广告系统基础及综述

Efficient and memory-optimized training methods for long-sequence modeling, including hybrid attention mechanisms (FlashAttention, Performer, Sparse Attention) for scalable long-context learning.

Efficiency and longer-sequence experiments on SASRec: vectorized negative sampling, full-softmax vs sampled loss, sequence-length scaling (Table V), and throughput profiling. PyTorch, CPU-friendly.

Customized linear attention for scaling long sequence in recommendation system

Tensor-core CUDA kernels for Nyström attention, linear-time forward and backward with exact autograd gradients. Faster than flash-attention at long sequence length.

Omni inference in C/C++

GenPark seasonal promo web page banner matching scheduler skill.

GenPark search autocomplete suggestions and semantic query expansion agent skill.

A query parser with operator support, spellcheck, and prefix/n-gram based autocomplete suggestions.

A small-scale implementation of Generative Modeling via Drifting

基于生成式检索（Generative Retrieval）范式的序列推荐系统。本项目在复现 TIGER 架构的基础上，围绕 Semantic ID tokenizer、Sinkhorn、逐层生成瓶颈和碰撞 suffix 展开系统性诊断与改进。

Personalized preference alignment from limited data

Timehash: Hierarchical temporal indexing for efficient "open now" search in large-scale POI systems.

Implementation of "Beyond Polarization: The Generative Constraint of Chain-of-Thought in Pointwise Reranking."

搜索、推荐、广告平台基础架构

CONGA: COntrastive Nested Graph Architecture for Continual Sequential Recommendation

Official implementation of our ACM RecSys 2026 paper.

【ICML2026 Spotlight】 T2PO: Uncertainty-Guided Exploration Control for Stable Multi-Turn Agentic Reinforcement Learning

rl-explainer

PCTD: Preference-Guided Counterfactual Task Decomposition for Agent Tool Retrieval

Examples for Recommenders - easy to train and deploy on accelerated infrastructure.

A hands-on lab implementing LLM inference kernels from scratch using Triton. Covers fused attention, Flash Attention, GQA, RoPE, INT8 quantization, KV cache optimization, and speculative decoding —…

Does quantizing the KV cache change the greedy output? Yes: q8_0 KV changes the generated text on 83% of prompts, q4_0 on 100% (often from the start) - with flash attention held constant, so KV pre…

gpt 2 forward pass CUDA kernels w/ optimizations like flash attention, kv cache, cublass/cutlass, split k , tensor cores

Flash Attention 2 inference with KV caching deployed on GPT-2