tenstorrent/tt-metal

Health v2 · maintenance only

Adoption not part of the score

1640 stars · 613 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

TT-Metal is Tenstorrent's open-source software stack containing TT-NN, a Python and C++ neural network operator library, and TT-Metalium, a low-level kernel programming model and SDK for Tenstorrent AI accelerators. It enables running and optimizing models like Llama, Qwen, and Stable Diffusion directly on Tenstorrent hardware such as Wormhole, Blackhole, and Galaxy systems.

Use cases

• run LLM inference on Tenstorrent hardware
• write custom kernels for Tensix cores
• port PyTorch models to Tenstorrent accelerators
• serve Llama 3.3 70B with tensor parallelism across devices
• optimize stable diffusion image generation on AI accelerators
• benchmark tokens-per-second throughput on Wormhole Galaxy clusters
• develop fused transformer operations with a PyTorch-like API

When to choose

• you own Tenstorrent hardware (Grayskull, Wormhole, Blackhole, Galaxy) and want to run or optimize models
• you need low-level, bare-metal access to Tenstorrent silicon for custom kernel development
• you want pre-optimized neural network operations via a PyTorch-familiar Python or C++ API
• you are deploying high-throughput LLM inference with vLLM on Tenstorrent systems

When to avoid

• you only have NVIDIA/AMD GPUs and no Tenstorrent hardware
• you want a framework-agnostic compiler path without custom kernel work (consider TT-Forge instead)
• you need a turnkey inference server UI (consider tt-inference-server or TT-Studio)
• you require broad OS support beyond Ubuntu 22.04, which is the recommended platform

Facets

library · maturity active

machine-learning deep-learning llm-inference gpu-computing sdk compiler deep-learning large-language-models gpu-computing hardware developer-tools cpp python tenstorrent ttnn tt-metalium neural-network-operators custom-kernels tensix wormhole blackhole tensor-parallelism hardware-accelerator linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="tenstorrent/tt-metal")