@flaxia/node 0.3.1 → 0.3.2

package/dist/assets/ai-inference.js

@@ -26,7 +26,9 @@ var r = [
 	"depth-estimation",
 	"feature-extraction",
 	"image-feature-extraction"
-], i = /* @__PURE__ */ new Map(), a = (e, t) => {
+], 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;
@@ -34,39 +36,43 @@ var r = [
 	return (e) => {
 		n += e, r ||= setTimeout(a, i);
 	};
-}, o = async (o, s) => {
-	let { task: c, model: l, input: u, options: d = {} } = o;
-	if (!r.includes(c)) throw Error(`Invalid or unsupported task: ${c}. Supported tasks are: ${r.join(", ")}`);
+}, c = async (o, c) => {
+	let { task: l, model: u, input: d, options: f = {} } = o;
+	if (!r.includes(l)) throw Error(`Invalid or unsupported task: ${l}. Supported tasks are: ${r.join(", ")}`);
 	if (self.crossOriginIsolated) {
-		let e = d.numThreads ?? navigator.hardwareConcurrency;
+		let e = f.numThreads ?? navigator.hardwareConcurrency;
 		t.backends.onnx?.wasm && (t.backends.onnx.wasm.numThreads = Math.max(1, Math.min(e, navigator.hardwareConcurrency || 4)));
 	}
-	let f = `${c}:${l}`, p = i.get(f);
-	if (!p) {
+	let p = `${l}:${u}`, m = i.get(p);
+	if (!m) {
+		if (i.size >= a) {
+			let e = i.keys().next().value;
+			i.delete(e);
+		}
 		let e = {
-			dtype: d.dtype ?? "q4f16",
-			device: d.device
+			dtype: f.dtype ?? "q4f16",
+			device: f.device
 		}, t = performance.now();
-		console.log(`[flaxia-node] ai-inference: loading pipeline task=${c} model=${l} dtype=${e.dtype} device=${e.device ?? "wasm"}`);
+		console.log(`[flaxia-node] ai-inference: loading pipeline task=${l} model=${u} dtype=${e.dtype} device=${e.device ?? "wasm"}`);
 		try {
-			p = await n(c, l, e);
+			m = await n(l, u, 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 (f.device && f.device !== "wasm") console.warn(`[AiInference] ${f.device} failed, falling back to wasm:`, t), e.device = "wasm", m = await n(l, u, e);
+			else throw console.error(`[flaxia-node] ai-inference: pipeline load FAILED task=${l} model=${u} error=${t instanceof Error ? t.message : String(t)}`), t;
 		}
-		console.log(`[flaxia-node] ai-inference: pipeline ready task=${c} model=${l} loadMs=${Math.round(performance.now() - t)}`), i.set(f, p);
+		console.log(`[flaxia-node] ai-inference: pipeline ready task=${l} model=${u} loadMs=${Math.round(performance.now() - t)}`), i.set(p, m);
 	}
-	let m = {
-		max_new_tokens: d.max_new_tokens ?? 128,
-		do_sample: d.do_sample ?? !1
+	let h = {
+		max_new_tokens: f.max_new_tokens ?? 128,
+		do_sample: f.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), s && p.tokenizer && (m.streamer = new e(p.tokenizer, {
+	f.temperature != null && (h.temperature = f.temperature), f.top_p != null && (h.top_p = f.top_p), f.top_k != null && (h.top_k = f.top_k), f.repetition_penalty != null && (h.repetition_penalty = f.repetition_penalty), f.src_lang != null && (h.src_lang = f.src_lang), f.tgt_lang != null && (h.tgt_lang = f.tgt_lang), c && m.tokenizer && (h.streamer = new e(m.tokenizer, {
 		skip_prompt: !0,
 		skip_special_tokens: !0,
-		callback_function: a(s, d)
+		callback_function: s(c, f)
 	}));
-	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 g = performance.now(), _ = await m(d, h);
+	return console.log(`[flaxia-node] ai-inference: done task=${l} model=${u} execMs=${Math.round(performance.now() - g)} inputLen=${String(d).length}`), { output: _ };
 };
 //#endregion
-export { o as handleAiInference };
+export { c as handleAiInference, o as releaseCache };

package/dist/assets/nudenet.js

@@ -14047,5 +14047,8 @@ var jd = async (e) => {
 		a.close();
 	}
 };
+function Md() {
+	Ed.clear();
+}
 //#endregion
-export { jd as handleNudeNet };
+export { jd as handleNudeNet, Md as releaseCache };

