Fixes the restart loop on the inference card
Rigs that also serve inference could hit a crash loop every 10–20 minutes:
CUDA_ERROR_ILLEGAL_ADDRESS → ggml_abort in llama_free → CUBLAS_STATUS_NOT_INITIALIZED → miner restart
On the serving GPU the PoM walk reads directly out of the resident model weights (zero-duplication —
that is what keeps VRAM use low). When the model was swapped, the walk on that card kept reading memory
that was being freed, which poisons the CUDA context for everything on the device, including the
inference engine.
The walk is now paused per GPU for the ~2 s swap: the serving card holds off, every other card in the
rig keeps mining at full rate. You will see one line per swap:
PoM[gpu3]: holding off — llama model swap in progress on this card (walk paused, ~2 s).
Validated on a 4-GPU mixed rig through live pool probes: zero illegal-address / abort / cuBLAS events,
all cards back to full hashrate after each swap.
Also in this release (carried from the 0.11.11 work)
- Proof build ~5.2× faster (28.6 ms → 5.5 ms at 8.4M chunks) via Merkle sibling memoization —
~4 MB of memo instead of a 12 GB resident tree. Byte-identical proofs. Disable withKERYX_POM_PATH_MEMO=0. - Pool-share proofs are built on a detached thread, off the grind loop, so share submission no longer
stalls hashing. Kill switch:KERYX_POM_PROOF_OVERLAP=0. - SM-count-derived batch sizing — smaller cards now stay inside the 100 ms window instead of overshooting.
- SIMD int8 dot product (AVX2 / SSE4.1 / NEON) and a zero-allocation re-walk for CPU-side verification.
- Chase sub-batching removed (+0.5%); cross-batch chase prefetch ships off by default (measured a loss).
Upgrading is recommended for anyone running inference alongside mining. Pure-mining rigs are unaffected by
the crash but still get the proof-build and batching improvements.