rig-llama-cpp

A Rig completion provider that runs GGUF models locally via llama.cpp and their Rust bindings llama-cpp-2.

Usage

use rig::client::CompletionClient;
use rig::completion::Prompt;
use rig_llama_cpp::Client;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    // The minimal form — every other knob has a sensible default. Chain
    // .n_ctx, .sampling, .fit, .kv_cache, .checkpoints, or (with the
    // `mtmd` feature) .mmproj to override.
    let client = Client::builder("path/to/model.gguf")
        .n_ctx(8192)
        .build()?;

    let agent = client
        .agent("local")
        .preamble("You are a helpful assistant.")
        .max_tokens(512)
        .build();

    let response = agent.prompt("Hello!").await?;
    println!("{response}");
    Ok(())
}

The legacy positional Client::from_gguf(...) constructor is still available for callers pinned to the 0.1.x API.

Features

• Local inference with any GGUF model
• Completion and streaming support
• Tool calling (for models with OpenAI-compatible chat templates)
• Reasoning / thinking output
• Vision (multimodal) inference for models with an mmproj projector — opt in via the mtmd feature
• Automatic GPU/CPU layer fitting — llama.cpp probes available device memory and picks n_gpu_layers for you, no manual tuning required
• Backend selection via Cargo feature flags
• Configurable sampling parameters (top-p, top-k, min-p, temperature, penalties)

Feature Flags

No default GPU backend — pick the one that matches your hardware:

cargo build --features vulkan
cargo build --features "cuda,mtmd"

Toolchain and runtime requirements per backend are documented upstream in llama.cpp's build guide. A successful build does not guarantee a successful run — if backend init fails at runtime, [LoadError::BackendInit] is returned rather than panicking, so the application can fall back gracefully.

Examples

MODEL_PATH=./model.gguf cargo run --example completion
MODEL_PATH=./model.gguf cargo run --example streaming
MODEL_PATH=./model.gguf cargo run --example stream_chat
MODEL_PATH=./model.gguf cargo run --example structured_output
MODEL_PATH=./model.gguf cargo run --example kv_cache
MODEL_PATH=./embedding-model.gguf cargo run --example embeddings

# Vision (requires mtmd feature + mmproj file)
MODEL_PATH=./vision-model.gguf MMPROJ_PATH=./mmproj.gguf IMAGE_PATH=./image.jpg \
    cargo run --features mtmd --example vision

# Hot-swap the loaded model on the same worker thread
RIG_MODEL_A=./model_a.gguf RIG_MODEL_B=./model_b.gguf cargo run --example reload

N_GPU_LAYERS=20 can be used to offload 20 layers to the GPU.

By default, llama.cpp backend logs are suppressed so streaming and test output stay readable. Set RIG_LLAMA_CPP_LOGS=1 to re-enable raw backend logs when debugging model startup or decode issues.

Testing

# Fast unit tests + doctests — no model required, run on every CI build.
cargo test --lib
cargo test --doc

The full integration suite (tests/e2e/) covers streaming completions, vision, tool roundtrips, structured output, KV-cache quantization, embedding, and sequential model reload. All tests are #[ignore]d and auto-download their fixtures via hf-hub into the standard HuggingFace cache (~/.cache/huggingface/hub) on first run — plan for ~20 GB. Backend compilation is already covered upstream by llama-cpp-rs, and the model fixtures are too large for hosted runners, so the e2e suite does not run in CI.

cargo test --test e2e --features mtmd -- --ignored --nocapture

Contributing

Issues and pull requests are welcome at github.com/camperking/rig-llama-cpp.

Before opening a PR, please run the same checks CI does (.github/workflows/ci.yml):

cargo fmt --all --check
cargo clippy --no-deps --all-targets -- -D warnings
cargo clippy --no-deps --all-targets --features mtmd -- -D warnings
cargo test --lib
cargo test --doc
RUSTDOCFLAGS="-D warnings" cargo doc --no-deps --features mtmd

If your change touches inference behaviour, validate it locally with cargo test --test e2e --features mtmd -- --ignored --nocapture — the fixtures auto-download on first run (~20 GB; see the Testing section).

For changes that affect the public API or the embedded llama-cpp-2 version, add an entry to CHANGELOG.md under [Unreleased]. The crate's pre-1.0 SemVer policy is documented at the top of that file.

License

See Cargo.toml for dependency details.