Lama Blanket

Lama Blanket is a desktop wrapper around llama.cpp's llama-server, built in Rust with egui.

It is aimed at local model serving with a simple GUI for picking GGUF models, estimating reasonable launch settings, saving presets, and managing the server without having to remember terminal commands.

How to run

1. Download Lama Blanket.
2. Download the llama.cpp binaries from the llama.cpp repository.
3. Put the llama.cpp folder next to lama-blanket.exe.
4. Run lama-blanket.exe.

Features

• GGUF model inspection and metadata parsing
• Hardware detection for CPU, RAM, and GPU-aware defaults
• Automatic launch argument generation for llama-server
• Resource estimates for RAM, VRAM, KV cache, and GPU layer offload
• Saved presets with JSON import support
• Automatic mmproj detection for multimodal models
• Light and dark theme

Current Scope

The project is currently stable and usable, only tested on Windows.
Some functionality, such as startup registration and the current tray behavior work, is Windows-focused.

How llama-server Is Found

Lama Blanket looks for the server executable in this order:

1. llama.cpp/llama-server.exe
2. ../llama.cpp/llama-server.exe
3. llama-server.exe

If you keep a local llama.cpp checkout beside this app, the first two options are the expected layout.

Build

To build Lama Blanket you need:

• Rust toolchain
• Cargo

cargo build

For a release build:

cargo build --release

To create a release folder and zip in dist/:

npm run dist

The generated zip contains Lama Blanket only. Users still need to download the llama.cpp binaries separately.

How Presets Are Estimated

The app uses heuristics from model metadata and detected hardware:

• Context size starts from the GGUF context length, capped at 128000, or falls back to 4096.
• GPU layers are estimated from available VRAM after subtracting a safety overhead and estimated KV cache usage.
• Threads default to physical CPU cores.
• Batch threads default to logical CPU cores.
• Flash attention defaults to auto for NVIDIA-backed llama models and off otherwise.
• mlock defaults on when available RAM is comfortably above model size.

These are starting points, not guaranteed best values. Larger context sizes, aggressive offload, and quantization choices can still require manual tuning.

Presets

Presets are stored in JSON and persist across launches.

• Save stores the current configuration as a named preset.
• Import... imports one preset or an array of presets from a JSON file.
• Imported presets are merged into the local preset list and remain available until deleted.
• If an imported preset has the same name as an existing preset, it replaces it.

Development

Useful commands:

cargo check
cargo test

License

MIT. See LICENSE.