@flaxia/node 0.5.0 → 0.5.1

package/dist/worker.js

@@ -1,13 +1,13 @@
-import { _ as e, a as t, b as n, c as r, g as i, h as a, i as o, l as s, n as c, o as l, s as u, t as d, v as f, x as p, y as m } from "./assets/dist.js";
+import { M as e, N as t, O as n, a as r, c as i, f as a, h as o, j as s, k as c, l, m as u, o as d, r as f, t as p, u as m, v as h, x as g, y as _ } from "./assets/webgpuProbe.js";
 //#region src/executor/throttle.ts
-function h(e) {
+function v(e) {
 	return new Promise((t) => setTimeout(t, e));
 }
-var g = 3e4, ee = 500, _ = class extends Error {
+var y = 3e4, ee = 500, b = class extends Error {
 	constructor(e, t, n) {
 		super(`device is over its CPU budget (load ${e.toFixed(2)} > max ${t.toFixed(2)}) after waiting ${Math.round(n)}ms; task not started`), this.load = e, this.maxLoad = t, this.waitedMs = n, this.retryable = !0, this.name = "ThrottleBusyError";
 	}
-}, v = class {
+}, te = class {
 	constructor(e = .15) {
 		this.windowStart = performance.now(), this.windowBusyMs = 0, this.lastLoad = 0, this.dutyStartedAt = null, this.maxLoad = Math.max(.05, Math.min(.3, e));
 	}
@@ -34,13 +34,13 @@ var g = 3e4, ee = 500, _ = class extends Error {
 	async shouldPause() {
 		return this.rollWindow(performance.now()), this.estimate(this.windowBusyMs) > this.maxLoad;
 	}
-	async waitForSlot(e = g) {
+	async waitForSlot(e = y) {
 		if (!Number.isFinite(e) || e < 0) throw Error(`waitForSlot: maxWaitMs must be a non-negative number, got ${String(e)}`);
 		let t = performance.now();
 		for (; await this.shouldPause();) {
 			let n = performance.now() - t;
-			if (n >= e) throw new _(this.lastLoad, this.maxLoad, n);
-			await h(Math.min(ee, Math.max(1, e - n)));
+			if (n >= e) throw new b(this.lastLoad, this.maxLoad, n);
+			await v(Math.min(ee, Math.max(1, e - n)));
 		}
 	}
 	async yieldIfOverloaded() {
@@ -49,7 +49,7 @@ var g = 3e4, ee = 500, _ = class extends Error {
 		let t = Math.max(0, e - this.dutyStartedAt);
 		if (t < 200) return 0;
 		let n = Math.ceil(t * (1 / this.maxLoad - 1));
-		return this.dutyStartedAt = null, n > 0 && await h(n), this.estimate(this.windowBusyMs), n;
+		return this.dutyStartedAt = null, n > 0 && await v(n), this.estimate(this.windowBusyMs), n;
 	}
 	get lastMeasuredLoad() {
 		return this.lastLoad;
@@ -57,13 +57,13 @@ var g = 3e4, ee = 500, _ = class extends Error {
 	get maxLoadValue() {
 		return this.maxLoad;
 	}
-}, y = 4096;
-function b(e) {
+}, x = 4096;
+function ne(e) {
 	let t = e.maxSeq;
-	return typeof t == "number" && Number.isInteger(t) && t > 0 ? t : y;
+	return typeof t == "number" && Number.isInteger(t) && t > 0 ? t : x;
 }
-function x(e, t, n) {
-	let { pos: r, tokens: i } = e.header, a = b(t);
+function S(e, t, n) {
+	let { pos: r, tokens: i } = e.header, a = ne(t);
 	if (!Number.isInteger(i) || i < 1) return {
 		kind: "corrupt",
 		reason: `carried ${i} token columns`
@@ -95,78 +95,78 @@ function x(e, t, n) {
 		hidden: o
 	};
 }
-function te(e, t, n, r) {
-	let i = s(e);
-	return i ? i.header.pos !== t || i.header.requestId !== t ? { kind: "stale" } : x(i, n, r) : {
+function re(e, t, n, r) {
+	let i = _(e);
+	return i ? i.header.pos !== t || i.header.requestId !== t ? { kind: "stale" } : S(i, n, r) : {
 		kind: "corrupt",
 		reason: "malformed frame header"
 	};
 }
-async function ne(t) {
-	let { engine: n, link: i, semantics: o, promptTokens: s, chainLength: c, maxNewTokens: l } = t;
-	if (s.length === 0) throw Error("swarm host requires at least one prompt token");
-	n.reset();
-	let u = null;
-	i.onFrame((t) => {
-		if (e(t)) return;
-		let r = u;
-		if (!r) return;
-		let i = te(t, r.pos, n, o);
+async function ie(n) {
+	let { engine: r, link: i, semantics: a, promptTokens: o, chainLength: s, maxNewTokens: c } = n;
+	if (o.length === 0) throw Error("swarm host requires at least one prompt token");
+	r.reset();
+	let l = null;
+	i.onFrame((e) => {
+		if (t(e)) return;
+		let n = l;
+		if (!n) return;