package/dist/assets/vector-embed.js

@@ -1,6 +1,8 @@
 //#region src/workloads/vector-embed.ts
-var e = null;
-async function t(t) {
+var e = null, t = () => {
+	e = null;
+};
+async function n(t) {
 	let n = performance.now();
 	if (!e) {
 		let t = performance.now();
@@ -25,4 +27,4 @@ async function t(t) {
 	};
 }
 //#endregion
-export { t as handleVectorEmbed };
+export { n as handleVectorEmbed, t as releaseCache };

package/dist/assets/vector-query.js

@@ -1,17 +1,19 @@
 import { t as e } from "./VectorStoreEngine.js";
 //#region src/workloads/vector-query.ts
-var t = null;
-async function n() {
+var t = null, n = () => {
+	t = null;
+};
+async function r() {
 	return t || (t = new e(), await t.initialize()), t;
 }
-async function r(e) {
-	let t = await n(), r = performance.now();
+async function i(e) {
+	let t = await r(), n = performance.now();
 	try {
-		let n = await t.query(e);
-		return console.log(`[flaxia-node] vector-query: done durationMs=${Math.round(performance.now() - r)}`), n;
+		let r = await t.query(e);
+		return console.log(`[flaxia-node] vector-query: done durationMs=${Math.round(performance.now() - n)}`), r;
 	} catch (e) {
 		throw console.error(`[flaxia-node] vector-query: failed error=${e instanceof Error ? e.message : String(e)}`), e;
 	}
 }
 //#endregion
-export { r as handleVectorQuery };
+export { i as handleVectorQuery, n as releaseCache };

package/dist/assets/vector-store.js

@@ -1,17 +1,19 @@
 import { t as e } from "./VectorStoreEngine.js";
 //#region src/workloads/vector-store.ts
-var t = null;
-async function n() {
+var t = null, n = () => {
+	t = null;
+};
+async function r() {
 	return t || (t = new e(), await t.initialize()), t;
 }
-async function r(e) {
-	let t = await n(), r = performance.now();
+async function i(e) {
+	let t = await r(), n = performance.now();
 	try {
-		let n = await t.store(e);
-		return console.log(`[flaxia-node] vector-store: done durationMs=${Math.round(performance.now() - r)}`), n;
+		let r = await t.store(e);
+		return console.log(`[flaxia-node] vector-store: done durationMs=${Math.round(performance.now() - n)}`), r;
 	} catch (e) {
 		throw console.error(`[flaxia-node] vector-store: failed error=${e instanceof Error ? e.message : String(e)}`), e;
 	}
 }
 //#endregion
-export { r as handleVectorStore };
+export { i as handleVectorStore, n as releaseCache };

package/dist/client/SignalingClient.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"SignalingClient.d.ts","sourceRoot":"","sources":["../../src/client/SignalingClient.ts"],"names":[],"mappings":"AAIA,OAAO,KAAK,EAAE,UAAU,EAAE,YAAY,EAAE,MAAM,aAAa,CAAC;AAK5D,MAAM,WAAW,WAAW;IAC1B,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,EAAE,MAAM,CAAC;IACf,QAAQ,EAAE,YAAY,CAAC;IACvB,OAAO,EAAE,OAAO,CAAC;IACjB,SAAS,CAAC,EAAE,MAAM,CAAC;CACpB;AAsPD,eAAO,MAAM,cAAc,GAAI,QAAQ,UAAU,SAehD,CAAC"}
+{"version":3,"file":"SignalingClient.d.ts","sourceRoot":"","sources":["../../src/client/SignalingClient.ts"],"names":[],"mappings":"AAIA,OAAO,KAAK,EAAE,UAAU,EAAE,YAAY,EAAE,MAAM,aAAa,CAAC;AAK5D,MAAM,WAAW,WAAW;IAC1B,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,EAAE,MAAM,CAAC;IACf,QAAQ,EAAE,YAAY,CAAC;IACvB,OAAO,EAAE,OAAO,CAAC;IACjB,SAAS,CAAC,EAAE,MAAM,CAAC;CACpB;AAqQD,eAAO,MAAM,cAAc,GAAI,QAAQ,UAAU,SAehD,CAAC"}

