Hidet: A compilation-based deep learning framework

Documentation

Hidet is an open-source DNN inference framework based on compilation. It supports end-to-end compilation of DNN models from PyTorch and ONNX to efficient cuda kernels. A series of graph-level and operator-level optimizations are applied to optimize the performance.

Getting Started

Installation

pip install hidet

See here for building from source.

Usage

Optimize a PyTorch model through hidet (require PyTorch 2.0):

import torch
import hidet

# Register hidet backends for pytorch dynamo, can be omitted if you import torch before hidet
hidet.torch.register_dynamo_backends()  

# Define pytorch model
model = torch.hub.load('pytorch/vision:v0.6.0', 'resnet18', pretrained=True).cuda().eval()
x = torch.rand(1, 3, 224, 224).cuda()

# Compile the model through Hidet
model_opt = torch.compile(model, backend='hidet')  

# Run the optimized model
y = model_opt(x)

See the following tutorials to learn other usgae:

• Quick Start
• Optimize PyTorch models
• Optimize ONNX models

Publication

Hidet originates from the following research work. If you used Hidet in your research, welcome to cite our paper.

• Hidet: Task-Mapping Programming Paradigm for Deep Learning Tensor Programs.
  Yaoyao Ding, Cody Hao Yu, Bojian Zheng, Yizhi Liu, Yida Wang, and Gennady Pekhimenko.

Development

Hidet is currently under active development by a team at CentML Inc.

Contributing

We welcome contributions from the community. Please see contribution guide for more details.

License

Hidet is released under the Apache 2.0 license.