ocminer/keryx-miner-supr v0.10.11

on GitHub

📦 Which file do I download?

Pick by GPU vendor, then your rig OS / manager. Most people want the modern NVIDIA build for their OS.

NVIDIA: modern vs legacy

• modern (no -legacy in the name) — the default. Built for recent drivers (~575+); covers Turing → Blackwell (RTX 20/30/40/50 series, incl. 5090). Use this unless it won't start.
• -legacy — for older drivers (~535) or older cards (Volta / some Pascal). Only use it if the modern build fails at startup with a driver/CUDA error.

What each format is

• …-0.10.11.tar.gz (plain version, no OS word) → HiveOS custom miner (unpack into HiveOS custom miners).
• …-mmpos_0.10.11.tar.gz → mmpOS.
• …-linux-x86_64.tar.gz → generic Linux — unpack and run ./run.sh (or ./keryx-miner-supr …). Needs only the NVIDIA driver; CUDA toolkit NOT required (runtime libs are bundled).
• …-windows-nvidia-pom.zip → Windows (NVIDIA), self-contained.
• …-windows-amd.zip → Windows (AMD/OpenCL).
• keryxcuda-windows.zip → the Windows CUDA plugin DLL only (advanced/component use — the full Windows miner is …-windows-nvidia-pom.zip).
• SHA256SUMS-0.10.11.txt (NVIDIA) / SHA256SUMS-amd.txt (AMD) → checksums to verify your download: sha256sum -c SHA256SUMS-0.10.11.txt.

NVIDIA and AMD are separate builds. AMD assets are versioned -amd-0.10.11.1 and maintained on the AMD/OpenCL backend.

v0.10.11 — fix "connected but not serving inference"

Rigs could authorize on the pool, pass the inference self-test, and mine a high tier while the pool still showed declared=[] / model=none for them — i.e. mining a tier they were never asked to serve, which accrues inference strikes. This release makes the OPoI serving declaration robust.

Root cause: the capability declaration (mining.declare_capabilities) was fire-once — sent only when the serveable set changed, as a notification with no pool acknowledgement. A single declaration lost to a startup race (pool not yet ready to bind it to the worker) or a pool-side registry reset left the pool showing declared=[] permanently.

Fixes

• Periodic re-declaration — the current serveable model set is re-announced every 90s even when unchanged, so a missed/lost declaration self-heals; declared no longer depends on one packet landing at exactly the right moment.
• Live served-model telemetry — mining.hello carries the model but is sent once at authorize, before the model is resident (so it reported model=none). Every periodic mining.telemetry tick now includes the live serveable model-id set, so the pool's view corrects the moment the model finishes loading.

Builds on v0.10.10 (loud, self-healing model staging — clear errors instead of a silent "preparing…" stall) and v0.10.9 (Gemma-4-12B fits 16 GB cards with inference via per-tensor gather). No consensus/algorithm change.