Lightning-AI/pytorch-lightning

enable loading `universal checkpointing` checkpoint in `DeepSpeedStrategy`

Open

#20,065 opened on Jul 9, 2024

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

Description

Description & Motivation

After I trained a model in some numbers of gpus, say, 8 gpus for a while, It's difficult to load the checkpoint to 16 gpus with optimizer and model states unchanged. The deepspeed has developed the universal checkpointing strategy to solve this problem, but I didn't see the pytorch-lightning has this feature.

Pitch

I want the pytorch-lightning could support this feature

Alternatives

try to add universal_checkpoint as a param of DeepSpeedStrategy and modify the class refering to https://www.deepspeed.ai/tutorials/universal-checkpointing/

Additional context

No response

cc @borda @awaelchli

Contributor guide

Tech stack
python
Domain
backendmachine learning
Issue type
feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
5
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
over 1 week
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
fresh
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
mostly clear
Prerequisites
PyTorchDeepSpeedcheckpointing basics
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
10
Research direction
The issue requests adding universal checkpointing support to DeepSpeedStrategy. The contributor should study the DeepSpeed universal checkpointing tutorial at https://www.deepspeed.ai/tutorials/universal checkpointing/. The implementation likely involves modifying the DeepSpeedStrategy class in pytorch lightning's source code (e.g., lightning fabric/strategies/deepspeed.py). The maintainer comments from @borda and @awaelchli may provide guidance. The work includes adding a parameter to enable universal checkpoint and adapting the save/load methods.