facebookresearch/fairscale

Question in offload.py: Moving activation to CPU does NOT reduce GPU memory.

Open

#948 opened on Mar 4, 2022

View on GitHubContributor guide
 (14 comments) (0 reactions) (0 assignees)Python (297 forks)auto 404
bughelp wantedoffload_modeltriaged

Repository metrics

Stars
 (3,411 stars)
PR merge metrics
 (PR metrics pending)

Description

I use my cuda_active_bytes function to measure the GPU memory before and after the code line below. I find moving activation to CPU does NOT reduce GPU memory. https://github.com/facebookresearch/fairscale/blob/9f347f373e32ee5cad11a40b70b8e28a74b5e2d4/fairscale/experimental/nn/offload.py#L524

def cuda_active_bytes():
    torch.cuda.synchronize()
    stats = torch.cuda.memory_stats()
    current_active_byte =  stats['active_bytes.all.current']
    return current_active_byte

So actually all the activations generated by forward is still in GPU memory? If so, I think the code line above is redundant.

Contributor guide

Research direction
Investigate the offload mechanism in offload.py, specifically the line that moves activations to CPU. Check if the tensor is actually detached or if a reference remains on GPU. Use memory profiling to verify GPU memory before and after. Trace the activation lifecycle during the forward pass.
Tech stack
pythonpytorch
Domain
backend
Issue type
Bug
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.
1-3 hours
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.
Mostly clear
Prerequisites
PyTorchCUDA
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
60