Lightning-AI / Lightning-AI/litgpt

GPU memory calculator

Open
#921 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

enhancement
Dominant language
Python
Stars
13.7k
Forks
1.5k
Avg merge
15h 37m
Merged PRs (30d)
1

Description

When scaling models to multiple GPUs, it can be difficult to estimate the memory requirement in advance, especially since it is a function of:

  • batch size
  • optimizer type
  • precision settings
  • number of parameters
  • model architecture / hyperparameters
  • FSDP settings
  • and more

We could provide a simple calculator tool that given a model config from lit-gpt, and a set of the parameters listed above, can compute the per GPU memory requirement or at least a rough estimation. Given that, the user would then be able to choose the appropriate machine type, number of GPUs, machines etc. to launch their job.

The memory usage is composed of:

  • parameters
  • activations
  • gradients
  • optimizer states

We need to compute these and sum them up to give the total estimated memory usage.

cc @lantiga @carmocca @rasbt

Contributor guide

Open the contributing guide

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

The issue names no files, tests, or entry points, so first locate the model configuration and FSDP-related code in litgpt. Clarify the calculator's inputs, estimation formulas, and supported configurations before implementation; done should include estimates for parameters, activations, gradients, and optimizer states per GPU.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
distributed-systems, machine-learning
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.