tt-metal

Why we included this project

Tenstorrent's AI accelerators don't run on CUDA, and that's the gap this repo fills. It bundles two layers: TT-NN, a Python and C++ operator library with tensor APIs that feel familiar if you've used PyTorch, and TT-Metalium, a lower-level kernel programming model for writing code that runs directly on the chip's cores. So you can start with high-level, framework-style code and drop to raw kernels when you need to tune performance. The repo also includes optimized implementations of real models, among them Llama, Qwen, Mixtral, Whisper, and Stable Diffusion, with published token throughput numbers for Wormhole and Blackhole boards. If you're sizing up Tenstorrent gear for LLM serving or generative workloads, this is where you'll see what the hardware can actually do and how to build on it.