TorchShard is a lightweight engine for slicing a PyTorch tensor into parallel shards

Last update: Nov 22, 2022

Overview

Documents | Projects | API References

TorchShard is a lightweight engine for slicing a PyTorch tensor into parallel shards. It can reduce GPU memory and scale up the training when the model has massive linear layers (e.g., ViT, BERT and GPT) or huge classes (millions). It has the same API design as PyTorch.

Installation

pip install torchshard

More options in INSTALL.md.

Usage

import torchshard as ts

ts.init_process_group(group_size=2)                       # init parallel groups

m = torch.nn.Sequential(
    torch.nn.Linear(20, 30, bias=True),               
    ts.nn.ParallelLinear(30, 30, bias=True, dim=None),    # equal to nn.Linear()
    ts.nn.ParallelLinear(30, 30, bias=True, dim=0),       # parallel in row dimension
    ts.nn.ParallelLinear(30, 30, bias=True, dim=1),       # parallel in column dimension
).cuda()

x = m(x)                                                  # forward
loss = ts.nn.functional.parallel_cross_entropy(x, y)      # parallel loss function
loss.backward()                                           # backward

torch.save(
  ts.collect_state_dict(m, m.state_dict()), 'm.pt')       # save model state

Performance

The following figure is a showcase of training ResNet-50 on 8 NVIDIA TITAN-XP (12196 MiB) GPUs with scaling up classes from 1000 → 1 Million. The input size is 224 x 224, and the batch size is 256. Parallelism is with 8-way data parallel and 8-way model parallel.

The following figure shows training minGPT on 8 NVIDIA TITAN-XP (12196 MiB) GPUs with scaling up parameters from 10 Million → 808 Million. The input size is 32 x 32, and the batch size is 16. Parallelism is with 1-way data parallel and 8-way model parallel.

Contributing

The TorchShard welcomes your expertise and enthusiasm!

If you are interested in torchshard, you are welcome to help

polish code and develop new features
develop high-quality tutorials, projects, and advanced materials

Direct pull requests are welcome. Contact: kaiyuyue [at] umd.edu.

Citing TorchShard

If you think TorchShard is helpful in your research and consider to cite it, please use the following BibTeX entry.

@misc{torchshard2021,
  author =       {Kaiyu Yue},
  title =        {TorchShard},
  howpublished = {\url{https://github.com/KaiyuYue/torchshard}},
  year =         {2021}
}

Comments

Future Planinig on this project.
Hello Kaiyu, I love this awesome project. The API design is elegant and simple and the software is lightweight and user-friendly. My understanding is that this project has realized a series of PyTorch wrappers for tensor slicing.

I am curious about the future planning of this project.

Is there some overlap in functionality between torchshard and N-D parallelism proposed in ColossalAI.

How is compatibility with ZeRO? According to am+zero example, the memory footprint has a little change after combination torchshard with ZeRO.
opened by feifeibear 2
Which one is faster?

Thanks for contributing this great lib. I have one question. Which one is faster (in speed) between dim=0and dim=1? The documentations seem to only contain accuracy results.

opened by NOBLES5E 2
8 gpus test example raise error.

When I do Unit Tests, it can pass when use two gpu devices, run command below: CUDA_VISIBLE_DEVICES=0,1 python3 -m unittest discover -v -s tests

But I do Unit Tests with eight gpu devices, it raise ncclSystemError. run command: CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m unittest discover -v -s tests raise error: RuntimeError: NCCL error in ../torch/lib/c10d/ProcessGroupNCCL.cpp:825, unhandled system error, NCCL version 2.7.8 ncclSystemError: System call (socket, malloc, munmap, etc) failed.

Is it necessary to pass unittest in eights gpu devices?

opened by JiaquanYe 1

Error?

Hi, thanks for the excellent job! When I install it from pip, and

import torchshard as ts
ts.init_process_group(group_size=2)

The AttributeError occurs:

AttributeError: module 'torchshard' has no attribute 'init_process_group'

opened by WangWenhao0716 1

Multi-node setting?

https://github.com/KaiyuYue/torchshard/blob/89e21def180bf6063ceb2e312a61631173abc7e7/projects/minGPT/main.py#L150

I have noticed that the group_size is set to world_size in examples, but in fact the group_size can be set to other numbers according to my understanding.

https://github.com/KaiyuYue/torchshard/blob/main/torchshard/distributed/core.py#L18

I have also found that the get_world_size() will return the number of all processes.

The two findings make me confused in a multi-node setting, say 2 nodes with each node with 2 processes.

If the group_size is 2, then there are 2 distinct groups besides the default group (w/ overlap). However, get_world_size() is used without specifying a group can make a layer be splitted to 4 parts, which is expected to be 2 in our case.

Correct me if I am wrong.
Good Issue

opened by GeneZC 1
Is it possible to collect state dict in cpu?

When I finish one epoch in trianing, the main_worker function will call ts.collect_state_dict(model, state_dict). But because the limit of GPU resource, it will raise Out of Memory in my machine, when call ts.collect_state_dict(model, state_dict). I found that will gather the state_dict in GPU, is it anyway to gather in CPU?
Good Issue

opened by JiaquanYe 2

Releases(v0.1)

v0.1(Apr 28, 2021)

TorchShard: Slicing a PyTorch Tensor Into Parallel Shards. It has the same API design as PyTorch.
Source code(tar.gz)
Source code(zip)

Owner

Kaiyu Yue

GitHub Repository

TorchShard is a lightweight engine for slicing a PyTorch tensor into parallel shards

Related tags

Overview

Installation

Usage

Performance

Contributing

Citing TorchShard

Comments

Future Planinig on this project.

Which one is faster?

8 gpus test example raise error.

Error?

Multi-node setting?

Is it possible to collect state dict in cpu?

Releases(v0.1)

v0.1(Apr 28, 2021)

Owner

Kaiyu Yue

Training RNNs as Fast as CNNs (https://arxiv.org/abs/1709.02755)

A simplified framework and utilities for PyTorch

Pretrained EfficientNet, EfficientNet-Lite, MixNet, MobileNetV3 / V2, MNASNet A1 and B1, FBNet, Single-Path NAS

Pytorch implementation of Distributed Proximal Policy Optimization

A Closer Look at Structured Pruning for Neural Network Compression

Tacotron 2 - PyTorch implementation with faster-than-realtime inference

OptNet: Differentiable Optimization as a Layer in Neural Networks

A PyTorch repo for data loading and utilities to be shared by the PyTorch domain libraries.

The easiest way to use deep metric learning in your application. Modular, flexible, and extensible. Written in PyTorch.

Official implementations of EigenDamage: Structured Pruning in the Kronecker-Factored Eigenbasis.

This is an differentiable pytorch implementation of SIFT patch descriptor.

Implementation of LambdaNetworks, a new approach to image recognition that reaches SOTA with less compute

Fast, general, and tested differentiable structured prediction in PyTorch

torch-optimizer -- collection of optimizers for Pytorch

PyGCL: Graph Contrastive Learning Library for PyTorch

A PyTorch implementation of L-BFGS.

PyTorch wrappers for using your model in audacity!

A tutorial on "Bayesian Compression for Deep Learning" published at NIPS (2017).

PyTorch implementation of TabNet paper : https://arxiv.org/pdf/1908.07442.pdf

Unofficial PyTorch implementation of DeepMind's Perceiver IO with PyTorch Lightning scripts for distributed training