Machine Learning Engineer — GPU Kernel

Key Responsibilities

• Understand, analyze, profile, optimize, and provide guidance to the team on deep learning workloads on state-of-the-art hardware and software platforms to improve their efficiency with different levels of optimization
• Design and implement performance benchmarks and testing methodologies to evaluate application performance
• Build tools to automate workload analysis, workload optimization, and other critical workflows
• Triage system issues and identify bottleneck and inefficiencies by analyzing the sources of issues and the impact on hardware, network and propose solutions to enhance GPU utilization
• Support the team to develop appropriate kernels and systems for new model architectures and algorithms
• Participate in, or lead design reviews with peers and stakeholders to decide amongst available technologies.
• Review code developed by other developers and provide feedback to ensure best practices (e.g., style guidelines, checking code in, accuracy, testability, and efficiency).
• Contribute to existing documentation or educational content and adapt content based on product/program updates and user feedback.
• Represent MBZUAI at industry conferences and events, showcasing the institution’s cutting-edge HPC and deep learning capabilities and establishing MBZUAI as a global leader in AI research and innovation.
• Perform all other duties as reasonably directed by the line manager that are commensurate with these functional objectives.
• Validate CUDA kernel outputs and gradients against reference implementations, and benchmark representative shapes, dtypes, and model workloads.

Technical Qualifications

Must-Haves:

• Strong C++ skills and hands-on CUDA kernel development and optimization for deep-learning workloads.
• Understanding of GPU memory hierarchy, warp/block execution, and compute-memory trade-offs, with demonstrated profiling-driven optimization.
• Strong Python skills and experience integrating kernels with PyTorch or an equivalent framework, including numerical and gradient validation where needed.

Nice-to-Haves:

• Experience with Triton, CUTLASS, or PTX/SASS analysis.
• Experience with multi-node distributed training or inference systems.
• Experience validating mixed-precision computations, such as BF16 or FP8.