ocminer/keryx-miner-supr v0.16.2

Keryx Miner Supr v0.16.2 — faster PoM walk, NUL-byte prompts, mixed-tier fix

on GitHub

Faster PoM walk on NVIDIA, NUL-byte AI prompts answered, and rigs that mine different tiers in one process fixed. Upgrade recommended for all NVIDIA rigs.

• Faster mining on NVIDIA (sm_80 and newer). A new walk kernel advances two or four nonces per warp in lock-step over one contiguous copy of the model and picks the faster width per card at startup (checked against the previous kernel before it is used). Inference cards keep a single resident model: the inference engine now lays the weights out in the walk's order. Measured on our rigs, same card and power limit, against 0.16.1: RTX 5090 at 450 W +32% (4.31 → 5.70 MH/s, tier 0; tier 4 4.37 → 5.70), RTX 3070 at 240 W +20% (1.31 → 1.57 MH/s), RTX 5080 at 360 W +2–3% (3.28–3.33 → 3.39–3.41 MH/s; this card was already close to its memory limit). KERYX_POM_V5_LOCKSTEP=0 returns to the previous kernel, KERYX_CANONICAL_WEIGHTS=0 to the previous weight layout.
• Fix: AI requests whose prompt contains a NUL byte are now answered. The prompt reaches the engine byte for byte with its length; nothing is removed, so the answer matches the prompt the network audits. Before, every miner refused such a request and the node counted a miss for each of them.
• Fix: rigs mining different tiers in one process no longer get every share of the main tier rejected (v5 verify: BadTilePath). Each card now builds the context pad from its own model.

H16 handling is unchanged (mainnet DAA 147,000,000, testnet 180,000, keryx-node 1.6.7). AMD (OpenCL/Vulkan) and Apple (Metal) mining kernels are unchanged; their inference engines were rebuilt for the NUL-prompt fix. The CUDA and Vulkan inference engines are rebuilt for this release.

Validation: host test suites (library, binary and the new mixed-tier proof test), startup checks on Ubuntu 20.04/22.04/24.04 for every Linux archive, and archive/license audits passed. On an RTX 5090 all five tier models passed the traced inference self-test with the new weight layout, and a single process mining different tiers on RTX 3070 and RTX 5080 cards had 317 shares accepted and 0 rejected on mainnet. The final packaged Linux miner on an RTX 5070 Ti passed its tier-0 inference self-test, ran the new lock-step walk on the engine's own weights and mined on mainnet for 12 minutes with 52 accepted and 0 rejected shares at 3.09 MH/s and 300 W.

Checksums: SHA256SUMS-*.txt. Keep the executable and its bundled libraries together. Existing project and third-party licenses are preserved.

The limitations listed for 0.16.0 still apply: Metal H15/H16 is unsupported, and Windows GPU hardware was not tested for this release.