bigscience-workshop/petals

Allocate all GPU memory on startup

Open

#308 opened on Apr 24, 2023

View on GitHub
 (0 comments) (0 reactions) (0 assignees)Python (424 forks)batch import
help wanted

Repository metrics

Stars
 (8,248 stars)
PR merge metrics
 (No merged PRs in 30d)

Description

Suggested by @slush0 on Discord.

It would be great to allocate all GPU memory on startup (including max space for attention cache and max space for intermediate tensors created during backward).

Right now, the memory is allocated on-demand, and people may see OOMs only after a few days of running the server. Allocating everything right away would allows to see OOMs right away and choose a correct number of blocks before leaving a server to run in the background.

Any other way to check if we have enough GPU memory on startup would be also good.

Contributor guide

Research direction
Investigate the current on demand GPU memory allocation in the codebase, identify where attention cache and intermediate tensors are allocated during backward pass, and implement a mechanism to pre allocate the maximum required memory on startup. Consider using CUDA memory pools or PyTorch tensor pre allocation.
Tech stack
python
Domain
backendperformance
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
PythonGPU memory management
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
50