Puiching-Memory/TAAC_2026

What it solves

This project provides an engineering workspace for large-scale recommendation systems, specifically targeting the challenge of unifying sequence modeling (user behavior over time) and feature interaction (high-dimensional sparse features). It aims to move away from fragmented architectures by providing a single framework to model both user behavior sequences and non-sequential multi-domain features for conversion rate prediction.

How it works

The workspace implements a modular pipeline for experimenting with various recommendation models. It uses a streaming data pipeline based on Parquet files with multiple caching strategies and supports various GPU kernels (Triton, TileLang, and PyTorch) for operations like RMSNorm and FlashAttention. The system allows users to quickly switch between different model architectures—such as Baseline, InterFormer, OneTrans, TokenFormer, Symbiosis, DualQ, and QueryFormer—using a shared data pipeline and evaluation process.

Who it’s for

It is designed for researchers and developers working on recommendation systems, specifically those participating in or studying the TAAC 2026 competition focused on unified sequence modeling and feature interaction.

Highlights

• Rapid Experimentation: New models can be integrated using just two files (__init__.py and model.py).
• High-Performance Kernels: Supports multiple GPU backends (torch, Triton, TileLang) for critical layers to optimize inference and training.
• Unified Architecture: Implements several state-of-the-art unified modeling approaches to bridge the gap between sequence and feature interaction.
• Engineering Rigor: Includes full CI coverage, layered testing (unit, contract, integration, benchmark, GPU), and automated bundling for deployment.