Agent 训推研究员
Core
Design and implement end-to-end architecture for Agent training and inference systems to support efficient training and inference of trillion-parameter models on large-scale clusters.
Role type
Senior IC machine-learning systems engineer (Agent training & inference)
Builds
Recursive self-improvement (RSI) systems, distributed training infrastructure, high-performance inference services
Domain
AI/ML infrastructure, large-scale model training, distributed systems
Deliverable
production ML models
Required skills
C++ or Python, distributed training frameworks (Megatron/DeepSpeed/FSDP/torchtitan), inference optimization (vLLM/SGLang/TensorRT-LLM), reinforcement learning (verl/slime/Agentic RL), hardware-aware optimization (Ascend/CANN or GPU/CUDA), system-level design
Preferred skills
Open source contributions to training/inference frameworks
Technologies
Megatron, DeepSpeed, FSDP, torchtitan, vLLM, SGLang, TensorRT-LLM, verl, slime, Ascend, CANN, CUDA, MindSpore, PyTorch
Responsibilities
Design and implement RSI system architecture for recursive self-improvement; Build distributed training systems for trillion-parameter models; Construct high-performance inference services with low latency and high concurrency.
