Note

Certain TT-NN tutorials currently work on Grayskull only. Please check the specific pages of tutorials below for more information.

Tutorials

This is a collection of tutorials written with Jupyter Notebooks to help you ramp up your skillset for using tt-metal. These notebooks can be found under https://github.com/tenstorrent/tt-metal/tree/main/ttnn/tutorials.

These tutorials assume you already have a machine set up with either a grayskull or wormhole device available and that you have successfully followed the instructions for installing and building the software from source.

From within the ttnn/tutorials directory, launch the notebooks with: jupyter lab --no-browser --port=8888 Hint: Be sure to always run the cells from top to bottom as the order of the cells are dependent.

• Tensor and Add Operation
  • Tensor and Add Operation
    • Creating a tensor
    • Host Storage: Borrowed vs Owned
    • Data Type
    • Layout
    • Device storage
    • Open the device
    • Initialize tensors a and b with random values using torch
    • Add tensor a and b
    • Inspect the output tensor of the add in ttnn
    • Convert to torch and inspect the attributes of the torch tensor
    • Close the device
• Matmul Operation
  • Matrix Multiplication
    • Enable program cache
  • Configuration
    • Initialize tensors a and b with random values using torch
    • Matrix multiply tensor a and b
    • Inspect the layout of matrix multiplication output
    • Inspect the result of the matrix multiplication
    • Matrix multiply tensor a and b by using more performant config
    • Close the device
• Multi-Head Attention
  • Multi-Head Attention
    • Enable program cache
    • 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
• ttnn Tracer
  • Tracing ttnn operations and torch modules/functions
    • Trace torch functions
    • Trace torch functions and ttnn operations
    • Trace torch functions, torch modules and ttnn operations
    • Trace models written using ttnn
• ttnn Profiling
  • Profiling ttnn operations
• Resnet Basic Block
  • Resnet Block
    • Torch Module (from torchvision)
    • Create torch module and preprocess it to get ttnn parameters
    • Display the parameters of the module
    • Display the traced torch graph
    • Implement ttnn version of the module. Pass in the parameters into the constructor.
    • Run ttnn module and display the traced graph
• Graphing Torch DiT_XL_2 With TTNN
  • Build a graph of a pytorch based model
    • Clone the library from https://github.com/facebookresearch/DiT.git
    • Download DiT-XL/2 Models
    • Sample from Pre-trained DiT Models and build the graph
    • Display the graph