+		let i = re(e, n.pos, r, a);
 		if (i.kind !== "stale") {
-			if (u = null, i.kind === "corrupt") {
-				r.reject(/* @__PURE__ */ Error(`swarm hop failed: ${i.reason}`));
+			if (l = null, i.kind === "corrupt") {
+				n.reject(/* @__PURE__ */ Error(`swarm hop failed: ${i.reason}`));
 				return;
 			}
-			r.resolve(i.hidden);
+			n.resolve(i.hidden);
 		}
 	});
-	let d = (e, t) => c <= 1 ? Promise.resolve(e) : new Promise((n, a) => {
-		u = {
+	let u = (e, t) => s <= 1 ? Promise.resolve(e) : new Promise((n, r) => {
+		l = {
 			pos: t,
 			resolve: n,
-			reject: a
-		}, i.send(r({
+			reject: r
+		}, i.send(h({
 			requestId: t,
 			pos: t,
 			tokens: 1
-		}, o.packHidden(e)));
-	}), f = async (e, t) => d(await n.embedRun(e, t), t), p = null;
-	for (let e = 0; e < s.length; e++) {
-		let t = await f(s[e], e);
-		e === s.length - 1 && (p = await n.headFromHidden(t));
-	}
-	if (!p) throw Error("swarm host produced no logits for the prompt");
-	let m = [], h = s.length, g = o.argmax(p);
-	for (; m.length < l && !(t.eosIds?.has(g) || (m.push(g), await t.onToken?.(g), m.length >= l));) {
-		let e = await f(g, h++);
-		p = await n.headFromHidden(e), g = o.argmax(p);
-	}
-	return c > 1 && i.send(a()), { tokens: m };
-}
-function re(t) {
-	let { engine: n, link: i, semantics: a } = t;
-	return n.reset(), new Promise((t, o) => {
-		let c = Promise.resolve(), l = !1;
-		i.onFrame((u) => {
-			if (!l) {
-				if (e(u)) {
-					l = !0, i.send(u), c = c.then(() => t());
+		}, a.packHidden(e)));
+	}), d = async (e, t) => u(await r.embedRun(e, t), t), f = null;
+	for (let e = 0; e < o.length; e++) {
+		let t = await d(o[e], e);
+		e === o.length - 1 && (f = await r.headFromHidden(t));
+	}
+	if (!f) throw Error("swarm host produced no logits for the prompt");
+	let p = [], m = o.length, g = a.argmax(f);
+	for (; p.length < c && !(n.eosIds?.has(g) || (p.push(g), await n.onToken?.(g), p.length >= c));) {
+		let e = await d(g, m++);
+		f = await r.headFromHidden(e), g = a.argmax(f);
+	}
+	return s > 1 && i.send(e()), { tokens: p };
+}
+function ae(e) {
+	let { engine: n, link: r, semantics: i } = e;
+	return n.reset(), new Promise((e, a) => {
+		let o = Promise.resolve(), s = !1;
+		r.onFrame((c) => {
+			if (!s) {
+				if (t(c)) {
+					s = !0, r.send(c), o = o.then(() => e());
 					return;
 				}
-				c = c.then(async () => {
-					if (l) return;
-					let e = s(u);
+				o = o.then(async () => {
+					if (s) return;
+					let e = _(c);
 					if (!e) throw Error("swarm frame has a malformed frame header");
-					let t = x(e, n, a);
+					let t = S(e, n, i);
 					if (t.kind === "corrupt") throw Error(`swarm frame is unusable: ${t.reason}`);
-					let o = await n.runHidden(t.hidden, e.header.pos);
-					i.send(r({
+					let a = await n.runHidden(t.hidden, e.header.pos);
+					r.send(h({
 						requestId: e.header.requestId,
 						pos: e.header.pos,
 						tokens: e.header.tokens
-					}, a.packHidden(o)));
+					}, i.packHidden(a)));
 				}).catch((e) => {
-					l = !0, o(e instanceof Error ? e : Error(String(e)));
+					s = !0, a(e instanceof Error ? e : Error(String(e)));
 				});
 			}
 		});
@@ -174,9 +174,9 @@ function re(t) {
 }
 //#endregion
 //#region src/swarm/controller.ts
-var S = class {
+var oe = class {
 	constructor(e) {
-		this.options = e, this.frameHandler = null, this.engine = null, this.semantics = null, this.sessionId = "", this.model = "", this.chainLength = 1, this.chain = [], this.slice = null, this.hostWarm = void 0, this.prompt = "", this.maxNewTokens = 128, this.sliceLoaded = !1, this.generationStarted = !1, this.settled = !1;
+		this.options = e, this.frameHandler = null, this.engine = null, this.semantics = null, this.sessionId = "", this.model = "", this.chainLength = 1, this.chain = [], this.slice = null, this.hostWarm = void 0, this.prompt = "", this.maxNewTokens = 128, this.sliceLoaded = !1, this.generationStarted = !1, this.settled = !1, this.flowPaused = !1, this.flowWaiters = /* @__PURE__ */ new Set();
