What is Felafax?

Felafax provides an enterprise-grade Artificial Intelligence platform engineered for simplicity, scalability, and openness. It supports running AI workloads on a variety of accelerators, including Google TPU, AWS Trainium, NVIDIA, and AMD hardware. The platform aims to deliver up to 2X cost-efficiency compared to standard setups without compromising on performance, achieving H100-level results at approximately 30% lower cost through its custom training platform built with XLA compiler and JAX.

Designed for enterprise needs, Felafax allows for on-premise deployment within a user's Virtual Private Cloud (VPC), ensuring data security and privacy. It simplifies the complexities of Machine Learning Operations (ML Ops) by handling tasks such as model partitioning for large models (e.g., Llama 3.1 405B), multi-controller training, and inference orchestration. Users benefit from effortless scaling, with one-click cluster spin-up from 8 to 1024 TPU chips, and high customizability through a no-code UI or direct Jupyter notebook access.

Features

Scale Effortlessly: One-click spin-up of clusters from 8 to 1024 TPU chips with seamless training orchestration.
Performance at Lower Cost: Custom training platform using XLA compiler and JAX achieves H100-level performance at reduced costs.
On-prem deployment: Deploy within your VPC, ensuring data security and privacy.
Highly Customizable: Offers both a no-code UI for fine-tuning and Jupyter notebook access for tailored training.
Managed ML Ops: Handles optimized model partitioning, multi-controller training, and inference.
Out-of-the-Box Templates: Pre-configured environments with PyTorch XLA or JAX and necessary dependencies.

Use Cases

Fine-tuning large language models like Llama3 within an enterprise VPC.
Deploying AI models cost-effectively on various hardware accelerators.
Scaling AI training infrastructure rapidly based on needs.
Managing complex ML Operations for large-scale model training and inference.
Developing and deploying custom AI solutions with flexible tooling (no-code or code).