ocminer/keryx-miner-supr v0.6.5.1

keryx-miner-supr v0.6.5.1 — --very-light/--very-high staging fix

on GitHub

Hotfix on v0.6.5.0. Fixes the model-staging tier for --very-light and --very-high:

• --very-light now downloads Qwen3-1.7B (~1 GB), not the 2.4 GB Gemma (it was hitting a pre-H2 staging fallback).
• --very-high now downloads the 32 GB-servable Llama-3.3-70B Q2_K_L, not the 48 GB Q4.

Carries the v0.6.5.0 per-device multi-GPU + inference-on-biggest-GPU (no CUDA_VISIBLE_DEVICES, no CPU inference). If a manually-placed model still re-downloads, put it at <miner>/models/<dir_name>/model.gguf with the full file set + a .ok marker in that dir.

Lines (pick your OS tarball; verify with SHA256SUMS): modern (CUDA 12.x, R575+), legacy (CUDA 12.4, R550+), pascal (sm_60, GTX 10-series). AMD/Windows separate.