GPU ML Infra Intern: Speed Up Training & Profiling

Plus 2; Santa Clara (CA); On-site; USD 150,000 - 210,000; Full time

Benefits offered by this job

Free lunch

Snacks & drinks

401(k) plan

Job summary

PlusAI in Silicon Valley seeks a highly skilled systems/performance engineer to optimize BEV model training pipelines and GPU kernels. You will profile, implement custom kernels with CUDA or Triton, and push performance through LLM-assisted code generation and profiling workflows.

Key focus areas include memory management, PyTorch custom ops, and GPU utilization, with opportunities to influence autonomous vehicle perception workloads and high-performance computing practices.

Responsibilities

• Identify Training Bottlenecks: Profile BEV model training pipelines to pinpoint computational and memory bottlenecks.
• Develop Custom Kernels: Design and implement high-performance kernels using CUDA, Triton, or C++ to accelerate training.
• Leverage LLMs for Optimization: Explore and integrate LLMs to generate high-performance code and optimize kernel logic.
• Automate Profiling Workflows: Build systems to automate performance profiling with NVIDIA Nsight and PyTorch Profiler.
• Iterative Performance Tuning: Analyze profiling data to maximize GPU utilization and reduce training times.
• Systems Programming: Strong proficiency in C++ with memory management and parallel processing principles.
• Deep Learning Frameworks: Hands-on PyTorch experience with custom operations and autograd.
• Performance-Oriented Mindset: Problem-solving with a focus on low-level optimization.
• GPU Programming Experience: Experience writing and optimizing custom GPU kernels with CUDA or Triton.
• Hardware Profiling Tools: Familiarity with profiling tools like Nsight and PyTorch Profiler.
• LLM for Code Generation: Experience with LLMs for code writing or refactoring.
• Autonomous Vehicle Perception: Understanding BEV models and 3D perception.

Job description

PlusAI in Silicon Valley seeks a highly skilled systems/performance engineer to optimize BEV model training pipelines and GPU kernels. You will profile, implement custom kernels with CUDA or Triton, and push performance through LLM-assisted code generation and profiling workflows.

Key focus areas include memory management, PyTorch custom ops, and GPU utilization, with opportunities to influence autonomous vehicle perception workloads and high-performance computing practices.