 	}
 	async start() {
 		if (this.settled) return;
@@ -200,7 +200,7 @@ var S = class {
 	}
 	async handleControl(e) {
 		if (!this.settled && e.sessionId === this.sessionId) try {
-			e.type === "swarm-slice" ? await this.loadSlice(e) : e.type === "swarm-start" ? await this.startHost() : e.type === "swarm-error" && this.fail(Error(e.error || "swarm session aborted by the coordinator"));
+			e.type === "swarm-slice" ? await this.loadSlice(e) : e.type === "swarm-start" ? await this.startHost() : e.type === "swarm-error" ? this.fail(Error(e.error || "swarm session aborted by the coordinator")) : e.type === "swarm-flow" && (this.flowPaused = e.paused, this.flowPaused || this.releaseFlowWaiters());
 		} catch (e) {
 			this.fail(e);
 		}
@@ -223,7 +223,7 @@ var S = class {
 		if (!Number.isInteger(e.chainLength) || e.chainLength < 1 || e.chainLength > 16) throw Error(`swarm slice has an invalid chain length: ${JSON.stringify(e.chainLength)}`);
 		if (!Number.isInteger(e.index) || e.index < 0 || e.index >= e.chainLength) throw Error(`swarm slice has an invalid index: ${JSON.stringify(e.index)}`);
 		if (e.role !== (e.index === 0 ? "host" : "worker")) throw Error(`swarm slice role ${e.role} does not match index ${e.index}`);
-		if (!l(e.slice)) throw Error(`swarm slice is out of bounds: ${JSON.stringify(e.slice)}`);
+		if (!u(e.slice)) throw Error(`swarm slice is out of bounds: ${JSON.stringify(e.slice)}`);
 		let t = e.index === 0;
 		if (e.slice.hasEmbed !== t || e.slice.hasHead !== t) throw Error("swarm slice embed/head ownership does not match the host position");
 	}
@@ -238,11 +238,11 @@ var S = class {
 		this.engine = t, this.semantics = n, this.hostWarm = r, this.sliceLoaded = !0, this.options.sendControl({
 			type: "swarm-ready",
 			sessionId: e.sessionId
-		}), e.role === "worker" && (this.generationStarted = !0, re({
+		}), e.role === "worker" && (this.generationStarted = !0, ae({
 			engine: t,
 			link: this.link(),
 			semantics: n
-		}).then(() => this.finish(C())).catch((e) => this.fail(e)));
+		}).then(() => this.finish(se())).catch((e) => this.fail(e)));
 	}
 	async startHost() {
 		if (this.generationStarted) return;
@@ -254,7 +254,7 @@ var S = class {
 		let r = t.encodePrompt(this.prompt), i = e.maxSeq ?? 4096;
 		if (r.length + n > i) throw Error(`swarm request needs ${r.length + n} tokens but the model context holds ${i}`);
 		this.generationStarted = !0;
-		let a = performance.now(), { tokens: o } = await ne({
+		let a = performance.now(), { tokens: o } = await ie({
 			chainLength: this.chainLength,
 			promptTokens: r,
 			maxNewTokens: n,
@@ -262,7 +262,10 @@ var S = class {
 			link: this.link(),
 			semantics: t,
 			eosIds: t.eosIds,
-			onToken: (e) => this.options.emitToken(t.decodeTokens([e]))
+			onToken: async (e) => {
+				if (await this.waitForFlow(), this.settled) throw Error("swarm session stopped while generation was paused");
+				await this.options.emitToken(t.decodeTokens([e]));
+			}
 		}), s = t.decodeTokens;
 		this.finish({
 			output: s(o),
@@ -294,13 +297,20 @@ var S = class {
 		} catch {}
 	}
 	finish(e) {
-		this.settled || (this.settled = !0, this.teardown(), this.options.onDone(e));
+		this.settled || (this.settled = !0, this.releaseFlowWaiters(), this.teardown(), this.options.onDone(e));
 	}
 	fail(e) {
-		this.settled || (this.settled = !0, this.teardown(), this.options.onError(e instanceof Error ? e.message : String(e)));
+		this.settled || (this.settled = !0, this.releaseFlowWaiters(), this.teardown(), this.options.onError(e instanceof Error ? e.message : String(e)));
+	}
+	waitForFlow() {
+		return !this.flowPaused || this.settled ? Promise.resolve() : new Promise((e) => this.flowWaiters.add(e));
+	}
+	releaseFlowWaiters() {
+		for (let e of this.flowWaiters) e();
+		this.flowWaiters.clear();
 	}
 };
-function C() {
+function se() {
 	return {
 		output: "",
 		tokens: [],
