Skip to main content

Custom Ops Development

Learn how to extend TT-NN by implementing custom neural network operators for your specific workloads. This guide covers creating device operations and composite operations in C++, registration, and exposing Python bindings for integration with the TT-NN framework.

Custom Operators

https://docs.tenstorrent.com/tt-metal/v0.55.0/ttnn/ttnn/adding_new_ttnn_operation.html

Create Custom TT-NN Operations​

The TT-NN library supports extending its built-in operator set by letting developers define custom neural network operations that can be invoked just like the standard ones.

You can implement a new operator in C++ either as:

  • A device operation (providing the logic and program to run on the accelerator)
  • A composite operation that builds on other existing ops

Once implemented, you register the operation so that it’s callable from the TT-NN API. You can also expose Python bindings for your custom operator so it integrates seamlessly with Python applications.

This extensibility makes it easy to tailor TT-NN to support new primitives or experimental kernels specific to your workloads.