Builds on v0.6.4 (the required H2 tier fix). Multi-GPU improvements — no flags needed:
- Per-device PoM: each GPU builds and mines its own model on its own device (no more
CUDA_VISIBLE_DEVICESworkaround, no GPU-0 OOM on multi-GPU rigs). - OPoI inference runs on the largest-VRAM GPU automatically (mixed rigs: the big card serves inference while small cards mine). CPU inference only as an emergency fallback.
Validated: mixed rigs (5090+170HX, 2×3070+2×5080) mine on every GPU with accepted tier-1 shares, 0 OOM. Reports version 0.6.5 to the pool.
Lines (pick your OS tarball; verify with SHA256SUMS): modern (CUDA 12.x, R575+), legacy (CUDA 12.4, R550+), pascal (sm_60, GTX 10-series — run with --cpu-inference). AMD/Windows separate.