keryx-miner-supr v0.9.5
Fix: engine-hosting GPU idling at 0 hash after a failed zero-dup start (AMD).
On rigs where the in-process llama engine's startup byte gate fails (log: byte gate fetch failed at chunk 0 — keeping the OpenCL blob), the fallback tried to install that card's OpenCL possession blob (~4.8 GB post-H5) while the engine still held the model (~5–6 GB) on the same card. On 8 GB GPUs both cannot fit — the install failed and the card idled at 0 hash while the rest of the rig mined.
v0.9.5 unloads the in-process engine (freeing its VRAM) when the blob install fails on the engine's own card, then retries the install: the card returns to full-rate mining, and OPoI inference falls back to the bundled llama-server / CPU. Cards with enough VRAM for both (e.g. 16 GB) are unaffected — the engine stays loaded there.
Also carries everything from the v0.9.4 asset updates: RX 580/Polaris single-allocation fix, per-device walk autotune (ILP×2/×4), download-progress + share-time-expectation logging, bundled Vulkan loader, -noavx engine variant.
Assets
keryx-miner-supr-amd-0.9.5.0.tar.gz— HiveOS custom miner (AMD/OpenCL).keryx-miner-supr-amd-mmpos_0.9.5.0.tar.gz— mmpOS package.keryx-miner-supr-amd-0.9.5.0.zip— portable Linux AMD build.- NVIDIA / Windows / macOS assets are built by CI and attached as they complete.
Verify downloads against SHA256SUMS-amd.txt.