Skip to content

Patch convolution to avoid large GPU memory usage of Conv2D

License

Notifications You must be signed in to change notification settings

mit-han-lab/patch_conv

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Patch Conv: Patch convolution to avoid large GPU memory usage of Conv2D [Blog]

patch_conv

Background

In current generative models, we usually apply convolutions over large-size activations to generate high-resolution content. However, PyTorch tends to use excessive memory for these operations, potentially leading to memory shortages even on 80GB A100 GPUs.

As shown in the figure below, memory demands for standard PyTorch convolutions drastically increase when the input size reaches 1B parameters (channel×height×width). Notably, with a kernel size of 7×7, the 80GB A100 GPUs would trigger Out of Memory (OOM) errors. Inputs exceeding 2B parameters can further cause 3×3 convolutions exhaust all the memory and that’s just for one layer! This memory bottleneck prevents users and the community from scaling up the models to produce high-quality images.

To bypass this issue and reduce memory consumption, we propose a simple and effective solution -- Patch Conv. As shown in the above figure, similar to SIGE, Patch Conv first divides the input into several smaller patches along the height dimension while keeping some overlap between them. These patches are then reorganized into the batch dimension and fed into the original convolution to produce output patches, which are then concatenated together to form the final output. Patch Conv can reduce memory usage by over 2.4×, providing a viable workaround for the limitations of current implementations.

Installation

After installing PyTorch, you can install PatchConv from PyPI:

pip install patch_conv

or via GitHub:

pip install git+https://github.com/mit-han-lab/patch_conv.git

or locally for development:

git clone [email protected]:mit-han-lab/patch_conv.git
cd patch_conv
pip install -e .

Usage

All you need to do is use convert_model to wrap all the Conv2d in your PyTorch model to our PatchConv. For example,

from patch_conv import convert_model

model = Model(...)  # Your PyTorch model
model = convert_model(model, splits=4)  # The only modification you need to make

with torch.no_grad():
    model(...)  # Run the model in the original way

Performance

performance

Patch Conv significantly reduces memory consumption by over 2.4× across various kernel sizes and input resolutions with a marginally slower inference speed compared to vanilla convolution.

Related Projects