Lightning-AI/lit-llama

Resume in the pretraining code

Open

#359 opened on Jun 2, 2023

View on GitHub
 (2 comments) (1 reaction) (0 assignees)Python (5,533 stars) (473 forks)batch import
help wanted

Description

I would like to request a new feature in the code: the ability to resume training from a checkpoint.

Currently, the code can save a checkpoint of the model's state at any point during training. However, there is no way to resume training from a checkpoint.

The code can save two things along with the model state_dict: 1)the optimizer, 2)the id of the last example it has seen (assuming the data is fed sequentially to the model not randomly)

Contributor guide

Tech stack
pythonpytorch
Domain
machine learningai
Issue type
feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
1-2 days
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.
clear
Prerequisites
Basic Python and PyTorch knowledgeUnderstanding of training loop checkpoints
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
60
Research direction
First, examine the current checkpoint saving logic in the pretraining code (likely in the `train.py` or similar file). The implementation should add a ` resume from checkpoint` argument that loads the model state dict, optimizer state dict, and the last processed example ID. Then, modify the data loading to skip already processed examples and continue training from that point. Refer to the existing saving logic and any comments in the code for guidance.
Resume in the pretraining code · Lightning-AI/lit-llama#359 | Good First Issue