- fix memory allocation on some low memory cards
- add --tier-mode auto to load the correct model on the cards
- first release for amd vram model load
keryx-miner-supr v0.6.3.4
Two build lines this release — pick by your NVIDIA driver version:
| Line | CUDA | Min driver | Use it if… |
|---|---|---|---|
| legacy | 12.2 | R535+ | older/standard rigs, HiveOS 535.x, RTX 20xx / 3060 / 3070, or GPU inference failed on prior builds. JITs forward to all newer cards incl. Blackwell. Pick this if unsure. |
| modern | 12.9 | R575+ | new drivers, RTX 50xx / Blackwell — native sm_120 cuBLAS (fastest inference on the newest cards). |
Each line ships four packages (same binary, different integration glue):
keryx-miner-supr-0.6.3.4-<line>.tar.gz— HiveOS custom minerkeryx-miner-supr-smos_0.6.3.4-<line>.tar.gz— SMOS (HiveOS layout)keryx-miner-supr-mmpos_0.6.3.4-<line>.tar.gz— mmpOSkeryx-miner-supr-0.6.3.4-<line>-linux-x86_64.tar.gz— generic Linux (dynamic;./run.sh)
All bundle the matching CUDA runtime in ./lib, so only the NVIDIA driver is required (no CUDA toolkit install). The bundled cuBLAS is what sets the driver floor (legacy 12.2 → 535, modern 12.9 → 575).
Why two lines (the "one build vs many" question)
The PoW walk uses the CUDA driver API (works on any R525+). OPoI inference uses cuBLAS (CUDA runtime), which enforces a minimum driver tied to the cuBLAS version we bundle. Linking 12.8/12.9 for native Blackwell speed raised that floor to 575 and broke older-driver rigs. The legacy line follows upstream's approach (build at 12.2 → floor 535, forward-JIT onto newer cards) so it runs on essentially every rig; modern keeps native Blackwell performance for new drivers.
Changes
- Auto GPU→CPU inference fallback (0.6.3.3): if the GPU model load fails (wrong-arch PTX, old driver, missing cuBLAS, OOM), the miner logs a loud warning, switches inference to CPU, and keeps mining (PoW walk stays on GPU) instead of crashing.
- Per-GPU PoM residency (AMD/OpenCL): each card mines independently (was funneling onto device 0).
- Orphan possession-tree sweep: auto-deletes ~5 GB
pom-tree-<pid>.binfiles left by killed/crashed miners (NVIDIA + AMD).
All binaries: glibc ≥ 2.30 (HiveOS-compatible), one process per GPU.