ocminer/keryx-miner-supr v0.13.4

keryx-miner-supr v0.13.4 — inference resilience fix

on GitHub

What's changed

• Inference resilience fix (all platforms, reported on Windows): a single failed live inference request on a llama.cpp-only model — for example Qwen3.6-27B, GLM-4, Gemma-4, EXAONE-4 or Kimi-Linear — no longer withdraws the model and suspends mining. While the in-process GPU engine stays loaded, a failed generation now fails only that one request; the model stays advertised and mining continues. Only three consecutive failures on the same card fall back to the previous recovery path, and the counter resets on the next successful answer.
• Split UTF-8 replies kept: an answer whose last character was cut in the middle of a multi-byte UTF-8 sequence is now trimmed to its valid text instead of being discarded as a failed generation. The native engine result code is logged when a generation does fail.
• llama.cpp warnings and errors visible in the dashboard: with the Matrix dashboard active, llama.cpp warnings and errors are now forwarded to the miner log/event panel instead of being dropped, so a failed model load or generation shows its actual reason. Informational llama.cpp output stays suppressed.

Windows users: if your miner passed the inference self-test but later reported no models ready — mining suspended (for example an RTX 3090 on --tier auto serving Qwen3.6-27B), please update to v0.13.4.

No other behaviour changes. Hashing/proof consensus rules, autotune, pool protocol and the dashboard design are unchanged from v0.13.3.

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:0.13.4 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

The v0.13.3 notes still apply: --inference-cards pinning, --low-ram, per-card CUDA inference pauses, and slow cold loads of large models such as Kimi-48B on low-RAM hosts. CPU inference remains a deprecated, explicit emergency override, not the normal inference route. GTX 1080 Ti users should select --tier very-light.

Integrity

Verify downloads against the matching SHA256SUMS-*.txt file. All packages were built from 521842378ed51700f1e8590d355d9e6e35d222a6 (tag v0.13.4), with Windows and macOS produced through GitHub Actions.

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

Workspace, CUDA and OpenCL regression suites passed, including new unit tests for split UTF-8 replies. The fix was exercised on an RTX 5090 with the dashboard active: the self-test passed, a live pool inference request was served and mining continued with accepted shares. All release packages were checked for layout, component identity and version against the v0.13.3 package set. These are compatibility checks, not a long-duration hashrate comparison, and the fix has not yet been tested on physical Windows hardware.

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