Run a program
Install libtt, check that JAX can see the card, and compare a computation against a CPU reference.
installation & device check →relativity.cc
Open-source software for bringing existing machine learning frameworks to Tenstorrent accelerators.
We want upstream frameworks and model code to just work. Instead of maintaining our own copies of models, we put the effort into compiler and runtime integration, so your code needs little or no change.
start hacking
Serve Qwen
Run SGLang-JAX's model and server code on the Tenstorrent backend. Serve Qwen3, Qwen3.5 or Qwen3.8 on one or more Blackhole chips.
inference recipe →Write a kernel
Follow a matrix multiply through the host code, circular buffers, and reader, compute and writer kernels.
hardware & kernel guides →Reproduce a failure
Cut a failing op down to fixed inputs, look at its StableHLO, and file a report with the result and your software versions.
correctness & timing example →how execution works
Tenstorrent
libtt packages the PJRT implementation from tt-xla with pinned compiler and runtime dependencies. Read the interfaces and debugging guide →
support & open work
TorchTPU
Once TorchTPU is open-sourced, we plan to build PyTorch support around it. Until then, PyTorch runs through torchax.
current support & planned direction →Multiple chips
Inference runs on up to four chips with tensor parallelism. Speculative decoding and training on more than one chip are still in progress.
track or report an issue ↗Qwen models
SGLang-JAX serves Qwen3, Qwen3.5 and Qwen3.8 on one to four chips, with DFlash speculative decoding on one chip.
models & performance →Training
Train a tiny TorchTitan Qwen3 model through TorchAX, watch the loss drop, and reload the checkpoint to check the weights.
train a small model →