From 978743a36ec37e286ea5f9754680ff69feb833ee Mon Sep 17 00:00:00 2001 From: Lucas Winther Date: Fri, 30 Oct 2026 14:36:51 +0100 Subject: [PATCH] Record that batched Opus-MT is still broken on Transformers.js 4.3.2 Co-Authored-By: Claude Opus 5.5 --- docs/WorkBreakdown.md | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/docs/WorkBreakdown.md b/docs/WorkBreakdown.md index 8122e31..3323afa 100644 --- a/docs/WorkBreakdown.md +++ b/docs/WorkBreakdown.md @@ -427,6 +427,13 @@ Found while building transcript translation (D-045), 2026-10-29: every row of a padded batch, so translation runs one sentence per call (~4× slower than batch 8 would be). Re-check after the 4.3 upgrade with `bun scripts/bench-translate.ts` ("output identical" across batch sizes). + **Re-checked on 4.3.2 (2026-10-30): still broken** — a batch of 4 Danish + sentences returned row 0's translation for all 4 (our `translateBatch` + passes the batch straight to the translation pipeline, so it is + upstream). Batch 8 would be ~4× faster (12 s vs 49 s for en_alien.txt). +- [ ] `src/worker/opusMt.ts` still forces `graphOptimizationLevel: "basic"` + (D-014) for Marian; D-043 dropped it for Whisper on ort-web 1.31. Check + whether the q8 Marian decoder now loads at the default level. - [x] The server API still accepts `task: "translate"` with large-v3-turbo (only the UI hides it, D-044). — Fixed: `POST /api/transcribe` and the live WebSocket answer 400 via `translateNotSupported` (src/server/transcribe.ts).