tenstorrent/tt-mlir

What it solves

It provides a way to compile and optimize machine learning models specifically for Tenstorrent's custom AI hardware (such as Wormhole and Blackhole), ensuring that high-level models can be efficiently translated into hardware-optimized binaries.

How it works

Built on the MLIR (Multi-Level Intermediate Representation) compiler infrastructure from LLVM, tt-mlir defines custom dialects and transformation passes. It translates models from front-end dialects, such as StableHLO, into binaries that target the TTNN (Tenstorrent Neural Network) backend.

Who it’s for

Developers and engineers working with Tenstorrent AI accelerators who need to deploy optimized ML models on this specific hardware architecture.

Highlights

• Broad Support: Designed to support a wide range of AI workloads, including training.
• Scalability: Includes first-class primitives for scaling models across multi-chip systems.
• Performance: Focuses on delivering high out-of-the-box performance for AI models.
• Open Source: The entire development process is conducted openly.