Skip to content

troubleshooting.md

Latest commit

 

History

History
31 lines (19 loc) · 2.69 KB

troubleshooting.md

File metadata and controls

31 lines (19 loc) · 2.69 KB

Troubleshooting Guide

This page collects common issues and setup tips encountered while working through the book.

 

Notebook Image Loading Issues

The chapter notebooks use Markdown image links hosted at https://sebastianraschka.com/images/LLMs-from-scratch-images/.... This keeps the repository download size manageable, but it also means the images depend on the image host and your network connection.

If images in the .ipynb notebooks do not render:

  • Open one of the image URLs directly in your browser, for example https://sebastianraschka.com/images/LLMs-from-scratch-images/ch02_compressed/02.webp.
  • If the URL does not load in the browser either, the issue is likely a temporary website, DNS, VPN, proxy, firewall, or local network problem rather than a notebook problem.
  • I recommend double-checking the URL on a different device or network (e.g., try opening the image on your phone); if the image loads fine on your phone, it likely points to a VPN or firewall issue on your computer.
  • If the images also don't load on your phone, please feel free to open GitHub Issue to help me debug this further.

 

Apple Silicon and MPS Support

Some notebooks and scripts use cuda when available and otherwise fall back to cpu, without selecting Apple's mps backend. This omission of mps support is intentional in many places because earlier PyTorch/MPS versions produced unstable or different results in several examples, especially during training and finetuning.

If you are using an Apple Silicon Mac and see diverging losses, sharp loss spikes, poor generated text, or results that do not match the book, rerun the example on cpu first. For faster training with book-matching behavior, I recommend using cuda on a local NVIDIA GPU or a cloud GPU.

Newer PyTorch versions may improve MPS behavior, and you can experiment with mps locally if you validate the results carefully. However, if you add mps support to a script yourself, keep in mind that CUDA-specific options such as pin_memory=True, torch.compile, and DDP/multi-GPU code may need separate guards.

For more context, see #977, #625, #644, #442, and #846.

 

Other Issues

For other issues, please feel free to open a new GitHub Issue.