ocminer/keryx-miner-supr v0.10.7

on GitHub

Inference-integrity + VRAM-fit release. Enforces the network's core rule — never declare or mine a tier this rig cannot actually serve — and fixes the 16 GB device miner build failed: OUT_OF_MEMORY spin.

Serve what you declare

• Inference self-test gate — a model must PASS one tiny real generation before it is declared to the pool or mined. A model that can't load/generate is withdrawn from ai:cap and its tier is not mined (mining a tier we can't serve would strike the pool). A live OPoI answer also counts as proof. The self-test runs on the serving GPU, so mixed rigs are covered (e.g. a 3070 mining Qwen while the big card serves GLM).
• Honest declaration — the pool is told only about proven-serveable models, never aspirationally, so it stops routing requests we can't answer.

Load a model that fits VRAM

• Auto-tier is VRAM-correct — Gemma-4-12B now needs ~20 GB (the walk gather + inference context need headroom beyond the raw weights), so 16 GB cards (5070 Ti / 5080) stay on the 9B tiers (t0/t1) instead of OOMing.
• OOM auto-recovery — if an auto-selected model still OOMs the walk build, the card demotes to the largest staged tier that fits instead of looping forever; if none fit, it halts with clear guidance.
• --force-model is honored exactly — no VRAM check, no demotion. You pick the model; the miner loads it. (Only auto-selection is VRAM-aware.)

Also in this line (from v0.10.6)

• Windows in-process llama.cpp engine (keryx-llama.dll) — Windows NVIDIA rigs can now serve H4/H6 models (previously the engine was stubbed on Windows, so they couldn't serve at all).
• CPU inference is OFF by default — a card that can't run GPU inference withdraws instead of doing futile CPU work. Opt in with --enable-cpu-inference.

Builds

HiveOS (*.tar.gz), MMPOS, and Linux (*-linux-x86_64.tar.gz) packages, each with AVX + noavx llama backends. AMD (*-amd-*) and Windows (*-windows-*) assets attached separately. Checksums: SHA256SUMS-0.10.7.txt.