@@ -310,13 +320,13 @@ function C() {
 }
 //#endregion
 //#region src/swarm/adapter.ts
-var w = class extends Error {
+var C = class extends Error {
 	constructor(e) {
 		super(e), this.name = "SwarmEngineUnavailableError";
 	}
 };
-async function T(e, t) {
-	if (!(await f()).webgpu) throw new w("WebGPU is not available in this context");
+async function w(e, t) {
+	if (!(await p()).webgpu) throw new C("WebGPU is not available in this context");
 	let n = t ?? ((e) => import(
 		/* @vite-ignore */
 		e
@@ -324,14 +334,18 @@ async function T(e, t) {
 	try {
 		r = await n(e);
 	} catch (e) {
-		throw new w(`failed to load swarm engine module: ${e instanceof Error ? e.message : String(e)}`);
+		throw new C(`failed to load swarm engine module: ${e instanceof Error ? e.message : String(e)}`);
 	}
 	let i = r?.Qwen35Engine;
-	if (!i || typeof i.create != "function") throw new w(`module at ${e} does not export Qwen35Engine.create`);
+	if (!i || typeof i.create != "function") throw new C(`module at ${e} does not export Qwen35Engine.create`);
 	return r;
 }
-function E(e, t) {
-	let n = !1;
+function ce(e, t, n, r) {
+	let i = !1, a = async (e) => {
+		r?.();
+		let t = await e();
+		return r?.(), t;
+	};
 	return {
 		get dim() {
 			return e.dims.dim;
@@ -342,61 +356,102 @@ function E(e, t) {
 		get maxSeq() {
 			return e.maxSeq;
 		},
-		reset: () => e.reset(),
-		embedRun: (t, n) => e.embedRun(t, n),
-		runHidden: (t, n) => e.runHidden(t, n),
-		headFromHidden: (t) => e.headFromHidden(t),
-		setHidden: (t) => e.setHidden(t),
+		reset: () => {
+			r?.(), e.reset();
+		},
+		embedRun: (t, n) => a(() => e.embedRun(t, n)),
+		runHidden: (t, n) => a(() => e.runHidden(t, n)),
+		headFromHidden: (t) => a(() => e.headFromHidden(t)),
+		setHidden: (t) => {
+			r?.(), e.setHidden(t);
+		},
 		dispose: () => {
-			if (!n) {
-				n = !0;
+			if (!i) {
+				i = !0;
 				try {
 					e.dispose?.();
 				} catch {}
 				try {
 					t?.destroy?.();
 				} catch {}
+				try {
+					n?.();
+				} catch {}
 			}
 		}
 	};
 }
 //#endregion
 //#region src/swarm/runtime.ts
-var D = {
+var T = {
 	"qwen3.5-2b": "https://huggingface.co/unsloth/Qwen3.5-2B-GGUF/resolve/main/Qwen3.5-2B-Q4_0.gguf",
 	"qwen3.8-27b": "https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/resolve/main/Qwen3.8-27B-Q4_0.gguf",
 	"qwen3.6-35b-moe": "https://huggingface.co/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF/resolve/main/Qwen_Qwen3.6-35B-A3B-Q4_0.gguf"
-}, ie = "flaxia-swarm-weights-v1", ae = 12 * 1024 * 1024, O = 256 * 1024 * 1024;
-function oe(e) {
-	let t = D[e];
+}, le = "flaxia-swarm-weights-v1", ue = 12 * 1024 * 1024, E = 256 * 1024 * 1024, D = /* @__PURE__ */ new Map();
+function O(e, t) {
+	D.set(e, {
+		status: t,
+		failedExpiresAt: t.state === "failed" ? Date.now() + 3e4 : void 0
+	});
+}
+function k(e, t, n, r) {
+	let i = j(e);
+	return {
+		modelId: e,
+		revision: /\/resolve\/([^/]+)\//.exec(new URL(i).pathname)?.[1] ?? "unknown",
+		dtype: /-(q\d+_\d+)\.gguf/i.exec(i)?.[1] ?? "unknown",
+		backend: "webgpu",
+		layerRange: [t.start, t.end],
+		state: n,
+		referenceCount: r
+	};
+}
+function A(e = Date.now()) {
+	let t = [];
+	for (let [n, r] of D) {
+		if (r.failedExpiresAt !== void 0 && r.failedExpiresAt <= e) {
+			D.delete(n);
+			continue;
+		}
+		t.push(r.status);
+	}
+	return {
+		version: 1,
+		sampledAt: e,
+		expiresAt: e + s,
+		entries: t.slice(0, 8)
+	};
+}
+function j(e) {
+	let t = Object.prototype.hasOwnProperty.call(T, e) ? T[e] : void 0;
 	if (t) return t;
-	throw new w(`unknown swarm model: ${e}`);
+	throw new C(`unknown swarm model: ${e}`);
 }
-var k = null, A = 0, j = 0;
-async function M() {
-	if (k !== null) return k;
+var M = null, N = 0, P = 0;
+async function de() {
+	if (M !== null) return M;
 	try {
-		k = typeof caches < "u" ? await caches.open(ie) : !1;
+		M = typeof caches < "u" ? await caches.open(le) : !1;
 	} catch {
-		k = !1;
+		M = !1;
 	}
-	return k;
