Unit 9.4 Compiling PyTorch Models

References

• Official PyTorch 2.0 Tutorial

Code

• Part 2: Code Demo, 9.4-compile/

What we covered in this video lecture

In this lecture, we covered torch.compile, which is a new feature in PyTorch 2.0 that allows you to speed up your PyTorch code by JIT-compiling it into optimized kernels. Torch.compile is a fully additive feature, which means that it does not require you to make any changes to your existing code.

As we’ve seen, to use torch.compile, we can simply use the torch.compile() function for an exisiting PyTorch model without making any further modifications. The torch.compile function will then compile the model into an optimized kernel the first time it is called. Subsequent calls to the function or module will be much faster, as they will be executed directly on the GPU.

Additional resources if you want to learn more

I highly recommend reading the official PyTorch 2.0 article, which delves into additional details explaining how torch.compile works under the hood.

Watch Video 1 Mark complete and go to Unit 9.5 →