Run a program
Install libtt, check that the accelerator is visible, and verify a computation against a host reference.
installation & device check →relativity.cc
Open-source software for bringing existing machine learning frameworks to Tenstorrent accelerators.
Our goal is to run upstream frameworks and model code out of the box. We focus on compiler and runtime integration so applications need little or no modification, rather than maintaining separate model implementations.
01 / start hacking
Serve Qwen3-8B
Use SGLang-JAX's model and serving code with the Tenstorrent backend. Launch Qwen3-8B on Blackhole and check its output.
inference recipe →Write a kernel
Follow a matrix multiply through host code, circular buffers, and reader, compute, and writer kernels.
hardware & kernel guides →Reproduce a failure
Reduce a failing operation to fixed inputs, inspect its StableHLO, and report the result and software versions.
correctness & timing example →02 / how execution works
01Tenstorrent
libtt packages the PJRT implementation from tt-xla with pinned compiler and runtime dependencies. Read the interfaces and debugging guide →
03 / support & open work
TorchTPU
TorchTPU is the planned focus for framework support once it is open-sourced. PyTorch support is available today through torchax.
current support & planned direction →Multiple accelerator cards
Multi-card execution is under development. The setup and inference recipes here target a single Blackhole card.
track or report an issue ↗Qwen3-8B
The documented serving path uses SGLang-JAX and the TT attention backend. Qwen 3.5 architecture support is in progress.
launch configuration & limits →Training
Train a tiny TorchTitan Qwen3 model through TorchAX, watch the loss fall, and reload its checkpoint to check the trained weights.
train a small model →