package/dist/executor/throttle.d.ts

@@ -1,31 +1,20 @@
-/**
- * Estimates current CPU contention and throttles work accordingly.
- *
- * Runs inside the compute Worker (where the actual workload executes) rather
- * than on the main thread, so the measurement reflects the thread that does
- * the work. Load is estimated with a calibrated busy-loop: a fixed-iteration
- * loop takes longer to finish when the system is busy, and the ratio to the
- * idle baseline is the estimated load.
- *
- * There is no direct API for CPU usage in the browser, so this is a heuristic
- * used to (a) avoid starting work while the machine is already loaded and
- * (b) space out task starts so average CPU stays near `maxLoad`.
- */
 export declare class CpuThrottle {
-    private baselineMs;
-    private lastLoad;
     private readonly maxLoad;
-    private readonly loopIterations;
-    constructor(maxLoad?: number, loopIterations?: number);
-    private busyLoop;
-    private sample;
-    calibrate(): Promise<void>;
-    getCurrentLoad(): Promise<number>;
+    private windowStart;
+    private windowBusyMs;
+    private lastLoad;
+    constructor(maxLoad?: number);
+    /**
+     * Charge a completed task's wall-clock duration (in ms) against the CPU budget.
+     * Call this from the worker once a workload has finished (incl. errors/timeouts).
+     */
+    markTaskComplete(durationMs: number): void;
+    private estimate;
     shouldPause(): Promise<boolean>;
     /**
-     * Waits until the system is calm enough to run a task.
-     * Bounded by `maxWaitMs` so a persistently busy machine cannot stall
-     * the queue forever.
+     * Waits until the rolling busy ratio is back under `maxLoad`.
+     * Bounded by `maxWaitMs` so a sustained heavy device cannot stall the queue
+     * forever.
      */
     waitForSlot(maxWaitMs?: number): Promise<void>;
     get lastMeasuredLoad(): number;

package/dist/executor/throttle.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"throttle.d.ts","sourceRoot":"","sources":["../../src/executor/throttle.ts"],"names":[],"mappings":"AAIA;;;;;;;;;;;;GAYG;AACH,qBAAa,WAAW;IACtB,OAAO,CAAC,UAAU,CAAuB;IACzC,OAAO,CAAC,QAAQ,CAAK;IACrB,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAS;IACjC,OAAO,CAAC,QAAQ,CAAC,cAAc,CAAS;gBAE5B,OAAO,SAAO,EAAE,cAAc,SAAY;IAKtD,OAAO,CAAC,QAAQ;IAQhB,OAAO,CAAC,MAAM;IAKR,SAAS,IAAI,OAAO,CAAC,IAAI,CAAC;IAM1B,cAAc,IAAI,OAAO,CAAC,MAAM,CAAC;IAWjC,WAAW,IAAI,OAAO,CAAC,OAAO,CAAC;IAIrC;;;;OAIG;IACG,WAAW,CAAC,SAAS,SAAQ,GAAG,OAAO,CAAC,IAAI,CAAC;IAQnD,IAAI,gBAAgB,IAAI,MAAM,CAE7B;IAED,IAAI,YAAY,IAAI,MAAM,CAEzB;CACF"}
+{"version":3,"file":"throttle.d.ts","sourceRoot":"","sources":["../../src/executor/throttle.ts"],"names":[],"mappings":"AAmBA,qBAAa,WAAW;IACtB,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAS;IACjC,OAAO,CAAC,WAAW,CAAqB;IACxC,OAAO,CAAC,YAAY,CAAK;IACzB,OAAO,CAAC,QAAQ,CAAK;gBAET,OAAO,SAAO;IAI1B;;;OAGG;IACH,gBAAgB,CAAC,UAAU,EAAE,MAAM,GAAG,IAAI;IAU1C,OAAO,CAAC,QAAQ;IAMV,WAAW,IAAI,OAAO,CAAC,OAAO,CAAC;IAIrC;;;;OAIG;IACG,WAAW,CAAC,SAAS,SAAmB,GAAG,OAAO,CAAC,IAAI,CAAC;IAQ9D,IAAI,gBAAgB,IAAI,MAAM,CAE7B;IAED,IAAI,YAAY,IAAI,MAAM,CAEzB;CACF"}

