Skip to content

Starmys/SparTA

 
 

Repository files navigation

SparTA Getting Started

This repo is under active development. We're trying to make it stable and user-friendly, but it is not surprising to meet problems in current phase. Please open issue or contact the authors when you need help.

SparTA is an end-to-end system to harvest the speeding up gain from the model sparsity.

Installation

SparTA depends on user's local CUDA environments. Here are some requirements

User could install through pip command as below (The PyPI install path is coming soon)

pip install git+https://github.com/microsoft/SparTA.git

or

git clone [email protected]:microsoft/SparTA.git
pip install SparTA

Please make sure that the CUDA version matches the version used to compile PyTorch binaries. If cuda and nvcc version issues met, the following commands may be helpful to verify the environments.

import os
import torch
import pycuda.driver

if torch.cuda.is_available():
    os.system('nvcc --version')
    print(torch.version.cuda)
    print(pycuda.driver.get_version())
    print(pycuda.driver.get_driver_version())

Usage

Tune a sparse operator

import torch
import sparta

batch_size, in_features, out_features = 1024, 1024, 1024
sparsity = 0.9
granularity = (8, 8)

# prepare data
x = torch.rand((batch_size, in_features), device='cuda')
weight = torch.rand((out_features, in_features), device='cuda')
bias = torch.rand((out_features, ), device='cuda')

# generate and apply weight mask
mask = sparta.testing.block_mask(weight.shape, granularity, sparsity, device='cuda')
weight = torch.mul(weight, mask)

# create a dense operator
dense_linear = torch.nn.Linear(in_features, out_features, device='cuda')
dense_linear.load_state_dict({'weight': weight, 'bias': bias})

# create a sparse operator
sparse_linear = sparta.nn.SparseLinear(dense_linear, weight_mask=mask)

# tune the sparse operator
best_config = sparta.nn.tune(sparse_linear, sample_inputs=[x], max_trials=10, algo='rand')

# check if the sparse operator runs correctly
torch.testing.assert_close(sparse_linear(x), dense_linear(x))

Build a sparse operator with specified config

# create a sparse operator
sparse_linear = sparta.nn.SparseLinear(dense_linear, weight_mask=mask)

# build the sparse operator with the `best_config` we got before.
sparse_linear.nn.build(best_config['root'], sample_inputs=[x])

Citing SparTA

If SparTA is helpful in your projects, please cite our paper as below

@inproceedings {SparTA2022,
    author = {Ningxin Zheng and Bin Lin and Quanlu Zhang and Lingxiao Ma and Yuqing Yang and Fan Yang and Yang Wang and Mao Yang and Lidong Zhou},
    title = {SparTA: Deep-Learning Model Sparsity via Tensor-with-Sparsity-Attribute},
    booktitle = {16th USENIX Symposium on Operating Systems Design and Implementation (OSDI 22)},
    year = {2022},
    isbn = {978-1-939133-28-1},
    address = {Carlsbad, CA},
    pages = {213--232},
    url = {https://www.usenix.org/conference/osdi22/presentation/zheng-ningxin},
    publisher = {USENIX Association},
    month = jul,
}

Contributor License Agreement

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.

Releases

No releases published

Packages

No packages published

Languages

  • Python 48.1%
  • Jinja 44.6%
  • Cuda 6.9%
  • C++ 0.4%