Llamanet is an open source, drop-in library/tool that lets you instantly turn any OpenAI-powered apps into llama.cpp apps, with just one line.
It works with zero setup and zero 3rd party dependency so you don't need to know anything, and just treat it like OpenAI.
Llamanet ships with an isomorphic API that works across:
- JavaScript
- Python
- CLI
- Want to instantly port an OpenAI API based LLM app to use local LLMs?
- Don't want to ask your users to download a 3rd party LLM app or server just to use your app?
- Want your app to take care of all the LLM management on its own, without relying on 3rd party systems?
Meet Llamanet.
Llamanet is an embeddable llama.cpp management system that "just works", automagically. It lets you instantly replace OpenAI with one line of code in your app. Because the engine is embedded in your app, you don't need to tell your users to install a 3rd party LLM app or server just to use your app.
- Ship your app without 3rd party LLMs: Llamanet is a self-contained library/cli tool that comes with everything you need to serve LLM, natively from your app. No need to download a 3rd party LLM app/server. Your app users only need to install your app. It just works.
- OpenAI API Compatible Server: Llamanet is a proxy server that can run and route to multiple Llama.cpp servers, which is OpenAI API Compatible. This compatibility means you can turn ANY existing OpenAI API powered app into Llama.cpp powered app, with just one line.
- Automagical Model Management System: The built-in model management system gets rid of the need to separately and manually download checkpoints. Just make an OpenAI compatible API request with ANY GGUF URL on Huggingface. Instead of specifying
"model": "gpt-4-turbo", you specify a huggingface url (for example:"model": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct-gguf/resolve/main/Phi-3-mini-4k-instruct-fp16.gguf") and should automatically download the checkpoint and launch an embedded llama.cpp server with the checkpoint, after which the request will be routed to the server.
Here's llamanet in action:
Note that it downloads llamacpp and the models on the fly the first time it's run.
Watch the video to see it in action:
https://github.com/pinokiocomputer/llamanet
Let's try the fastest way to get going. Start the llamanet server:
On Linux/Mac:
LLAMANET_DEBUG=true npx llamanet@latest
On Windows:
set LLAMANET_DEBUG=true && npx llamanet@latest
Then open another terminal and run a curl command:
curl --request POST \
--url http://127.0.0.1:42424/v1/chat/completions \
--header "Content-Type: application/json" \
--data '{
"model": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct-gguf/resolve/main/Phi-3-mini-4k-instruct-q4.gguf",
"messages": [
{ "role": "system", "content": "You are a helpful assistant." },
{ "role": "user", "content": "Hello!" }
]
}'
Here's the full process:
Let's first start by using the openai npm package. First install llamanet and openai:
npm install llamanet openai
Then create a file named index.js:
// 1. Require llamanet
const llamanet = require("llamanet");
const OpenAI = require('openai');
(async () => {
// 2. Start llamanet
await llamanet.run()
const openai = new OpenAI()
const stream = await openai.chat.completions.create({
// 3. Specify the Huggingface URL of any GGUF model
model: "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct-gguf/resolve/main/Phi-3-mini-4k-instruct-fp16.gguf",
messages: [{
role: 'user',
content: "what comes after 1 2 3 5 8? give me as many variations as possible with reasons"
}],
stream: true,
});
for await (const chunk of stream) {
process.stdout.write(chunk.choices[0]?.delta?.content || '');
}
})();And run it:
node index
When you run it for the first time, it will download llamacpp and the requested model before making the request:
If you try running the same code again, it will not require the downloads therefore the request will go through immediately:
The openai package is simply an HTTP client that makes HTTP requests in certain format. Let's try WITHOUT using the openai package and directly make fetch requests:
const llamanet = require("llamanet");
const url = 'http://127.0.0.1:42424/v1/chat/completions';
(async () => {
await llamanet.run();
const result = await fetch(url, {
method: "post",
headers: {
"Content-Type": "application/json"
},
body: JSON.stringify({
model: "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct-gguf/resolve/main/Phi-3-mini-4k-instruct-fp16.gguf",
messages: [
{ role: "system", content: "You are a helpful assistant." },
{ role: "user", content: "Hello!" }
]
})
}).then((res) => {
return res.json()
});
console.log(JSON.stringify(result, null, 2));
})();Let's start by installing the openai and llamanet dependencies:
pip install openai llamanet
Now create a file named app.py
from openai import OpenAI
import llamanet
llamanet.run()
client = OpenAI()
stream = client.chat.completions.create(
model="gpt-3.5-turbo",
messages=[{"role": "user", "content": "why did chicken cross the road?"}],
stream=True,
)
for chunk in stream:
if chunk.choices[0].delta.content is not None:
print(chunk.choices[0].delta.content, end="")Run the file:
python app.py
Here's the full flow:
Note that it's taking an existing OpenAI powered app and just adding two lines (import llamanet and llamanet.run()), and it just works.
