Minerva Travel Plan — deployed 2026-08-26 (hardware verified 2026-08-26)

The offline inference plan for the laptop minerva.lan (minerva.lan).

Verified Hardware (2026-08-26)

Fact	Value
Host	`minerva` (Ubuntu 24.04.4 LTS)
CPU	AMD Ryzen 7 7840U (16 threads)
GPU	AMD Radeon 780M (RDNA3 iGPU, shared memory; Phoenix1)
Vulkan	Mesa RADV installed (`/usr/share/vulkan/icd.d/radeon_icd.json` present); `vulkaninfo` (vulkan-tools) not yet installed
RAM	14 GiB total, ~12 GiB available (iGPU VRAM comes out of this)
Disk	935 GB, ~678 GB free
llama.cpp	installed 2026-08-26: 0.3.0-dev (build 10642), Vulkan backend, `llama.service` active on `127.0.0.1:2001`
SSH	`ssh -i ~/.ssh/keys/key-athena owner@minerva.lan` (key-athena works; no dedicated key)

Travel Model Selection

The 780M with the shared 14 GiB pool comfortably runs Q4/Q8 models up to ~7B. All candidate GGUFs are already on minerva (copied from athena 2026-08-26; exact files and sizes verified):

Model (exact GGUF on disk)	Size	Expected ctx	Use
`qwen2.5-coder-3b-instruct-q8_0.gguf`	3.37 GiB	8192	recommended default: coding + editing; FIM-capable
`Qwen3.5-4B-Q4_K_M.gguf`	2.55 GiB	8192	general chat alternative
`gemma-4-E2B-it-qat-UD-Q4_K_XL.gguf`	2.44 GiB	8192	light chat, battery-friendly
`qwen2.5-coder-1.5b-instruct-q8_0.gguf`	1.76 GiB	8192	minimal fallback; fastest

Also on disk (bonus quants that came with the repos): qwen2.5-coder-3b-instruct-q4_k_m.gguf (1.96 GiB) and qwen2.5-coder-1.5b-instruct-q4_k_m.gguf (1.04 GiB) — usable as lighter fallbacks.

Rationale: the 3B coder Q8 is the sweet spot — full coding capability, FIM support, ~2.5× the size of the 1.5B for far better output quality, and expected ~30–50 t/s on the 780M. Keep at or below 4B for battery life and thermals.

Deployment Shape — executed 2026-08-26

Build notes: Ubuntu 24.04 needs glslc and spirv-headers (not shaderc, which is beast’s package); run sudo ldconfig after cmake --install.

Build llama.cpp from source with the Vulkan backend (same process as athena/beast):

sudo apt install -y mesa-vulkan-drivers libvulkan-dev vulkan-tools cmake git
    git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
    cmake -B build -DGGML_VULKAN=ON -DCMAKE_BUILD_TYPE=Release
    cmake --build build -j && sudo cmake --install build

Verify the GPU is seen: vulkaninfo --summary shows AMD Radeon 780M (RADV PHOENIX1). Set GGML_VK_VISIBLE_DEVICES=0 if more than one device appears.
Serve one small preset on 127.0.0.1:2001 (no router needed) — a config.ini with the default model:
```
[*]
    ctx-size = 8192
    flash-attn = true
    [Qwen/Qwen2.5-Coder-3B-Instruct-GGUF:Q8_0]
    ngl = 99
    temp = 0.2
    top-p = 0.95
```
launched as llama-server --models-preset config.ini --host 127.0.0.1 --port 2001 --tools all, ideally under a systemd unit (pattern from sources/models/install.sh).
Point the tools at the local server (aider / llama.vscode / optional DSH minerva route):
- aider: aider --openai-api-base http://127.0.0.1:2001/v1 --openai-api-key sk-local --model Qwen/Qwen2.5-Coder-3B-Instruct-GGUF:Q8_0 --edit-format whole
- llama.vscode: server URL http://127.0.0.1:2001/v1, completion/chat on the 3B coder.
- Optional: a minerva DSH provider route (llm-pi-ai) if a DSH session ever runs on the laptop; same shape as the beast route in sources/config/dsh-settings-beast.example.yaml.
The GGUFs are already predownloaded into minerva’s HuggingFace cache — nothing to fetch while travelling.

Offline Checklist — status 2026-08-26

vulkaninfo --summary shows the 780M (RADV) — verified.
curl http://127.0.0.1:2001/v1/models lists the models — verified (6 listed).
Chat completion with the network disabled — verified via coder3 on the live service; the truly-offline test remains for the user (disabling WiFi would drop the SSH session).
aider and llama.vscode answer from the local server with WiFi off — pending the user’s offline test.

Provisioning

minerva.lan is in the homelab Ansible inventory (minerva.lan) with SSH. A Host minerva.lan block in ~/.ssh/config (User owner, IdentityFile ~/.ssh/keys/key-athena) was added 2026-08-26; ssh minerva.lan verified. Ansible targets of minerva now resolve the key through the SSH config.