package/dist/index.js

@@ -144,7 +144,7 @@ var e = class {
 	}
 }, c = (...e) => console.log("[flaxia-node]", ...e), l = (...e) => console.error("[flaxia-node]", ...e), u = "flaxia_node_id", d = "flaxia_node_token", f = class {
 	constructor(e, t, n) {
-		this.config = e, this.workerPool = t, this.nodeId = n, this.ws = null, this.reconnectAttempts = 0, this.MAX_RECONNECT_DELAY = 3e4, this.destroyed = !1, this.suspended = !1, this.visibilityHandler = null;
+		this.config = e, this.workerPool = t, this.nodeId = n, this.ws = null, this.reconnectAttempts = 0, this.MAX_RECONNECT_DELAY = 3e4, this.destroyed = !1, this.suspended = !1, this.visibilityHandler = null, this.inflightTasks = /* @__PURE__ */ new Set();
 	}
 	async connect() {
 		if (this.destroyed) return;
@@ -241,7 +241,8 @@ var e = class {
 				body: JSON.stringify({
 					siteId: this.config.siteId,
 					nodeId: this.nodeId,
-					capabilities: this.config.capabilities ?? ["ai-inference", "image-process"]
+					capabilities: this.config.capabilities ?? ["ai-inference", "image-process"],
+					deviceMemory: typeof navigator.deviceMemory == "number" ? navigator.deviceMemory : null
 				})
 			});
 			if (!n.ok) return l(`node register failed HTTP ${n.status} (${t}/crowd/nodes/register)`), null;
@@ -260,18 +261,25 @@ var e = class {
 		this.ws && this.ws.readyState === WebSocket.OPEN && this.ws.send(JSON.stringify(e));
 	}
 	async handleTask(e) {
-		let t = e.timeoutMs ? e.timeoutMs + 3e4 : void 0, n = performance.now(), r = await this.workerPool.run(e.taskId, e.workload, e.payload, t, (t) => {
+		if (this.inflightTasks.has(e.taskId)) return;
+		this.inflightTasks.add(e.taskId);
+		let t = performance.now();
+		try {
+			let n = e.timeoutMs ? e.timeoutMs + 3e4 : void 0, r = await this.workerPool.run(e.taskId, e.workload, e.payload, n, (t) => {
+				this.send({
+					type: "progress",
+					taskId: e.taskId,
+					token: t
+				});
+			}, { maxCpuLoad: this.config.maxCpuLoad });
 			this.send({
-				type: "progress",
+				type: "result",
 				taskId: e.taskId,
-				token: t
-			});
-		}, { maxCpuLoad: this.config.maxCpuLoad });
-		this.send({
-			type: "result",
-			taskId: e.taskId,
-			payload: r
-		}), c(`task result sent taskId=${e.taskId} workload=${e.workload} durationMs=${Math.round(performance.now() - n)}`);
+				payload: r
+			}), c(`task result sent taskId=${e.taskId} workload=${e.workload} durationMs=${Math.round(performance.now() - t)}`);
+		} finally {
+			this.inflightTasks.delete(e.taskId);
+		}
 	}
 	setupVisibilityHandler() {
 		this.removeVisibilityHandler(), this.visibilityHandler = () => {

