Backend Engineer , AML Engine Orchestration

Team Introduction The mission of our AML team is to push next-generation machine learning algorithms and platforms for the recommendation system, ads ranking and search ranking in our company. We also drive substantial impact on core businesses of the company.

Responsibilities

1. Resource Efficiency Optimization in Distributed Orchestration and Scheduling： - Develop and extend distributed orchestration frameworks within the Kubernetes/Godel ecosystem. Select appropriate frameworks based on different business scenarios, and optimize cluster utilization and load balancing strategies according to the specific characteristics of each scenario; - Integrate and expand AutoScaling and automatic parallelization capabilities for various models and tasks. Employ load modeling and analytic methods for different models to automatically optimize resource requests, achieving large-scale improvements in resource usage efficiency and global optimality; - Responsible for preemption and re-scheduling mechanisms for services with different prioritties, and manage automatic resource multiplexing across different clusters and resource types; handle scheduling and load adaptation across multi-datacenter, multi-region, and multi-cloud environments. 2. Building Training System Architecture for Next-Generation Ultra-Large and Ultra-Deep Recommendation Models: - Develop a flexible, elastic and robust distributed training runtime focused on hyper-scaled embeddings and large-scale GPU training; - Design and optimize distributed computing APIs and runtimes geared towards future recommendation and ads model paradigms (e.g., reinforcement learning, fine-tuning and/or distillation); - Collaborate with platform teams to enhance the diagnosability and usability of distributed training systems. 3. Constructing Online Orchestration Architecture for Next-Generation Recommendation Systems: - Build a robust distributed model inference architecture for online learning scenarios involving hyper-scaled embeddings; - Optimize the usability of online recommendation and ads model architectures and MLops workflows.