Tutorials

This section provides a collection of Python tutorials designed to help you get started with TT-NN for tasks such as tensor operations, model conversion, and inference.

To run these tutorials smoothly, we recommend using a Python virtual environment with the necessary dependencies installed. You can set this up in one of two ways:

• Full Development Environment: Follow the instructions in the TT-NN / TT-Metal Installation Guide to set up the complete development environment.

• Lightweight Tutorial Environment: Use the provided Python Environment Setup Script to create a minimal virtual environment specifically for running the tutorials.

Each tutorial also has an equivalent standalone Python script that you can run locally. These scripts are located in the ttnn/tutorials/basic_python/ directory of the TT-Metal repository.

With your virtual environment activated, you can run the tutorials directly:

$ python3 --version
Python 3.10.12
$ python3 example.py
...

Available tutorials:

• Add Tensors
  • Import the necessary libraries
  • Open Tenstorrent device
  • Tensor Creation
  • Perform the addition operation and convert back
  • Close The Device
  • Full example and output
• More Basic Tensor Operations
  • Import the necessary libraries
  • Open Tenstorrent device
  • Host Tensor Creation
  • Convert Host Tensors to TT-NN Tiled Tensors or Create Natively on Device
  • Tile-Based Arithmetic Operations
  • Simulated Broadcasting (Row Vector Expansion)
  • Close The Device
  • Full example and output
• MLP Inference
  • Import the necessary libraries
  • Open Tenstorrent device
  • Load MNIST Test Data
  • Load Pretrained MLP Weights
  • Basic accuracy tracking, inference, loop, and image flattening
  • Close The Device
  • Full example and output
• Multi-Head Attention
  • Write Multi-Head Attention using ttnn
  • Configuration
  • Initialize activations and weights using torch
  • Convert activations and weights to ttnn
  • Run the first iteration of Multi-Head Attention
  • Run a subsequent iteration of Multi-Head Attention
  • Write optimized version of Multi-Head Attention
  • Pre-process the parameters of the optimized model
  • Run the first iteration of the optimized Multi-Head Attention
  • Run a subsequent iteration of the optimized Multi-Head Attention
  • Check that the output of the optimized version matches the output of the original implementation
  • Close the device
  • Full example and output
• Basic Convolution
  • Import the necessary libraries
  • Set manual seed for reproducibility
  • Open Tenstorrent device
  • Create forward method
  • Set input and convolution parameters
  • Create tensors
  • Run convolution operation
  • Close The Device
  • Full example and output
• Running a Simple CNN Inference on CIFAR-10
  • Setup and Imports
  • Open Tenstorrent device
  • Load CIFAR-10 Dataset
  • Load or Initialize Weights
  • Define Convolution + Pooling Stage
  • Run Inference on Test Samples
  • Close The Device
  • Full example and output
• TT-NN Visualizer
  • Prerequisites
  • Running TTNN Visualizer
  • Model Profiling
    • Generating the Memory Report
    • Generating the Performance Report
  • Result Analysis
    • Uploading Reports
    • Operations Tab
    • Tensors Tab
    • Buffers Tab
    • Graph Tab
    • Performance Tab
  • Recap