Text Generation
Transformers
Chinese
English
character-level
chinese
gru
mamba
Mixture of Experts
titans
Instructions to use masterofthestars/tianxuan with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use masterofthestars/tianxuan with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="masterofthestars/tianxuan")# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("masterofthestars/tianxuan", device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use masterofthestars/tianxuan with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "masterofthestars/tianxuan" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "masterofthestars/tianxuan", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/masterofthestars/tianxuan
- SGLang
How to use masterofthestars/tianxuan with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "masterofthestars/tianxuan" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "masterofthestars/tianxuan", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "masterofthestars/tianxuan" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "masterofthestars/tianxuan", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }' - Docker Model Runner
How to use masterofthestars/tianxuan with Docker Model Runner:
docker model run hf.co/masterofthestars/tianxuan
File size: 19,425 Bytes
db7bc06 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 | import { Engine, GenerateOptions, GenerateResult, CharRoute } from "../ml/engine";
import type { GRUState } from "../ml/engine";
import { createStation } from "../station";
import { Task, TaskResult } from "../station/types";
import { chunkText } from "../ml/data";
import { reflect } from "../ml/reflection";
import { mulberry32 } from "../ml/rng";
import { ContextWindow } from "../ml/tool_system";
import { ConfidenceFloors } from "../ml/floors";
import { TitansMemory } from "../ml/titans";
import { L0Consistency } from "../station/l0probe";
// SessionState holds all mutable state per session
interface SessionState {
// Core generation state
state: GRUState;
recentChars: string[];
memory: Map<string, string>;
titansMem: TitansMemory;
floors: ConfidenceFloors;
l0Probe: L0Consistency;
ctxWinWindow: ContextWindow;
habit: () => number;
options: { temperature: number; topK: number };
ctxWin: number[];
}
/**
* SessionEngine - 每个会话独立的引擎实例
* 包装一个共享的 Engine 实例(只读部分),维护每个会话自己的 mutable state
* 解决并发请求共享可变引擎状态问题
*
* [P0-1 修复 2026-10-13 第一性审核] 原实现存在 9 处编译断裂:
* 1. import GRUState 自 engine (未导出) → 现 engine 已 re-export
* 2. import BeautyOfWords/Characterizer 自不存在的 ../ml/beauty_of_words / characterizer → 删 (死引用)
* 3. import TitansMemoryConfig (titans.ts 只导出 TitansConfig) → 现 titans.ts 补别名
* 4. import { l0Probe } (l0probe 模块无实例导出) → 现 l0probe 补共享实例; 会话侧用 per-session L0Consistency
* 5. class 字段在 baseEngine 之前初始化 (TS2729) → 现改 getter 延迟读取
* 6. 直接访问 Engine 私有成员 (store/cnStore/mathStore/worldModel/bigramStats/resetSpike) → 现走公开 getter / multiNeuro 直调
* 7. advance(id, advDepth, ss) 参数顺序错 (签名 advance(chId, ss, depth)) → 现 advance(id, ss, advDepth)
* 8. routes 类型与 CharRoute.memorySource 窄化不符 → 现直用 CharRoute[]
* 9. baseEngine.resetSpike 私有 → 现 baseEngine.multiNeuro?.resetSpikeSeq() 直调
*
* [P0-2 修复 2026-10-13] 会话工作站单元注入:
* 旧实现会话 station 只有 SEED 种子 (text 链零单元) → executor 恒「零单元」分支,
* MoD 路由/审计/经济性在服务器路径死代码 (真实输出全来自 peekLogits(4) 兜底, 审计恒 depth=0)。
* 现构造时 baseEngine.registerUnitsTo(station.registry, station.economics),
