rig-llama-cpp 0.1.2

Rig completion provider for local GGUF models via llama.cpp, with streaming, tool calling, reasoning, and multimodal (mtmd) support.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31

use rig::client::CompletionClient;
use rig::completion::Prompt;
use rig_llama_cpp::{Client, KvCacheParams, KvCacheType};

#[tokio::main]
async fn main() -> Result<(), anyhow::Error> {
    let model_path =
        std::env::var("MODEL_PATH").expect("Set MODEL_PATH env var to your GGUF model file path");

    // Quantize both K and V caches to Q8_0 to roughly halve KV-cache VRAM usage
    // at long `n_ctx`, at a small accuracy cost. Try `Q4_0` for ~1/4 VRAM.
    let kv_cache = KvCacheParams::default()
        .with_type_k(KvCacheType::Q8_0)
        .with_type_v(KvCacheType::Q8_0);

    let client = Client::builder(&model_path)
        .n_ctx(32_768)
        .kv_cache(kv_cache)
        .build()?;

    let response = client
        .agent("local")
        .preamble("You are a helpful assistant.")
        .max_tokens(256)
        .build()
        .prompt("In one sentence, what is KV-cache quantization?")
        .await?;

    println!("{response}");
    Ok(())
}