ocminer/keryx-miner-supr v0.7.1

v0.7.1 — H4 + AMD auto-tier

on GitHub

keryx-miner-supr v0.7.1

H4 hardfork support + real AMD auto-tier.

H4 (coin-age verification, activation DAA 54,766,000)

• Full H4 proof pipeline (proof v2: recompute-from-chunks, all 256 walk chunks Merkle-proven) with era-aware tier selection and the new H4 model lineup (EXAONE-4.0-1.2B / Mistral-7B-v0.3 / GLM-4-9B / Qwen3.6-27B / Kimi-Linear-48B).
• Includes the latest upstream correctness fixes: model-lineup ordering, tokenizer handling for H4 GGUFs, and pre-staged-GGUF adoption / 416-resume on download.

AMD / OpenCL

• Real auto-tier (--tier auto, now the default): picks the largest H4 model tier that fits card 0's VRAM (via CL_DEVICE_GLOBAL_MEM_SIZE) with a conservative AMD memory margin — an 8 GB card mines EXAONE (the floor), a 16 GB card reaches Mistral-7B, larger cards go higher, for more reward without OOM. Pin a tier explicitly with --light / --high / --very-light / --force-model.
• Zero-dup in-process Vulkan engine bundled (libkeryx-llama-vk.so): on RDNA cards the model is hosted once in VRAM and serves both OPoI inference and the PoM walk, with the Vulkan llama-server and CPU inference as fallbacks.

Assets

• keryx-miner-supr-amd-0.7.1.0.tar.gz — HiveOS custom miner (AMD/OpenCL).
• keryx-miner-supr-amd-mmpos_0.7.1.0.tar.gz — mmpOS package.
• keryx-miner-supr-amd-0.7.1.0.zip — portable Linux AMD build (run directly).
• NVIDIA / Windows / macOS assets are built by CI and attached to this release as they complete.

Verify downloads against SHA256SUMS-amd.txt.