PersiaML/PERSIA

High performance distributed framework for training deep learning recommendation models based on PyTorch.

What it solves

PERSIA addresses the challenge of training massive deep learning recommendation models that exceed the capacity of single-device memory. It enables the training of models with up to 100 trillion parameters on commodity hardware, overcoming the scale and efficiency bottlenecks of traditional training systems.

How it works

It is a PyTorch-based system that utilizes hybrid acceleration to provide a parallel recommendation training environment. By leveraging heterogeneous parallelization, it allows recommendation models to be scaled across multiple hardware components to handle extreme parameter counts while maintaining high efficiency and robustness.

Who it’s for

Machine learning engineers and researchers focusing on large-scale recommendation systems, particularly those dealing with datasets and models that require distributed training across commodity hardware.

Highlights

  • Extreme Scale: Capable of training models with up to 100 trillion parameters.
  • PyTorch Integration: The first public PyTorch-based system specifically designed for large-scale recommendation training.
  • PyTorch-based: Built on PyTorch, making it it easier for researchers to integrate with the same ecosystem.
  • Industry Validated: Proven effective in production environments with 100 million level Daily Active Users (DAU) at Kuaishou.
  • High Efficiency: Demonstrated significant performance advantages over existing recommendation training systems.

Related

  • Project
    alibaba/TorchEasyRec

    A PyTorch-based recommendation system framework that provides a library of production-ready deep learning models for matching, ranking, and generative recommendation.

  • Project
    meta-pytorch/torchrec

    A PyTorch domain library providing sparsity and parallelism primitives for training and deploying large-scale recommender systems with massive embedding tables sharded across multiple GPUs.

  • Project
    pytorch/torchtitan

    torchtitan is a PyTorch‑native library for large‑scale generative‑AI training. It bundles modern parallelism (FSDP2, tensor, pipeline, context), memory‑saving tricks (activation checkpointing, Float8/MXFP8 quantization), distributed checkpointing, and extensive logging. Ready‑to‑use for Llama 3.1 (8 B‑405 B) and extensible to new models, it runs on single‑node or multi‑node clusters via simple CLI/config files, and is released under a BSD‑3‑Clause license.

  • Project
    Lightning-AI/pytorch-lightning

    A deep learning framework that abstracts away the engineering boilerplate of PyTorch, enabling researchers to pretrain and finetune AI models at scale across various hardware.

  • Project
    pytorch/pytorch

    A Python-based machine learning framework that provides GPU-accelerated tensor computation and a dynamic autograd system for building deep neural networks.