+	return M;
 }
-function N(e, t, n) {
+function fe(e, t, n) {
 	return `https://flaxia-weights.local/${encodeURIComponent(e)}/${t}-${n}`;
 }
-async function P(e, t, n) {
-	let r = await F(e, t, n), i = new Uint8Array(await r.arrayBuffer()), a = n - t + 1;
+async function F(e, t, n) {
+	let r = await I(e, t, n), i = new Uint8Array(await r.arrayBuffer()), a = n - t + 1;
 	if (i.byteLength !== a) throw Error(`short range ${i.byteLength}/${a}`);
 	return i;
 }
-async function F(e, t, n) {
-	let r = n - t + 1, i = await M(), a = N(e, t, n);
+async function I(e, t, n) {
+	let r = n - t + 1, i = await de(), a = fe(e, t, n);
 	if (i) try {
 		let e = await i.match(a);
 		if (e) {
 			let t = new Uint8Array(await e.arrayBuffer());
-			if (t.byteLength === r) return A++, new Response(t, { status: 200 });
+			if (t.byteLength === r) return N++, new Response(t, { status: 200 });
 			await i.delete(a).catch(() => {});
 		}
 	} catch {}
@@ -405,7 +460,7 @@ async function F(e, t, n) {
 		try {
 			let s = await fetch(e, { headers: { Range: `bytes=${t}-${n}` } });
 			if (s.status !== 206) o = /* @__PURE__ */ Error(`model host refused range requests (${s.status})`);
-			else return j++, i && s.clone().arrayBuffer().then((e) => {
+			else return P++, i && s.clone().arrayBuffer().then((e) => {
 				if (e.byteLength === r) return i.put(a, new Response(e, {
 					status: 200,
 					headers: { "content-type": "application/octet-stream" }
@@ -418,29 +473,29 @@ async function F(e, t, n) {
 	}
 	throw Error(`range fetch ${t}-${n} failed: ${o instanceof Error ? o.message : String(o)}`);
 }
-async function I(e, t) {
-	let n = ae;
+async function pe(e, t) {
+	let n = ue;
 	for (;;) {
-		let r = await P(t, 0, n - 1), i = r.buffer.slice(r.byteOffset, r.byteOffset + r.byteLength);
+		let r = await F(t, 0, n - 1), i = r.buffer.slice(r.byteOffset, r.byteOffset + r.byteLength);
 		try {
 			return e.parseGGUFHeader(i);
 		} catch (e) {
-			if (n >= O) throw e;
-			n = Math.min(n * 2, O);
+			if (n >= E) throw e;
+			n = Math.min(n * 2, E);
 		}
 	}
 }
-async function L() {
+async function me() {
 	let e = navigator.gpu;
-	if (!e) throw new w("WebGPU is not available in this context");
+	if (!e) throw new C("WebGPU is not available in this context");
 	let t = await e.requestAdapter({ powerPreference: "high-performance" });
-	if (!t) throw new w("no WebGPU adapter");
+	if (!t) throw new C("no WebGPU adapter");
 	return t.requestDevice({ requiredLimits: {
 		maxBufferSize: t.limits.maxBufferSize,
 		maxStorageBufferBindingSize: t.limits.maxStorageBufferBindingSize
 	} });
 }
-function R(e, t) {
+function he(e, t) {
 	let n = t.vocab, r = n["<|im_start|>"], i = n["<|im_end|>"], a = n["<think>"], o = n["</think>"], s = (e) => [
 		r,
 		...t.encode("user\n" + e),
@@ -470,32 +525,32 @@ function R(e, t) {
 		eosIds: new Set(c)
 	};
 }
-function z() {
-	let e = null, t = null, n = () => {
-		if (!e) {
-			let t = new URL("./swarm-engine.js", "" + import.meta.url).href;
-			e = T(t).then((e) => e);
+function L(e) {
+	let t = null, n = null, r = () => {
+		if (!t) {
+			let n = new URL("./swarm-engine.js", "" + import.meta.url).href;
+			t = e ? e() : w(n).then((e) => e);
 		}
-		return e;
-	}, r = async (e) => {
-		let r = await n(), i = oe(e);
-		(!t || t.url !== i) && (t = {
+		return t;
+	}, i = async (e) => {
+		let t = await r(), i = j(e);
+		(!n || n.url !== i) && (n = {
 			url: i,
-			gguf: await I(r, i)
+			gguf: await pe(t, i)
 		});
-		let a = Number(t.gguf.meta["qwen35.block_count"]) - Number(t.gguf.meta["qwen35.nextn_predict_layers"] || 0);
-		if (!Number.isInteger(a) || a < 1) throw new w(`model ${e} has no qwen35 trunk layers`);
+		let a = Number(n.gguf.meta["qwen35.block_count"]) - Number(n.gguf.meta["qwen35.nextn_predict_layers"] || 0);
+		if (!Number.isInteger(a) || a < 1) throw new C(`model ${e} has no qwen35 trunk layers`);
 		return {
-			mod: r,
+			mod: t,
 			url: i,
-			gguf: t.gguf,
