torch-snapkit 0.1.4

Generate PyTorch CUDA memory snapshots for torch-memory-visualizer

Torch Snapkit

This is a light-weight wrapper around PyTorch's memory recording utils, that additionally attaches the source code of your files for later analysis. The snapshots are designed to be visualized with Torch Memory Visualizer, which helps tracking memory usage back to individual lines and shows you how much memory which of your modules used.

Example Usage

import torch
from torch import nn
from torch_snapkit import memory_snapshot

def main():
    model = nn.Sequential(nn.Linear(32, 10), nn.ReLU(), nn.Linear(10, 32))
    model.to(device='cuda')
    x = torch.randn((64, 32)).to(device='cuda')
    y = model(x)
    y.sum().backward()
 

if __name__=='__main__':
    with memory_snapshot('test_run', save_path='.'):
        main()

In addition, the memory_snapshot manager provides some convenient options:

• on_oom: Configures PyTorch to save memory snapshots after OOM occured.
• share: Enables you to quickly download the generated snapshots by using croc, a peer-to-peer filesharing tool. This flag is for situations where you generate your snapshots on a cloud instance, but want to inspect the stacktraces on your local machine.

Here is an example usage of the snapshot utility with share=True:

import torch
from torch import nn
from torch_snapkit import memory_snapshot

def main():
    model = nn.Sequential(nn.Linear(32, 10), nn.ReLU(), nn.Linear(10, 32))
    model.to(device='cuda')
    x = torch.randn((64, 32)).to(device='cuda')
    y = model(x)
    y.sum().backward()
 

if __name__=='__main__':
    with memory_snapshot('test_run_shared', share=True, share_code='your_secret_code'):
        main()

Then, on your local machine, you can either run

CROC_SECRET="your_secret_code" croc

after the remote run completed, or you can run

snapkit-receive your_secret_code

which just calls croc in a loop and downloads all files that are being send under that code.