Getting Started w TT-Blacksmith

Setup

In order to make use of TT-Blacksmith experiements, you need to clone GitHub repository first:

git clone https://github.com/tenstorrent/tt-blacksmith.git

After cloning the repository you must build and activate one of the Python frontend environments (for TT-Forge-FE or TT-XLA) in order to make use of Tenstorrent hardware. In the following sections we will provide step by step instructions on how to build and activate your frontend environments.

Build Frontend environment

TT-Forge-FE

To build the TT-Forge-FE frontend, you need to run:

./scripts/build_frontends.sh --ffe

TT-XLA

Since TT-XLA frontend depends on the MLIR environment, link to your toolchain directory needs to be stored in TTMLIR_TOOLCHAIN_DIR. If you do not specify toolchain directory, scripts will default to:

/opt/ttmlir-toolchain

If you’re setting up for the first time (or don’t have the MLIR environment installed), you need to perform a full build:

./scripts/build_frontends.sh --xla --full

For subsequent any builds, a regular rebuild is enough:

./scripts/build_frontends.sh --xla

Activating Frontend Environment

To activate the previously built Python environments for specific frontends, you need to run:

TT-Forge-FE:

source ./scripts/activate_frontend.sh --ffe

TT-XLA:

source ./scripts/activate_frontend.sh --xla

Cleaning Build Files

If at run into problems with building TT-XLA environment, try cleaning the previous build with:

source ./scripts/activate_frontend.sh --clean [--full]

Second parameter --full is optional and defines if you would like to remove TT-MLIR third party repository, so that the next build can start with fetching latest version of TT-MLIR and building it from scratch.

Running Experiments

This section guides you through the process of running experiments included in this project, allowing you to reproduce results and explore different configurations.

Explore Available Experiments: Browse the experiments documentation to find a list of all available experiments.
Understand Experiment Details: Before running an experiment, review its dedicated README file for high-level description and specific instructions.
Execute the Experiment: Follow the detailed steps outlined in the experiment’s README file to run it successfully.
Experiment with Configurations: Feel free to modify the experiment configurations (e.g., parameters) as described in the README to observe their impact on the results.