How to install all necessary libraries/software to write and debug Gaudi-oriented code on a local PC?

System Setup
1

Is there a way to configure my own PC with x86 CPU to run the code made for Habana’s Gaudi 1? I don’t care if all the operations run slowly on CPU, I just need to check if all the steps (htcore.mark_step(), graph compiler, etc.) work correctly on a proxy task before renting AWS DL1 and paying money for it. I’m reading the docs, and debugging locally is still confusing.
I’m writing a custom training pipeline with PyTorch, not just downloading a model from Hugging Face.

Current setup:
Threadripper 1950x or i7-8700
Ubuntu 22.04

2

You can write regular, normal Pytorch code and test it on CPU. However to run HPU specific code like htcore.mark_step(), you will need HPU access.

For migrating pytorch code to HPU, I suggest looking at the “hello world” example and also the migration guide. Later you can consider the more advanced features like mixed precision, multi card, and other optimizations etc

1 Like