What is TorchEasyRec?

TorchEasyRec implements state-of-the-art deep learning models for recommendation tasks: candidate generation (matching), scoring (ranking), multi-task learning, and generative recommendation. It enables efficient development of high-performance models through simple configuration and easy customization.

Key Features

Data Sources

• MaxCompute/ODPS - Native Alibaba Cloud data warehouse integration
• Parquet - High-performance columnar file format when using Local | OSS | NAS storage, with built-in auto-rebalancing capabilities
• CSV - Standard tabular file format
• Streaming - Kafka message queue integration, also compatible with Alibaba Datahub
• Checkpointable - Resume training from exact data position

Scalability

• Distributed Training - Hybrid data/model parallelism via TorchRec
• Large Embeddings - Row-wise, column-wise, table-wise sharding
• Zero-Collision Hash - Large scale Dynamic embedding with eviction policies (LFU/LRU)
• Mixed Precision - FP16/BF16 training support

Production

• Run Everywhere - Local, PAI-DLC, PAI-DSW
• Feature Generation - Consistent FG between training and serving
• EAS Deployment - Auto-scaling model serving on Alibaba Cloud
• TensorRT/AOTInductor - Model acceleration for inference

Features & Models

• 20+ Models - Battle-tested algorithms powering real-world recommendation: DSSM, TDM, DeepFM, DIN, MMoE, PLE, PEPNet, DLRM-HSTU and more
• 10+ Feature Types - IdFeature, RawFeature, ComboFeature, LookupFeature, ExprFeature, SequenceFeature, CustomFeature, and more
• Custom Model - Easy to implement customized models
• Custom Feature - Easy to implement customized features

Supported Models

Matching (Candidate Generation)

Model	Description
DSSM	Two-tower deep semantic matching model
MIND	Multi-interest network with dynamic routing
TDM	Tree-based deep model for large-scale retrieval
DAT	Dual augmented two-tower model

Ranking (Scoring)

Model	Description
DeepFM	Factorization-machine based neural network
WideAndDeep	Wide & Deep learning for recommendations
MultiTower	Flexible multi-tower architecture
DIN	Deep Interest Network with attention mechanism
DLRM	Deep Learning Recommendation Model
DCN	Deep & Cross Network
DCN-V2	Improved Deep & Cross Network
MaskNet	Instance-guided mask for feature interaction
xDeepFM	Compressed interaction network
WuKong	Dense scaling with high-order interactions
RocketLaunching	Knowledge distillation framework

Multi-Task Learning

Model	Description
MMoE	Multi-gate Mixture-of-Experts
PLE	Progressive Layered Extraction
DBMTL	Deep Bayesian Multi-task Learning
PEPNet	Personalized Embedding and Parameter Network

Generative Recommendation

Model	Description
DLRM-HSTU	Hierarchical Sequential Transduction Units
ULTRA-HSTU	HSTU with Semi-Local Attention, Attention Truncation, and Mixture of Transducers
HSTU-Match	HSTU-based two-tower retrieval model

Semantic ID (SID) Generation

Model	Description
SID RQVAE	RQ-VAE residual-quantized semantic-ID generation for generative retrieval
SID RQKMeans	FAISS residual K-Means semantic-ID generation

Documentation

Get started with TorchEasyRec in minutes:

Tutorial	Description
Local Training	Train models on your local machine or single server
PAI-DLC Training	Distributed training on Alibaba Cloud PAI-DLC
PAI-DLC + MaxCompute Table	Train with MaxCompute (ODPS) tables on PAI-DLC

For the complete documentation, please refer to https://torcheasyrec.readthedocs.io/

Community & Support

• GitHub Issues - Report bugs or Request features

• DingTalk Groups

  • DingDing Group: 32260796 - Join
  • DingDing Group2: 37930014162 - Join
  

• If you have any questions about how to use TorchEasyRec, please join the DingTalk group and contact us.

• If you have enterprise service needs or need to purchase Alibaba Cloud services to build a recommendation system, please join the DingTalk group to contact us.

Contributing

Any contributions you make are greatly appreciated!

• Please report bugs by submitting an issue
• Please submit contributions using pull requests
• Please refer to the Development Guide for more details

Citation

If you use TorchEasyRec in your research, please cite:

@software{torcheasyrec2024,
  title = {TorchEasyRec: An Easy-to-Use Framework for Recommendation},
  author = {Alibaba PAI Team},
  year = {2024},
  url = {https://github.com/alibaba/TorchEasyRec}
}

License

TorchEasyRec is released under Apache License 2.0. Please note that third-party libraries may not have the same license as TorchEasyRec.