package/dist/worker.js

@@ -2,36 +2,24 @@
 function e(e) {
 	return new Promise((t) => setTimeout(t, e));
 }
-var t = class {
-	constructor(e = .15, t = 2e6) {
-		this.baselineMs = null, this.lastLoad = 0, this.maxLoad = Math.max(.05, Math.min(.3, e)), this.loopIterations = t;
+var t = 3e4, n = 3e4, r = class {
+	constructor(e = .15) {
+		this.windowStart = performance.now(), this.windowBusyMs = 0, this.lastLoad = 0, this.maxLoad = Math.max(.05, Math.min(.3, e));
 	}
-	busyLoop() {
-		let e = this.loopIterations, t = performance.now(), n = 0;
-		for (; n < e;) n++;
-		return performance.now() - t;
-	}
-	sample() {
-		return [
-			this.busyLoop(),
-			this.busyLoop(),
-			this.busyLoop()
-		].sort((e, t) => e - t)[1];
-	}
-	async calibrate() {
-		this.busyLoop(), this.baselineMs = this.sample(), this.lastLoad = 0;
+	markTaskComplete(e) {
+		let n = performance.now();
+		n - this.windowStart > t && (this.windowStart = n, this.windowBusyMs = 0), this.windowBusyMs += Math.max(0, e), this.lastLoad = this.estimate();
 	}
-	async getCurrentLoad() {
-		if (this.baselineMs === null) return await this.calibrate(), 0;
-		let e = this.sample(), t = Math.min(1, Math.max(0, e / this.baselineMs - 1));
-		return this.lastLoad = t, t;
+	estimate() {
+		let e = performance.now() - this.windowStart;
+		return e <= 0 ? 0 : Math.min(1, Math.max(0, this.windowBusyMs / e));
 	}
 	async shouldPause() {
-		return await this.getCurrentLoad() > this.maxLoad;
+		return this.estimate() > this.maxLoad;
 	}
-	async waitForSlot(t = 3e4) {
-		let n = performance.now();
-		for (; await this.shouldPause() && !(performance.now() - n > t);) await e(500);
+	async waitForSlot(t = n) {
+		let r = performance.now();
+		for (; await this.shouldPause() && !(performance.now() - r > t);) await e(500);
 	}
 	get lastMeasuredLoad() {
 		return this.lastLoad;
@@ -39,71 +27,89 @@ var t = class {
 	get maxLoadValue() {
 		return this.maxLoad;
 	}
-}, n = null;
+}, i = null, a = new Set([
+	"ai-inference",
+	"vector-embed",
+	"vector-query",
+	"nudenet",
+	"image-process",
+	"container"
+]), o = 6e4;
+function s() {
+	let e = navigator.deviceMemory;
+	return typeof e == "number" && e >= 4;
+}
+var c = /* @__PURE__ */ new Map(), l = null;
+function u() {
+	for (let e of c.values()) try {
+		e();
+	} catch {}
+	c.clear();
+}
+function d() {
+	l && clearTimeout(l), l = setTimeout(() => {
+		l = null, u();
+	}, o);
+}
+async function f(e, t, n) {
+	switch (e) {
+		case "ai-inference":
+			let r = await import("./assets/ai-inference.js");
+			return c.set(e, r.releaseCache), await r.handleAiInference(t, n);
+		case "image-process": return await (await import("./assets/image-process.js")).handleImageProcess(t);
+		case "container": return await (await import("./assets/container.js")).handleContainer(t);
+		case "vector-embed":
+			let i = await import("./assets/vector-embed.js");
+			return c.set(e, i.releaseCache), await i.handleVectorEmbed(t);
+		case "vector-store":
+			let a = await import("./assets/vector-store.js");
+			return c.set(e, a.releaseCache), await a.handleVectorStore(t);
+		case "vector-query":
+			let o = await import("./assets/vector-query.js");
+			return c.set(e, o.releaseCache), await o.handleVectorQuery(t);
+		case "nudenet":
+			let s = await import("./assets/nudenet.js");
+			return c.set(e, s.releaseCache), await s.handleNudeNet(t);
+		default: throw Error(`Unknown workload type: ${e}`);
+	}
+}
 self.onmessage = async (e) => {
-	let { id: r, workload: i, payload: a, config: o } = e.data, s = performance.now();
+	let { id: t, workload: n, payload: o, config: c } = e.data, l = performance.now();
 	try {
-		n ||= new t(o?.maxCpuLoad), console.log(`[flaxia-node:worker] task start id=${r} workload=${i}`), await n.waitForSlot();
+		if (i ||= new r(c?.maxCpuLoad), console.log(`[flaxia-node:worker] task start id=${t} workload=${n}`), a.has(n) && !s()) {
+			let e = navigator.deviceMemory;
+			throw Error(`${n}: insufficient device memory (deviceMemory=${e ?? "unknown"})`);
+		}
+		await i.waitForSlot();
 		let e = setInterval(() => {
 			self.postMessage({
-				id: r,
+				id: t,
 				type: "heartbeat",
-				cpuLoad: n?.lastMeasuredLoad ?? 0
+				cpuLoad: i?.lastMeasuredLoad ?? 0
 			});
-		}, 1e4);
+		}, 1e4), u = performance.now();
 		try {
-			let e, t = (e) => {
+			let e = await f(n, o, (e) => {
 				self.postMessage({
-					id: r,
+					id: t,
 					type: "token",
 					token: e
 				});
-			};
-			switch (i) {
-				case "ai-inference":
-					let { handleAiInference: n } = await import("./assets/ai-inference.js");
-					e = await n(a, t);
-					break;
-				case "image-process":
-					let { handleImageProcess: r } = await import("./assets/image-process.js");
-					e = await r(a);
-					break;
-				case "container":
-					let { handleContainer: o } = await import("./assets/container.js");
-					e = await o(a);
-					break;
-				case "vector-embed":
-					let { handleVectorEmbed: s } = await import("./assets/vector-embed.js");
-					e = await s(a);
-					break;
-				case "vector-store":
-					let { handleVectorStore: c } = await import("./assets/vector-store.js");
-					e = await c(a);
-					break;
-				case "vector-query":
-					let { handleVectorQuery: l } = await import("./assets/vector-query.js");
-					e = await l(a);
-					break;
-				case "nudenet":
-					let { handleNudeNet: u } = await import("./assets/nudenet.js");
-					e = await u(a);
-					break;
-				default: throw Error(`Unknown workload type: ${i}`);
-			}
+			});
 			self.postMessage({
-				id: r,
+				id: t,
 				type: "done",
 				result: e
-			}), console.log(`[flaxia-node:worker] task done id=${r} workload=${i} durationMs=${Math.round(performance.now() - s)}`);
+			}), console.log(`[flaxia-node:worker] task done id=${t} workload=${n} durationMs=${Math.round(performance.now() - l)}`);
 		} finally {
-			clearInterval(e);
+			clearInterval(e), i.markTaskComplete(performance.now() - u), d();
 		}
 	} catch (e) {
-		let t = e instanceof Error ? e.message : String(e);
-		console.error(`[flaxia-node:worker] task error id=${r} workload=${i} error=${t}`), self.postMessage({
-			id: r,
+		let r = e instanceof Error ? e.message : String(e);
+		console.error(`[flaxia-node:worker] task error id=${t} workload=${n} error=${r}`), self.postMessage({
+			id: t,
 			type: "error",
-			error: t
+			error: r
 		});
 	}
 };

