Research MLE (Training Optimization) (AI)

TL;DR

Research MLE (AI Engineering): Designing and optimizing large-scale training systems for multimodal and foundation models with an accent on GPU utilization, memory efficiency, and communication layers. Focus on implementing custom CUDA/Triton kernels and scaling distributed training using frameworks like Megatron-LM and NeMo.

Location: Beijing, China

Company

Canva is building an AI-powered future for design, developing a creative intelligence engine to power tools like smart editing and AI video.

What you will do

• Design, implement, and optimize large-scale machine learning systems for training.
• Improve GPU utilization, communication overhead, and memory efficiency across the stack.
• Partner with research and modeling teams to align systems with algorithmic needs.
• Apply and evaluate best practices for distributed training using Megatron-LM, NeMo, and FSDP.
• Develop low-level optimizations, including custom CUDA or Triton kernels.
• Debug, profile, and fine-tune training workflows to maximize scalability.

Requirements

• Strong background in LLMs, multimodal AI, or diffusion models.
• Proficiency in Python; familiarity with C++ or Rust is a plus.
• Deep knowledge of PyTorch or JAX and libraries such as Megatron-LM, NeMo, or DeepSpeed.
• Experience with optimization techniques like FSDP/ZeRO, gradient checkpointing, and low-precision data types.
• Hands-on experience writing custom GPU kernels in CUDA or Triton.
• Full proficiency in English required.