Files
tonemark/src/lib/server/whisper.ts
Giancarmine Salucci 1072679360
All checks were successful
Build & Push Docker Image / test (push) Successful in 12s
Build & Push Docker Image / build-and-push (push) Successful in 52s
fix(whisper): handle model warmup events
- Ignore backend model lifecycle webhooks so model warmup does not
  mark jobs done early
- Parse batched SSE messages and relay model load states during
  submit retries

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-05-15 00:08:32 +02:00

239 lines
7.5 KiB
TypeScript

import type { ModelStateTag, ModelStatus } from '$lib/types.js';
const MODEL_STATES = new Set<ModelStateTag>(['unloaded', 'loading', 'waiting_for_gpu', 'ready']);
function isModelStateTag(value: unknown): value is ModelStateTag {
return typeof value === 'string' && MODEL_STATES.has(value as ModelStateTag);
}
function extractSseMessages(buffer: string): { messages: { eventType: string; data: string }[]; rest: string } {
const normalized = buffer.replace(/\r/g, '');
const chunks = normalized.split('\n\n');
const rest = chunks.pop() ?? '';
const messages = chunks
.map((chunk) => {
let eventType = '';
const dataLines: string[] = [];
for (const line of chunk.split('\n')) {
if (line.startsWith('event:')) {
eventType = line.slice(6).trim();
} else if (line.startsWith('data:')) {
dataLines.push(line.slice(5).trim());
}
}
return { eventType, data: dataLines.join('\n') };
})
.filter((message) => message.data.length > 0);
return { messages, rest };
}
function whisperUrl() {
return process.env.WHISPER_URL ?? 'http://localhost:8080';
}
/** Get the current model state from whisper-rtx2080. */
export async function getModelStatus(): Promise<ModelStatus> {
const { default: fetch } = await import('node-fetch');
const res = await fetch(`${whisperUrl()}/model/status`, {
signal: AbortSignal.timeout(5000)
});
if (!res.ok) throw new Error(`/model/status returned ${res.status}`);
return res.json() as Promise<ModelStatus>;
}
/**
* Wait for the whisper model to become ready.
*
* Subscribes to /model/events SSE and resolves as soon as a payload with
* state:"ready" arrives. Falls back to a plain timeout (`timeoutMs`) if the
* SSE connection fails or closes without that event, so the retry loop can
* try again without hanging indefinitely.
*/
async function waitForModelReady(
timeoutMs: number,
onStateChange?: (state: ModelStateTag) => void
): Promise<void> {
const { default: fetch } = await import('node-fetch');
const ac = new AbortController();
return new Promise((resolve) => {
let done = false;
const finish = () => {
if (!done) {
done = true;
ac.abort();
resolve();
}
};
const timer = setTimeout(finish, timeoutMs);
fetch(`${whisperUrl()}/model/events`, { signal: ac.signal as AbortSignal })
.then(async (res) => {
if (!res.body) {
clearTimeout(timer);
return finish();
}
let buf = '';
for await (const chunk of res.body) {
if (ac.signal.aborted) break;
buf += chunk.toString();
const { messages, rest } = extractSseMessages(buf);
buf = rest;
for (const message of messages) {
try {
const payload = JSON.parse(message.data) as { state?: unknown };
if (!isModelStateTag(payload.state)) continue;
if (payload.state === 'ready') {
clearTimeout(timer);
finish();
return;
}
onStateChange?.(payload.state);
} catch { /* ignore parse errors */ }
}
}
// Stream ended without model_ready → proceed to retry immediately
clearTimeout(timer);
finish();
})
.catch(() => {
// SSE unreachable — fallback timer will fire eventually
});
});
}
/**
* Submit an audio file to whisper-rtx2080. Returns the whisper job id.
*
* Handles 503 (model not ready) transparently: retries after subscribing to
* /model/events SSE — proceeds as soon as state:"ready" arrives, or after the
* Retry-After timeout elapses (whichever comes first).
* Calls `onModelWaiting` on each 503 so the caller can surface the wait to the user.
*/
export async function submitJob(
