A guard for a crash reported on v1.1.0, and clearer advice on the DFlash2 drafter.
Update: the dashboard's About > Update, or git pull, then ./setup.sh (Windows: START-MAYA.bat). The engine doesn't change, so nothing recompiles and nothing is downloaded again.
What's new
- The setup no longer offers the experimental EXL3 quant on Windows (#129). There an EXL3 model stopped at its first prompt with an illegal memory access, after a 125 GB download.
--model EXL3-3.05bpwstill asks for it, with a warning, while we look into it. Linux: nothing changes. Maya's own models are not affected. - DFlash2: measure yours. Where most experts come from RAM, the MTP block can stay faster than the drafter: an RTX 3090 + 3060 with Maya-M decoded faster without it (#127). The setup and the docs say so;
--benchtells.
Checked
The installer tests, with a new one for the Windows model list; the GitHub checks pass. The engine is v1.1.0's.