Lightning-AI/pytorch-lightning

Update deepspeed activation checkpointing docs

Open

#17621 opened on May 12, 2023

View on GitHub
 (1 comment) (0 reactions) (0 assignees)Python (26,687 stars) (3,233 forks)batch import
docshelp wantedstrategy: deepspeed

Description

📚 Documentation

In your documentation, you refer to the function deepspeed.checkpointing.checkpoint, but it looks like it does not exist (anymore?). Can you update that section?

While we're at it, can you provide a more common use-case as an example? The guide warns against wrapping an entire model, but having a pretrained language model from transformers is probably the most common use case. What if someone is just using GPT-2 or T5 with no further mathematical layers? What should get wrapped then?

cc @borda @awaelchli

Contributor guide

Tech stack
pythonpytorch
Domain
documentationmachine learning
Issue type
documentation
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
2
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
under 1 hour
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
stale
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
clear
Prerequisites
PyTorch LightningDeepSpeed basics
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
80
Research direction
First, verify the correct DeepSpeed activation checkpointing function by checking DeepSpeed documentation or source code. The issue reports that `deepspeed.checkpointing.checkpoint` does not exist. Update the docs to reference the correct function. Then, add a practical example using a pretrained transformer model (e.g., GPT 2 from Hugging Face transformers) to demonstrate wrapping only specific layers rather than the whole model. The example should clarify what to wrap when no custom layers are added.