Lightning-AI / Lightning-AI/lit-llama

Group data by length to reduce wasted computation

Open
#396 3 comments 2 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
6.1k
Forks
517
PR merge metrics
No merged PRs in 30d

Description

Informed by [this post](https://discuss.huggingface.co/t/fine-tuning-pre-training-tips/15367), I implemented the `group_by_length` feature in finetuning scripts and found on single V100 GPU `finetune/lora.py` LLaMA-7B with the same hyper parameters:

- Without `group_by_length`: 8 hours, 37 minutes, 1.353 seconds
- With `group_by_length`: 6 hours, 15 minutes, 24.636 seconds

So the time saving is about 27%. I can submit a PR if anyone is interested.

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Begin with finetune/lora.py, the only file named in the issue, then inspect the other finetuning scripts for their corresponding entry points. Use the reported single-V100 comparison and unchanged hyperparameters as the validation reference. Done means length grouping is available in the intended finetuning scripts and its behavior can be checked against the reported training result.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning
Issue type
Feature
Difficulty
3/5
Estimated time
1-2 days
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
38/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.