Yes. Tenstorrent's documentation describes TT-Forge as an "end-to-end MLIR compiler" that compiles models from PyTorch, JAX and ONNX.
How the pieces fit
Tenstorrent's documentation separates the stack into layers. TT-Forge takes a PyTorch model and compiles it. TT-NN is a high-level Python API of pre-optimized neural-network operations, and TT-Metalium is a low-level C++ SDK for custom kernels on Tensix cores. TT-MLIR is the compiler infrastructure underneath that lowers models onto the hardware.
The documentation does not go into per-operator PyTorch coverage on that page, so test your model before assuming it compiles cleanly.
Related
The stack is open, as explained in whether Tenstorrent's software is open source. For the chips against NVIDIA, see Tenstorrent Blackhole vs NVIDIA. The field overview is NVIDIA's main competitors in AI.
