github ocminer/keryx-miner-supr v0.6.10.0
v0.6.10.0 — llama.cpp inference engine (new default), candle now fallback

latest releases: v0.12.7, v0.12.6, v0.12.5...
one month ago

v0.6.10.0 — llama.cpp becomes the default AI engine

The NVIDIA Linux packages now bundle libkeryx-llama.so — an in-process llama.cpp engine (pinned b10015, CUDA) that replaces candle as the default for everything AI:

  • Single VRAM copy (zero-dup): llama.cpp hosts the model on the inference GPU; the PoM walk gathers directly over its resident tensors. The 70B very-high tier runs on a 32 GB card with ~4.5 GB headroom (measured: 27.6 GB total on an RTX 5090, model load ~9 s).
  • OPoI inference runs through llama.cpp — faster quantized inference than candle, and new GGUF model architectures are supported as soon as llama.cpp supports them (no longer blocked on candle implementing an arch).
  • candle is now a dormant fallback, still compiled in: if the bundled engine can't load (or you delete the .so), the miner transparently falls back to the exact previous candle behavior. candle will be removed entirely in a future release.
  • Walk/consensus untouched and guarded: gather is byte-gated against the possession index at every startup (129 sampled chunks) plus the existing N-guard; the full-model byte-identity of llama-resident tensors vs the on-disk GGUF was verified with a 4.9 GB memcmp.
  • Env knobs: KERYX_LLAMA_SO (engine path override), KERYX_LLAMA_CTX (context size, default 4096; use 2048 on tight 70B/32GB setups), KERYX_LLAMA_CUDA_DEVICE / KERYX_LLAMA_SERVER+KERYX_LLAMA_PORT (subprocess fallback).

Validated live: 70B-on-32GB + mixed multi-GPU rig, packaged-tarball runs, 140+ shares 0 rejects, per-card walk rates identical to previous releases (the walk is memory-bound — the wins are VRAM and inference speed).

Windows: the in-process engine is Linux-only for now — Windows stays on candle (unchanged); a Windows DLL build is a follow-up. AMD: unchanged (Vulkan llama-server as before); AMD Linux packages follow separately.

NVIDIA Linux assets

12 tarballs (modern sm_75+ / legacy sm_70+ / pascal sm_60-61) + per-line SHA256SUMS. Package sizes grew by the bundled engine (~200 MB–1 GB per line). Windows zips + macOS tarball attached by CI.

Package versions: CUSTOM_VERSION 0.6.10.0 / BUILD 27; banner/user-agent now 0.6.10.

Don't miss a new keryx-miner-supr release

NewReleases is sending notifications on new releases.