› relativity.cc / docs
docs
Install libtt, run JAX on a Tenstorrent card, and debug what goes wrong.
Choose a starting point
- Setting up a new machine? Getting started installs the driver and the Python plugin, then checks that JAX can use the card.
- JAX already working? First experiment runs a matrix multiply, checks it against NumPy, times it, and dumps its StableHLO.
- Want to serve a model? Inference has launch commands for one or more chips and a quick MMLU check. The models page lists the supported models and how fast they run.
- Want to train? Training trains a tiny Qwen3 model with TorchTitan and TorchAX, saves a checkpoint, and checks it in PyTorch.
- Working on the compiler or runtime? Software stack explains how the pieces fit together and what to collect when something breaks.
Most examples use a single Blackhole p150a. Inference also runs across several chips; training currently covers a tiny model on one chip.
01 getting started Install the system driver and libtt plugin, then run a computation on the TT backend. → 02 software stack How PyTorch via torchax and JAX reach Tenstorrent through libtt, and where TorchTPU fits in the roadmap. → 03 first experiment Check a matrix multiply against NumPy, measure warm execution, and save its StableHLO. → 04 inference Serve Qwen models with SGLang-JAX and libtt on one or more Blackhole chips. → 05 training Train a tiny TorchTitan Qwen3 model on a p150a, save its weights, and check the checkpoint with PyTorch. →