ocminer/keryx-miner-supr v0.14.0

keryx-miner-supr v0.14.0 — H14 hard fork support

on GitHub

Keryx H14 hard fork — update required before DAA 121,985,000 (≈ Friday 2026-10-09 14:00 UTC)

v0.14.0 mines normally before the fork and switches to the H14 rules by itself when the chain crosses the gate. No restart is needed if you leave it running across the fork. Older versions (v0.13.x and earlier) produce invalid shares after the gate.

What's changed

• H14 proof-of-model rules: from the gate the PoM walk uses the new H14 seed, and the miner logs H14 GATE CROSSED once when it switches. Proofs were cross-checked against the keryx-node v1.6.4 consensus code: the node accepts proofs from both eras and rejects wrong-era proofs.
• Tier 3 becomes Qwen3.8-27B: after the gate, tier 3 (--tier high and --tier auto on cards with enough VRAM) uses Qwen3.8-27B instead of Qwen3.6-27B.
  • Before the gate, a running miner downloads Qwen3.8 in the background, about 16.8 GB.
  • At the gate, NVIDIA cards switch to it in place.
  • A miner started after the fork goes straight to Qwen3.8.
  • AMD tier-3 rigs: restart the miner once after the fork.
• Private inference: after the gate, inference requests are sealed. In pool mining the pool opens them and sends the prompt with mining.ai_request. The miner answers with mining.ai_response. Plaintext inference tasks from the pool are ignored after the gate, and mining continues normally.
• Longer context with an 8-bit KV cache and flash attention:
  • Context size: at least 32k tokens for the large models, and up to each model's limit when VRAM allows.
  • KV cache: q8_0, with an automatic fallback to f16 for models that can't use it.
  • Mining VRAM: about 1.5 GB of VRAM stays reserved for mining, so the larger context never takes memory from it. Change it with KERYX_LLAMA_WALK_RESERVE_MB; KERYX_LLAMA_CTX forces a fixed size.
  • Platforms: CUDA, Vulkan (AMD) and Metal.
• Solo mining on a quiet chain: a block template that arrived while the model was still loading was dropped and not requested again until the node announced a new one. On a slow or stalled chain the miner could wait indefinitely after warm-up. It now requests a fresh template as soon as a model is ready.
• Upstream system prompt: answers use the upstream system prompt. AMD/Vulkan answers now go through the model's chat template and accept prompts up to 64 KiB.

No changes to hashing kernels, autotune or the dashboard design. Before the fork, behaviour and shares are the same as v0.13.4.

Pool miners: your pool must also be upgraded for H14. Until it is, shares after the fork are rejected regardless of miner version.

Low-bandwidth rigs: each PoM share uploads a proof of about 0.9 MB. On very low-difficulty ports a slow uplink can fall behind. Use a higher-difficulty port if your miner reports stalls or pending shares.

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:0.14.0 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.

Integrity and testing

Verify downloads against the matching SHA256SUMS-*.txt file. Packages were built from 5d1623bb1bbbb687a4381fa23825ade8b6dadc87 (tag v0.14.0), with Windows and macOS produced through GitHub Actions.

Testing:

• Regression suites: workspace, CUDA and OpenCL, 586 tests passed and 0 failed.
• Live tests: on a CMP 170HX + RTX 3070 rig:
  • the release package mined before the fork with 30 shares accepted and 0 rejected;
  • a simulated gate crossing switched without a restart;
  • Qwen3.8-27B tier 3 passed its self-test after the gate with a 64k context.
  • Solo mining on the public Keryx testnet, which is already past its H14 gate: 1,398 blocks submitted and 0 rejected. A separate node syncing only from the public peer received them.
• Not tested here: real private-inference traffic needs an H14-upgraded pool and chain. The Windows builds are not GPU-tested here.

Build records: Windows CI, macOS CI.