AI: 게임 시작 때 모델 미리 올리기, 2시간 유지

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
EJClaw
2026-10-11 03:40:08 +09:00
parent c7dab689ca
commit 119d10fac6
2 changed files with 11 additions and 2 deletions

View File

@@ -207,6 +207,8 @@ export class BotBrain {
onGameStart(): void {
this.happened = [];
// A game with computers and people: make sure the model is loaded before anyone talks.
if (this.llm?.available() && this.room.humans().length && this.room.bots().length) void this.llm.warm();
if (!this.enabled() || this.room.gameId() === 'mafia' || this.random() > 0.7) return;
const bots = this.room.bots();
const bot = bots[Math.floor(this.random() * bots.length)];

View File

@@ -83,7 +83,7 @@ export class Llm {
model: this.cfg.model,
messages,
stream: false,
keep_alive: this.cfg.keepAlive ?? '30m',
keep_alive: this.cfg.keepAlive ?? '2h',
...(opts.format ? { format: opts.format } : {}),
options: { temperature: opts.temperature ?? 0.9, num_predict: opts.maxTokens ?? 80 },
}),
@@ -115,8 +115,15 @@ export class Llm {
}
}
/** Loads the model into GPU memory ahead of the first real request (the first load takes ~20 s). */
private lastWarm = -Infinity;
/**
* Loads the model into GPU memory ahead of the first real request (a cold load takes ~20 s; Ollama unloads it after
* `keepAlive` without use). Cheap when it is already loaded; at most once a minute.
*/
async warm(): Promise<boolean> {
if (this.now() - this.lastWarm < 60_000) return true;
this.lastWarm = this.now();
const r = await this.chat([{ role: 'user', content: '안녕' }], { maxTokens: 1, timeoutMs: 60_000 });
return r !== null;
}