+			gguf: n.gguf,
 			layers: a
 		};
 	};
 	return {
 		async plan(e, t) {
-			let { layers: n } = await r(t);
-			return d({
+			let { layers: n } = await i(t);
+			return g({
 				sessionId: "",
 				taskId: "",
 				model: t,
@@ -507,60 +562,103 @@ function z() {
 			}).chain;
 		},
 		async load(e, t) {
-			A = 0, j = 0;
-			let { mod: n, url: i, gguf: a } = await r(t), o = await L(), s = n.makeTokenizer(n.tokenizerFromGGUF(a.meta)), c = (e) => P(i, e.byteOffset, e.byteOffset + e.byteLength - 1);
-			a.streamEntry = (e) => n.streamEntryToGPU(o, e, (e) => F(i, e.byteOffset, e.byteOffset + e.byteLength - 1), { staging: 4 * 1024 * 1024 });
-			let l = await n.qwen35Weights(a, c, {
-				lo: e.start,
-				hi: e.end,
-				hasEmbed: e.hasEmbed,
-				hasHead: e.hasHead,
-				mtp: e.hasHead
-			}, void 0, (e, t) => n.gpuUploadEntry(o, e, t === n.GGML_EMBED)), u = a.tensors[n.GGML_EMBED]?.shape[0] ?? 0, d = await n.Qwen35Engine.create({
-				device: o,
-				meta: a.meta,
-				layerRange: [e.start, e.end],
-				hasEmbed: e.hasEmbed,
-				hasHead: e.hasHead,
-				vocab: u,
-				maxSeq: Math.min(y, Number(a.meta["qwen35.context_length"]) || 4096),
-				batchCols: 4,
-				weights: l
-			}), f = A > 0 && j === 0;
-			return {
-				engine: E(d, o),
-				semantics: R(n, s),
-				warm: f
-			};
+			N = 0, P = 0;
+			let n = JSON.stringify([
+				t,
+				e.start,
+				e.end
+			]);
+			O(n, k(t, e, "loading", 1));
+			let r, a = !1;
+			try {
+				let { mod: o, url: s, gguf: c } = await i(t);
+				r = await me();
+				let l = null, u = r;
+				u.addEventListener?.("uncapturederror", () => {
+					l = "swarm GPU validation failed";
+				}), u.lost?.then(() => {
+					l = "swarm GPU device lost";
+				});
+				let d = () => {
+					if (l) throw new C(l);
+				}, f = he(o, o.makeTokenizer(o.tokenizerFromGGUF(c.meta))), p = (e) => F(s, e.byteOffset, e.byteOffset + e.byteLength - 1);
+				c.streamEntry = (e) => o.streamEntryToGPU(r, e, (e) => I(s, e.byteOffset, e.byteOffset + e.byteLength - 1), { staging: 4 * 1024 * 1024 });
+				let m = await o.qwen35Weights(c, p, {
+					lo: e.start,
+					hi: e.end,
+					hasEmbed: e.hasEmbed,
+					hasHead: e.hasHead,
+					mtp: e.hasHead
+				}, void 0, (e, t) => o.gpuUploadEntry(r, e, t === o.GGML_EMBED)), h = c.tensors[o.GGML_EMBED]?.shape[0] ?? 0, g = await o.Qwen35Engine.create({
+					device: r,
+					meta: c.meta,
+					layerRange: [e.start, e.end],
+					hasEmbed: e.hasEmbed,
+					hasHead: e.hasHead,
+					vocab: h,
+					maxSeq: Math.min(x, Number(c.meta["qwen35.context_length"]) || 4096),
+					batchCols: 4,
+					weights: m
+				});
+				d();
+				let _ = ce(g, r, () => {
+					O(n, k(t, e, "evicting", 0)), queueMicrotask(() => D.delete(n));
+				}, d);
+				return a = !0, O(n, k(t, e, "ready", 1)), {
+					engine: _,
+					semantics: f,
+					warm: N > 0 && P === 0
+				};
+			} catch (i) {
+				if (O(n, k(t, e, "failed", 0)), !a) try {
+					r?.destroy?.();
+				} catch {}
+				throw i;
+			}
 		}
 	};
 }
 //#endregion
 //#region src/swarm/messages.ts
-function B(e) {
-	return o(e) || c(e) || i(e) || t(e);
+function ge(e) {
+	return a(e) || m(e) || o(e) || c(e) || i(e) || n(e);
 }
-function V(e) {
+function _e(e) {
 	if (!e || typeof e != "object") return !1;
 	let t = e;
-	return t.type === "swarm-control" && typeof t.id == "string" && B(t.message);
+	return t.type === "swarm-control" && typeof t.id == "string" && ge(t.message);
 }
 //#endregion
 //#region src/worker/main.worker.ts
-var H = null, U = g, W = 1, G = 15e3, K = /* @__PURE__ */ new Map(), q = null, J = null, Y = null, se = 6e4;
-async function ce() {
+var R = null, ve = y, ye = 1, z = 15e3, B = /* @__PURE__ */ new Map(), V = /* @__PURE__ */ new Set(), H = /* @__PURE__ */ new Map();
+function U(e) {
+	V.delete(e);
+	let t = H.get(e);
+	if (t) {
+		H.delete(e);
+		for (let e of t) e();
+	}
+}
+function be(e) {
+	return V.has(e) ? new Promise((t) => {
+		let n = H.get(e) ?? /* @__PURE__ */ new Set();
