- Shipped
- October 22, 2026 at 9:37 PM UTC
- Author
- Kamo
- Commit
- 8691751
Measured on the nodes' CPU class with the dictionary sync's own traffic (15-key chunks): one stream of eight threads did 1.5 keys/s, two of eight 1.8, four of four 2.9 with four chunks in flight. Small batches do not use eight threads well, so the engine now runs four streams of four (same 16 cores) and accepts eight requests. Requests marked priority 'low' (a sync, a sweep) are held to MAX_LOW_CONCURRENT=3 of the four streams, so a chat message still finds one free. Beam widths are unchanged: this does not trade quality for speed.