* 把与共享 Engine 完全相同的计算单元 (gru/mamba/attn/cnn/snn/titans/world/vision/speech/output)
* 注入会话工作站 → 服务器路径多神经协同真跑 (宪章第1/2条)。
* 会话黑板键: 带 task.id (= `${i}:${ctxFp}`) 与 mem:${ctxFp}, 每会话独立编号+上下文指纹, 无串读。
*/
export class SessionEngine {
/** 会话 ID */
public readonly sessionId: string;
/**
* [P0-1 修复 2026-10-13] 每会话专用 Station — 旧实现为类字段初始化器
* `station = createStation()` (TS 字段初始化器在 baseEngine 参数赋值前执行,
* 且与旧字段 `model = this.baseEngine.model` 同列, 构造即崩溃 "reading 'model' of undefined")。
* 现改延迟 getter: 首次访问时创建, 构造时 registerUnitsTo 注入共享引擎单元。
*/
private _station: import("../station").Station | null = null;
public get station(): import("../station").Station {
if (!this._station) this._station = createStation();
return this._station;
}
/** 每会话可变状态 */
private sessionState: SessionState;
// [P0-1] 共享只读资源改 getter 延迟读取 (旧 class 字段在 baseEngine 之前初始化 → TS2729)
private get model() { return this.baseEngine.model; }
private get tokenizer() { return this.baseEngine.tokenizer; }
private get bigramStats() { return this.baseEngine.bigramStatsGetter; }
constructor(private baseEngine: Engine, sessionId: string) {
this.sessionId = sessionId;
// [P0-2] 注入工作站单元 (titans/gru/mamba/attn/cnn/snn/world/vision/speech/output + 经济学上界)
this.baseEngine.registerUnitsTo(this.station.registry, this.station.economics);
this.sessionState = this.createFreshState();
}
private createFreshState(): SessionState {
const cfg = this.model.cfg;
return {
state: this.model.newState(),
recentChars: [],
memory: new Map(),
titansMem: new TitansMemory({ dim: cfg.emb, slots: 256 }),
floors: new ConfidenceFloors(),
l0Probe: new L0Consistency(),
ctxWinWindow: new ContextWindow(cfg.ctx * 10),
habit: mulberry32(Date.now() % 2147483647),
options: { temperature: 0.1, topK: 3 },
ctxWin: [],
};
}
/** 重置会话状态(用于超时清理) */
reset(): void {
this.sessionState = this.createFreshState();
}
// ==================== 核心生成逻辑 - 复制 Engine.generate 并适配 per-session ====================
async generate(opts: GenerateOptions): Promise<GenerateResult> {
const { prompt, maxLen, budget } = opts;
const ss = this.sessionState;
// 重置探针、选项
ss.l0Probe.reset();
ss.options = { temperature: opts.temperature, topK: opts.topK };
if (opts.seed !== undefined) ss.habit = mulberry32(opts.seed);
// 重置状态
ss.state = this.model.newState();
// SNN 膜电位重置(如果有)
// [P0-1] baseEngine.resetSpike 为私有 → 直接走 multiNeuro.resetSpikeSeq
const spike = this.baseEngine.multiNeuro?.spike;
if (spike) {
this.baseEngine.multiNeuro!.resetSpikeSeq();
this.baseEngine.multiNeuro!.setSpikeInferenceMode(true);
}
const promptIds = this.tokenizer.encode(prompt);
for (const id of promptIds) this.advance(id, ss);
ss.ctxWin = promptIds.slice(-this.model.cfg.ctx);
const routes: CharRoute[] = [];
let text = prompt;
for (let i = 0; i < maxLen; i++) {
const ctx = ss.ctxWin.slice();
const ctxFp = SessionEngine.ctxFingerprint(ctx, this.model.cfg.vocabSize);
const task: Task = { id: `${i}:${ctxFp}`, type: "text", input: ctx, complexity: Math.min(1, 0.4 + (i / Math.max(maxLen, 1)) * 0.4), budget };
// [P0-2] 会话侧召回: 重复缓存 (会话内) + 共享向量库 (只读) — 与 Engine.recallAt 同契约,
// 命中写 mem:${ctxFp} → 会话 executor L0 分支真触发 (旧会话 recallAt 只查内存缓存, 向量库召回缺失)
const recalled = this.recallAt(ctx, ss);
if (recalled) this.station.blackboard.write(`mem:${ctxFp}`, recalled.char);
ss.l0Probe.noteTask();
if (recalled) {
const deep = this.deepTop1At(ss);
ss.l0Probe.record({
ctxPreview: this.tokenizer.decode(ctx).slice(-4),
chL0: recalled.char,
chDeep: deep.char,
deepP: deep.p,
source: recalled.source,
});
}