+		n.add(t), H.set(e, n);
+	}) : Promise.resolve();
+}
+var W = null, G = null, K = null, xe = 6e4;
+async function Se() {
 	for (let e = 0;; e++) try {
-		await H.waitForSlot(U);
+		await R.waitForSlot(ve);
 		return;
 	} catch (t) {
-		if (!(t instanceof _)) throw t;
-		if (e >= W) throw Error(`device CPU busy: ${t.message}; task not started, retry on an idle node (retryable)`);
-		console.warn(`[flaxia-node:worker] ${t.message}; cooling down ${G}ms before retrying`), await new Promise((e) => setTimeout(e, G));
+		if (!(t instanceof b)) throw t;
+		if (e >= ye) throw Error(`device CPU busy: ${t.message}; task not started, retry on an idle node (retryable)`);
+		console.warn(`[flaxia-node:worker] ${t.message}; cooling down ${z}ms before retrying`), await new Promise((e) => setTimeout(e, z));
 	}
 }
-function le(e) {
+function Ce(e) {
 	let t = (t) => {
-		let n = K.get(e);
+		let n = B.get(e);
 		if (n !== void 0) throw Error(n);
 		self.postMessage({
 			id: e,
@@ -568,54 +666,71 @@ function le(e) {
 			token: t
 		});
 	};
-	return async (e) => {
-		t(e), await H?.yieldIfOverloaded();
+	return async (n) => {
+		await be(e), t(n), await R?.yieldIfOverloaded();
 	};
 }
-var X = p(), Z = X >= m;
-console.log(`[flaxia-node:worker] startup memory probe wasmMemoryBytes=${X} capable=${Z}`);
-function ue() {
-	return n();
+var q = d(), J = q >= f;
+console.log(`[flaxia-node:worker] startup memory probe wasmMemoryBytes=${q} capable=${J}`);
+function Y() {
+	return r();
 }
-var Q = /* @__PURE__ */ new Map(), $ = null;
-function de() {
-	for (let e of Q.values()) try {
+var X = /* @__PURE__ */ new Map(), Z = /* @__PURE__ */ new Map();
+Z.set("swarm-inference", A);
+var Q = null;
+function we() {
+	for (let e of X.values()) try {
 		e();
 	} catch {}
-	Q.clear();
+	X.clear(), Z.clear(), Z.set("swarm-inference", A);
 }
-function fe() {
-	$ && clearTimeout($), $ = setTimeout(() => {
-		$ = null, de();
-	}, se);
+function $() {
+	let e = Date.now(), t = [...Z.values()].flatMap((e) => {
+		try {
+			return e().entries;
+		} catch {
+			return [];
+		}
+	}).slice(0, 8);
+	return {
+		version: 1,
+		sampledAt: e,
+		expiresAt: e + 6e4,
+		entries: t
+	};
+}
+function Te() {
+	Q && clearTimeout(Q), Q = setTimeout(() => {
+		Q = null, we();
+	}, xe);
 }
-async function pe(e, t, n, r) {
+async function Ee(e, t, n, r) {
 	switch (e) {
 		case "ai-inference":
 			let i = await import("./assets/ai-inference.js");
-			return Q.set(e, i.releaseCache), await i.handleAiInference(t, n);
+			return X.set(e, i.releaseCache), Z.set(e, i.getModelCacheSnapshot), await i.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, r?.fileSourceOrigins ?? [], r?.containerImageOrigins ?? []);
 		case "vector-embed":
 			let a = await import("./assets/vector-embed.js");
-			return Q.set(e, a.releaseCache), await a.handleVectorEmbed(t);
+			return X.set(e, a.releaseCache), await a.handleVectorEmbed(t);
 		case "vector-store":
 			let o = await import("./assets/vector-store.js");
-			return Q.set(e, o.releaseCache), await o.handleVectorStore(t);
+			return X.set(e, o.releaseCache), await o.handleVectorStore(t);
 		case "vector-query":
 			let s = await import("./assets/vector-query.js");
-			return Q.set(e, s.releaseCache), await s.handleVectorQuery(t);
+			return X.set(e, s.releaseCache), await s.handleVectorQuery(t);
 		case "nudenet":
 			let c = await import("./assets/nudenet.js");
-			return Q.set(e, c.releaseCache), await c.handleNudeNet(t);
+			return X.set(e, c.releaseCache), await c.handleNudeNet(t);
 		default: throw Error(`Unknown workload type: ${e}`);
 	}
 }
-function me(e, t, n) {
+function De(e, t, n) {
 	return new Promise((r, i) => {
-		let a = new S({
+		let a = new oe({
 			initial: t,
-			runtime: z(),
+			runtime: L(),
 			sendControl: (t) => self.postMessage({
 				id: e,
 				type: "swarm-message",
@@ -628,66 +743,74 @@ function me(e, t, n) {
 			}, [t]),
 			emitToken: n,
