ocminer/keryx-miner-supr v0.13.3

keryx-miner-supr v0.13.3 — GPU inference isolation

on GitHub

What's changed

• NVIDIA multi-GPU inference isolation: only the GPU serving an inference request pauses its PoM walk. Other cards keep hashing and receiving fresh pool or solo jobs.
• Stable inference pinning: --inference-cards now constrains placement and failover, including on rigs with identical GPUs. Busy installations and overlapping probes defer without falsely withdrawing a model or demoting its tier.
• Low-RAM coordination: --low-ram serializes model preparation and implies --wait-ready. Slow model installation no longer trips the unrelated 30-second GPU walk-drain guard.
• Responsive pool handling: buffered GGUF metadata reads and duplicate-read suppression prevent model-readiness checks from delaying share acknowledgements and inference requests.
• Accurate dashboard state: per-GPU inference pauses are shown without marking unrelated CUDA cards as paused.
• Large-rig autotune fixes included in every applicable build: rigs above four GPUs admit up to four distinct tuning profiles; identical cards reuse the same profile, and experimental Blackwell candidates are no longer benchmarked by default.
• Linux legacy/Pascal packaging corrected: CUDA 12.4, compute_61 PoM/Candle targets and Pascal-capable GPU inference engines are now checked before packaging.
• Complete platform packages: Windows AMD includes OpenCL mining plus Vulkan GPU inference. HiveOS AMD uses the correct fixed install directory and manifest. Docker now requires both CUDA inference-engine CPU variants and generates unique SSH host keys at container startup instead of baking private host keys into public image layers.

This release brings the previously split v0.13.2 miner artifacts back onto one common miner-source revision. Hashing/proof consensus rules and the Matrix dashboard design are unchanged.

Choose your package

For HiveOS, use the bare versioned archive, not the linux-x86_64 standalone archive. Keep the entire extracted package together: the inference engines and runtime libraries are required. --no-tui is available for headless use; HiveOS/mmpOS launchers handle this automatically.

Windows requires the appropriate GPU driver and the Microsoft Visual C++ v14 Redistributable (x64). Install/update it if Windows reports a missing VCRUNTIME140.dll or MSVCP140.dll; do not download individual DLLs from third-party sites.

Docker: ocminersupr/keryx-miner-supr:0.13.3 and :latest (NVIDIA modern line). Requires a suitable NVIDIA driver and NVIDIA Container Toolkit; GPU selection can use --runtime=nvidia with NVIDIA_VISIBLE_DEVICES.

The Docker entrypoint starts an SSH service. Set your own ROOT_PASSWORD and restrict access before exposing port 22; do not expose it publicly with the default password. Fresh containers generate new host keys; existing mounted SSH keys are retained.

Inference and memory notes

• On an affected low-RAM NVIDIA rig, start with --low-ram --no-resident-tree --inference-cards 2 to restrict inference to CUDA ordinal 2. Check visible-device numbering, especially when using CUDA_VISIBLE_DEVICES.
• The allowed inference GPU(s) must support every advertised model. Pinning is supported on identical-GPU rigs too.
• The serving CUDA card still pauses during generation/model changes to protect shared model buffers. OpenCL and Metal retain their existing backend-wide pause behaviour.
• Low-RAM mode does not make cold model loading fast. Kimi-48B's roughly 28 GiB GGUF can exceed system RAM; storage and cache pressure still affect startup and reload latency. Use a smaller tier if the hardware cannot reliably serve the selected model.
• CPU inference remains a deprecated, explicit emergency override, not the normal inference route.
• GTX 1080 Ti users should select --tier very-light; the legacy line does not support compute_60/P100.

Integrity

Verify downloads against the matching SHA256SUMS-*.txt file. All miner packages were built from 9bb2f9d01ae5128ca55fd9bec9060baa6bbefb1d, with Windows and macOS produced through GitHub Actions. Release tag v0.13.3 points to 6840f1dc591f681382a77bf5eb4bb464cde9723d, which adds only Docker recipe/entrypoint SSH-key hardening: every miner, inference-engine, CI and other packaging source file is unchanged. The Docker image records both revisions in its labels and contains the same validated modern miner/engine bytes.

Build records: Windows AMD and NVIDIA CI, macOS CI.

Workspace, CUDA and OpenCL regression suites passed. Six bounded Linux release-package/Docker smoke tests on RTX 5090, GTX 1080 Ti and RX 7900 XTX completed GPU inference and submitted 191 accepted shares with no rejects. The final hardened Docker image passed its own pool/inference repeat. These are compatibility checks, not a long-duration hashrate comparison.

Windows hosted CI validates builds and packaging; it is not a physical Windows GPU/pool certification. The macOS backend remains experimental.