- Shipped
- October 22, 2026 at 4:28 PM UTC
- Author
- Kamo
- Commit
- 41f32f3
Replaces LibreTranslate (Argos), Bergamot and Kamo IndicTrans, whose language sets did not overlap and whose gaps came back as English. MADLAD-400-3B-MT (Apache-2.0, int8 CTranslate2 by Santhosh Thottingal) translates between any two of ~400 languages, so inbound chat in any language now reaches English too. - 8 cores / 16 GiB reserved, up to 16 cores / 32 GiB, as asked. Two streams of eight threads: sixteen threads on one stream measured slower (83 vs 123 chars/s) on the nodes' CPU class (Haswell, AVX2). - The API takes lists: 16 short strings in one call took 9.7 s against 25.8 s one by one. - Every input ends with </s>; without it a short label ran on (Save came back three times, and in Arabic as 'Save 10% of the order value'). The beam follows the length of the piece. - Placeholders go to the model as {0}, not KMPH0: lost in 0 of 60 strings in each of ar, hi, sw and zh against 11, 6, 6 and 4. A reply that still loses one is asked again wider, then translated around it. - Same pattern as indictrans/: hash-pinned model and wheels fetched by a Job, serving pod has no egress, auth proxy in front, read-only root, k3m1. The old engines keep running and are removed in the next commit, once TranslateService speaks to this one.
