Keryx-miner-v0.6.1-PoM

Performance

• About +40% overall versus 0.6.0 on an RTX 3090.
• Faster PoM v5 walk kernel, checked against the reference kernel at startup: 1.94 → 2.15 MH/s on an RTX 3090 (tier 3). Native SASS for sm_103.
• New lock-step walk on top of it: several nonces share a warp and weights are read in contiguous blocks. The width is picked per card at startup, after the same check. 2.13 → 2.72 MH/s on the same card.
• Model weights are loaded in a canonical layout, so the walk reads them in place with no extra copy.

Multi-GPU rigs

• During an inference, only the GPU serving it pauses; the other GPUs keep mining.
• A GPU that hits a CUDA fault is dropped on its own when another GPU still mines its model. The rest of the rig keeps going.

Stratum / pools

• Traced inference and audit openings served through the pool, with the pool deciding when a request is traced.
• PoM proofs sent in base64 when the pool supports it (hex otherwise).
• Trace records uploaded to the pool when it asks for them.
• The miner announces its models to the pool at connect and keeps it updated.
• Sturdier pool connection: unknown messages and oversized lines are skipped instead of dropping the link, longer job timeout, lower latency on the socket, pool password option.
• In pool mode, the chain height is read from the public API, so models of past eras are no longer downloaded.

Fixes

• Some prompts that could not be served before are now answered normally.
• Audit state is kept across restarts. Transient inference errors are retried, and broken UTF-8 in model output is handled.
• Rigs sharing a payout address take over sooner when the rig ahead of them stays silent. The log line of a rig that was waiting its turn no longer blames it for the strike.

Credits

• Lock-step walk and canonical weight layout ported from keryx-miner-supr by @ocminer.
• Prompt serving issue spotted by @rogersmitha51.
• Silent-rig takeover issue spotted by GiovannieR.