hybridbackend-tf115-cu121 1.0.0

A high-performance framework for training wide-and-deep recommender systems on heterogeneous cluster

HybridBackend

HybridBackend is a high-performance framework for training wide-and-deep recommender systems on heterogeneous cluster.

Features

• Memory-efficient loading of categorical data
• GPU-efficient orchestration of embedding layers
• Communication-efficient training and evaluation at scale
• Easy to use with existing AI workflows

Usage

A minimal example:

import tensorflow as tf
import hybridbackend.tensorflow as hb

ds = hb.data.Dataset.from_parquet(filenames)
ds = ds.batch(batch_size)
# ...

with tf.device('/gpu:0'):
  embs = tf.nn.embedding_lookup_sparse(weights, input_ids)
  # ...

Please see documentation for more information.

Install

Method 1: Install from PyPI

pip install {PACKAGE}

{PACKAGE}	Dependency	Python	CUDA	GLIBC	Data Opt.	Embedding Opt.	Parallelism Opt.
hybridbackend-tf115-cu121	TensorFlow 1.15	3.8	12.1	>=2.31	✓	✓	✓
hybridbackend-tf115-cu100	TensorFlow 1.15	3.6	10.0	>=2.27	✓	✓	✗
hybridbackend-tf115-cpu	TensorFlow 1.15	3.6	-	>=2.24	✓	✗	✗

Method 2: Build from source

See Building Instructions.

We also provide built docker images for latest DeepRec: registry.cn-shanghai.aliyuncs.com/pai-dlc/hybridbackend:1.0.0-deeprec-py3.6-cu114-ubuntu18.04

License

HybridBackend is licensed under the Apache 2.0 License.

Community

• Please see Contributing Guide before your first contribution.

• Please register as an adopter if your organization is interested in adoption. We will discuss RoadMap with registered adopters in advance.

• Please cite HybridBackend in your publications if it helps:

  @inproceedings{zhang2022picasso,
    title={PICASSO: Unleashing the Potential of GPU-centric Training for Wide-and-deep Recommender Systems},
    author={Zhang, Yuanxing and Chen, Langshi and Yang, Siran and Yuan, Man and Yi, Huimin and Zhang, Jie and Wang, Jiamang and Dong, Jianbo and Xu, Yunlong and Song, Yue and others},
    booktitle={2022 IEEE 38th International Conference on Data Engineering (ICDE)},
    year={2022},
    organization={IEEE}
  }
  

Contact Us

If you would like to share your experiences with others, you are welcome to contact us in DingTalk: