keryx-miner-supr 0.14.1

Maintenance release for the H14 lineup — update recommended for every rig that mines or serves the Gemma-4-12B tier (the default tier, 16 GB cards), and required for Tesla V100 / Volta rigs.

Fixed

• Gemma-4 answers were generated without the chat template. The Gemma-4 GGUF uses a new turn format (<|turn>role … <turn|>) that the bundled llama.cpp build cannot detect, so the model received the raw prompt — no system prompt and no turn boundary. Very short requests (such as the pool's one-word health probe) produced garbage, and on some GPUs the garbage was reduced to an empty answer, which after three attempts withdrew the model and suspended mining. The engine now renders the Gemma-4 format itself (system turn, user turn, model turn with thinking off), on CUDA, Vulkan and Metal. Answers start immediately and stop cleanly.
• Tesla V100 / Volta: inference failures that ended in "candle" errors and stopped mining. Caused by the issue above (Volta hit the empty-answer path every time). Verified live on a Tesla V100 with the legacy line: probes answered, ~1.2 MH/s on Gemma-4-12B, shares accepted.
• Wrong build line no longer takes the miner down. The CUDA inference engine now checks the GPU's compute capability against the kernels it was built with and refuses the load with a clear message naming the right download, instead of aborting the whole process on the first kernel launch (what happened with a V100 on the modern line).

Which download

No consensus, protocol or model changes; 0.14.0 and 0.14.1 miners can run side by side. H14 gate (DAA 121,985,000) handling is unchanged from 0.14.0.