const llamanet = require('llamanet');
await llamanet.run();You can pass commands to the await llamanet.run() call to run commands:
await llamanet.run(COMMAND_ARRAY)For example:
awati llamanet.run([
"start",
"https://huggingface.co/bartowski/Phi-3-medium-128k-instruct-GGUF/resolve/main/Phi-3-medium-128k-instruct-IQ2_M.gguf?download=true"
"-c",
128000,
"--verbose"
})Above call is equivalent to running:
npx llamanet start https://huggingface.co/bartowski/Phi-3-medium-128k-instruct-GGUF/resolve/main/Phi-3-medium-128k-instruct-IQ2_M.gguf?download=true -c 128000 --verbose
Which starts a llama.cpp server with the model from https://huggingface.co/bartowski/Phi-3-medium-128k-instruct-GGUF/resolve/main/Phi-3-medium-128k-instruct-IQ2_M.gguf?download=true
import llamanet
llamanet.run()You can pass commands to the llamanet.run() call to run commands:
llamanet.run(["start", "https://huggingface.co/bartowski/Phi-3-medium-128k-instruct-GGUF/resolve/main/Phi-3-medium-128k-instruct-IQ2_M.gguf", "-c", 128000, "--verbose"])If you want to use llamanet with non-node.js apps, you can also do so easily by using the CLI commands:
npx llamanet start
This will start llamanet locally.
windows
set OPENAI_BASE_URL=http://localhost:42424/v1
set OPENAI_API_KEY=llamanet
<YOUR COMMAND HERE>
linux/mac:
export OPENAI_BASE_URL=http://localhost:42424/v1
export OPENAI_API_KEY=llamanet
<YOUR COMMAND HERE>
For example:
export OPENAI_BASE_URL=http://localhost:42424/v1
export OPENAI_API_KEY=llamanet
python app.py
This will run app.py, and if the app.py is powered by OpenAI, all requests will be routed to the locally running llamanet we started in step 1.
The simplest way to integrate Llamanet into your app is to use the npm package.
But you can also use terminal commands, which also means you can integrate into any programming language simply by spawning a process with the commands.
start a llamanet server if it's not already running. This will start the llamanet daemon, which acts as a proxy and a management system for starting/stopping/routing incoming requests to llama.cpp servers.
llamanet server is NOT llama.cpp server.
llamanet is a management server that automatically launches and routes one or more llama.cpp servers
npx llamanet on
const llamanet = require('llamanet');
await llamanet.run(["on"])Turn off the llamanet server.
npx llamanet off
const llamanet = require('llamanet');
await llamanet.run(["off"])start a llama.cpp server for any GGUF model by specifying a huggingface url. If llamanet is not already running, launch llamanet first.
npx llamanet start <HUGGINGFACE_URL>
Custom llama.cpp launch
npx llamanet start <HUGGINGFACE_URL> -c 128000 --verbose
const llamanet = require('llamanet');
await llamanet.run([
"start",
<HUGGINGFACE_URL>,
"-c",
128000
"--verbose"
])stop a llama.cpp server if it's running
Stop a specific llama.cpp endpoint:
npx llamanet stop <HUGGINGFACE_URL>
stop all endpoints:
npx llamanet stop
Stop a specific llama.cpp endpoint:
const llamanet = require('llamanet');
await llamanet.run(["stop", <HUGGINGFACE_URL>"])Stop all endpoints:
const llamanet = require('llamanet');
await llamanet.run(["stop"])print status
npx llamanet status
const llamanet = require('llamanet');
await llamanet.run(["status"])display models
npx llamanet models
const llamanet = require('llamanet');
const models = await llamanet.run(["models"])Here's what happens when you run await llamanet.run():
- Downloads
llamacppprebuilt binaries to~/llamanetif it doesn't already exist. - Starts a llamanet proxy server at port 42424 if it's not already running
- Sets the
OPENAI_BASE_URLenvironment variable to http://localhost:42424/v1 so you don't have to manually change thebaseUrlattribute for OpenAI clients.
You don't have to change ANYTHING in your existing OpenAI powered code, because the OPENAI_BASE_URL is already set to http://localhost:42424/v1 so all requests will be routed to the llamanet proxy, which will spawn and route requests to the embedded llama.cpp servers.
Let's take a look at an example:
const response = await openai.chat.completions.create({
model: "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct-gguf/resolve/main/Phi-3-mini-4k-instruct-fp16.gguf",
messages: [{ role: 'user', content: 'what is 1+3' }],
});When you make the request,
- If the model
https://huggingface.co/microsoft/Phi-3-mini-4k-instruct-gguf/resolve/main/Phi-3-mini-4k-instruct-fp16.ggufhas not been downloaded, it downloads the model to the~/llamanetfolder. - Then llamanet checks if the llamacpp server for the model is already running. If not already running, it first starts a llama.cpp server for the model.
- Then routes the request to the llamacpp server.
You can customize the behavior of Llamanet with environment variables.
LLAMANET_PATH: Where to store everything. (Default:~/llamanet)LLAMANET_DEFAULT_MODEL: Every API request contains amodelattribute (example: https://huggingface.co/NousResearch/Hermes-2-Pro-Llama-3-8B-GGUF/resolve/main/Hermes-2-Pro-Llama-3-8B-F16.gguf). If the specified model url fails to download, theLLAMANET_DEFAULT_MODELenvironment variable will be used to decide the default model to load from (The default is https://huggingface.co/microsoft/Phi-3-mini-4k-instruct-gguf/resolve/main/Phi-3-mini-4k-instruct-fp16.gguf)LLAMANET_OFFLINE: Ensure everything runs offline. SettingLLAMANET_OFFLINEtotruehas the effect of disabling remote HTTP requests (such as downloading checkpoints from huggingface, downloading the llama.cpp binaries, etc.)LLAMANET_DEBUG: set totrueto display the entire llamanet log in realtime. The default isfalse.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent