feat(worker): parallel per-model rating batches with resume via state file
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Showing 1 changed file with +32 and −8
modified
apps/worker/src/scripts/rate.ts
+32 −8
@@ -10,8 +10,13 @@ import { ratingJobId } from "../raters/job-id"; | ||
| 10 | 10 | |
| 11 | 11 | // One-shot rating run: submit a Message Batch per model for every task that |
| 12 | 12 | // lacks ratings under the current prompt version, poll to completion, ingest. |
| 13 | +// Models run in PARALLEL (batches are independent server-side queues). | |
| 13 | 14 | // Usage: tsx src/scripts/rate.ts [--limit N] [--poll-seconds S] |
| 14 | −// Idempotent: already-rated tasks are excluded; re-running resumes cleanly. | |
| 15 | +// Idempotent & resumable: already-rated tasks are excluded, and an in-flight | |
| 16 | +// batch per model is remembered in a state file so a relaunch resumes polling | |
| 17 | +// instead of resubmitting (a resubmit would double the spend). | |
| 18 | + | |
| 19 | +import { readFileSync, writeFileSync } from "node:fs"; | |
| 15 | 20 | |
| 16 | 21 | function argValue(flag: string): string | undefined { |
| 17 | 22 | const index = process.argv.indexOf(flag); |
@@ -20,6 +25,19 @@ function argValue(flag: string): string | undefined { | ||
| 20 | 25 | |
| 21 | 26 | const limit = Number(argValue("--limit") ?? "0") || undefined; |
| 22 | 27 | const pollSeconds = Number(argValue("--poll-seconds") ?? "60") || 60; |
| 28 | +const STATE_PATH = process.env.RATE_STATE ?? "/tmp/aix-batch-state.json"; | |
| 29 | + | |
| 30 | +function readState(): Record<string, string> { | |
| 31 | + try { | |
| 32 | + return JSON.parse(readFileSync(STATE_PATH, "utf8")) as Record<string, string>; | |
| 33 | + } catch { | |
| 34 | + return {}; | |
| 35 | + } | |
| 36 | +} | |
| 37 | + | |
| 38 | +function writeState(state: Record<string, string>): void { | |
| 39 | + writeFileSync(STATE_PATH, JSON.stringify(state, null, 2)); | |
| 40 | +} | |
| 23 | 41 | |
| 24 | 42 | const sleep = (ms: number) => new Promise((resolve) => setTimeout(resolve, ms)); |
| 25 | 43 | |
@@ -48,9 +66,15 @@ async function runForModel(model: string): Promise<void> { | ||
| 48 | 66 | ratingTasks.map((task) => [ratingJobId(task.taskId, model, RATER_PROMPT_VERSION), task.taskId]), |
| 49 | 67 | ); |
| 50 | 68 | |
| 51 | − console.log(`[${model}] submitting batch of ${ratingTasks.length} tasks…`); | |
| 52 | − const batchId = await submitRatingBatch(ratingTasks, model); | |
| 53 | − console.log(`[${model}] batch ${batchId} submitted; polling every ${pollSeconds}s`); | |
| 69 | + let batchId = readState()[model]; | |
| 70 | + if (batchId) { | |
| 71 | + console.log(`[${model}] resuming in-flight batch ${batchId}; polling every ${pollSeconds}s`); | |
| 72 | + } else { | |
| 73 | + console.log(`[${model}] submitting batch of ${ratingTasks.length} tasks…`); | |
| 74 | + batchId = await submitRatingBatch(ratingTasks, model); | |
| 75 | + writeState({ ...readState(), [model]: batchId }); | |
| 76 | + console.log(`[${model}] batch ${batchId} submitted; polling every ${pollSeconds}s`); | |
| 77 | + } | |
| 54 | 78 | |
| 55 | 79 | for (;;) { |
| 56 | 80 | await sleep(pollSeconds * 1000); |
@@ -59,6 +83,9 @@ async function runForModel(model: string): Promise<void> { | ||
| 59 | 83 | } |
| 60 | 84 | |
| 61 | 85 | const { ingested, failed } = await ingestBatchResults(batchId, model, jobIdToTaskId); |
| 86 | + const state = readState(); | |
| 87 | + delete state[model]; | |
| 88 | + writeState(state); | |
| 62 | 89 | console.log(`[${model}] batch ${batchId} done: ${ingested} dimension ratings ingested, ${failed} requests failed`); |
| 63 | 90 | } |
| 64 | 91 | |
@@ -69,10 +96,7 @@ async function main(): Promise<void> { | ||
| 69 | 96 | console.log( |
| 70 | 97 | `rating run — prompt ${RATER_PROMPT_VERSION}, models: ${RATER_MODELS.join(", ")}${limit ? `, limit ${limit}` : " (all unrated tasks)"}`, |
| 71 | 98 | ); |
| 72 | − // Sequential per model keeps memory/log output simple; batches themselves are parallel server-side. | |
| 73 | − for (const model of RATER_MODELS) { | |
| 74 | − await runForModel(model); | |
| 75 | − } | |
| 99 | + await Promise.all(RATER_MODELS.map((model) => runForModel(model))); | |
| 76 | 100 | console.log("rating run complete — next: pnpm score:recompute"); |
| 77 | 101 | } |
| 78 | 102 | |
| 79 | 103 | |