Overview

Tenstorrent’s software stack takes models from popular frameworks all the way down to Tensix cores. Most users start at the top with the TT-Forge compiler; the lower layers are available when you need finer control.

TT-Forge TT-Forge™ End-to-end MLIR compiler. Compile models from PyTorch, JAX, and ONNX — the recommended entry point. TT-XLA TT-XLA PJRT-based frontend for running JAX and OpenXLA models on Tenstorrent hardware. TT-NN TT-NN™ High-level Python API of pre-optimized neural-network operations. TT-Lang TT-Lang™ Python DSL for authoring fused custom operations. TT-MLIR TT-MLIR™ MLIR-based backend compiler infrastructure that lowers models onto Tenstorrent hardware. TT-Metalium TT-Metalium™ Low-level C++ SDK for writing custom kernels on Tensix cores. Cloud-Native Support Cloud-Native Support Run Tenstorrent accelerators on Kubernetes — install, configure, and operate the device stack.

Choosing your entry point

  • Compile a model from PyTorch, JAX, or ONNX → TT-Forge (recommended start).

  • Build networks in Python from pre-optimized ops → TT-NN.

  • Author custom fused operations → TT-Lang.

  • Write kernels directly on Tensix cores → TT-Metalium.

  • Work on compiler internals or custom backends → TT-MLIR.