ocminer/keryx-miner-supr v0.15.8 on GitHub

keryx-miner-supr 0.15.8 — Tesla V100 / Pascal H15 traces and per-card tier pinning

Tesla V100 (Volta) and GTX 10xx (Pascal), legacy package: H15 traced inference works. These cards failed every H15 self-test with trace read failed (-5), N positions traced for M fed (issue #29), or with point 0: Mismatch on Q4_K/Q5_K models such as Qwen3.5-9B. They were withdrawn and did not mine. The bundled inference engine now always records the full prompt trace and computes K-quant products exactly as the verifier recomputes them. On a Tesla V100 (driver 580), tier 0 and tier 1 pass the self-test and mine with 0 rejected shares.
Multi-card rigs no longer withdraw healthy tiers. When several cards loaded different models in one process, a card's self-check could fail with point 0: Mismatch while the other cards were starting up. The tier was then withdrawn or moved to another card, which swapped models on every request. Each card now checks against its own model.
--force-model pins one tier per card. --force-model t0,t1,t2,t3,t4 loads tier 0 on GPU0, tier 1 on GPU1 and so on in one process. With --cuda-device, the list maps onto the selected cards in order. A pinned card keeps its tier. If its self-check fails, it is retried on the same card instead of moving to another card.
KERYX_LLAMA_CTX is capped at the largest context a trace can cover.

Tier names t0–t4 are accepted everywhere a tier name is (--tier, --force-model). --help shows the model behind every tier: t0 Qwen3.5-9B, t1 GLM-4-9B, t2 Gemma-4-12B, t3 Qwen3.8-27B, t4 Kimi-Linear-48B.
With the interactive dashboard, the miner writes a plain log file keryx-miner.log (rotated at 64 MiB). Use --plain-log-file <path> or KERYX_PLAIN_LOG_FILE for another path, or an empty value to disable it. Classic mode (HiveOS, mmpOS, --no-tui) writes a file only when asked.
When the miner stops itself, the reason is printed on the console after the dashboard closes.
Stale analysis: stale, duplicate and low-difficulty shares are logged with the age of their job at submission and the pool's reply time. Every 200 answered shares, a summary of the job age is logged.
The dashboard no longer highlights healthy error: false status text as an alert.
The AMD standalone package's RUN.txt correctly describes H15 support.

RTX 5090 (modern package), tiers 0–4: traced generations are byte-identical to 0.15.6, and every audit opening is verified. Hashrate is unchanged (5.06 MH/s at tier 3), with 0 rejected shares.
RTX 3070 (legacy package): traced generation is byte-identical to 0.15.6. Hashrate is unchanged (1.229 MH/s), with 0 rejected shares.
RX 7900 XTX (AMD package): traced generation is byte-identical to 0.15.6. Hashrate is unchanged within run-to-run variation, with 0 rejected shares.
5 × CMP 170HX, one process, --force-model t0,t1,t2,t3,t4 --only-inference: all five tiers pass and are declared on every run, with one model load per card.
Tesla V100: tiers 0 and 1 pass, and live mining had 0 rejected shares.

0.15.7 was an internal build and was not released.