keryx-miner-supr v0.6.8.2
Patch release: OPoI inference fixes (ported from upstream Keryx-Labs/keryx-miner v0.3.6) + tokio worker cap.
Inference fixes — stop the "repeated-answer loop" on two tiers
Both were the same bug: a served model never emitted its stop token, so it looped the same answer to max_tokens — a failed OPoI inference (garbage/timeout response). This directly cost inference rewards on the affected tiers.
- Llama-3.3-70B (
--very-high, the top inference tier). It's abliterated/ChatML on stock LLaMA-3 vocab, so it writes<|im_end|>as plain text and never emits<|eot_id|>. Now stops on the ChatML markers too. (upstreamfaee090) - Qwen3-1.7B (
--very-light). Wasn't handled in the stop-config at all → fell through to generic</s>stops that Qwen3 never emits. Now shares the Qwen3 ChatML stop-config,/no_thinkprompt, and think-strip. (upstreama033620)
If you mine --very-high or --very-light, upgrading fixes broken/looping inference responses.
Runtime — cap tokio async workers
#[tokio::main] spawned one async worker per logical CPU; the miner's async workload is tiny (one connection + timers — heavy work runs on spawn_blocking/dedicated threads). On a many-core rig (cloud H100/A100/B200) that was dozens of idle executor threads = pure overhead. Now capped (default 2 workers; KERYX_ASYNC_WORKERS to override, KERYX_BLOCKING_THREADS optional). (upstream 9bb0d55, main.rs part)
Notes
- No consensus/PoM change — the walk, proofs, and the solo/pool paths are byte-identical to v0.6.8.1. Cargo user-agent stays
0.6.8. - Packages: modern / legacy / pascal × {hiveos, linux-x86_64, mmpos, smos} + SHA256SUMS; Windows attached by CI. Same driver-floor split as v0.6.8.x. CMP 100-210 / Volta = legacy line.