// L0 记忆直出 / 深链路由 — 会话 executor 真跑 (P0-2 注入单元后)
let ch = "";
let memoryHit = false;
const res: TaskResult = await this.station.executor.execute(task);
memoryHit = res.memoryHit ?? false;
if (res.memoryHit && res.depth === 0 && res.output !== undefined && res.output !== "") {
ch = String(res.output);
const hitId = this.tokenizer.idOf(ch);
if (hitId !== undefined) {
this.advance(hitId, ss);
ss.ctxWin.push(hitId);
ss.ctxWin = ss.ctxWin.slice(-this.model.cfg.ctx);
ss.ctxWinWindow.add({ type: "char", char: ch, id: hitId }, 1);
ss.titansMem.write(Float64Array.from(this.model.embedAvg(ctx)), ch);
}
}
if (ch === "" && res.depth > 0) {
const zLogits = this.station.blackboard.read(`z:logits:${task.id}`) as Float64Array | undefined;
if (zLogits) {
const spikeLogits = this.station.blackboard.read(`z:spike:${task.id}`) as Float64Array | undefined;
if (spikeLogits && spikeLogits.length === zLogits.length) {
for (let k = 0; k < zLogits.length; k++) zLogits[k] += spikeLogits[k];
}
const probs = this.model.logitsToProbs(zLogits);
ch = this.tokenizer.charOf(this.sampleToken(this.penalized(probs), ss.options.topK, ss.options.temperature, ss.habit));
}
}
if (ch === "") {
const fallbackLogits = this.peekLogits(4, ss);
const probs = this.model.logitsToProbs(fallbackLogits);
ch = this.tokenizer.charOf(this.sampleToken(this.penalized(probs), ss.options.topK, ss.options.temperature, ss.habit));
}
// 路由深度 → 真实状态推进深度 (MoD 算力与路由结果挂钩, 同 Engine.generate P0-D 纪律)
const advDepth = (res.depth >= 1 && res.depth <= 4 ? res.depth : 4) as 1 | 2 | 3 | 4;
// 语义反思层
const reflection = reflect(ch, text, {
minSentenceLen: 8,
maxSentenceLen: 30,
maxRepeatCount: 2,
bigramThreshold: 0.001,
punctWeight: 5.0,
enabled: true,
}, this.bigramStats ?? undefined);
if (reflection.action === 'force_end') {
ch = '。';
} else if (reflection.action === 'reject') {
ch = reflection.char;
}
const id = this.tokenizer.idOf(ch);
if (id !== undefined) {
// [P0-1] 旧 advance(id, advDepth, ss) 参数顺序错 → advance(chId, ss, depth)
this.advance(id, ss, advDepth);
ss.ctxWin.push(id);
ss.ctxWin = ss.ctxWin.slice(-this.model.cfg.ctx);
ss.ctxWinWindow.add({ type: "char", char: ch, id }, 1);
ss.titansMem.write(Float64Array.from(this.model.embedAvg(ctx)), ch);
}
// [P0-1] routes 直用 CharRoute[] (旧自定义类型 memorySource: string 与窄化联合不符)
routes.push({
char: ch,
depth: res.depth,
confidence: res.confidence,
units: res.units,
memoryHit,
memorySource: recalled?.source,
});
this.propose(ch, ss);
text += ch;
if (!memoryHit && id !== undefined) ss.memory.set(ctx.join(","), ch);
if (ss.memory.size > 600) {
const first = ss.memory.keys().next().value;
if (first !== undefined) ss.memory.delete(first);
}
}
// 记住生成内容 - 写入到共享向量库 (带 session 前缀防自举循环; 宪章第10条: 共享库只写本会话命名空间)
const sharedStore = this.baseEngine.vectorStore;
if (sharedStore) {
this.remember(text.substring(prompt.length));
}
// SNN 膜电位恢复 (生成结束)
if (spike) {
this.baseEngine.multiNeuro!.setSpikeInferenceMode(false);
}
// 写入探针和经济性 (会话各自 30s 写盘节律)
this.station.probe.writeToFile();
this.station.economics.report();
return { text, routes, budget };
}
// ==================== 辅助方法 ====================
private peekLogits(depth: 1 | 2 | 3 | 4, ss: SessionState): Float64Array {
const lastId = ss.ctxWin.length > 0 ? ss.ctxWin[ss.ctxWin.length - 1] : 0;
const copy: GRUState = {
h1: ss.state.h1.slice(),
h2: ss.state.h2.slice(),
hist: ss.state.hist.map((h) => h.slice()),
ids: [...ss.state.ids],
};
return this.model.step(lastId, copy, depth);
}