package/dist/workloads/ai-inference.d.ts

@@ -1,3 +1,9 @@
 import type { AiInferencePayload, AiInferenceResult } from '@flaxia/sdk';
+/**
+ * Drops all cached model pipelines so the (potentially hundreds of MB of)
+ * transformers/onnxruntime memory can be garbage-collected. Called by the
+ * worker after an idle period so a node does not pin every model it ever saw.
+ */
+export declare const releaseCache: () => void;
 export declare const handleAiInference: (payload: AiInferencePayload, onToken?: (token: string) => void) => Promise<AiInferenceResult>;
 //# sourceMappingURL=ai-inference.d.ts.map

package/dist/workloads/ai-inference.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"ai-inference.d.ts","sourceRoot":"","sources":["../../src/workloads/ai-inference.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,kBAAkB,EAAE,iBAAiB,EAAsB,MAAM,aAAa,CAAC;AAwC7F,eAAO,MAAM,iBAAiB,GAC5B,SAAS,kBAAkB,EAC3B,UAAU,CAAC,KAAK,EAAE,MAAM,KAAK,IAAI,KAChC,OAAO,CAAC,iBAAiB,CAsE3B,CAAC"}
+{"version":3,"file":"ai-inference.d.ts","sourceRoot":"","sources":["../../src/workloads/ai-inference.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,kBAAkB,EAAE,iBAAiB,EAAsB,MAAM,aAAa,CAAC;AAe7F;;;;GAIG;AACH,eAAO,MAAM,YAAY,QAAO,IAE/B,CAAC;AA4BF,eAAO,MAAM,iBAAiB,GAC5B,SAAS,kBAAkB,EAC3B,UAAU,CAAC,KAAK,EAAE,MAAM,KAAK,IAAI,KAChC,OAAO,CAAC,iBAAiB,CA2E3B,CAAC"}

package/dist/workloads/nudenet.d.ts

@@ -1,3 +1,4 @@
 import type { NudeNetPayload, NudeNetResult } from '@flaxia/sdk';
 export declare const handleNudeNet: (payload: NudeNetPayload) => Promise<NudeNetResult>;
+export declare function releaseCache(): void;
 //# sourceMappingURL=nudenet.d.ts.map

package/dist/workloads/nudenet.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"nudenet.d.ts","sourceRoot":"","sources":["../../src/workloads/nudenet.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAoB,cAAc,EAAE,aAAa,EAAE,MAAM,aAAa,CAAC;AA+KnF,eAAO,MAAM,aAAa,GAAU,SAAS,cAAc,KAAG,OAAO,CAAC,aAAa,CA6BlF,CAAC"}
+{"version":3,"file":"nudenet.d.ts","sourceRoot":"","sources":["../../src/workloads/nudenet.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAoB,cAAc,EAAE,aAAa,EAAE,MAAM,aAAa,CAAC;AA+KnF,eAAO,MAAM,aAAa,GAAU,SAAS,cAAc,KAAG,OAAO,CAAC,aAAa,CA6BlF,CAAC;AAEF,wBAAgB,YAAY,IAAI,IAAI,CAEnC"}

