v0.10.10 — loud, specific, self-healing model staging
Fixes rigs that sat for up to an hour at preparing models (staging/verifying files) with no indication of what was wrong. The mining-tier model download was fire-once — one "will retry" line, then the task ended and nothing retried, while the reassuring "preparing…" line repeated every 10s and buried the real reason.
What changed
- Downloads retry until they succeed — the mining-tier prefetch now loops with 60s backoff, logging a LOUD
ERRORwith the specific reason each round, so a transient IPFS-gateway / disk problem self-heals and stays visible. - Disk-space preflight — before downloading, the miner checks free space and, if short, aborts naming the exact path, MB needed vs MB free (instead of streaming to ENOSPC and dying on an opaque OS error).
- Errors stay in view — when a rig is stuck with no model ready, the concrete failure is now printed at
ERRORevery cycle (always in the last log lines), not hidden behind the generic "preparing…" spinner. - Corrupt / wrong on-disk model — a downloaded or hand-placed GGUF that is still invalid now tells you to DELETE the exact path and restart so the correct file re-downloads.
- GPU load / self-test failures are surfaced — "can't serve this tier", out-of-memory with no smaller tier, and load-onto-card errors now report the reason with a fix (force a smaller tier
--very-light/--light, or use a card with more VRAM), and clear once loading succeeds. - Download failures also point at
KERYX_IPFS_GATEWAYas the override to try.
No consensus/algorithm change from v0.10.9 — this is purely operator-visibility and robustness.
Assets: linux-x86_64 (dynamic), HiveOS, and MMPOS tarballs, each in modern (575-class driver) and legacy (-legacy, 535-class) lines, plus SHA256SUMS-0.10.10.txt. AMD and Windows builds are attached separately.