private advance(chId: number, ss: SessionState, depth: 1 | 2 | 3 | 4 = 4): void {
this.model.step(chId, ss.state, depth);
if (this.baseEngine.multiNeuro?.spike) {
this.baseEngine.multiNeuro.spikeAdvance(chId);
}
}
private sampleToken(probs: Float64Array, topK: number, temperature: number, rand: () => number): number {
for (let i = 0; i < probs.length; i++) if (Number.isNaN(probs[i])) return -1;
const v = probs.length;
const scaled = new Float64Array(v);
let sum = 0;
for (let k = 0; k < v; k++) {
scaled[k] = Math.exp(Math.log(probs[k] + 1e-12) / Math.max(temperature, 1e-3));
sum += scaled[k];
}
const k = Math.min(topK, v);
const idxs = Array.from({ length: v }, (_, i) => i).sort((a, b) => scaled[b] - scaled[a]).slice(0, k);
let mass = 0;
for (const i of idxs) mass += scaled[i];
let r = rand() * mass;
for (const i of idxs) {
r -= scaled[i];
if (r <= 0) return i;
}
return idxs[idxs.length - 1];
}
private penalized(probs: Float64Array): Float64Array {
const out = probs.slice();
out[out.length - 1] = 0;
const freq = new Map<string, number>();
for (const ch of this.sessionState.recentChars) freq.set(ch, (freq.get(ch) ?? 0) + 1);
for (const [ch, count] of freq) {
if (count < 2) continue;
const id = this.tokenizer.idOf(ch);
if (id !== undefined && id < out.length) out[id] = 0;
}
let sum = 0;
for (let i = 0; i < out.length; i++) sum += out[i];
for (let i = 0; i < out.length; i++) out[i] /= sum;
return out;
}
private propose(ch: string, ss: SessionState): void {
ss.recentChars.push(ch);
if (ss.recentChars.length > 16) ss.recentChars.shift();
}
private isDegenerate(ch: string): boolean {
const n = this.sessionState.recentChars.length;
if (n >= 3) {
const last3 = this.sessionState.recentChars.slice(-3).join("");
if (last3 === ch.repeat(3)) return true;
}
if (n >= 16) {
const a = this.sessionState.recentChars.slice(-8).join("");
const b = this.sessionState.recentChars.slice(-16, -8).join("");
if (a === b) return true;
}
return false;
}
/**
* [P0-2] 会话侧召回: 重复缓存 (会话内) + 共享向量库 (只读, MEMORY_TAG 前缀隔离防自举)。
* 命中写点火标记 z:l0:act → 会话 executor L0 分支读 z:l0:act 推导置信度 (与 Engine.recallAt 同契约)。
*/
private recallAt(ctxIds: number[], ss: SessionState): { char: string; source: "vector" | "repeat" | "titans" } | null {
const cached = ss.memory.get(ctxIds.join(","));
if (cached !== undefined && !this.isDegenerate(cached)) {
this.station.blackboard.write("z:l0:act", 0.5);
return { char: cached, source: "repeat" };
}
// 共享向量库召回 (只读; MEMORY_TAG + session 前缀 → 检索跳过防自举循环)
const store = this.baseEngine.vectorStore;
if (store) {
const q = this.tokenizer.decode(ctxIds);
const hits = store.search(this.model.embedAvg(ctxIds), 8);
for (const hit of hits) {
if (hit.text.startsWith("\uE000")) continue;
const at = hit.text.indexOf(q);
if (at >= 0 && at + q.length < hit.text.length) {
const ch = hit.text[at + q.length];
if (!this.isDegenerate(ch)) {
this.station.blackboard.write("z:l0:act", 0.5);
return { char: ch, source: "vector" };
}
}
}
}
// 会话内 Titans 在线记忆召回 (与 Engine.recallAt 第三路同机制)
const q = Float64Array.from(this.model.embedAvg(ctxIds));
const cache = { query: new Float64Array(0), keys: new Float64Array(0), values: new Float64Array(0), weights: new Float64Array(0) };
const mem = ss.titansMem.read(q, cache);
if (mem.hitSlots > 0 && cache.weights.length > 0) {
let best = 0;
for (let i = 1; i < cache.weights.length; i++) if (cache.weights[i] > cache.weights[best]) best = i;
if (cache.weights[best] > 0.5) {
