HiveOS/SMOS launcher fix + docs. Rust binary unchanged from v0.6.9.0.
- HiveOS/SMOS now default to per-card AUTO. With blank flight-sheet extra args, each GPU loads the heaviest model its own VRAM can hold (mixed rigs get a different model per card = max reward). Previously the launcher forced
--light(Gemma) on every card. Pass a single tier flag (--very-high, …) to instead pin all cards to one tier. --force-model <csv>supported on HiveOS/SMOS — set a model per GPU in CUDA device order (e.g.--force-model very-high,light,default); unlisted cards fall back to AUTO. Previously it got a spurious--lightprepended.- Docs updated (README / FLIGHT_SHEET / mmpos) with the current model lineup (
very-lightQwen3-1.7B →very-highLlama-3.3-70B), AUTO per-card selection, and--force-model. See the README → Model tiers.
mmpos already passed args verbatim (AUTO by default), so no change there. Modern/legacy/pascal × hiveos/linux/smos/mmpos = 12 tarballs.