package/dist/workloads/vector-embed.d.ts

@@ -1,3 +1,4 @@
 import type { VectorEmbedPayload, VectorEmbedResult } from '@flaxia/sdk';
+export declare const releaseCache: () => void;
 export declare function handleVectorEmbed(payload: VectorEmbedPayload): Promise<VectorEmbedResult>;
 //# sourceMappingURL=vector-embed.d.ts.map

package/dist/workloads/vector-embed.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"vector-embed.d.ts","sourceRoot":"","sources":["../../src/workloads/vector-embed.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,kBAAkB,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAIzE,wBAAsB,iBAAiB,CAAC,OAAO,EAAE,kBAAkB,GAAG,OAAO,CAAC,iBAAiB,CAAC,CA2C/F"}
+{"version":3,"file":"vector-embed.d.ts","sourceRoot":"","sources":["../../src/workloads/vector-embed.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,kBAAkB,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAIzE,eAAO,MAAM,YAAY,QAAO,IAE/B,CAAC;AAEF,wBAAsB,iBAAiB,CAAC,OAAO,EAAE,kBAAkB,GAAG,OAAO,CAAC,iBAAiB,CAAC,CA2C/F"}

package/dist/workloads/vector-query.d.ts

@@ -1,3 +1,4 @@
 import type { VectorQueryPayload, VectorQueryResult } from '@flaxia/sdk';
+export declare const releaseCache: () => void;
 export declare function handleVectorQuery(payload: VectorQueryPayload): Promise<VectorQueryResult>;
 //# sourceMappingURL=vector-query.d.ts.map

package/dist/workloads/vector-query.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"vector-query.d.ts","sourceRoot":"","sources":["../../src/workloads/vector-query.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,kBAAkB,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAazE,wBAAsB,iBAAiB,CAAC,OAAO,EAAE,kBAAkB,GAAG,OAAO,CAAC,iBAAiB,CAAC,CAW/F"}
+{"version":3,"file":"vector-query.d.ts","sourceRoot":"","sources":["../../src/workloads/vector-query.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,kBAAkB,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAKzE,eAAO,MAAM,YAAY,QAAO,IAE/B,CAAC;AAUF,wBAAsB,iBAAiB,CAAC,OAAO,EAAE,kBAAkB,GAAG,OAAO,CAAC,iBAAiB,CAAC,CAW/F"}

package/dist/workloads/vector-store.d.ts

@@ -1,4 +1,5 @@
 import type { VectorStorePayload, VectorStoreResult } from '@flaxia/sdk';
+export declare const releaseCache: () => void;
 export declare function handleVectorStore(payload: VectorStorePayload): Promise<VectorStoreResult>;
 export declare function handleVectorStoreAssignShard(rangeStart: number, rangeEnd: number): Promise<void>;
 //# sourceMappingURL=vector-store.d.ts.map

package/dist/workloads/vector-store.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"vector-store.d.ts","sourceRoot":"","sources":["../../src/workloads/vector-store.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,kBAAkB,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAazE,wBAAsB,iBAAiB,CAAC,OAAO,EAAE,kBAAkB,GAAG,OAAO,CAAC,iBAAiB,CAAC,CAW/F;AAED,wBAAsB,4BAA4B,CAAC,UAAU,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAGtG"}
+{"version":3,"file":"vector-store.d.ts","sourceRoot":"","sources":["../../src/workloads/vector-store.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,kBAAkB,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAKzE,eAAO,MAAM,YAAY,QAAO,IAE/B,CAAC;AAUF,wBAAsB,iBAAiB,CAAC,OAAO,EAAE,kBAAkB,GAAG,OAAO,CAAC,iBAAiB,CAAC,CAW/F;AAED,wBAAsB,4BAA4B,CAAC,UAAU,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAGtG"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@flaxia/node",
-  "version": "0.3.1",
+  "version": "0.3.2",
   "private": false,
   "main": "dist/index.js",
   "types": "./dist/index.d.ts",