keryx-miner-supr 0.15.1 — H15 hard fork (updated)

Mandatory update before the Keryx H15 hard fork (mainnet DAA 140,115,000, expected around 2026-10-30 15:00 UTC). Everyone running 0.15.0 must move to 0.15.1 before the gate.

keryx-node 1.6.6 retired the H15 block task: the block-task root is no longer part of the block header and nodes ignore it. 0.15.0 still commits that root into every post-gate header, so every block it finds after the gate would be invalid. 0.15.1 hashes headers without the task root, exactly like the node. The PoM v5 context pad still activates at H15, unchanged.

As before, the switch is driven by the block DAA in each job, never by wall-clock time: start 0.15.1 now and it keeps mining normally until the gate, then changes over on its own. No restart, pool, wallet or model change is needed.

What changes

• H15 follows keryx-node 1.6.6. Post-gate headers are hashed without the block-task root. The per-card block-task pass, its self-check and the block-candidate trace opening are gone, so cards start mining post-gate jobs immediately. The PoM v5 context pad (per-tier pad next to the model weights, generated and walked on the GPU) still activates at H15.
• Smaller share submits. When the pool supports it, proofs are sent base64-encoded instead of hex: about a third less upload per share (an H15 proof shrinks from ~975 KB to ~650 KB). Pools that do not advertise it keep receiving hex. If a pool refuses a base64 submit, that share counts as rejected, the connection stays up and the miner sticks to hex for that pool.
• More robust pool connection. A pool reply the miner cannot match to a pending share (for example a result without an id) no longer drops the connection; a negative reply for a pending share counts as a rejected share.
• --tier auto keeps a VRAM safety margin. On GPUs shared with other processes, auto-tier now leaves about 2 GiB of free VRAM instead of picking a model that only barely fits and then fails to load. Load failures now report the VRAM headroom and how to pin a smaller tier with --force-model.
• Correct per-GPU telemetry with CUDA_VISIBLE_DEVICES. Health readings and the API now match CUDA devices to the physical card by PCI bus id, so a miner pinned to a subset of GPUs reports the readings of its own cards.

Platform notes

• NVIDIA (Linux modern/legacy, Windows, Docker): full H15 support.
• AMD (OpenCL/Vulkan) and Apple Silicon (Metal): these walks have no PoM v5 context pad yet. They mine normally until the gate and then idle with a warning instead of producing work that can never become a valid block. H15 support for these lines follows in a later release.
• Linux packages run on Ubuntu 20.04 and newer (glibc 2.31+), including current HiveOS, mmpos and SMOS images.

Downloads

Use the matching SHA256SUMS file to verify downloads.