const ch = ss.titansMem.chars[best % ss.titansMem.cfg.slots];
if (ch !== undefined && ch !== "" && !this.isDegenerate(ch)) {
const l0Act = mem.hitSlots / ss.titansMem.cfg.slots;
this.station.blackboard.write("z:l0:act", l0Act);
return { char: ch, source: "repeat" };
}
}
}
return null;
}
private deepTop1At(ss: SessionState): { char: string; p: number } {
const logits = this.peekLogits(4, ss);
const probs = this.model.logitsToProbs(logits);
let best = 0;
for (let i = 1; i < probs.length; i++) if (probs[i] > probs[best]) best = i;
return { char: this.tokenizer.charOf(best), p: probs[best] };
}
private remember(text: string): void {
const store = this.baseEngine.vectorStore;
if (!store) return;
for (const chunk of chunkText(text, 64, 32)) {
if (chunk.length < 8) continue;
const ids = this.tokenizer.encode(chunk);
// [P0-1] 向量库 add 接受 number[] (vectorstore.ts 签名); multiNeuro 在则用其 embedAvg, 否则空向量占位
const vec: number[] = this.baseEngine.multiNeuro
? this.baseEngine.multiNeuro.embedAvg(ids)
: [];
store.add(`\uE000${this.sessionId}:${chunk}`, vec);
}
}
// 静态工具方法
static ctxFingerprint(ctx: number[], vocabSize: number): string {
let h = 2166136261;
for (let i = 0; i < ctx.length; i++) {
h ^= ctx[i] % vocabSize;
h = Math.imul(h, 16777619);
}
return (h >>> 0).toString(16).padStart(8, "0");
}
// ==================== 查询方法 (透传共享引擎) ====================
retrieve(q: string, k = 4): Array<{ text: string; score: number }> {
return this.baseEngine.retrieve(q, k);
}
retrieveCN(q: string, k = 4, category?: string): Array<{ text: string; score: number; category?: string }> {
return this.baseEngine.retrieveCN(q, k, category);
}
retrieveMath(q: string, k = 6, category?: string): Array<{ expr: string; description: string; score: number; matchType: string }> {
return this.baseEngine.retrieveMath(q, k, category);
}
reason(q: string) {
return this.baseEngine.reason(q);
}
retrieveCount(): number {
return this.baseEngine.retrieveCount();
}
totalParamCount(): number {
return this.baseEngine.totalParamCount();
}
}
/** 会话管理器:管理多个 SessionEngine,处理会话生命周期和队列 */
export class SessionManager {
private sessions = new Map<string, SessionEngine>();
private static readonly SESSION_TIMEOUT = 30 * 60 * 1000; // 30 分钟
private cleanupTimer: NodeJS.Timeout | null = null;
constructor(private defaultEngine: Engine) {}
getSession(sessId?: string): SessionEngine {
if (!sessId) sessId = this.generateSessionId();
let session = this.sessions.get(sessId);
if (!session) {
session = new SessionEngine(this.defaultEngine, sessId);
this.sessions.set(sessId, session);
this.startCleanup();
}
return session;
}
endSession(sessId: string): void {
const session = this.sessions.get(sessId);
if (!session) return;
session.reset();
this.sessions.delete(sessId);
if (this.sessions.size === 0 && this.cleanupTimer) {
clearInterval(this.cleanupTimer);
this.cleanupTimer = null;
}
}
/** [P0-1] 旧 private 但 server/index.ts /api/session 直接调用 → 放宽为 public */
generateSessionId(): string {
return "session-" + Math.random().toString(36).substring(2, 15) + Math.random().toString(36).substring(2, 15);
}
private startCleanup(): void {
if (this.cleanupTimer) return;
this.cleanupTimer = setInterval(() => this.cleanupStale(), SessionManager.SESSION_TIMEOUT / 2);
}
private cleanupStale(): void {
for (const id of this.sessions.keys()) {
this.endSession(id);
}
}
getStats(): { active: number; totalParams: number } {
return { active: this.sessions.size, totalParams: this.defaultEngine.totalParamCount() };
}
}
|