Skip to main content

Alibaba Intelligent Engine Algorithm Platform Team

The Algorithm Platform team of Alibaba Holding Group's Intelligent Engine Business Unit builds Alibaba Group's model training infrastructure and is responsible for data and training Infra for the HappyHorse and HappyOyster model families. The team maintains industry-leading pre-training and post-training frameworks for large language models, multimodal models, and generative models, together with sample storage and compute systems. Open source projects include Megatron-LLaMA, ROLL, and RecIS. The team has published multiple works at top conferences including NSDI, OSDI, and SIGMOD, and received the 2026 NSDI Outstanding Paper Award. Through distributed optimization, software-hardware co-design, and model-Infra codesign, the team optimizes large model iteration efficiency from data processing through training, expands the ceiling of model quality, and builds frontier infrastructure for large models.

Mission & Responsibilities

We are responsible for building Alibaba's large-scale training infrastructure, covering LLM training Infra, multi-modal large model training Infra, recommendation algorithm model training Infra, feature computation & processing Infra, and algorithm platform construction.

Selected Publications

Open Source Frameworks

Megatron-LLaMA

An open-source LLM training framework based on Megatron, supporting efficient distributed training.

LLMPyTorch
GitHub

X-DeepLearning (XDL)

Alibaba's open-source sparse model training framework for large-scale recommendation and advertising.

SparseRecSys
GitHub

ROLL

An open-source RL training framework for efficient distributed RL post-training at scale.

RLDistributed
GitHub

Euler

Alibaba's open-source distributed graph learning engine for large-scale GNN training.

GNNGraph
GitHub

RecIS

A large model training framework for recommendation and advertising at industrial scale.

RecSysTraining
GitHub

Open Roles

Select a role or use the apply button to open the Alibaba Talent application page.