/nitro

An inference server on top of llama.cpp. OpenAI-compatible API, queue, & scaling. Embed a prod-ready, local inference engine in your apps. Powers Jan

Primary LanguageC++GNU Affero General Public License v3.0AGPL-3.0

Nitro - Embeddable AI

nitrologo

Documentation - API Reference - Changelog - Bug reports - Discord

⚠️ Nitro is currently in Development: Expect breaking changes and bugs!

Features

  • Fast Inference: Built on top of the cutting-edge inference library llama.cpp, modified to be production ready.
  • Lightweight: Only 3MB, ideal for resource-sensitive environments.
  • Easily Embeddable: Simple integration into existing applications, offering flexibility.
  • Quick Setup: Approximately 10-second initialization for swift deployment.
  • Enhanced Web Framework: Incorporates drogon cpp to boost web service efficiency.

About Nitro

Nitro is a high-efficiency C++ inference engine for edge computing, powering Jan. It is lightweight and embeddable, ideal for product integration.

The binary of nitro after zipped is only ~3mb in size with none to minimal dependencies (if you use a GPU need CUDA for example) make it desirable for any edge/server deployment 👍.

Read more about Nitro at https://nitro.jan.ai/

Repo Structure

.
├── controllers
├── docs 
├── llama.cpp -> Upstream llama C++
├── nitro_deps -> Dependencies of the Nitro project as a sub-project
└── utils

Quickstart

Step 1: Install Nitro

  • For Linux and MacOS

    curl -sfL https://raw.githubusercontent.com/janhq/nitro/main/install.sh | sudo /bin/bash -
  • For Windows

    powershell -Command "& { Invoke-WebRequest -Uri 'https://raw.githubusercontent.com/janhq/nitro/main/install.bat' -OutFile 'install.bat'; .\install.bat; Remove-Item -Path 'install.bat' }"

Step 2: Downloading a Model

mkdir model && cd model
wget -O llama-2-7b-model.gguf https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/resolve/main/llama-2-7b-chat.Q5_K_M.gguf?download=true

Step 3: Run Nitro server

nitro

Step 4: Load model

curl http://localhost:3928/inferences/llamacpp/loadmodel \
  -H 'Content-Type: application/json' \
  -d '{
    "llama_model_path": "/model/llama-2-7b-model.gguf",
    "ctx_len": 512,
    "ngl": 100,
  }'

Step 5: Making an Inference

curl http://localhost:3928/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      {
        "role": "user",
        "content": "Who won the world series in 2020?"
      },
    ]
  }'

Table of parameters

Parameter Type Description
llama_model_path String The file path to the LLaMA model.
ngl Integer The number of GPU layers to use.
ctx_len Integer The context length for the model operations.
embedding Boolean Whether to use embedding in the model.
n_parallel Integer The number of parallel operations.
cont_batching Boolean Whether to use continuous batching.
user_prompt String The prompt to use for the user.
ai_prompt String The prompt to use for the AI assistant.
system_prompt String The prompt to use for system rules.
pre_prompt String The prompt to use for internal configuration.
cpu_threads Integer The number of threads to use for inferencing (CPU MODE ONLY)
n_batch Integer The batch size for prompt eval step
caching_enabled Boolean To enable prompt caching or not
clean_cache_threshold Integer Number of chats that will trigger clean cache action
grp_attn_n Integer Group attention factor in self-extend
grp_attn_w Integer Group attention width in self-extend
mlock Boolean Prevent system swapping of the model to disk in macOS
grammar_file String You can constrain the sampling using GBNF grammars by providing path to a grammar file

OPTIONAL: You can run Nitro on a different port like 5000 instead of 3928 by running it manually in terminal

./nitro 1 127.0.0.1 5000 ([thread_num] [host] [port] [uploads_folder_path])
  • thread_num : the number of thread that nitro webserver needs to have
  • host : host value normally 127.0.0.1 or 0.0.0.0
  • port : the port that nitro got deployed onto
  • uploads_folder_path: custom path for file uploads in Drogon.

Nitro server is compatible with the OpenAI format, so you can expect the same output as the OpenAI ChatGPT API.

Compile from source

To compile nitro please visit Compile from source

Download

Version Type Windows MacOS Linux
Stable (Recommended) CPU CUDA Intel M1/M2 CPU CUDA
Experimental (Nighlty Build) GitHub action artifactory

Download the latest version of Nitro at https://nitro.jan.ai/ or visit the GitHub Releases to download any previous release.

Nightly Build

Nightly build is a process where the software is built automatically every night. This helps in detecting and fixing bugs early in the development cycle. The process for this project is defined in .github/workflows/build.yml

You can join our Discord server here and go to channel github-nitro to monitor the build process.

The nightly build is triggered at 2:00 AM UTC every day.

The nightly build can be downloaded from the url notified in the Discord channel. Please access the url from the browser and download the build artifacts from there.

Manual Build

Manual build is a process where the software is built manually by the developers. This is usually done when a new feature is implemented or a bug is fixed. The process for this project is defined in .github/workflows/build.yml

It is similar to the nightly build process, except that it is triggered manually by the developers.

Contact

  • For support, please file a GitHub ticket.
  • For questions, join our Discord here.
  • For long-form inquiries, please email hello@jan.ai.

Star History

Star History Chart