A Memory-saving Training Framework for Transformers


This is the official PyTorch implementation for Mesa: A Memory-saving Training Framework for Transformers.

By Zizheng Pan, Peng Chen, Haoyu He, Jing Liu, Jianfei Cai and Bohan Zhuang.


  1. Create a virtual environment with anaconda.

    conda create -n mesa python=3.7 -y
    conda activate mesa
    # Install PyTorch, we use PyTorch 1.7.1 with CUDA 10.1 
    pip install torch==1.7.1+cu101 torchvision==0.8.2+cu101 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html
    # Install ninja
    pip install ninja
  2. Build and install Mesa.

    # clone this repo
    git clone https://github.com/zhuang-group/Mesa
    # build
    cd Mesa/
    # You need to have an NVIDIA GPU
    python setup.py develop


  1. Prepare your policy and save as a text file, e.g. policy.txt.

    on gelu: # layer tag, choices: fc, conv, gelu, bn, relu, softmax, matmul, layernorm
        by_index: all # layer index
        enable: True # enable for compressing
        level: 256 # we adopt 8-bit quantization by default
        ema_decay: 0.9 # the decay rate for running estimates
        by_index: 1 2 # e.g. exluding GELU layers that indexed by 1 and 2.
        enable: False
  2. Next, you can wrap your model with Mesa by:

    import mesa as ms
    ms.policy.convert_by_num_groups(model, 3)
    # or convert by group size with ms.policy.convert_by_group_size(model, 64)
    # setup compression policy
    ms.policy.deploy_on_init(model, '[path to policy.txt]', verbose=print, override_verbose=False)

    That’s all you need to use Mesa for memory saving.

    Note that convert_by_num_groups and convert_by_group_size only recognize nn.XXX, if your code has functional operations, such as [email protected] and F.Softmax, you may need to manually setup these layers. For example:

    import mesa as ms
    # matrix multipcation (before)
    out = Q@K.transpose(-2, -1)
    # with Mesa
    self.mm = ms.MatMul(quant_groups=3)
    out = self.mm(q, k.transpose(-2, -1))
    # sofmtax (before)
    attn = attn.softmax(dim=-1)
    # with Mesa
    self.softmax = ms.Softmax(dim=-1, quant_groups=3)
    attn = self.softmax(attn)
  3. You can also target one layer by:

    import mesa as ms
    # previous 
    self.act = nn.GELU()
    # with Mesa
    self.act = ms.GELU(quant_groups=[num of quantization groups])

Demo projects for DeiT and Swin

We provide demo projects to replicate our results of training DeiT and Swin with Mesa, please refer to DeiT-Mesa and Swin-Mesa.

Results on ImageNet

Model Param (M) FLOPs (G) Train Memory (MB) Top-1 (%)
DeiT-Ti 5 1.3 4,171 71.9
DeiT-Ti w/ Mesa 5 1.3 1,858 72.1
DeiT-S 22 4.6 8,459 79.8
DeiT-S w/ Mesa 22 4.6 3,840 80.0
DeiT-B 86 17.5 17,691 81.8
DeiT-B w/ Mesa 86 17.5 8,616 81.8
Swin-Ti 29 4.5 11,812 81.3
Swin-Ti w/ Mesa 29 4.5 5,371 81.3
PVT-Ti 13 1.9 7,800 75.1
PVT-Ti w/ Mesa 13 1.9 3,782 74.9

Memory footprint at training time is measured with a batch size of 128 and an image resolution of 224×224 on a single GPU.


If you find our work interesting or helpful to your research, please consider citing Mesa.

      title={Mesa: A Memory-saving Training Framework for Transformers}, 
      author={Zizheng Pan and Peng Chen and Haoyu He and Jing Liu and Jianfei Cai and Bohan Zhuang},
      journal={arXiv preprint arXiv:2111.11124}


This repository is released under the Apache 2.0 license as found in the LICENSE file.


This repository has adopted part of the quantization codes from ActNN, we thank the authors for their open-sourced code.


View Github