All Tools
M
Fine-tuningFreeOpen Source
MAXTEXT
Simple, scalable JAX LLM training
Apache-2.0
ABOUT
JAX LLM training stacks are often research mazes with poor out-of-box throughput. MaxText is a relatively small, config-driven trainer tuned for high model-flop utilization on Cloud TPU and GPU, so labs can pretrain or finetune decoder models without assembling Pax and MaxText forks themselves.
INTEGRATION GUIDE
1. Pretrain a decoder LLM on Cloud TPU with a documented high-MFU config
2. Finetune a JAX transformer without standing up a full Pax stack
3. Compare TPU training efficiency against a PyTorch trainer on the same model
4. Fork a compact JAX trainer to test architecture or data changes
TAGS
pythonjaxtpullmtraininggoogleopen-source