Geniusrise is a modular, loosely-coupled MLOps framework designed for the era of Large Language Models, offering flexibility and standardization in designing networks of AI agents.

It defines components and orchestrates them providing observability, state management and data handling, all while supporting diverse infrastructures. With its modular and unopinonated architecture, Geniusrise empowers teams to build, share, and deploy AI across various platforms.

Geniusrise is powered by its components:

• geniusrise-text: Text components offerring:
  • Inference APIs
  • Bulk inference
  • Fine-tuning
• geniusrise-audio: Audio components offerring:
  • Inference APIs
  • Bulk inference
  • Fine-tuning
• geniusrise-vision: Vision components offerring:
  • Inference APIs
  • Bulk inference
  • Fine-tuning
• geniusrise-listeners: Streaming data ingestion
• geniusrise-databases: Bulk data ingestion

The easiest way to use geniusrise is to host an API over a desired model. Use one of the examples from text, vision or audio.

Say, we are interested in running an API over a vision / multi-modal model such as bakLlava:

pip install torch
pip install geniusrise
pip install geniusrise-vision # vision multi-modal models
# pip install geniusrise-text # text models, LLMs
# pip install geniusrise-audio # audio models

Create a YAML file named genius.yml:

version: '1'

bolts:
    my_bolt:
        name: VisualQAAPI
        state:
            type: none
        input:
            type: batch
            args:
                input_folder: ./input
        output:
            type: batch
            args:
                output_folder: ./output
        method: listen
        args:
            model_name: 'llava-hf/bakLlava-v1-hf'
            model_class: 'LlavaForConditionalGeneration'
            processor_class: 'AutoProcessor'
            device_map: 'cuda:0'
            use_cuda: True
            precision: 'bfloat16'
            quantization: 0
            max_memory: None
            torchscript: False
            compile: False
            flash_attention: False
            better_transformers: False
            endpoint: '*'
            port: 3000
            cors_domain: 'http://localhost:3000'
            username: 'user'
            password: 'password'

Navigate to the directory and do:

genius rise

MY_IMAGE=/path/to/test/image

(base64 -w 0 $MY_IMAGE | awk '{print "{\"image_base64\": \""$0"\", \"question\": \"<image>\nUSER: Whats the content of the image?\nASSISTANT:\", \"do_sample\": false, \"max_new_tokens\": 128}"}' > /tmp/image_payload.json)
curl -X POST http://localhost:3000/api/v1/answer_question \
    -H "Content-Type: application/json" \
    -u user:password \
    -d @/tmp/image_payload.json | jq

See usage for more advanced examples. There are over 150 examples for text, vision and audio.

• Website: geniusrise.ai

This is the base geniusrise framework, also the genius CLI. It provides the base structure and orchestration primitives for Spouts and Bolts, and also covers operations.

• Docs: docs.geniusrise.ai
• Examples: geniusrise/examples
• Cloud: geniusrise.com