wavPath: string,
webhookUrl: string,
language?: string,
onModelWaiting?: (state: ModelStateTag, retryAfterSecs: number) => void,
maxAttempts = 20
): Promise<string> {
const FormData = (await import('form-data')).default;
const { createReadStream } = await import('fs');
const { default: fetch } = await import('node-fetch');
for (let attempt = 1; attempt <= maxAttempts; attempt++) {
// Recreate form with a fresh readable stream on every attempt.
// A consumed ReadStream cannot be rewound, so reusing it across retries
// would send an empty body to whisper after the first 503.
const form = new FormData();
form.append('audio', createReadStream(wavPath));
form.append('task', 'transcribe');
form.append('webhook_url', webhookUrl);
if (language) form.append('language', language);
const res = await fetch(`${whisperUrl()}/jobs`, {
method: 'POST',
body: form,
headers: form.getHeaders()
});
if (res.status === 202) {
const json = (await res.json()) as { job_id: string };
return json.job_id;
}
if (res.status === 503) {
const body = (await res.json().catch(() => ({}))) as {
state?: string;
retry_after_secs?: number;
};
const state = isModelStateTag(body.state) ? body.state : 'unloaded';
const waitSecs = body.retry_after_secs ?? parseInt(res.headers.get('Retry-After') ?? '15');
onModelWaiting?.(state, waitSecs);
let lastState = state;
await waitForModelReady((waitSecs + 1) * 1000, (nextState) => {
if (nextState === lastState) return;
lastState = nextState;
onModelWaiting?.(nextState, waitSecs);
});
continue;
}
const text = await res.text();
throw new Error(`whisper /jobs returned ${res.status}: ${text}`);
}
throw new Error(`Whisper model did not become ready after ${maxAttempts} attempts`);
}
/** Unload the model from VRAM. Throws if the whisper server returns non-ok. */
export async function unloadModel(): Promise<{ ok: boolean }> {
const { default: fetch } = await import('node-fetch');
const res = await fetch(`${whisperUrl()}/model/unload`, {
method: 'POST',
signal: AbortSignal.timeout(10000)
});
if (!res.ok) throw new Error(`/model/unload returned ${res.status}`);
return res.json() as Promise<{ ok: boolean }>;
}
/**
* Cancel a queued or running job on the whisper server (best-effort).
* Errors are silently ignored — local job status is already set to cancelled.
*/
export async function cancelJob(whisperJobId: string): Promise<void> {
try {
const { default: fetch } = await import('node-fetch');
await fetch(`${whisperUrl()}/jobs/${whisperJobId}`, {
method: 'DELETE',
signal: AbortSignal.timeout(5000)
});
} catch { /* best-effort */ }
}
/** Open an SSE stream from whisper and call onProgress/onDone callbacks. */
export async function streamJob(
whisperJobId: string,
onProgress: (percent: number, chunk: number, total: number) => void,
onDone: () => void,
onError: (msg: string) => void
): Promise<void> {
const { default: fetch } = await import('node-fetch');
const res = await fetch(`${whisperUrl()}/jobs/${whisperJobId}/stream`);
if (!res.ok || !res.body) throw new Error(`SSE stream returned ${res.status}`);
let buf = '';
for await (const chunk of res.body) {
buf += chunk.toString();
const { messages, rest } = extractSseMessages(buf);
buf = rest;
for (const message of messages) {
try {
const payload = JSON.parse(message.data);
if (payload.type === 'progress') {
onProgress(payload.percent ?? 0, payload.chunk ?? 0, payload.total ?? 0);
} else if (payload.type === 'done') {
onDone();
return;
} else if (payload.type === 'error') {
onError(payload.message ?? 'unknown error');
return;
}
} catch { /* ignore parse errors */ }
}
}
}
/** Check if the whisper server is healthy. */
export async function checkHealth(): Promise<boolean> {
try {
const { default: fetch } = await import('node-fetch');
const res = await fetch(`${whisperUrl()}/health`, { signal: AbortSignal.timeout(3000) });
return res.ok;
} catch {
return false;
}
}