ocminer/keryx-miner-supr v0.6.11.0

v0.6.11.0 — AMD: zero-dup in-process inference engine

on GitHub

v0.6.11.0 — AMD: zero-dup in-process inference engine

A new optimization pass gives the AMD build an in-process llama.cpp engine (bundled libkeryx-llama-vk.so): llama.cpp hosts the model once in VRAM, serves the OPoI inference in-process (no subprocess), and — on RDNA GPUs — hosts the Proof-of-Model possession walk over the same resident weights, so that card needs no separate weight copy.

What changes on a rig

• Inference card (RDNA, e.g. RX 7600 XT / 7900 XT): holds ONE model copy instead of two. On our reference rig the inference card dropped ~1.5 GB VRAM, the external inference subprocess is gone, and the walk on that card is byte-identical and slightly faster (measured +2% vs the OpenCL path).
• Other cards / GCN (Instinct MI50/MI60, Vega, Polaris): keep their existing full-speed OpenCL walk unchanged. The engine is used only for in-process inference there (still dropping the subprocess).
• Net: never slower than before on any card, with a VRAM saving where it is free.

Safety

Every start runs a byte gate: the engine's resident weights are sampled through the exact gather path the walk uses and compared byte-for-byte against the on-disk model; any mismatch refuses the zero-dup path and the card falls back to its own weight copy. The possession/consensus math is unchanged and byte-exact (verified end-to-end against the pinned tier root and a fixed winning nonce).

Controls

• Default: the zero-dup walk is used only on RDNA cards (where it is a strict win).
• KERYX_ZERO_DUP=force — use it on any inference card (trade VRAM for the walk even on non-RDNA).
• KERYX_ZERO_DUP=off — never; every card keeps its own weight copy.
• KERYX_LLAMA_VK_DEVICE=N — which GPU hosts the engine (default 0).

Absent the .so (or with no Vulkan driver), the build behaves exactly as v0.6.10 — llama-server subprocess for inference, per-card OpenCL walk.

Validation (live pool, 3-GPU rig)

Byte gate passed, RDNA card walking zero-dup, both Instinct cards on OpenCL, 29.36 Mh/s total (= the all-OpenCL baseline), 13 accepted / 0 rejected.

AMD assets

Windows AMD build follows via CI.