node-llama-cpp

Run AI models locally on your machine

_{Pre-built bindings are provided with a fallback to building from source with cmake}

✨ New! Try the beta of version 3.0.0 ✨ (included: function calling, automatic chat wrapper detection, embedding support, and more)

Features

• Run a text generation model locally on your machine
• Metal, CUDA and Vulkan support
• Pre-built binaries are provided, with a fallback to building from source without node-gyp or Python
• Chat with a model using a chat wrapper
• Use the CLI to chat with a model without writing any code
• Up-to-date with the latest version of llama.cpp. Download and compile the latest release with a single CLI command.
• Force a model to generate output in a parseable format, like JSON, or even force it to follow a specific JSON schema

Documentation

• Cuda Support
• Getting started guide
• API reference
• CLI help
• Changelog
• Roadmap

Installation

npm install --save node-llama-cpp@beta.34

This package comes with pre-built binaries for macOS, Linux and Windows.

If binaries are not available for your platform, it'll fallback to download the latest version of llama.cpp and build it from source with cmake. To disable this behavior set the environment variable NODE_LLAMA_CPP_SKIP_DOWNLOAD to true.

Note the prev --cuda flag got replaced by --gpu cuda --gpu amd

npx --no node-llama-cpp download --gpu cuda

const model = new LlamaModel({
    modelPath,
    gpuLayers: 64 // or any other number of layers you want
});

You'll see logs like these in the console when the model loads:

llm_load_tensors: ggml ctx size =    0.09 MB
llm_load_tensors: using CUDA for GPU acceleration
llm_load_tensors: mem required  =   41.11 MB (+ 2048.00 MB per state)
llm_load_tensors: offloading 32 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloading v cache to GPU
llm_load_tensors: offloading k cache to GPU
llm_load_tensors: offloaded 35/35 layers to GPU
llm_load_tensors: VRAM used: 4741 MB

On Linux, you can monitor GPU usage with this command:

watch -d nvidia-smi

Usage

import {fileURLToPath} from "url";
import path from "path";
import {getLlama, LlamaChatSession} from "node-llama-cpp";

const __dirname = path.dirname(fileURLToPath(import.meta.url));

const llama = await getLlama();
const model = await llama.loadModel({
    modelPath: path.join(__dirname, "models", "dolphin-2.1-mistral-7b.Q4_K_M.gguf")
});
const context = await model.createContext();
const session = new LlamaChatSession({
    contextSequence: context.getSequence()
});

const q1 = "Hi there, how are you?";
console.log("User: " + q1);

const a1 = await session.prompt(q1);
console.log("AI: " + a1);

const q2 = "Summarize what you said";
console.log("User: " + q2);

const a2 = await session.prompt(q2);
console.log("AI: " + a2);

For more examples, see the getting started guide

Contributing

To contribute to node-llama-cpp read the contribution guide.

Acknowledgements

• llama.cpp: ggerganov/llama.cpp

If you like this repo, star it ✨