Record: Minerva deployed; athena service deduped — 2026-08-26

Outcome

• Deployed llama.cpp to minerva (task 01): built 0.3.0-dev (build 10642) from source with the Vulkan backend, installed, llama.service enabled and active on 127.0.0.1:2001, four presets (coder3 Q8 default, qwen4, gemma-e2, coder15), RADV device pinned via GGML_VK_VISIBLE_DEVICES=0.
• Deduped athena’s llama.service ExecStart (task 04): the repeated --chat-template llama3 (4x) reduced to one; service restarted, 15 models up.

Decisions

• Ubuntu 24.04 build deps: glslc + spirv-headers (beast’s distro uses the shaderc package instead); sudo ldconfig required after cmake --install for the shared-library loader.
• Minerva serves on 127.0.0.1:2001 only (travel laptop; no LAN exposure), matching documents/11-minerva-travel-plan.md.

Verification (2026-08-26)

• vulkaninfo: AMD Radeon 780M Graphics (RADV PHOENIX) (+ llvmpipe fallback, pinned away).
• curl http://127.0.0.1:2001/v1/models on minerva: 6 models listed (4 presets + 2 cache quants).
• Chat completion via coder3 alias on minerva: answered “OK”.
• Athena router: 15 models up after the dedup restart; service file now has one --chat-template llama3.

Open items

• Minerva offline test (WiFi disabled) is left for the user — disabling the interface would drop the SSH session used for deployment.
• TODO 02 (Open WebUI RAG) deferred by the owner.