📦 Which file do I download?
| Your setup | Download |
|---|---|
| NVIDIA — HiveOS | keryx-miner-supr-0.10.12.tar.gz
|
| NVIDIA — mmpOS | keryx-miner-supr-mmpos_0.10.12.tar.gz
|
| NVIDIA — plain Linux | keryx-miner-supr-0.10.12-linux-x86_64.tar.gz
|
| NVIDIA — Windows | keryx-miner-supr-windows-nvidia-pom.zip
|
| AMD — HiveOS / mmpOS / Linux / Windows | keryx-miner-supr-amd-* / keryx-miner-supr-windows-amd.zip
|
-legacy = older drivers (~535) / older cards; the plain (modern) build is for drivers 575+ (RTX 20–50 incl. 5090). Verify with SHA256SUMS-0.10.12.txt.
v0.10.12 — fix Windows "silent stall / no model download"
Windows NVIDIA rigs connected to the pool and detected the GPU but then sat forever at "no model lineup installed yet (waiting for chain tip)" / 0 h/s "preparing", never downloading a model — while Linux worked fine.
Root cause: lineup staging read GPU VRAM from two inconsistent sources — select_tier_auto() used the nvidia-smi -i 0 CLI, filter_specs_by_vram() used the CUDA driver. On Windows the nvidia-smi -i 0 form fails, so auto-select fell back to the Light tier (GLM-4-9B, 12 GB); the CUDA-driver filter then correctly saw 10 GB on a 3080 and dropped every model → empty lineup → no model → no download → silent stall.
Fixes:
select_tier_autonow reads the CUDA-driver VRAM (the same source the filter uses), so Windows cards resolve the correct tier; the no-VRAM fallback is now VeryLight (smallest, fits any card).filter_specs_by_vramnever returns an empty lineup — if the VRAM gate would drop every model it keeps the lineup so the model still downloads and mining can start.
Builds on v0.10.11 (periodic re-declare so the pool sees served models), v0.10.10 (loud/self-healing staging), v0.10.9 (Gemma fits 16 GB). No consensus/algorithm change.
Assets: HiveOS, mmpOS, and Linux tarballs in modern + legacy lines, Windows (NVIDIA + AMD), and SHA256SUMS. AMD builds attached separately.