 			onDone: (e) => {
-				q === a && (q = null, J = null), r(e);
+				W === a && (W = null, G = null), r(e);
 			},
 			onError: (e) => {
-				q === a && (q = null, J = null), i(Error(e));
+				W === a && (W = null, G = null), i(Error(e));
 			}
 		});
-		q = a, J = e, a.start();
+		W = a, G = e, a.start();
 	});
 }
 self.onmessage = async (e) => {
 	let t = e.data;
+	if (t?.type === "token-flow") {
+		let e = typeof t.id == "string" ? t.id : "";
+		e && K === e && typeof t.paused == "boolean" && (t.paused ? V.add(e) : U(e));
+		return;
+	}
 	if (t?.type === "swarm-control") {
-		if (!V(t)) {
+		if (!_e(t)) {
 			console.warn("[flaxia-node:worker] dropped malformed swarm-control message");
 			return;
 		}
-		q && q.handleControl(t.message);
+		W && W.handleControl(t.message);
 		return;
 	}
 	if (t?.type === "swarm-frame") {
-		t.frame instanceof ArrayBuffer ? q?.handleFrame(t.frame) : console.warn("[flaxia-node:worker] dropped malformed swarm-frame message");
+		t.frame instanceof ArrayBuffer ? W?.handleFrame(t.frame) : console.warn("[flaxia-node:worker] dropped malformed swarm-frame message");
 		return;
 	}
 	if (t?.type === "abort") {
 		let e = t.id, n = typeof t.reason == "string" && t.reason ? t.reason : "aborted by the coordinator";
-		q && J === e ? q.abort(n) : Y === e && K.set(e, n);
+		U(e), W && G === e ? W.abort(n) : K === e && B.set(e, n);
 		return;
 	}
 	let { id: n, workload: r, payload: i, config: a } = t, o = performance.now();
-	Y = n;
+	K = n, U(n);
 	try {
-		if (H ||= new v(a?.maxCpuLoad), console.log(`[flaxia-node:worker] task start id=${n} workload=${r}`), !Z) throw Error(`node incapable: insufficient WASM memory (${X} bytes < ${m}); rejecting all tasks`);
-		if (u.has(r) && !ue()) throw Error(`${r}: insufficient WASM memory (${X} bytes < ${m})`);
-		await ce(), H.beginWork();
+		if (R ||= new te(a?.maxCpuLoad), console.log(`[flaxia-node:worker] task start id=${n} workload=${r}`), !J) throw Error(`node incapable: insufficient WASM memory (${q} bytes < ${f}); rejecting all tasks`);
+		if (l.has(r) && !Y()) throw Error(`${r}: insufficient WASM memory (${q} bytes < ${f})`);
+		await Se(), R.beginWork();
 		let e = setInterval(() => {
 			self.postMessage({
 				id: n,
 				type: "heartbeat",
-				cpuLoad: H?.lastMeasuredLoad ?? 0
+				cpuLoad: R?.lastMeasuredLoad ?? 0,
+				modelCache: $()
 			});
 		}, 1e4), t = performance.now();
 		try {
-			let e = le(n), t = Array.isArray(a?.fileSourceOrigins) ? a.fileSourceOrigins.filter((e) => typeof e == "string") : [], s = Array.isArray(a?.containerImageOrigins) ? a.containerImageOrigins.filter((e) => typeof e == "string") : [], c = r === "swarm-inference" ? await me(n, i, e) : await pe(r, i, e, {
+			let e = Ce(n), t = Array.isArray(a?.fileSourceOrigins) ? a.fileSourceOrigins.filter((e) => typeof e == "string") : [], s = Array.isArray(a?.containerImageOrigins) ? a.containerImageOrigins.filter((e) => typeof e == "string") : [], c = r === "swarm-inference" ? await De(n, i, e) : await Ee(r, i, e, {
 				fileSourceOrigins: t,
 				containerImageOrigins: s
 			});
 			self.postMessage({
 				id: n,
 				type: "done",
-				result: c
+				result: c,
+				modelCache: $()
 			}), console.log(`[flaxia-node:worker] task done id=${n} workload=${r} durationMs=${Math.round(performance.now() - o)}`);
 		} finally {
-			Y = null, K.delete(n), clearInterval(e), H.markTaskComplete(performance.now() - t), fe();
+			K = null, B.delete(n), U(n), clearInterval(e), R.markTaskComplete(performance.now() - t), Te();
 		}
 	} catch (e) {
 		let t = e instanceof Error ? e.message : String(e);
-		console.error(`[flaxia-node:worker] task error id=${n} workload=${r} error=${t}`), Y === n && (Y = null), self.postMessage({
+		console.error(`[flaxia-node:worker] task error id=${n} workload=${r} error=${t}`), K === n && (K = null), U(n), self.postMessage({
 			id: n,
 			type: "error",
-			error: t
+			error: t,
+			modelCache: $()
 		});
 	}
 };