Puiching-Memory/TAAC_2026

What it solves

This project provides an out-of-the-box, extensible experimental workspace for large-scale recommendation systems. It specifically targets the challenge of unifying sequence modeling (capturing user behavior over time) and feature interaction (crossing high-dimensional sparse features), which are traditionally handled by separate architectures.

How it works

The workspace implements a unified modeling paradigm using a single architecture to handle both user behavior sequences and non-sequential multi-domain features for conversion rate prediction. It provides a streamlined pipeline for:

• Model Integration: New models can be added via two files (__init__.py and model.py) and switched with a single line of code.
• Data Pipeline: A step-level streaming pipeline using Parquet for efficient reading, supporting five caching strategies and various data augmentation techniques (sequence clipping, domain-level dropout, feature masking).
• Hardware Optimization: Support for multiple GPU kernels (Triton, TileLang, torch) for operations like RMSNorm, LayerNorm, and FlashAttention.
• Deployment: Tools to package training and inference bundles for platform submission.

Who it’s for

Researchers and developers working on recommendation systems, specifically those participating in or studying the TAAC 2026 competition focusing on unified sequence modeling and feature interaction.

Highlights

• Multi-Model Support: Includes implementations of Baseline, InterFormer, OneTrans, TokenFormer, Symbiosis, DualQ, and QueryFormer.
• High-Performance Kernels: Integrated support for Triton and TileLang to optimize GPU performance.
• HuggingFace Integration: Easy downloading of sample datasets via huggingface-cli.
• Comprehensive Tooling: Includes CI gates, layered testing (unit, contract, integration, benchmark, GPU), and automated documentation generation via Zensical.