github ocminer/keryx-miner-supr v0.14.4
keryx-miner-supr v0.14.4 — long private prompts served, no disconnect on large requests, AMD INF display

3 hours ago

keryx-miner-supr 0.14.4

Inference fixes for the H14 private-inference era, and an AMD dashboard fix. Recommended for every rig that serves inference, especially cards on the larger tiers (Qwen3.8-27B, Kimi-Linear-48B), which can now answer long private requests.

Fixed

  • Private prompts larger than 4 KiB were refused. Before H14 an inference request was limited to 4 KiB and so was the miner (prompt exceeds 4096 bytes). Since the H14 gate, private requests can carry up to 64 KiB, and a compressed prompt can expand to more. Every valid larger request was refused and counted as a miss. The miner now accepts prompts up to 1 MiB, and each card decides by its model's context size whether the prompt fits. A prompt too large for a card's context is refused for that request only: the card keeps its model, stays serveable and keeps mining.
  • A large inference request dropped the pool connection. The miner read pool messages with a 64 KiB line limit. A request carrying a prompt above about 48 KB exceeded it, and the miner disconnected and reconnected 30 s later, losing the request and any shares in flight. The limit is now 2 MiB, and a pool message that is still too large is skipped without disconnecting.
  • AMD: two cards shown as INF. On AMD rigs where one card is reserved for inference, the dashboard also marked a second card INF whenever a request was served, although that card was mining (it numbered the serving card differently from the dashboard). Now only the reserved card shows INF; during a request the other cards show as paused, then mining again.

Tested

On an RTX 5090 serving Qwen3.8-27B (64k-token context) after the H14 gate, against a test pool with the node's share verifier:

Prompt 0.14.2 0.14.4
3 KB answered answered (1.4 s)
30 KB refused answered (3.1 s)
60 KB connection dropped answered (5.2 s)
200 KB — answered (17.9 s)
400 KB (above the model context) — refused cleanly; the next request was answered normally

Every answer recalled a code word placed at the very start of the prompt, so the whole context was read. 982 shares, 0 rejected. Also verified on the live pool.

Which download

GPU Download
NVIDIA Turing (RTX 20-series) and newer, Linux, driver 575+ modern line (standalone, HiveOS, mmpOS, SMOS, Docker)
NVIDIA Tesla V100 / Titan V / Quadro GV100 (Volta), GTX 10-series/Pascal, CMP 100–210, or driver 550–574 legacy line (standalone, HiveOS, mmpOS, SMOS)
Windows NVIDIA, Turing and newer keryx-miner-supr-windows-nvidia-pom.zip
AMD (Linux Vulkan / Windows) -amd packages

No consensus, PoM or model changes; mining speed is the same as 0.14.2. H14 gate handling is unchanged. (0.14.3 was not released; its fixes are included here.)

Don't miss a new keryx-miner-supr release

NewReleases is sending notifications on new releases.