Lightning-AI/pytorch-lightning

Integrating pytorch XLA when using multiple GPUs

Open

#16,130 opened on Dec 20, 2022

View on GitHub
 (13 comments) (4 reactions) (0 assignees)Python (3,233 forks)batch import
accelerator: cudafeaturehelp wantedstrategy: xla

Repository metrics

Stars
 (26,687 stars)
PR merge metrics
 (Avg merge 9d 15h) (3 merged PRs in 30d)

Description

Description & Motivation

I've experienced with pytorch XLA using multitple NVIDIA A100 GPU and I observed that in most cases training is faster. So it would be really nice to have the option to use XLA for training in pytorch lightning.

The main advantage is faster training.

Additional context

Here is a code link : https://github.com/Dhouib-med/Test-XLA/blob/17e5b6bd6c77fffa67818462856277a57877ff3b/test_xla.py to train a simple CNN on the MNIST dataset using XLA (on 2 GPUS). The main parts where taken from https://github.com/pytorch/xla. This wheel needs to be installed along with adequate pytorch and torchvision versions (1.11 and 0.14) https://storage.googleapis.com/tpu-pytorch/wheels/cuda/112/torch_xla-1.13-cp37-cp37m-linux_x86_64.whl @justusschock

cc @borda @justusschock @awaelchli @carmocca @JackCaoG @steventk-g @Liyang90

Contributor guide

Research direction
Study PyTorch Lightning's multi GPU strategy implementation and XLA device integration. Examine existing XLA support in Lightning and identify where to add multi GPU XLA handling.
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.
Active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Clear
Prerequisites
PythonPyTorchPyTorch LightningCUDA
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
45