Five tweaks take Strata local inference from 35 to 60+ tps on dual RTX 3060

Dreeew84 · reddit · 2026-10-15

The author shares hands-on tuning results for the relatively new Strata local inference framework (docs are hard to navigate — use your agent):

--pcie-frac auto-picking is often wrong; his best value was a counterintuitive 0.2.
Tuning --pool-workers gave an easy +10% decode.
--trim-stage-weights added +8% (may increase swaps later in a session).
Raising --pipeline-windows to 6 gave nearly +15%, at the cost of higher CPU usage.
Multi-GPU users: the 0/1 card order is basically random — his GPU 0 sat on the slower PCIe lane and acted only as a memory pool. Forcing the fast card with "gpu": [1,0] and "gpuorder": "asgiven" yielded another +15%.

Cumulatively, Qwen Flash Next IQ3S went from 35 tps peak decode to 60 tps average code generation with 60-70 tps peaks, and full-context/compaction sessions still run 15-20% above the previous baseline.

Original post →

Five tweaks take Strata local inference from 35 to 60+ tps on dual RTX 3060

More from coding & agent