => {
+	let { task: i, model: a, input: o, options: s = {} } = t;
+	if (!g.includes(i)) throw Error(`Invalid or unsupported task: ${i}. Supported tasks are: ${g.join(", ")}`);
+	let c = d(a), l = f(s.device), u = p(s.dtype), y = m(o), x = {};
+	h(x, s);
+	let S = `${i}:${c.repo}:${u ?? "q4f16"}:${l ?? "wasm"}`, C = _.get(S);
+	if (!C) {
+		if (_.size >= v) {
+			let e = _.keys().next().value;
+			_.delete(e);
 		}
 		let e = {
-			dtype: d.dtype ?? "q4f16",
-			device: d.device
+			dtype: u ?? "q4f16",
+			device: l
 		}, t = performance.now();
-		console.log(`[flaxia-node] ai-inference: loading pipeline task=${c} model=${l} dtype=${e.dtype} device=${e.device ?? "wasm"}`);
+		console.log(`[flaxia-node] ai-inference: loading pipeline task=${i} model=${c.repo} revision=${c.revision} dtype=${e.dtype} device=${l ?? "wasm"}`);
 		try {
-			p = await n(c, l, e);
+			C = await n(i, c.repo, e);
 		} catch (t) {
-			if (d.device && d.device !== "wasm") console.warn(`[AiInference] ${d.device} failed, falling back to wasm:`, t), e.device = "wasm", p = await n(c, l, e);
-			else throw console.error(`[flaxia-node] ai-inference: pipeline load FAILED task=${c} model=${l} error=${t instanceof Error ? t.message : String(t)}`), t;
+			if (l && l !== "wasm") console.warn(`[AiInference] ${l} failed, falling back to wasm:`, t), e.device = "wasm", C = await n(i, c.repo, e);
+			else throw console.error(`[flaxia-node] ai-inference: pipeline load FAILED task=${i} model=${c.repo} error=${t instanceof Error ? t.message : String(t)}`), t;
 		}
-		console.log(`[flaxia-node] ai-inference: pipeline ready task=${c} model=${l} loadMs=${Math.round(performance.now() - t)}`), i.set(f, p);
+		console.log(`[flaxia-node] ai-inference: pipeline ready task=${i} model=${c.repo} loadMs=${Math.round(performance.now() - t)}`), _.set(S, C);
 	}
-	let m = {
-		max_new_tokens: d.max_new_tokens ?? 128,
-		do_sample: d.do_sample ?? !1
+	let w = {
+		max_new_tokens: x.max_new_tokens ?? 128,
+		do_sample: x.do_sample ?? !1
 	};
-	d.temperature != null && (m.temperature = d.temperature), d.top_p != null && (m.top_p = d.top_p), d.top_k != null && (m.top_k = d.top_k), d.repetition_penalty != null && (m.repetition_penalty = d.repetition_penalty), d.src_lang != null && (m.src_lang = d.src_lang), d.tgt_lang != null && (m.tgt_lang = d.tgt_lang), o && p.tokenizer && (m.streamer = new e(p.tokenizer, {
+	for (let [e, t] of Object.entries(x)) e === "max_new_tokens" || e === "do_sample" || (w[e] = t);
+	let T = r && C.tokenizer ? b(r, s) : null;
+	T && (w.streamer = new e(C.tokenizer, {
 		skip_prompt: !0,
 		skip_special_tokens: !0,
-		callback_function: s(o, d)
+		callback_function: T.callback
 	}));
-	let h = performance.now(), g = await p(u, m);
-	return console.log(`[flaxia-node] ai-inference: done task=${c} model=${l} execMs=${Math.round(performance.now() - h)} inputLen=${String(u).length}`), { output: g };
+	let E = performance.now(), D;
+	try {
+		D = await C(y, w);
+	} catch (e) {
+		throw T && await T.cancel(), e;
+	}
+	return await T?.finish(), console.log(`[flaxia-node] ai-inference: done task=${i} model=${c.repo} execMs=${Math.round(performance.now() - E)} inputLen=${String(y).length}`), { output: D };
 };
 //#endregion
-export { c as handleAiInference, o as releaseCache };
+export { S as handleAiInference, y as releaseCache };