SUPER-GIANT is my end-to-end framework for custom large language models, covering data preparation, resumable training, fast inference, and Anchor-TiDAR speculative decoding. I built it in JAX to stay adjustable, scalable, and easy to extend as the architecture evolves.

• Data
  Data preparation pipeline HuggingFace dataset and tokenizer integration, plus custom tokenizer training and data curation models tuned to your domain.
• Train
  Training loop Fully resumable runs with S3 sync, CI/CD automation, and multi-GPU training in v3 without the usual Docker and VPN friction.
• Serve
  Inference pipeline Fast decoding, optimized KV-cache layouts, and chat-oriented interactions for practical local or hosted use.
• Speed
  Anchor-TiDAR speculative decoding Roughly 5x to 10x faster GPU inference, so existing hardware can handle a lot more useful work.

Watch my YouTube introduction and overview of the original GIANT implementation