magnet

📖 docs | 💻 examples | 📓 substack

the small distributed language model toolkit

⚡️ fine-tune state-of-the-art LLMs anywhere, rapidly ⚡️

🧬 Installation

pip install llm-magnet

or

python3 setup.py install

🎉 usage

check out the example notebooks

a snippet to get you started

from magnet.base import Magnet
from magnet.base import EmbeddedMagnet

cluster = EmbeddedMagnet()
cluster.start()
magnet = cluster.create_magnet()
await magnet.align()

config = {
    "host": "127.0.0.1",
    "credentials": None,
    "domain": None,
    "name": "my_stream",
    "category": "my_category",
    "kv_name": "my_kv",
    "session": "my_session",
    "os_name": "my_object_store",
    "index": {
        "milvus_uri": "127.0.0.1",
        "milvus_port": 19530,
        "milvus_user": "test",
        "milvus_password": "test",
        "dimension": 1024,
        "model": "BAAI/bge-large-en-v1.5",
        "name": "test",
        "options": {
            'metric_type': 'COSINE',
            'index_type':'HNSW',
            'params': {
                "efConstruction": 40
                , "M": 48
            }
        }
    }
}

magnet = Magnet(config)
await magnet.align()

🔮 features

• ⚡️ It's Fast
  • fast on consumer hardware
  • very fast on Apple Silicon
  • extremely fast on ROCm/CUDA
• 🫵 Automatic or your way
  • rely on established transformer patterns to let magnet do the work
  • keep your existing data processing functions, bring them to magnet!
• 🛰️ 100% Distributed
  • processing, embedding, storage, retrieval, querying, or inference from anywhere
  • as much or as little compute as you need
• 🧮 Choose Inference Method
  • HuggingFace
  • vLLM node
  • GPU
  • mlx
• 🌎 Huge Volumes
  • handle gigantic amounts of data inexpensively
  • fault-tolerant by design
  • decentralized workloads
• 🔐 Secure
  • JWT
  • Basic
• 🪵 World-Class Comprehension
  • magnet optionally logs its own code as it's executed (yes, really)
  • build a self-aware system and allow it to learn from itself
  • emojis are the future

🧲 why

• build a distributed LLM research node with any hardware, from Rasbperry Pi to the expensive cloud
• Apple silicon first-class citizen with mlx
• embed & index to vector db with milvus
• distributed processing with NATS
• upload to S3
• ideal cyberpunk vision of LLM power users in vectorspace